git.sesse.net Git - ffmpeg/blob - libavutil/tx_template.c

   1 /*
   2  * Copyright (c) 2019 Lynne <dev@lynne.ee>
   3  * Power of two FFT:
   4  * Copyright (c) 2008 Loren Merritt
   5  * Copyright (c) 2002 Fabrice Bellard
   6  * Partly based on libdjbfft by D. J. Bernstein
   7  *
   8  * This file is part of FFmpeg.
   9  *
  10  * FFmpeg is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * FFmpeg is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with FFmpeg; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  23  */
  24
  25 /* All costabs for a type are defined here */
  26 COSTABLE(16);
  27 COSTABLE(32);
  28 COSTABLE(64);
  29 COSTABLE(128);
  30 COSTABLE(256);
  31 COSTABLE(512);
  32 COSTABLE(1024);
  33 COSTABLE(2048);
  34 COSTABLE(4096);
  35 COSTABLE(8192);
  36 COSTABLE(16384);
  37 COSTABLE(32768);
  38 COSTABLE(65536);
  39 COSTABLE(131072);
  40 DECLARE_ALIGNED(32, FFTComplex, TX_NAME(ff_cos_53))[4];
  41
  42 static FFTSample * const cos_tabs[18] = {
  43     NULL,
  44     NULL,
  45     NULL,
  46     NULL,
  47     TX_NAME(ff_cos_16),
  48     TX_NAME(ff_cos_32),
  49     TX_NAME(ff_cos_64),
  50     TX_NAME(ff_cos_128),
  51     TX_NAME(ff_cos_256),
  52     TX_NAME(ff_cos_512),
  53     TX_NAME(ff_cos_1024),
  54     TX_NAME(ff_cos_2048),
  55     TX_NAME(ff_cos_4096),
  56     TX_NAME(ff_cos_8192),
  57     TX_NAME(ff_cos_16384),
  58     TX_NAME(ff_cos_32768),
  59     TX_NAME(ff_cos_65536),
  60     TX_NAME(ff_cos_131072),
  61 };
  62
  63 static av_always_inline void init_cos_tabs_idx(int index)
  64 {
  65     int m = 1 << index;
  66     double freq = 2*M_PI/m;
  67     FFTSample *tab = cos_tabs[index];
  68     for(int i = 0; i <= m/4; i++)
  69         tab[i] = RESCALE(cos(i*freq));
  70     for(int i = 1; i < m/4; i++)
  71         tab[m/2 - i] = tab[i];
  72 }
  73
  74 #define INIT_FF_COS_TABS_FUNC(index, size)                                     \
  75 static av_cold void init_cos_tabs_ ## size (void)                              \
  76 {                                                                              \
  77     init_cos_tabs_idx(index);                                                  \
  78 }
  79
  80 INIT_FF_COS_TABS_FUNC(4, 16)
  81 INIT_FF_COS_TABS_FUNC(5, 32)
  82 INIT_FF_COS_TABS_FUNC(6, 64)
  83 INIT_FF_COS_TABS_FUNC(7, 128)
  84 INIT_FF_COS_TABS_FUNC(8, 256)
  85 INIT_FF_COS_TABS_FUNC(9, 512)
  86 INIT_FF_COS_TABS_FUNC(10, 1024)
  87 INIT_FF_COS_TABS_FUNC(11, 2048)
  88 INIT_FF_COS_TABS_FUNC(12, 4096)
  89 INIT_FF_COS_TABS_FUNC(13, 8192)
  90 INIT_FF_COS_TABS_FUNC(14, 16384)
  91 INIT_FF_COS_TABS_FUNC(15, 32768)
  92 INIT_FF_COS_TABS_FUNC(16, 65536)
  93 INIT_FF_COS_TABS_FUNC(17, 131072)
  94
  95 static av_cold void ff_init_53_tabs(void)
  96 {
  97     TX_NAME(ff_cos_53)[0] = (FFTComplex){ RESCALE(cos(2 * M_PI / 12)), RESCALE(cos(2 * M_PI / 12)) };
  98     TX_NAME(ff_cos_53)[1] = (FFTComplex){ RESCALE(cos(2 * M_PI /  6)), RESCALE(cos(2 * M_PI /  6)) };
  99     TX_NAME(ff_cos_53)[2] = (FFTComplex){ RESCALE(cos(2 * M_PI /  5)), RESCALE(sin(2 * M_PI /  5)) };
 100     TX_NAME(ff_cos_53)[3] = (FFTComplex){ RESCALE(cos(2 * M_PI / 10)), RESCALE(sin(2 * M_PI / 10)) };
 101 }
 102
 103 static CosTabsInitOnce cos_tabs_init_once[] = {
 104     { ff_init_53_tabs, AV_ONCE_INIT },
 105     { NULL },
 106     { NULL },
