Alexandre Lision | 744f742 | 2013-09-25 11:39:37 -0400 | [diff] [blame] | 1 | /* Copyright (c) 2007-2008 CSIRO |
| 2 | Copyright (c) 2007-2009 Xiph.Org Foundation |
| 3 | Copyright (c) 2008-2009 Gregory Maxwell |
| 4 | Written by Jean-Marc Valin and Gregory Maxwell */ |
| 5 | /* |
| 6 | Redistribution and use in source and binary forms, with or without |
| 7 | modification, are permitted provided that the following conditions |
| 8 | are met: |
| 9 | |
| 10 | - Redistributions of source code must retain the above copyright |
| 11 | notice, this list of conditions and the following disclaimer. |
| 12 | |
| 13 | - Redistributions in binary form must reproduce the above copyright |
| 14 | notice, this list of conditions and the following disclaimer in the |
| 15 | documentation and/or other materials provided with the distribution. |
| 16 | |
| 17 | THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 18 | ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 19 | LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR |
| 20 | A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER |
| 21 | OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, |
| 22 | EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, |
| 23 | PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR |
| 24 | PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF |
| 25 | LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING |
| 26 | NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS |
| 27 | SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 28 | */ |
| 29 | |
| 30 | #ifdef HAVE_CONFIG_H |
| 31 | #include "config.h" |
| 32 | #endif |
| 33 | |
| 34 | #include <math.h> |
| 35 | #include "bands.h" |
| 36 | #include "modes.h" |
| 37 | #include "vq.h" |
| 38 | #include "cwrs.h" |
| 39 | #include "stack_alloc.h" |
| 40 | #include "os_support.h" |
| 41 | #include "mathops.h" |
| 42 | #include "rate.h" |
| 43 | |
| 44 | opus_uint32 celt_lcg_rand(opus_uint32 seed) |
| 45 | { |
| 46 | return 1664525 * seed + 1013904223; |
| 47 | } |
| 48 | |
| 49 | /* This is a cos() approximation designed to be bit-exact on any platform. Bit exactness |
| 50 | with this approximation is important because it has an impact on the bit allocation */ |
| 51 | static opus_int16 bitexact_cos(opus_int16 x) |
| 52 | { |
| 53 | opus_int32 tmp; |
| 54 | opus_int16 x2; |
| 55 | tmp = (4096+((opus_int32)(x)*(x)))>>13; |
| 56 | celt_assert(tmp<=32767); |
| 57 | x2 = tmp; |
| 58 | x2 = (32767-x2) + FRAC_MUL16(x2, (-7651 + FRAC_MUL16(x2, (8277 + FRAC_MUL16(-626, x2))))); |
| 59 | celt_assert(x2<=32766); |
| 60 | return 1+x2; |
| 61 | } |
| 62 | |
| 63 | static int bitexact_log2tan(int isin,int icos) |
| 64 | { |
| 65 | int lc; |
| 66 | int ls; |
| 67 | lc=EC_ILOG(icos); |
| 68 | ls=EC_ILOG(isin); |
| 69 | icos<<=15-lc; |
| 70 | isin<<=15-ls; |
| 71 | return (ls-lc)*(1<<11) |
| 72 | +FRAC_MUL16(isin, FRAC_MUL16(isin, -2597) + 7932) |
| 73 | -FRAC_MUL16(icos, FRAC_MUL16(icos, -2597) + 7932); |
| 74 | } |
| 75 | |
| 76 | #ifdef FIXED_POINT |
| 77 | /* Compute the amplitude (sqrt energy) in each of the bands */ |
| 78 | void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int M) |
| 79 | { |
| 80 | int i, c, N; |
| 81 | const opus_int16 *eBands = m->eBands; |
| 82 | N = M*m->shortMdctSize; |
| 83 | c=0; do { |
| 84 | for (i=0;i<end;i++) |
| 85 | { |
| 86 | int j; |
| 87 | opus_val32 maxval=0; |
| 88 | opus_val32 sum = 0; |
| 89 | |
| 90 | j=M*eBands[i]; do { |
| 91 | maxval = MAX32(maxval, X[j+c*N]); |
| 92 | maxval = MAX32(maxval, -X[j+c*N]); |
| 93 | } while (++j<M*eBands[i+1]); |
| 94 | |
| 95 | if (maxval > 0) |
| 96 | { |
| 97 | int shift = celt_ilog2(maxval)-10; |
| 98 | j=M*eBands[i]; do { |
| 99 | sum = MAC16_16(sum, EXTRACT16(VSHR32(X[j+c*N],shift)), |
| 100 | EXTRACT16(VSHR32(X[j+c*N],shift))); |
| 101 | } while (++j<M*eBands[i+1]); |
| 102 | /* We're adding one here to ensure the normalized band isn't larger than unity norm */ |
| 103 | bandE[i+c*m->nbEBands] = EPSILON+VSHR32(EXTEND32(celt_sqrt(sum)),-shift); |
| 104 | } else { |
| 105 | bandE[i+c*m->nbEBands] = EPSILON; |
| 106 | } |
| 107 | /*printf ("%f ", bandE[i+c*m->nbEBands]);*/ |
| 108 | } |
| 109 | } while (++c<C); |
| 110 | /*printf ("\n");*/ |
| 111 | } |
| 112 | |
| 113 | /* Normalise each band such that the energy is one. */ |
| 114 | void normalise_bands(const CELTMode *m, const celt_sig * OPUS_RESTRICT freq, celt_norm * OPUS_RESTRICT X, const celt_ener *bandE, int end, int C, int M) |
| 115 | { |
| 116 | int i, c, N; |
| 117 | const opus_int16 *eBands = m->eBands; |
| 118 | N = M*m->shortMdctSize; |
| 119 | c=0; do { |
| 120 | i=0; do { |
| 121 | opus_val16 g; |
| 122 | int j,shift; |
| 123 | opus_val16 E; |
| 124 | shift = celt_zlog2(bandE[i+c*m->nbEBands])-13; |
| 125 | E = VSHR32(bandE[i+c*m->nbEBands], shift); |
| 126 | g = EXTRACT16(celt_rcp(SHL32(E,3))); |
| 127 | j=M*eBands[i]; do { |