 107     { NULL },
 108     { init_cos_tabs_16, AV_ONCE_INIT },
 109     { init_cos_tabs_32, AV_ONCE_INIT },
 110     { init_cos_tabs_64, AV_ONCE_INIT },
 111     { init_cos_tabs_128, AV_ONCE_INIT },
 112     { init_cos_tabs_256, AV_ONCE_INIT },
 113     { init_cos_tabs_512, AV_ONCE_INIT },
 114     { init_cos_tabs_1024, AV_ONCE_INIT },
 115     { init_cos_tabs_2048, AV_ONCE_INIT },
 116     { init_cos_tabs_4096, AV_ONCE_INIT },
 117     { init_cos_tabs_8192, AV_ONCE_INIT },
 118     { init_cos_tabs_16384, AV_ONCE_INIT },
 119     { init_cos_tabs_32768, AV_ONCE_INIT },
 120     { init_cos_tabs_65536, AV_ONCE_INIT },
 121     { init_cos_tabs_131072, AV_ONCE_INIT },
 122 };
 123
 124 static av_cold void init_cos_tabs(int index)
 125 {
 126     ff_thread_once(&cos_tabs_init_once[index].control,
 127                     cos_tabs_init_once[index].func);
 128 }
 129
 130 static av_always_inline void fft3(FFTComplex *out, FFTComplex *in,
 131                                   ptrdiff_t stride)
 132 {
 133     FFTComplex tmp[2];
 134 #ifdef TX_INT32
 135     int64_t mtmp[4];
 136 #endif
 137
 138     BF(tmp[0].re, tmp[1].im, in[1].im, in[2].im);
 139     BF(tmp[0].im, tmp[1].re, in[1].re, in[2].re);
 140
 141     out[0*stride].re = in[0].re + tmp[1].re;
 142     out[0*stride].im = in[0].im + tmp[1].im;
 143
 144 #ifdef TX_INT32
 145     mtmp[0] = (int64_t)TX_NAME(ff_cos_53)[0].re * tmp[0].re;
 146     mtmp[1] = (int64_t)TX_NAME(ff_cos_53)[0].im * tmp[0].im;
 147     mtmp[2] = (int64_t)TX_NAME(ff_cos_53)[1].re * tmp[1].re;
 148     mtmp[3] = (int64_t)TX_NAME(ff_cos_53)[1].re * tmp[1].im;
 149     out[1*stride].re = in[0].re - (mtmp[2] + mtmp[0] + 0x40000000 >> 31);
 150     out[1*stride].im = in[0].im - (mtmp[3] - mtmp[1] + 0x40000000 >> 31);
 151     out[2*stride].re = in[0].re - (mtmp[2] - mtmp[0] + 0x40000000 >> 31);
 152     out[2*stride].im = in[0].im - (mtmp[3] + mtmp[1] + 0x40000000 >> 31);
 153 #else
 154     tmp[0].re = TX_NAME(ff_cos_53)[0].re * tmp[0].re;
 155     tmp[0].im = TX_NAME(ff_cos_53)[0].im * tmp[0].im;
 156     tmp[1].re = TX_NAME(ff_cos_53)[1].re * tmp[1].re;
 157     tmp[1].im = TX_NAME(ff_cos_53)[1].re * tmp[1].im;
 158     out[1*stride].re = in[0].re - tmp[1].re + tmp[0].re;
 159     out[1*stride].im = in[0].im - tmp[1].im - tmp[0].im;
 160     out[2*stride].re = in[0].re - tmp[1].re - tmp[0].re;
 161     out[2*stride].im = in[0].im - tmp[1].im + tmp[0].im;
 162 #endif
 163 }
 164
 165 #define DECL_FFT5(NAME, D0, D1, D2, D3, D4)                                                       \
 166 static av_always_inline void NAME(FFTComplex *out, FFTComplex *in,                                \
 167                                   ptrdiff_t stride)                                               \
 168 {                                                                                                 \
 169     FFTComplex z0[4], t[6];                                                                       \
 170                                                                                                   \
 171     BF(t[1].im, t[0].re, in[1].re, in[4].re);                                                     \
 172     BF(t[1].re, t[0].im, in[1].im, in[4].im);                                                     \
 173     BF(t[3].im, t[2].re, in[2].re, in[3].re);                                                     \