| 128 | X[j+c*N] = MULT16_16_Q15(VSHR32(freq[j+c*N],shift-1),g); |
| 129 | } while (++j<M*eBands[i+1]); |
| 130 | } while (++i<end); |
| 131 | } while (++c<C); |
| 132 | } |
| 133 | |
| 134 | #else /* FIXED_POINT */ |
| 135 | /* Compute the amplitude (sqrt energy) in each of the bands */ |
| 136 | void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int M) |
| 137 | { |
| 138 | int i, c, N; |
| 139 | const opus_int16 *eBands = m->eBands; |
| 140 | N = M*m->shortMdctSize; |
| 141 | c=0; do { |
| 142 | for (i=0;i<end;i++) |
| 143 | { |
| 144 | int j; |
| 145 | opus_val32 sum = 1e-27f; |
| 146 | for (j=M*eBands[i];j<M*eBands[i+1];j++) |
| 147 | sum += X[j+c*N]*X[j+c*N]; |
| 148 | bandE[i+c*m->nbEBands] = celt_sqrt(sum); |
| 149 | /*printf ("%f ", bandE[i+c*m->nbEBands]);*/ |
| 150 | } |
| 151 | } while (++c<C); |
| 152 | /*printf ("\n");*/ |
| 153 | } |
| 154 | |
| 155 | /* Normalise each band such that the energy is one. */ |
| 156 | void normalise_bands(const CELTMode *m, const celt_sig * OPUS_RESTRICT freq, celt_norm * OPUS_RESTRICT X, const celt_ener *bandE, int end, int C, int M) |
| 157 | { |
| 158 | int i, c, N; |
| 159 | const opus_int16 *eBands = m->eBands; |
| 160 | N = M*m->shortMdctSize; |
| 161 | c=0; do { |
| 162 | for (i=0;i<end;i++) |
| 163 | { |
| 164 | int j; |
| 165 | opus_val16 g = 1.f/(1e-27f+bandE[i+c*m->nbEBands]); |
| 166 | for (j=M*eBands[i];j<M*eBands[i+1];j++) |
| 167 | X[j+c*N] = freq[j+c*N]*g; |
| 168 | } |
| 169 | } while (++c<C); |
| 170 | } |
| 171 | |
| 172 | #endif /* FIXED_POINT */ |
| 173 | |
| 174 | /* De-normalise the energy to produce the synthesis from the unit-energy bands */ |
| 175 | void denormalise_bands(const CELTMode *m, const celt_norm * OPUS_RESTRICT X, celt_sig * OPUS_RESTRICT freq, const celt_ener *bandE, int end, int C, int M) |
| 176 | { |
| 177 | int i, c, N; |
| 178 | const opus_int16 *eBands = m->eBands; |
| 179 | N = M*m->shortMdctSize; |
| 180 | celt_assert2(C<=2, "denormalise_bands() not implemented for >2 channels"); |
| 181 | c=0; do { |
| 182 | celt_sig * OPUS_RESTRICT f; |
| 183 | const celt_norm * OPUS_RESTRICT x; |
| 184 | f = freq+c*N; |
| 185 | x = X+c*N; |
| 186 | for (i=0;i<end;i++) |
| 187 | { |
| 188 | int j, band_end; |
| 189 | opus_val32 g = SHR32(bandE[i+c*m->nbEBands],1); |
| 190 | j=M*eBands[i]; |
| 191 | band_end = M*eBands[i+1]; |
| 192 | do { |
| 193 | *f++ = SHL32(MULT16_32_Q15(*x, g),2); |
| 194 | x++; |
| 195 | } while (++j<band_end); |
| 196 | } |
| 197 | for (i=M*eBands[end];i<N;i++) |
| 198 | *f++ = 0; |
| 199 | } while (++c<C); |
| 200 | } |
| 201 | |
| 202 | /* This prevents energy collapse for transients with multiple short MDCTs */ |
| 203 | void anti_collapse(const CELTMode *m, celt_norm *X_, unsigned char *collapse_masks, int LM, int C, int size, |
| 204 | int start, int end, opus_val16 *logE, opus_val16 *prev1logE, |
| 205 | opus_val16 *prev2logE, int *pulses, opus_uint32 seed) |
| 206 | { |
| 207 | int c, i, j, k; |
| 208 | for (i=start;i<end;i++) |
| 209 | { |
| 210 | int N0; |
| 211 | opus_val16 thresh, sqrt_1; |
| 212 | int depth; |
| 213 | #ifdef FIXED_POINT |
| 214 | int shift; |
| 215 | opus_val32 thresh32; |
| 216 | #endif |
| 217 | |
| 218 | N0 = m->eBands[i+1]-m->eBands[i]; |
| 219 | /* depth in 1/8 bits */ |
| 220 | depth = (1+pulses[i])/((m->eBands[i+1]-m->eBands[i])<<LM); |
| 221 | |
| 222 | #ifdef FIXED_POINT |
| 223 | thresh32 = SHR32(celt_exp2(-SHL16(depth, 10-BITRES)),1); |
| 224 | thresh = MULT16_32_Q15(QCONST16(0.5f, 15), MIN32(32767,thresh32)); |
| 225 | { |
| 226 | opus_val32 t; |
| 227 | t = N0<<LM; |
| 228 | shift = celt_ilog2(t)>>1; |
| 229 | t = SHL32(t, (7-shift)<<1); |
| 230 | sqrt_1 = celt_rsqrt_norm(t); |
| 231 | } |
| 232 | #else |
| 233 | thresh = .5f*celt_exp2(-.125f*depth); |
| 234 | sqrt_1 = celt_rsqrt(N0<<LM); |
| 235 | #endif |
| 236 | |
| 237 | c=0; do |
| 238 | { |
| 239 | celt_norm *X; |
| 240 | opus_val16 prev1; |
| 241 | opus_val16 prev2; |
| 242 | opus_val32 Ediff; |
| 243 | opus_val16 r; |
| 244 | int renormalize=0; |
| 245 | prev1 = prev1logE[c*m->nbEBands+i]; |
| 246 | prev2 = prev2logE[c*m->nbEBands+i]; |
| 247 | if (C==1) |
| 248 | { |
| 249 | prev1 = MAX16(prev1,prev1logE[m->nbEBands+i]); |
| 250 | prev2 = MAX16(prev2,prev2logE[m->nbEBands+i]); |
| 251 | } |
| 252 | Ediff = EXTEND32(logE[c*m->nbEBands+i])-EXTEND32(MIN16(prev1,prev2)); |
| 253 | Ediff = MAX32(0, Ediff); |
| 254 | |
| 255 | #ifdef FIXED_POINT |
| 256 | if (Ediff < 16384) |
| 257 | { |
| 258 | opus_val32 r32 = SHR32(celt_exp2(-EXTRACT16(Ediff)),1); |
| 259 | r = 2*MIN16(16383,r32); |
| 260 | } else { |
| 261 | r = 0; |
| 262 | } |
| 263 | if (LM==3) |
| 264 | r = MULT16_16_Q14(23170, MIN32(23169, r)); |
| 265 | r = SHR16(MIN16(thresh, r),1); |
| 266 | r = SHR32(MULT16_16_Q15(sqrt_1, r),shift); |
| 267 | #else |
| 268 | /* r needs to be multiplied by 2 or 2*sqrt(2) depending on LM because |
| 269 | short blocks don't have the same energy as long */ |