 174     BF(t[3].re, t[2].im, in[2].im, in[3].im);                                                     \
 175                                                                                                   \
 176     out[D0*stride].re = in[0].re + t[0].re + t[2].re;                                             \
 177     out[D0*stride].im = in[0].im + t[0].im + t[2].im;                                             \
 178                                                                                                   \
 179     SMUL(t[4].re, t[0].re, TX_NAME(ff_cos_53)[2].re, TX_NAME(ff_cos_53)[3].re, t[2].re, t[0].re); \
 180     SMUL(t[4].im, t[0].im, TX_NAME(ff_cos_53)[2].re, TX_NAME(ff_cos_53)[3].re, t[2].im, t[0].im); \
 181     CMUL(t[5].re, t[1].re, TX_NAME(ff_cos_53)[2].im, TX_NAME(ff_cos_53)[3].im, t[3].re, t[1].re); \
 182     CMUL(t[5].im, t[1].im, TX_NAME(ff_cos_53)[2].im, TX_NAME(ff_cos_53)[3].im, t[3].im, t[1].im); \
 183                                                                                                   \
 184     BF(z0[0].re, z0[3].re, t[0].re, t[1].re);                                                     \
 185     BF(z0[0].im, z0[3].im, t[0].im, t[1].im);                                                     \
 186     BF(z0[2].re, z0[1].re, t[4].re, t[5].re);                                                     \
 187     BF(z0[2].im, z0[1].im, t[4].im, t[5].im);                                                     \
 188                                                                                                   \
 189     out[D1*stride].re = in[0].re + z0[3].re;                                                      \
 190     out[D1*stride].im = in[0].im + z0[0].im;                                                      \
 191     out[D2*stride].re = in[0].re + z0[2].re;                                                      \
 192     out[D2*stride].im = in[0].im + z0[1].im;                                                      \
 193     out[D3*stride].re = in[0].re + z0[1].re;                                                      \
 194     out[D3*stride].im = in[0].im + z0[2].im;                                                      \
 195     out[D4*stride].re = in[0].re + z0[0].re;                                                      \
 196     out[D4*stride].im = in[0].im + z0[3].im;                                                      \
 197 }
 198
 199 DECL_FFT5(fft5,     0,  1,  2,  3,  4)
 200 DECL_FFT5(fft5_m1,  0,  6, 12,  3,  9)
 201 DECL_FFT5(fft5_m2, 10,  1,  7, 13,  4)
 202 DECL_FFT5(fft5_m3,  5, 11,  2,  8, 14)
 203
 204 static av_always_inline void fft15(FFTComplex *out, FFTComplex *in,
 205                                    ptrdiff_t stride)
 206 {
 207     FFTComplex tmp[15];
 208
 209     for (int i = 0; i < 5; i++)
 210         fft3(tmp + i, in + i*3, 5);
 211
 212     fft5_m1(out, tmp +  0, stride);
 213     fft5_m2(out, tmp +  5, stride);
 214     fft5_m3(out, tmp + 10, stride);
 215 }
 216
 217 #define BUTTERFLIES(a0,a1,a2,a3) {\
 218     BF(t3, t5, t5, t1);\
 219     BF(a2.re, a0.re, a0.re, t5);\
 220     BF(a3.im, a1.im, a1.im, t3);\
 221     BF(t4, t6, t2, t6);\
 222     BF(a3.re, a1.re, a1.re, t4);\
 223     BF(a2.im, a0.im, a0.im, t6);\
 224 }
 225
 226 // force loading all the inputs before storing any.