| 270 | r = 2.f*celt_exp2(-Ediff); |
| 271 | if (LM==3) |
| 272 | r *= 1.41421356f; |
| 273 | r = MIN16(thresh, r); |
| 274 | r = r*sqrt_1; |
| 275 | #endif |
| 276 | X = X_+c*size+(m->eBands[i]<<LM); |
| 277 | for (k=0;k<1<<LM;k++) |
| 278 | { |
| 279 | /* Detect collapse */ |
| 280 | if (!(collapse_masks[i*C+c]&1<<k)) |
| 281 | { |
| 282 | /* Fill with noise */ |
| 283 | for (j=0;j<N0;j++) |
| 284 | { |
| 285 | seed = celt_lcg_rand(seed); |
| 286 | X[(j<<LM)+k] = (seed&0x8000 ? r : -r); |
| 287 | } |
| 288 | renormalize = 1; |
| 289 | } |
| 290 | } |
| 291 | /* We just added some energy, so we need to renormalise */ |
| 292 | if (renormalize) |
| 293 | renormalise_vector(X, N0<<LM, Q15ONE); |
| 294 | } while (++c<C); |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | static void intensity_stereo(const CELTMode *m, celt_norm *X, celt_norm *Y, const celt_ener *bandE, int bandID, int N) |
| 299 | { |
| 300 | int i = bandID; |
| 301 | int j; |
| 302 | opus_val16 a1, a2; |
| 303 | opus_val16 left, right; |
| 304 | opus_val16 norm; |
| 305 | #ifdef FIXED_POINT |
| 306 | int shift = celt_zlog2(MAX32(bandE[i], bandE[i+m->nbEBands]))-13; |
| 307 | #endif |
| 308 | left = VSHR32(bandE[i],shift); |
| 309 | right = VSHR32(bandE[i+m->nbEBands],shift); |
| 310 | norm = EPSILON + celt_sqrt(EPSILON+MULT16_16(left,left)+MULT16_16(right,right)); |
| 311 | a1 = DIV32_16(SHL32(EXTEND32(left),14),norm); |
| 312 | a2 = DIV32_16(SHL32(EXTEND32(right),14),norm); |
| 313 | for (j=0;j<N;j++) |
| 314 | { |
| 315 | celt_norm r, l; |
| 316 | l = X[j]; |
| 317 | r = Y[j]; |
| 318 | X[j] = MULT16_16_Q14(a1,l) + MULT16_16_Q14(a2,r); |
| 319 | /* Side is not encoded, no need to calculate */ |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | static void stereo_split(celt_norm *X, celt_norm *Y, int N) |
| 324 | { |
| 325 | int j; |
| 326 | for (j=0;j<N;j++) |
| 327 | { |
| 328 | celt_norm r, l; |
| 329 | l = MULT16_16_Q15(QCONST16(.70710678f,15), X[j]); |
| 330 | r = MULT16_16_Q15(QCONST16(.70710678f,15), Y[j]); |
| 331 | X[j] = l+r; |
| 332 | Y[j] = r-l; |
| 333 | } |
| 334 | } |
| 335 | |
| 336 | static void stereo_merge(celt_norm *X, celt_norm *Y, opus_val16 mid, int N) |
| 337 | { |
| 338 | int j; |
| 339 | opus_val32 xp=0, side=0; |
| 340 | opus_val32 El, Er; |
| 341 | opus_val16 mid2; |
| 342 | #ifdef FIXED_POINT |
| 343 | int kl, kr; |
| 344 | #endif |
| 345 | opus_val32 t, lgain, rgain; |
| 346 | |
| 347 | /* Compute the norm of X+Y and X-Y as |X|^2 + |Y|^2 +/- sum(xy) */ |
| 348 | for (j=0;j<N;j++) |
| 349 | { |
| 350 | xp = MAC16_16(xp, X[j], Y[j]); |
| 351 | side = MAC16_16(side, Y[j], Y[j]); |
| 352 | } |
| 353 | /* Compensating for the mid normalization */ |
| 354 | xp = MULT16_32_Q15(mid, xp); |
| 355 | /* mid and side are in Q15, not Q14 like X and Y */ |
| 356 | mid2 = SHR32(mid, 1); |
| 357 | El = MULT16_16(mid2, mid2) + side - 2*xp; |
| 358 | Er = MULT16_16(mid2, mid2) + side + 2*xp; |
| 359 | if (Er < QCONST32(6e-4f, 28) || El < QCONST32(6e-4f, 28)) |
| 360 | { |
| 361 | for (j=0;j<N;j++) |
| 362 | Y[j] = X[j]; |
| 363 | return; |
| 364 | } |
| 365 | |
| 366 | #ifdef FIXED_POINT |
| 367 | kl = celt_ilog2(El)>>1; |
| 368 | kr = celt_ilog2(Er)>>1; |
| 369 | #endif |
| 370 | t = VSHR32(El, (kl-7)<<1); |
| 371 | lgain = celt_rsqrt_norm(t); |
| 372 | t = VSHR32(Er, (kr-7)<<1); |
| 373 | rgain = celt_rsqrt_norm(t); |
| 374 | |
| 375 | #ifdef FIXED_POINT |
| 376 | if (kl < 7) |
| 377 | kl = 7; |
| 378 | if (kr < 7) |
| 379 | kr = 7; |
| 380 | #endif |
| 381 | |
| 382 | for (j=0;j<N;j++) |
| 383 | { |
| 384 | celt_norm r, l; |
| 385 | /* Apply mid scaling (side is already scaled) */ |
| 386 | l = MULT16_16_Q15(mid, X[j]); |
| 387 | r = Y[j]; |
| 388 | X[j] = EXTRACT16(PSHR32(MULT16_16(lgain, SUB16(l,r)), kl+1)); |
| 389 | Y[j] = EXTRACT16(PSHR32(MULT16_16(rgain, ADD16(l,r)), kr+1)); |
| 390 | } |
| 391 | } |
| 392 | |
| 393 | /* Decide whether we should spread the pulses in the current frame */ |
| 394 | int spreading_decision(const CELTMode *m, celt_norm *X, int *average, |
| 395 | int last_decision, int *hf_average, int *tapset_decision, int update_hf, |
| 396 | int end, int C, int M) |
| 397 | { |
| 398 | int i, c, N0; |
| 399 | int sum = 0, nbBands=0; |
| 400 | const opus_int16 * OPUS_RESTRICT eBands = m->eBands; |
| 401 | int decision; |
| 402 | int hf_sum=0; |
| 403 | |
| 404 | celt_assert(end>0); |
| 405 | |
| 406 | N0 = M*m->shortMdctSize; |
| 407 | |
| 408 | if (M*(eBands[end]-eBands[end-1]) <= 8) |
| 409 | return SPREAD_NONE; |
| 410 | c=0; do { |
| 411 | for (i=0;i<end;i++) |
| 412 | { |
| 413 | int j, N, tmp=0; |
| 414 | int tcount[3] = {0,0,0}; |
| 415 | celt_norm * OPUS_RESTRICT x = X+M*eBands[i]+c*N0; |
| 416 | N = M*(eBands[i+1]-eBands[i]); |
| 417 | if (N<=8) |
| 418 | continue; |
| 419 | /* Compute rough CDF of |x[j]| */ |
| 420 | for (j=0;j<N;j++) |
| 421 | { |
| 422 | opus_val32 x2N; /* Q13 */ |
| 423 | |
| 424 | x2N = MULT16_16(MULT16_16_Q15(x[j], x[j]), N); |
| 425 | if (x2N < QCONST16(0.25f,13)) |
| 426 | tcount[0]++; |
| 427 | if (x2N < QCONST16(0.0625f,13)) |
| 428 | tcount[1]++; |
| 429 | if (x2N < QCONST16(0.015625f,13)) |
| 430 | tcount[2]++; |
| 431 | } |
| 432 | |
| 433 | /* Only include four last bands (8 kHz and up) */ |
| 434 | if (i>m->nbEBands-4) |
| 435 | hf_sum += 32*(tcount[1]+tcount[0])/N; |
| 436 | tmp = (2*tcount[2] >= N) + (2*tcount[1] >= N) + (2*tcount[0] >= N); |
| 437 | sum += tmp*256; |
| 438 | nbBands++; |
| 439 | } |
| 440 | } while (++c<C); |
| 441 | |
| 442 | if (update_hf) |
| 443 | { |
| 444 | if (hf_sum) |
| 445 | hf_sum /= C*(4-m->nbEBands+end); |
| 446 | *hf_average = (*hf_average+hf_sum)>>1; |
| 447 | hf_sum = *hf_average; |
| 448 | if (*tapset_decision==2) |
| 449 | hf_sum += 4; |
| 450 | else if (*tapset_decision==0) |
| 451 | hf_sum -= 4; |
| 452 | if (hf_sum > 22) |
| 453 | *tapset_decision=2; |
| 454 | else if (hf_sum > 18) |
| 455 | *tapset_decision=1; |
| 456 | else |
| 457 | *tapset_decision=0; |
| 458 | } |
| 459 | /*printf("%d %d %d\n", hf_sum, *hf_average, *tapset_decision);*/ |
| 460 | celt_assert(nbBands>0); /*M*(eBands[end]-eBands[end-1]) <= 8 assures this*/ |
| 461 | sum /= nbBands; |
| 462 | /* Recursive averaging */ |
| 463 | sum = (sum+*average)>>1; |
| 464 | *average = sum; |
| 465 | /* Hysteresis */ |
| 466 | sum = (3*sum + (((3-last_decision)<<7) + 64) + 2)>>2; |
| 467 | if (sum < 80) |
| 468 | { |
| 469 | decision = SPREAD_AGGRESSIVE; |
| 470 | } else if (sum < 256) |
| 471 | { |
| 472 | decision = SPREAD_NORMAL; |
| 473 | } else if (sum < 384) |
| 474 | { |
| 475 | decision = SPREAD_LIGHT; |
| 476 | } else { |
| 477 | decision = SPREAD_NONE; |
| 478 | } |
| 479 | #ifdef FUZZING |
| 480 | decision = rand()&0x3; |
| 481 | *tapset_decision=rand()%3; |
| 482 | #endif |
| 483 | return decision; |
| 484 | } |
| 485 | |
| 486 | #ifdef MEASURE_NORM_MSE |
| 487 | |
| 488 | float MSE[30] = {0}; |
| 489 | int nbMSEBands = 0; |
| 490 | int MSECount[30] = {0}; |
| 491 | |
| 492 | void dump_norm_mse(void) |
| 493 | { |
| 494 | int i; |
| 495 | for (i=0;i<nbMSEBands;i++) |
| 496 | { |
| 497 | printf ("%g ", MSE[i]/MSECount[i]); |
| 498 | } |
| 499 | printf ("\n"); |
| 500 | } |
| 501 | |
| 502 | void measure_norm_mse(const CELTMode *m, float *X, float *X0, float *bandE, float *bandE0, int M, int N, int C) |
| 503 | { |
| 504 | static int init = 0; |
| 505 | int i; |
| 506 | if (!init) |
| 507 | { |
| 508 | atexit(dump_norm_mse); |
| 509 | init = 1; |
| 510 | } |
| 511 | for (i=0;i<m->nbEBands;i++) |
| 512 | { |
| 513 | int j; |
| 514 | int c; |
| 515 | float g; |
| 516 | if (bandE0[i]<10 || (C==2 && bandE0[i+m->nbEBands]<1)) |
| 517 | continue; |
| 518 | c=0; do { |
| 519 | g = bandE[i+c*m->nbEBands]/(1e-15+bandE0[i+c*m->nbEBands]); |
| 520 | for (j=M*m->eBands[i];j<M*m->eBands[i+1];j++) |
| 521 | MSE[i] += (g*X[j+c*N]-X0[j+c*N])*(g*X[j+c*N]-X0[j+c*N]); |
| 522 | } while (++c<C); |
| 523 | MSECount[i]+=C; |
| 524 | } |
| 525 | nbMSEBands = m->nbEBands; |
| 526 | } |
| 527 | |
| 528 | #endif |
| 529 | |
| 530 | /* Indexing table for converting from natural Hadamard to ordery Hadamard |
| 531 | This is essentially a bit-reversed Gray, on top of which we've added |
| 532 | an inversion of the order because we want the DC at the end rather than |
| 533 | the beginning. The lines are for N=2, 4, 8, 16 */ |
| 534 | static const int ordery_table[] = { |
| 535 | 1, 0, |
| 536 | 3, 0, 2, 1, |
| 537 | 7, 0, 4, 3, 6, 1, 5, 2, |
| 538 | 15, 0, 8, 7, 12, 3, 11, 4, 14, 1, 9, 6, 13, 2, 10, 5, |
| 539 | }; |
| 540 | |
| 541 | static void deinterleave_hadamard(celt_norm *X, int N0, int stride, int hadamard) |
| 542 | { |
| 543 | int i,j; |
| 544 | VARDECL(celt_norm, tmp); |
| 545 | int N; |
| 546 | SAVE_STACK; |
| 547 | N = N0*stride; |
| 548 | ALLOC(tmp, N, celt_norm); |
| 549 | celt_assert(stride>0); |
| 550 | if (hadamard) |
| 551 | { |
| 552 | const int *ordery = ordery_table+stride-2; |
| 553 | for (i=0;i<stride;i++) |
| 554 | { |
| 555 | for (j=0;j<N0;j++) |
| 556 | tmp[ordery[i]*N0+j] = X[j*stride+i]; |
| 557 | } |
| 558 | } else { |
| 559 | for (i=0;i<stride;i++) |
| 560 | for (j=0;j<N0;j++) |
| 561 | tmp[i*N0+j] = X[j*stride+i]; |
| 562 | } |
| 563 | for (j=0;j<N;j++) |
| 564 | X[j] = tmp[j]; |
| 565 | RESTORE_STACK; |
| 566 | } |
| 567 | |
| 568 | static void interleave_hadamard(celt_norm *X, int N0, int stride, int hadamard) |
| 569 | { |
| 570 | int i,j; |
| 571 | VARDECL(celt_norm, tmp); |
| 572 | int N; |
| 573 | SAVE_STACK; |
| 574 | N = N0*stride; |
| 575 | ALLOC(tmp, N, celt_norm); |
| 576 | if (hadamard) |
| 577 | { |
| 578 | const int *ordery = ordery_table+stride-2; |
| 579 | for (i=0;i<stride;i++) |
| 580 | for (j=0;j<N0;j++) |
| 581 | tmp[j*stride+i] = X[ordery[i]*N0+j]; |
| 582 | } else { |
| 583 | for (i=0;i<stride;i++) |
| 584 | for (j=0;j<N0;j++) |
| 585 | tmp[j*stride+i] = X[i*N0+j]; |
| 586 | } |
| 587 | for (j=0;j<N;j++) |