 227 // this is slightly slower for small data, but avoids store->load aliasing
 228 // for addresses separated by large powers of 2.
 229 #define BUTTERFLIES_BIG(a0,a1,a2,a3) {\
 230     FFTSample r0=a0.re, i0=a0.im, r1=a1.re, i1=a1.im;\
 231     BF(t3, t5, t5, t1);\
 232     BF(a2.re, a0.re, r0, t5);\
 233     BF(a3.im, a1.im, i1, t3);\
 234     BF(t4, t6, t2, t6);\
 235     BF(a3.re, a1.re, r1, t4);\
 236     BF(a2.im, a0.im, i0, t6);\
 237 }
 238
 239 #define TRANSFORM(a0,a1,a2,a3,wre,wim) {\
 240     CMUL(t1, t2, a2.re, a2.im, wre, -wim);\
 241     CMUL(t5, t6, a3.re, a3.im, wre,  wim);\
 242     BUTTERFLIES(a0,a1,a2,a3)\
 243 }
 244
 245 #define TRANSFORM_ZERO(a0,a1,a2,a3) {\
 246     t1 = a2.re;\
 247     t2 = a2.im;\
 248     t5 = a3.re;\
 249     t6 = a3.im;\
 250     BUTTERFLIES(a0,a1,a2,a3)\
 251 }
 252
 253 /* z[0...8n-1], w[1...2n-1] */
 254 #define PASS(name)\
 255 static void name(FFTComplex *z, const FFTSample *wre, unsigned int n)\
 256 {\
 257     FFTSample t1, t2, t3, t4, t5, t6;\
 258     int o1 = 2*n;\
 259     int o2 = 4*n;\
 260     int o3 = 6*n;\
 261     const FFTSample *wim = wre+o1;\
 262     n--;\
 263 \
 264     TRANSFORM_ZERO(z[0],z[o1],z[o2],z[o3]);\
 265     TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
 266     do {\
 267         z += 2;\
 268         wre += 2;\
 269         wim -= 2;\
 270         TRANSFORM(z[0],z[o1],z[o2],z[o3],wre[0],wim[0]);\
 271         TRANSFORM(z[1],z[o1+1],z[o2+1],z[o3+1],wre[1],wim[-1]);\
 272     } while(--n);\
 273 }
 274
 275 PASS(pass)
 276 #undef BUTTERFLIES
 277 #define BUTTERFLIES BUTTERFLIES_BIG
 278 PASS(pass_big)
 279
 280 #define DECL_FFT(n,n2,n4)\
 281 static void fft##n(FFTComplex *z)\
 282 {\
 283     fft##n2(z);\
 284     fft##n4(z+n4*2);\
 285     fft##n4(z+n4*3);\
 286     pass(z,TX_NAME(ff_cos_##n),n4/2);\
 287 }
 288
 289 static void fft2(FFTComplex *z)
 290 {
 291     FFTComplex tmp;
 292     BF(tmp.re, z[0].re, z[0].re, z[1].re);
 293     BF(tmp.im, z[0].im, z[0].im, z[1].im);
 294     z[1] = tmp;
 295 }
 296
 297 static void fft4(FFTComplex *z)
 298 {
 299     FFTSample t1, t2, t3, t4, t5, t6, t7, t8;
 300
 301     BF(t3, t1, z[0].re, z[1].re);
 302     BF(t8, t6, z[3].re, z[2].re);
 303     BF(z[2].re, z[0].re, t1, t6);
 304     BF(t4, t2, z[0].im, z[1].im);
 305     BF(t7, t5, z[2].im, z[3].im);
 306     BF(z[3].im, z[1].im, t4, t8);
 307     BF(z[3].re, z[1].re, t3, t7);
 308     BF(z[2].im, z[0].im, t2, t5);
 309 }
 310
 311 static void fft8(FFTComplex *z)
 312 {
 313     FFTSample t1, t2, t3, t4, t5, t6;
 314
 315     fft4(z);
 316
 317     BF(t1, z[5].re, z[4].re, -z[5].re);
 318     BF(t2, z[5].im, z[4].im, -z[5].im);
 319     BF(t5, z[7].re, z[6].re, -z[7].re);
 320     BF(t6, z[7].im, z[6].im, -z[7].im);
 321
 322     BUTTERFLIES(z[0],z[2],z[4],z[6]);
 323     TRANSFORM(z[1],z[3],z[5],z[7],RESCALE(M_SQRT1_2),RESCALE(M_SQRT1_2));
 324 }
 325
 326 static void fft16(FFTComplex *z)
 327 {
 328     FFTSample t1, t2, t3, t4, t5, t6;
 329     FFTSample cos_16_1 = TX_NAME(ff_cos_16)[1];
 330     FFTSample cos_16_3 = TX_NAME(ff_cos_16)[3];
 331
 332     fft8(z);
 333     fft4(z+8);
 334     fft4(z+12);
 335
 336     TRANSFORM_ZERO(z[0],z[4],z[8],z[12]);
 337     TRANSFORM(z[2],z[6],z[10],z[14],RESCALE(M_SQRT1_2),RESCALE(M_SQRT1_2));
 338     TRANSFORM(z[1],z[5],z[9],z[13],cos_16_1,cos_16_3);
 339     TRANSFORM(z[3],z[7],z[11],z[15],cos_16_3,cos_16_1);
 340 }
 341
 342 DECL_FFT(32,16,8)
 343 DECL_FFT(64,32,16)
 344 DECL_FFT(128,64,32)
 345 DECL_FFT(256,128,64)
 346 DECL_FFT(512,256,128)
 347 #define pass pass_big
 348 DECL_FFT(1024,512,256)
 349 DECL_FFT(2048,1024,512)
 350 DECL_FFT(4096,2048,1024)
 351 DECL_FFT(8192,4096,2048)
 352 DECL_FFT(16384,8192,4096)
 353 DECL_FFT(32768,16384,8192)
 354 DECL_FFT(65536,32768,16384)
 355 DECL_FFT(131072,65536,32768)
 356
 357 static void (* const fft_dispatch[])(FFTComplex*) = {
 358     NULL, fft2, fft4, fft8, fft16, fft32, fft64, fft128, fft256, fft512,
 359     fft1024, fft2048, fft4096, fft8192, fft16384, fft32768, fft65536, fft131072
 360 };
 361
 362 #define DECL_COMP_FFT(N)                                                       \
 363 static void compound_fft_##N##xM(AVTXContext *s, void *_out,                   \
 364                                  void *_in, ptrdiff_t stride)                  \
 365 {                                                                              \
 366     const int m = s->m, *in_map = s->pfatab, *out_map = in_map + N*m;          \
 367     FFTComplex *in = _in;                                                      \
 368     FFTComplex *out = _out;                                                    \
 369     FFTComplex fft##N##in[N];                                                  \
 370     void (*fftp)(FFTComplex *z) = fft_dispatch[av_log2(m)];                    \
 371                                                                                \
 372     for (int i = 0; i < m; i++) {                                              \
 373         for (int j = 0; j < N; j++)                                            \
 374             fft##N##in[j] = in[in_map[i*N + j]];                               \
 375         fft##N(s->tmp + s->revtab[i], fft##N##in, m);                          \
 376     }                                                                          \
 377                                                                                \
 378     for (int i = 0; i < N; i++)                                                \
 379         fftp(s->tmp + m*i);                                                    \
 380                                                                                \
 381     for (int i = 0; i < N*m; i++)                                              \
 382         out[i] = s->tmp[out_map[i]];                                           \
 383 }
 384
 385 DECL_COMP_FFT(3)
 386 DECL_COMP_FFT(5)
 387 DECL_COMP_FFT(15)
 388
 389 static void monolithic_fft(AVTXContext *s, void *_out, void *_in,
 390                            ptrdiff_t stride)
 391 {
 392     FFTComplex *in = _in;
 393     FFTComplex *out = _out;
 394     int m = s->m, mb = av_log2(m);
 395     for (int i = 0; i < m; i++)
 396         out[s->revtab[i]] = in[i];
 397     fft_dispatch[mb](out);
 398 }
 399
 400 #define DECL_COMP_IMDCT(N)                                                     \
 401 static void compound_imdct_##N##xM(AVTXContext *s, void *_dst, void *_src,     \
 402                                    ptrdiff_t stride)                           \
 403 {                                                                              \
 404     FFTComplex fft##N##in[N];                                                  \
 405     FFTComplex *z = _dst, *exp = s->exptab;                                    \
 406     const int m = s->m, len8 = N*m >> 1;                                       \
 407     const int *in_map = s->pfatab, *out_map = in_map + N*m;                    \
 408     const FFTSample *src = _src, *in1, *in2;                                   \
 409     void (*fftp)(FFTComplex *) = fft_dispatch[av_log2(m)];                     \
 410                                                                                \
 411     stride /= sizeof(*src); /* To convert it from bytes */                     \
 412     in1 = src;                                                                 \
 413     in2 = src + ((N*m*2) - 1) * stride;                                        \
 414                                                                                \