| 588 | X[j] = tmp[j]; |
| 589 | RESTORE_STACK; |
| 590 | } |
| 591 | |
| 592 | void haar1(celt_norm *X, int N0, int stride) |
| 593 | { |
| 594 | int i, j; |
| 595 | N0 >>= 1; |
| 596 | for (i=0;i<stride;i++) |
| 597 | for (j=0;j<N0;j++) |
| 598 | { |
| 599 | celt_norm tmp1, tmp2; |
| 600 | tmp1 = MULT16_16_Q15(QCONST16(.70710678f,15), X[stride*2*j+i]); |
| 601 | tmp2 = MULT16_16_Q15(QCONST16(.70710678f,15), X[stride*(2*j+1)+i]); |
| 602 | X[stride*2*j+i] = tmp1 + tmp2; |
| 603 | X[stride*(2*j+1)+i] = tmp1 - tmp2; |
| 604 | } |
| 605 | } |
| 606 | |
| 607 | static int compute_qn(int N, int b, int offset, int pulse_cap, int stereo) |
| 608 | { |
| 609 | static const opus_int16 exp2_table8[8] = |
| 610 | {16384, 17866, 19483, 21247, 23170, 25267, 27554, 30048}; |
| 611 | int qn, qb; |
| 612 | int N2 = 2*N-1; |
| 613 | if (stereo && N==2) |
| 614 | N2--; |
| 615 | /* The upper limit ensures that in a stereo split with itheta==16384, we'll |
| 616 | always have enough bits left over to code at least one pulse in the |
| 617 | side; otherwise it would collapse, since it doesn't get folded. */ |
| 618 | qb = IMIN(b-pulse_cap-(4<<BITRES), (b+N2*offset)/N2); |
| 619 | |
| 620 | qb = IMIN(8<<BITRES, qb); |
| 621 | |
| 622 | if (qb<(1<<BITRES>>1)) { |
| 623 | qn = 1; |
| 624 | } else { |
| 625 | qn = exp2_table8[qb&0x7]>>(14-(qb>>BITRES)); |
| 626 | qn = (qn+1)>>1<<1; |
| 627 | } |
| 628 | celt_assert(qn <= 256); |
| 629 | return qn; |
| 630 | } |
| 631 | |
| 632 | /* This function is responsible for encoding and decoding a band for both |
| 633 | the mono and stereo case. Even in the mono case, it can split the band |
| 634 | in two and transmit the energy difference with the two half-bands. It |
| 635 | can be called recursively so bands can end up being split in 8 parts. */ |
| 636 | static unsigned quant_band(int encode, const CELTMode *m, int i, celt_norm *X, celt_norm *Y, |
| 637 | int N, int b, int spread, int B, int intensity, int tf_change, celt_norm *lowband, ec_ctx *ec, |
| 638 | opus_int32 *remaining_bits, int LM, celt_norm *lowband_out, const celt_ener *bandE, int level, |
| 639 | opus_uint32 *seed, opus_val16 gain, celt_norm *lowband_scratch, int fill) |
| 640 | { |
| 641 | const unsigned char *cache; |
| 642 | int q; |
| 643 | int curr_bits; |
| 644 | int stereo, split; |
| 645 | int imid=0, iside=0; |
| 646 | int N0=N; |
| 647 | int N_B=N; |
| 648 | int N_B0; |
| 649 | int B0=B; |
| 650 | int time_divide=0; |
| 651 | int recombine=0; |
| 652 | int inv = 0; |
| 653 | opus_val16 mid=0, side=0; |
| 654 | int longBlocks; |
| 655 | unsigned cm=0; |
| 656 | #ifdef RESYNTH |
| 657 | int resynth = 1; |
| 658 | #else |
| 659 | int resynth = !encode; |
| 660 | #endif |
| 661 | |
| 662 | longBlocks = B0==1; |
| 663 | |
| 664 | N_B /= B; |
| 665 | N_B0 = N_B; |
| 666 | |
| 667 | split = stereo = Y != NULL; |
| 668 | |
| 669 | /* Special case for one sample */ |
| 670 | if (N==1) |
| 671 | { |
| 672 | int c; |
| 673 | celt_norm *x = X; |
| 674 | c=0; do { |
| 675 | int sign=0; |
| 676 | if (*remaining_bits>=1<<BITRES) |
| 677 | { |
| 678 | if (encode) |
| 679 | { |
| 680 | sign = x[0]<0; |
| 681 | ec_enc_bits(ec, sign, 1); |
| 682 | } else { |
| 683 | sign = ec_dec_bits(ec, 1); |
| 684 | } |
| 685 | *remaining_bits -= 1<<BITRES; |
| 686 | b-=1<<BITRES; |
| 687 | } |
| 688 | if (resynth) |
| 689 | x[0] = sign ? -NORM_SCALING : NORM_SCALING; |
| 690 | x = Y; |
| 691 | } while (++c<1+stereo); |
| 692 | if (lowband_out) |
| 693 | lowband_out[0] = SHR16(X[0],4); |
| 694 | return 1; |
| 695 | } |
| 696 | |
| 697 | if (!stereo && level == 0) |
| 698 | { |
| 699 | int k; |
| 700 | if (tf_change>0) |
| 701 | recombine = tf_change; |
| 702 | /* Band recombining to increase frequency resolution */ |
| 703 | |
| 704 | if (lowband && (recombine || ((N_B&1) == 0 && tf_change<0) || B0>1)) |
| 705 | { |
| 706 | int j; |
| 707 | for (j=0;j<N;j++) |
| 708 | lowband_scratch[j] = lowband[j]; |
| 709 | lowband = lowband_scratch; |
| 710 | } |
| 711 | |
| 712 | for (k=0;k<recombine;k++) |
| 713 | { |
| 714 | static const unsigned char bit_interleave_table[16]={ |
| 715 | 0,1,1,1,2,3,3,3,2,3,3,3,2,3,3,3 |
| 716 | }; |
| 717 | if (encode) |
| 718 | haar1(X, N>>k, 1<<k); |
| 719 | if (lowband) |
| 720 | haar1(lowband, N>>k, 1<<k); |
| 721 | fill = bit_interleave_table[fill&0xF]|bit_interleave_table[fill>>4]<<2; |
| 722 | } |
| 723 | B>>=recombine; |
| 724 | N_B<<=recombine; |
| 725 | |
| 726 | /* Increasing the time resolution */ |
| 727 | while ((N_B&1) == 0 && tf_change<0) |
| 728 | { |
| 729 | if (encode) |
| 730 | haar1(X, N_B, B); |
| 731 | if (lowband) |
| 732 | haar1(lowband, N_B, B); |
| 733 | fill |= fill<<B; |
| 734 | B <<= 1; |
| 735 | N_B >>= 1; |
| 736 | time_divide++; |
| 737 | tf_change++; |
| 738 | } |
| 739 | B0=B; |
| 740 | N_B0 = N_B; |
| 741 | |
| 742 | /* Reorganize the samples in time order instead of frequency order */ |
| 743 | if (B0>1) |
| 744 | { |
| 745 | if (encode) |
| 746 | deinterleave_hadamard(X, N_B>>recombine, B0<<recombine, longBlocks); |
| 747 | if (lowband) |
| 748 | deinterleave_hadamard(lowband, N_B>>recombine, B0<<recombine, longBlocks); |
| 749 | } |
| 750 | } |
| 751 | |
| 752 | /* If we need 1.5 more bit than we can produce, split the band in two. */ |
| 753 | cache = m->cache.bits + m->cache.index[(LM+1)*m->nbEBands+i]; |
| 754 | if (!stereo && LM != -1 && b > cache[cache[0]]+12 && N>2) |
| 755 | { |
| 756 | N >>= 1; |
| 757 | Y = X+N; |
| 758 | split = 1; |
| 759 | LM -= 1; |
| 760 | if (B==1) |
| 761 | fill = (fill&1)|(fill<<1); |
| 762 | B = (B+1)>>1; |
| 763 | } |
| 764 | |
| 765 | if (split) |
| 766 | { |
| 767 | int qn; |
| 768 | int itheta=0; |
| 769 | int mbits, sbits, delta; |
| 770 | int qalloc; |
| 771 | int pulse_cap; |
| 772 | int offset; |
| 773 | int orig_fill; |
| 774 | opus_int32 tell; |
| 775 | |
| 776 | /* Decide on the resolution to give to the split parameter theta */ |
| 777 | pulse_cap = m->logN[i]+LM*(1<<BITRES); |
| 778 | offset = (pulse_cap>>1) - (stereo&&N==2 ? QTHETA_OFFSET_TWOPHASE : QTHETA_OFFSET); |
| 779 | qn = compute_qn(N, b, offset, pulse_cap, stereo); |
| 780 | if (stereo && i>=intensity) |
| 781 | qn = 1; |
| 782 | if (encode) |
| 783 | { |
| 784 | /* theta is the atan() of the ratio between the (normalized) |
| 785 | side and mid. With just that parameter, we can re-scale both |
| 786 | mid and side because we know that 1) they have unit norm and |
| 787 | 2) they are orthogonal. */ |
| 788 | itheta = stereo_itheta(X, Y, stereo, N); |
| 789 | } |
| 790 | tell = ec_tell_frac(ec); |
| 791 | if (qn!=1) |
| 792 | { |
| 793 | if (encode) |
| 794 | itheta = (itheta*qn+8192)>>14; |
| 795 | |
| 796 | /* Entropy coding of the angle. We use a uniform pdf for the |
| 797 | time split, a step for stereo, and a triangular one for the rest. */ |
| 798 | if (stereo && N>2) |
| 799 | { |
| 800 | int p0 = 3; |
| 801 | int x = itheta; |
| 802 | int x0 = qn/2; |
| 803 | int ft = p0*(x0+1) + x0; |
| 804 | /* Use a probability of p0 up to itheta=8192 and then use 1 after */ |
| 805 | if (encode) |
| 806 | { |
| 807 | ec_encode(ec,x<=x0?p0*x:(x-1-x0)+(x0+1)*p0,x<=x0?p0*(x+1):(x-x0)+(x0+1)*p0,ft); |
| 808 | } else { |
| 809 | int fs; |
| 810 | fs=ec_decode(ec,ft); |
| 811 | if (fs<(x0+1)*p0) |
| 812 | x=fs/p0; |
| 813 | else |
| 814 | x=x0+1+(fs-(x0+1)*p0); |
| 815 | ec_dec_update(ec,x<=x0?p0*x:(x-1-x0)+(x0+1)*p0,x<=x0?p0*(x+1):(x-x0)+(x0+1)*p0,ft); |
| 816 | itheta = x; |
| 817 | } |
| 818 | } else if (B0>1 || stereo) { |
| 819 | /* Uniform pdf */ |
| 820 | if (encode) |
| 821 | ec_enc_uint(ec, itheta, qn+1); |
| 822 | else |
| 823 | itheta = ec_dec_uint(ec, qn+1); |
| 824 | } else { |
| 825 | int fs=1, ft; |
| 826 | ft = ((qn>>1)+1)*((qn>>1)+1); |
| 827 | if (encode) |
| 828 | { |
| 829 | int fl; |
| 830 | |
| 831 | fs = itheta <= (qn>>1) ? itheta + 1 : qn + 1 - itheta; |
| 832 | fl = itheta <= (qn>>1) ? itheta*(itheta + 1)>>1 : |
| 833 | ft - ((qn + 1 - itheta)*(qn + 2 - itheta)>>1); |
| 834 | |
| 835 | ec_encode(ec, fl, fl+fs, ft); |
| 836 | } else { |
| 837 | /* Triangular pdf */ |
| 838 | int fl=0; |
| 839 | int fm; |
| 840 | fm = ec_decode(ec, ft); |
| 841 | |
| 842 | if (fm < ((qn>>1)*((qn>>1) + 1)>>1)) |
| 843 | { |
| 844 | itheta = (isqrt32(8*(opus_uint32)fm + 1) - 1)>>1; |
| 845 | fs = itheta + 1; |
| 846 | fl = itheta*(itheta + 1)>>1; |
| 847 | } |
| 848 | else |
| 849 | { |
| 850 | itheta = (2*(qn + 1) |
| 851 | - isqrt32(8*(opus_uint32)(ft - fm - 1) + 1))>>1; |
| 852 | fs = qn + 1 - itheta; |
| 853 | fl = ft - ((qn + 1 - itheta)*(qn + 2 - itheta)>>1); |
| 854 | } |
| 855 | |
| 856 | ec_dec_update(ec, fl, fl+fs, ft); |
| 857 | } |
| 858 | } |
| 859 | itheta = (opus_int32)itheta*16384/qn; |
| 860 | if (encode && stereo) |
| 861 | { |
| 862 | if (itheta==0) |
| 863 | intensity_stereo(m, X, Y, bandE, i, N); |
| 864 | else |
| 865 | stereo_split(X, Y, N); |
| 866 | } |
| 867 | /* NOTE: Renormalising X and Y *may* help fixed-point a bit at very high rate. |
| 868 | Let's do that at higher complexity */ |
| 869 | } else if (stereo) { |
| 870 | if (encode) |
| 871 | { |
| 872 | inv = itheta > 8192; |
| 873 | if (inv) |
| 874 | { |
| 875 | int j; |
| 876 | for (j=0;j<N;j++) |
| 877 | Y[j] = -Y[j]; |
| 878 | } |
| 879 | intensity_stereo(m, X, Y, bandE, i, N); |
| 880 | } |
| 881 | if (b>2<<BITRES && *remaining_bits > 2<<BITRES) |
| 882 | { |
| 883 | if (encode) |
| 884 | ec_enc_bit_logp(ec, inv, 2); |
| 885 | else |
| 886 | inv = ec_dec_bit_logp(ec, 2); |
| 887 | } else |
| 888 | inv = 0; |
| 889 | itheta = 0; |
| 890 | } |
| 891 | qalloc = ec_tell_frac(ec) - tell; |
| 892 | b -= qalloc; |
| 893 | |
| 894 | orig_fill = fill; |
| 895 | if (itheta == 0) |
| 896 | { |
| 897 | imid = 32767; |
| 898 | iside = 0; |
| 899 | fill &= (1<<B)-1; |
| 900 | delta = -16384; |
| 901 | } else if (itheta == 16384) |
| 902 | { |
| 903 | imid = 0; |
| 904 | iside = 32767; |