 415     for (int i = 0; i < m; i++) {                                              \
 416         for (int j = 0; j < N; j++) {                                          \
 417             const int k = in_map[i*N + j];                                     \
 418             FFTComplex tmp = { in2[-k*stride], in1[k*stride] };                \
 419             CMUL3(fft##N##in[j], tmp, exp[k >> 1]);                            \
 420         }                                                                      \
 421         fft##N(s->tmp + s->revtab[i], fft##N##in, m);                          \
 422     }                                                                          \
 423                                                                                \
 424     for (int i = 0; i < N; i++)                                                \
 425         fftp(s->tmp + m*i);                                                    \
 426                                                                                \
 427     for (int i = 0; i < len8; i++) {                                           \
 428         const int i0 = len8 + i, i1 = len8 - i - 1;                            \
 429         const int s0 = out_map[i0], s1 = out_map[i1];                          \
 430         FFTComplex src1 = { s->tmp[s1].im, s->tmp[s1].re };                    \
 431         FFTComplex src0 = { s->tmp[s0].im, s->tmp[s0].re };                    \
 432                                                                                \
 433         CMUL(z[i1].re, z[i0].im, src1.re, src1.im, exp[i1].im, exp[i1].re);    \
 434         CMUL(z[i0].re, z[i1].im, src0.re, src0.im, exp[i0].im, exp[i0].re);    \
 435     }                                                                          \
 436 }
 437
 438 DECL_COMP_IMDCT(3)
 439 DECL_COMP_IMDCT(5)
 440 DECL_COMP_IMDCT(15)
 441
 442 #define DECL_COMP_MDCT(N)                                                      \
 443 static void compound_mdct_##N##xM(AVTXContext *s, void *_dst, void *_src,      \
 444                                   ptrdiff_t stride)                            \
 445 {                                                                              \
 446     FFTSample *src = _src, *dst = _dst;                                        \
 447     FFTComplex *exp = s->exptab, tmp, fft##N##in[N];                           \
 448     const int m = s->m, len4 = N*m, len3 = len4 * 3, len8 = len4 >> 1;         \
 449     const int *in_map = s->pfatab, *out_map = in_map + N*m;                    \
 450     void (*fftp)(FFTComplex *) = fft_dispatch[av_log2(m)];                     \
 451                                                                                \
 452     stride /= sizeof(*dst);                                                    \
 453                                                                                \
 454     for (int i = 0; i < m; i++) { /* Folding and pre-reindexing */             \
 455         for (int j = 0; j < N; j++) {                                          \
 456             const int k = in_map[i*N + j];                                     \
 457             if (k < len4) {                                                    \
 458                 tmp.re = FOLD(-src[ len4 + k],  src[1*len4 - 1 - k]);          \
 459                 tmp.im = FOLD(-src[ len3 + k], -src[1*len3 - 1 - k]);          \
 460             } else {                                                           \
 461                 tmp.re = FOLD(-src[ len4 + k], -src[5*len4 - 1 - k]);          \
 462                 tmp.im = FOLD( src[-len4 + k], -src[1*len3 - 1 - k]);          \
 463             }                                                                  \
 464             CMUL(fft##N##in[j].im, fft##N##in[j].re, tmp.re, tmp.im,           \
 465                  exp[k >> 1].re, exp[k >> 1].im);                              \
 466         }                                                                      \
 467         fft##N(s->tmp + s->revtab[i], fft##N##in, m);                          \
 468     }                                                                          \
 469                                                                                \
 470     for (int i = 0; i < N; i++)                                                \
 471         fftp(s->tmp + m*i);                                                    \
 472                                                                                \
 473     for (int i = 0; i < len8; i++) {                                           \
 474         const int i0 = len8 + i, i1 = len8 - i - 1;                            \
 475         const int s0 = out_map[i0], s1 = out_map[i1];                          \
 476         FFTComplex src1 = { s->tmp[s1].re, s->tmp[s1].im };                    \
 477         FFTComplex src0 = { s->tmp[s0].re, s->tmp[s0].im };                    \
 478                                                                                \
 479         CMUL(dst[2*i1*stride + stride], dst[2*i0*stride], src0.re, src0.im,    \
 480              exp[i0].im, exp[i0].re);                                          \
 481         CMUL(dst[2*i0*stride + stride], dst[2*i1*stride], src1.re, src1.im,    \
 482              exp[i1].im, exp[i1].re);                                          \
 483     }                                                                          \
 484 }
 485
 486 DECL_COMP_MDCT(3)
 487 DECL_COMP_MDCT(5)
 488 DECL_COMP_MDCT(15)
 489
 490 static void monolithic_imdct(AVTXContext *s, void *_dst, void *_src,
 491                              ptrdiff_t stride)
 492 {
 493     FFTComplex *z = _dst, *exp = s->exptab;
 494     const int m = s->m, len8 = m >> 1;
 495     const FFTSample *src = _src, *in1, *in2;
 496     void (*fftp)(FFTComplex *) = fft_dispatch[av_log2(m)];
 497
 498     stride /= sizeof(*src);
 499     in1 = src;
 500     in2 = src + ((m*2) - 1) * stride;
 501
 502     for (int i = 0; i < m; i++) {
 503         FFTComplex tmp = { in2[-2*i*stride], in1[2*i*stride] };
 504         CMUL3(z[s->revtab[i]], tmp, exp[i]);
 505     }
 506
 507     fftp(z);
 508
 509     for (int i = 0; i < len8; i++) {
 510         const int i0 = len8 + i, i1 = len8 - i - 1;
 511         FFTComplex src1 = { z[i1].im, z[i1].re };
 512         FFTComplex src0 = { z[i0].im, z[i0].re };
 513
 514         CMUL(z[i1].re, z[i0].im, src1.re, src1.im, exp[i1].im, exp[i1].re);
 515         CMUL(z[i0].re, z[i1].im, src0.re, src0.im, exp[i0].im, exp[i0].re);
 516     }
 517 }
 518
 519 static void monolithic_mdct(AVTXContext *s, void *_dst, void *_src,
 520                             ptrdiff_t stride)
 521 {
 522     FFTSample *src = _src, *dst = _dst;
 523     FFTComplex *exp = s->exptab, tmp, *z = _dst;
 524     const int m = s->m, len4 = m, len3 = len4 * 3, len8 = len4 >> 1;
 525     void (*fftp)(FFTComplex *) = fft_dispatch[av_log2(m)];
 526
 527     stride /= sizeof(*dst);
 528
 529     for (int i = 0; i < m; i++) { /* Folding and pre-reindexing */
 530         const int k = 2*i;
 531         if (k < len4) {
 532             tmp.re = FOLD(-src[ len4 + k],  src[1*len4 - 1 - k]);
 533             tmp.im = FOLD(-src[ len3 + k], -src[1*len3 - 1 - k]);
 534         } else {
 535             tmp.re = FOLD(-src[ len4 + k], -src[5*len4 - 1 - k]);
 536             tmp.im = FOLD( src[-len4 + k], -src[1*len3 - 1 - k]);
 537         }
 538         CMUL(z[s->revtab[i]].im, z[s->revtab[i]].re, tmp.re, tmp.im,