| 905 | fill &= ((1<<B)-1)<<B; |
| 906 | delta = 16384; |
| 907 | } else { |
| 908 | imid = bitexact_cos((opus_int16)itheta); |
| 909 | iside = bitexact_cos((opus_int16)(16384-itheta)); |
| 910 | /* This is the mid vs side allocation that minimizes squared error |
| 911 | in that band. */ |
| 912 | delta = FRAC_MUL16((N-1)<<7,bitexact_log2tan(iside,imid)); |
| 913 | } |
| 914 | |
| 915 | #ifdef FIXED_POINT |
| 916 | mid = imid; |
| 917 | side = iside; |
| 918 | #else |
| 919 | mid = (1.f/32768)*imid; |
| 920 | side = (1.f/32768)*iside; |
| 921 | #endif |
| 922 | |
| 923 | /* This is a special case for N=2 that only works for stereo and takes |
| 924 | advantage of the fact that mid and side are orthogonal to encode |
| 925 | the side with just one bit. */ |
| 926 | if (N==2 && stereo) |
| 927 | { |
| 928 | int c; |
| 929 | int sign=0; |
| 930 | celt_norm *x2, *y2; |
| 931 | mbits = b; |
| 932 | sbits = 0; |
| 933 | /* Only need one bit for the side */ |
| 934 | if (itheta != 0 && itheta != 16384) |
| 935 | sbits = 1<<BITRES; |
| 936 | mbits -= sbits; |
| 937 | c = itheta > 8192; |
| 938 | *remaining_bits -= qalloc+sbits; |
| 939 | |
| 940 | x2 = c ? Y : X; |
| 941 | y2 = c ? X : Y; |
| 942 | if (sbits) |
| 943 | { |
| 944 | if (encode) |
| 945 | { |
| 946 | /* Here we only need to encode a sign for the side */ |
| 947 | sign = x2[0]*y2[1] - x2[1]*y2[0] < 0; |
| 948 | ec_enc_bits(ec, sign, 1); |
| 949 | } else { |
| 950 | sign = ec_dec_bits(ec, 1); |
| 951 | } |
| 952 | } |
| 953 | sign = 1-2*sign; |
| 954 | /* We use orig_fill here because we want to fold the side, but if |
| 955 | itheta==16384, we'll have cleared the low bits of fill. */ |
| 956 | cm = quant_band(encode, m, i, x2, NULL, N, mbits, spread, B, intensity, tf_change, lowband, ec, remaining_bits, LM, lowband_out, NULL, level, seed, gain, lowband_scratch, orig_fill); |
| 957 | /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse), |
| 958 | and there's no need to worry about mixing with the other channel. */ |
| 959 | y2[0] = -sign*x2[1]; |
| 960 | y2[1] = sign*x2[0]; |
| 961 | if (resynth) |
| 962 | { |
| 963 | celt_norm tmp; |
| 964 | X[0] = MULT16_16_Q15(mid, X[0]); |
| 965 | X[1] = MULT16_16_Q15(mid, X[1]); |
| 966 | Y[0] = MULT16_16_Q15(side, Y[0]); |
| 967 | Y[1] = MULT16_16_Q15(side, Y[1]); |
| 968 | tmp = X[0]; |
| 969 | X[0] = SUB16(tmp,Y[0]); |
| 970 | Y[0] = ADD16(tmp,Y[0]); |
| 971 | tmp = X[1]; |
| 972 | X[1] = SUB16(tmp,Y[1]); |
| 973 | Y[1] = ADD16(tmp,Y[1]); |
| 974 | } |
| 975 | } else { |
| 976 | /* "Normal" split code */ |
| 977 | celt_norm *next_lowband2=NULL; |
| 978 | celt_norm *next_lowband_out1=NULL; |
| 979 | int next_level=0; |
| 980 | opus_int32 rebalance; |
| 981 | |
| 982 | /* Give more bits to low-energy MDCTs than they would otherwise deserve */ |
| 983 | if (B0>1 && !stereo && (itheta&0x3fff)) |
| 984 | { |
| 985 | if (itheta > 8192) |
| 986 | /* Rough approximation for pre-echo masking */ |
| 987 | delta -= delta>>(4-LM); |
| 988 | else |
| 989 | /* Corresponds to a forward-masking slope of 1.5 dB per 10 ms */ |
| 990 | delta = IMIN(0, delta + (N<<BITRES>>(5-LM))); |
| 991 | } |
| 992 | mbits = IMAX(0, IMIN(b, (b-delta)/2)); |
| 993 | sbits = b-mbits; |
| 994 | *remaining_bits -= qalloc; |
| 995 | |
| 996 | if (lowband && !stereo) |
| 997 | next_lowband2 = lowband+N; /* >32-bit split case */ |
| 998 | |
| 999 | /* Only stereo needs to pass on lowband_out. Otherwise, it's |
| 1000 | handled at the end */ |
| 1001 | if (stereo) |
| 1002 | next_lowband_out1 = lowband_out; |
| 1003 | else |
| 1004 | next_level = level+1; |
| 1005 | |
| 1006 | rebalance = *remaining_bits; |
| 1007 | if (mbits >= sbits) |
| 1008 | { |
| 1009 | /* In stereo mode, we do not apply a scaling to the mid because we need the normalized |
| 1010 | mid for folding later */ |
| 1011 | cm = quant_band(encode, m, i, X, NULL, N, mbits, spread, B, intensity, tf_change, |
| 1012 | lowband, ec, remaining_bits, LM, next_lowband_out1, |
| 1013 | NULL, next_level, seed, stereo ? Q15ONE : MULT16_16_P15(gain,mid), lowband_scratch, fill); |
| 1014 | rebalance = mbits - (rebalance-*remaining_bits); |
| 1015 | if (rebalance > 3<<BITRES && itheta!=0) |
| 1016 | sbits += rebalance - (3<<BITRES); |
| 1017 | |
| 1018 | /* For a stereo split, the high bits of fill are always zero, so no |
| 1019 | folding will be done to the side. */ |
| 1020 | cm |= quant_band(encode, m, i, Y, NULL, N, sbits, spread, B, intensity, tf_change, |
| 1021 | next_lowband2, ec, remaining_bits, LM, NULL, |
| 1022 | NULL, next_level, seed, MULT16_16_P15(gain,side), NULL, fill>>B)<<((B0>>1)&(stereo-1)); |
| 1023 | } else { |
| 1024 | /* For a stereo split, the high bits of fill are always zero, so no |
| 1025 | folding will be done to the side. */ |
| 1026 | cm = quant_band(encode, m, i, Y, NULL, N, sbits, spread, B, intensity, tf_change, |
| 1027 | next_lowband2, ec, remaining_bits, LM, NULL, |