 539              exp[i].re, exp[i].im);
 540     }
 541
 542     fftp(z);
 543
 544     for (int i = 0; i < len8; i++) {
 545         const int i0 = len8 + i, i1 = len8 - i - 1;
 546         FFTComplex src1 = { z[i1].re, z[i1].im };
 547         FFTComplex src0 = { z[i0].re, z[i0].im };
 548
 549         CMUL(dst[2*i1*stride + stride], dst[2*i0*stride], src0.re, src0.im,
 550              exp[i0].im, exp[i0].re);
 551         CMUL(dst[2*i0*stride + stride], dst[2*i1*stride], src1.re, src1.im,
 552              exp[i1].im, exp[i1].re);
 553     }
 554 }
 555
 556 static int gen_mdct_exptab(AVTXContext *s, int len4, double scale)
 557 {
 558     const double theta = (scale < 0 ? len4 : 0) + 1.0/8.0;
 559
 560     if (!(s->exptab = av_malloc_array(len4, sizeof(*s->exptab))))
 561         return AVERROR(ENOMEM);
 562
 563     scale = sqrt(fabs(scale));
 564     for (int i = 0; i < len4; i++) {
 565         const double alpha = M_PI_2 * (i + theta) / len4;
 566         s->exptab[i].re = RESCALE(cos(alpha) * scale);
 567         s->exptab[i].im = RESCALE(sin(alpha) * scale);
 568     }
 569
 570     return 0;
 571 }
 572
 573 int TX_NAME(ff_tx_init_mdct_fft)(AVTXContext *s, av_tx_fn *tx,
 574                                  enum AVTXType type, int inv, int len,
 575                                  const void *scale, uint64_t flags)
 576 {
 577     const int is_mdct = ff_tx_type_is_mdct(type);
 578     int err, n = 1, m = 1, max_ptwo = 1 << (FF_ARRAY_ELEMS(fft_dispatch) - 1);
 579
 580     if (is_mdct)
 581         len >>= 1;
 582
 583 #define CHECK_FACTOR(DST, FACTOR, SRC)                                         \
 584     if (DST == 1 && !(SRC % FACTOR)) {                                         \
 585         DST = FACTOR;                                                          \
 586         SRC /= FACTOR;                                                         \
 587     }
 588     CHECK_FACTOR(n, 15, len)
 589     CHECK_FACTOR(n,  5, len)
 590     CHECK_FACTOR(n,  3, len)
 591 #undef CHECK_FACTOR
 592
 593     /* len must be a power of two now */
 594     if (!(len & (len - 1)) && len >= 2 && len <= max_ptwo) {
 595         m = len;
 596         len = 1;
 597     }
 598
 599     s->n = n;
 600     s->m = m;
 601     s->inv = inv;
 602     s->type = type;
 603
 604     /* Filter out direct 3, 5 and 15 transforms, too niche */
 605     if (len > 1 || m == 1) {
 606         av_log(NULL, AV_LOG_ERROR, "Unsupported transform size: n = %i, "
 607                "m = %i, residual = %i!\n", n, m, len);
 608         return AVERROR(EINVAL);
 609     } else if (n > 1 && m > 1) { /* 2D transform case */
 610         if ((err = ff_tx_gen_compound_mapping(s)))
 611             return err;
 612         if (!(s->tmp = av_malloc(n*m*sizeof(*s->tmp))))
 613             return AVERROR(ENOMEM);
 614         *tx = n == 3 ? compound_fft_3xM :
 615               n == 5 ? compound_fft_5xM :
 616                        compound_fft_15xM;
 617         if (is_mdct)
 618             *tx = n == 3 ? inv ? compound_imdct_3xM  : compound_mdct_3xM :
 619                   n == 5 ? inv ? compound_imdct_5xM  : compound_mdct_5xM :
 620                            inv ? compound_imdct_15xM : compound_mdct_15xM;
 621     } else { /* Direct transform case */
 622         *tx = monolithic_fft;
 623         if (is_mdct)
 624             *tx = inv ? monolithic_imdct : monolithic_mdct;
 625     }
 626
 627     if (n != 1)
 628         init_cos_tabs(0);
 629     if (m != 1) {
 630         ff_tx_gen_ptwo_revtab(s);
 631         for (int i = 4; i <= av_log2(m); i++)
 632             init_cos_tabs(i);
 633     }
 634
 635     if (is_mdct)
 636         return gen_mdct_exptab(s, n*m, *((SCALE_TYPE *)scale));
 637
 638     return 0;
 639 }