| 1028 | NULL, next_level, seed, MULT16_16_P15(gain,side), NULL, fill>>B)<<((B0>>1)&(stereo-1)); |
| 1029 | rebalance = sbits - (rebalance-*remaining_bits); |
| 1030 | if (rebalance > 3<<BITRES && itheta!=16384) |
| 1031 | mbits += rebalance - (3<<BITRES); |
| 1032 | /* In stereo mode, we do not apply a scaling to the mid because we need the normalized |
| 1033 | mid for folding later */ |
| 1034 | cm |= quant_band(encode, m, i, X, NULL, N, mbits, spread, B, intensity, tf_change, |
| 1035 | lowband, ec, remaining_bits, LM, next_lowband_out1, |
| 1036 | NULL, next_level, seed, stereo ? Q15ONE : MULT16_16_P15(gain,mid), lowband_scratch, fill); |
| 1037 | } |
| 1038 | } |
| 1039 | |
| 1040 | } else { |
| 1041 | /* This is the basic no-split case */ |
| 1042 | q = bits2pulses(m, i, LM, b); |
| 1043 | curr_bits = pulses2bits(m, i, LM, q); |
| 1044 | *remaining_bits -= curr_bits; |
| 1045 | |
| 1046 | /* Ensures we can never bust the budget */ |
| 1047 | while (*remaining_bits < 0 && q > 0) |
| 1048 | { |
| 1049 | *remaining_bits += curr_bits; |
| 1050 | q--; |
| 1051 | curr_bits = pulses2bits(m, i, LM, q); |
| 1052 | *remaining_bits -= curr_bits; |
| 1053 | } |
| 1054 | |
| 1055 | if (q!=0) |
| 1056 | { |
| 1057 | int K = get_pulses(q); |
| 1058 | |
| 1059 | /* Finally do the actual quantization */ |
| 1060 | if (encode) |
| 1061 | { |
| 1062 | cm = alg_quant(X, N, K, spread, B, ec |
| 1063 | #ifdef RESYNTH |
| 1064 | , gain |
| 1065 | #endif |
| 1066 | ); |
| 1067 | } else { |
| 1068 | cm = alg_unquant(X, N, K, spread, B, ec, gain); |
| 1069 | } |
| 1070 | } else { |
| 1071 | /* If there's no pulse, fill the band anyway */ |
| 1072 | int j; |
| 1073 | if (resynth) |
| 1074 | { |
| 1075 | unsigned cm_mask; |
| 1076 | /*B can be as large as 16, so this shift might overflow an int on a |
| 1077 | 16-bit platform; use a long to get defined behavior.*/ |
| 1078 | cm_mask = (unsigned)(1UL<<B)-1; |
| 1079 | fill &= cm_mask; |
| 1080 | if (!fill) |
| 1081 | { |
| 1082 | for (j=0;j<N;j++) |
| 1083 | X[j] = 0; |
| 1084 | } else { |
| 1085 | if (lowband == NULL) |
| 1086 | { |
| 1087 | /* Noise */ |
| 1088 | for (j=0;j<N;j++) |
| 1089 | { |
| 1090 | *seed = celt_lcg_rand(*seed); |
| 1091 | X[j] = (celt_norm)((opus_int32)*seed>>20); |
| 1092 | } |
| 1093 | cm = cm_mask; |
| 1094 | } else { |
| 1095 | /* Folded spectrum */ |
| 1096 | for (j=0;j<N;j++) |
| 1097 | { |
| 1098 | opus_val16 tmp; |
| 1099 | *seed = celt_lcg_rand(*seed); |
| 1100 | /* About 48 dB below the "normal" folding level */ |
| 1101 | tmp = QCONST16(1.0f/256, 10); |
| 1102 | tmp = (*seed)&0x8000 ? tmp : -tmp; |
| 1103 | X[j] = lowband[j]+tmp; |
| 1104 | } |
| 1105 | cm = fill; |
| 1106 | } |
| 1107 | renormalise_vector(X, N, gain); |
| 1108 | } |
| 1109 | } |
| 1110 | } |
| 1111 | } |
| 1112 | |
| 1113 | /* This code is used by the decoder and by the resynthesis-enabled encoder */ |
| 1114 | if (resynth) |
| 1115 | { |
| 1116 | if (stereo) |
| 1117 | { |
| 1118 | if (N!=2) |
| 1119 | stereo_merge(X, Y, mid, N); |
| 1120 | if (inv) |
| 1121 | { |
| 1122 | int j; |
| 1123 | for (j=0;j<N;j++) |
| 1124 | Y[j] = -Y[j]; |
| 1125 | } |
| 1126 | } else if (level == 0) |
| 1127 | { |
| 1128 | int k; |
| 1129 | |
| 1130 | /* Undo the sample reorganization going from time order to frequency order */ |
| 1131 | if (B0>1) |
| 1132 | interleave_hadamard(X, N_B>>recombine, B0<<recombine, longBlocks); |
| 1133 | |
| 1134 | /* Undo time-freq changes that we did earlier */ |
| 1135 | N_B = N_B0; |
| 1136 | B = B0; |
| 1137 | for (k=0;k<time_divide;k++) |
| 1138 | { |
| 1139 | B >>= 1; |
| 1140 | N_B <<= 1; |
| 1141 | cm |= cm>>B; |
| 1142 | haar1(X, N_B, B); |
| 1143 | } |
| 1144 | |
| 1145 | for (k=0;k<recombine;k++) |
| 1146 | { |
| 1147 | static const unsigned char bit_deinterleave_table[16]={ |
| 1148 | 0x00,0x03,0x0C,0x0F,0x30,0x33,0x3C,0x3F, |
| 1149 | 0xC0,0xC3,0xCC,0xCF,0xF0,0xF3,0xFC,0xFF |
| 1150 | }; |
| 1151 | cm = bit_deinterleave_table[cm]; |
| 1152 | haar1(X, N0>>k, 1<<k); |
| 1153 | } |
| 1154 | B<<=recombine; |
| 1155 | |
| 1156 | /* Scale output for later folding */ |
| 1157 | if (lowband_out) |
| 1158 | { |
| 1159 | int j; |
| 1160 | opus_val16 n; |
| 1161 | n = celt_sqrt(SHL32(EXTEND32(N0),22)); |
| 1162 | for (j=0;j<N0;j++) |
| 1163 | lowband_out[j] = MULT16_16_Q15(n,X[j]); |
| 1164 | } |
| 1165 | cm &= (1<<B)-1; |
| 1166 | } |
| 1167 | } |
| 1168 | return cm; |
| 1169 | } |
| 1170 | |
| 1171 | void quant_all_bands(int encode, const CELTMode *m, int start, int end, |
| 1172 | celt_norm *X_, celt_norm *Y_, unsigned char *collapse_masks, const celt_ener *bandE, int *pulses, |
| 1173 | int shortBlocks, int spread, int dual_stereo, int intensity, int *tf_res, |
| 1174 | opus_int32 total_bits, opus_int32 balance, ec_ctx *ec, int LM, int codedBands, opus_uint32 *seed) |
| 1175 | { |
| 1176 | int i; |
| 1177 | opus_int32 remaining_bits; |
| 1178 | const opus_int16 * OPUS_RESTRICT eBands = m->eBands; |
| 1179 | celt_norm * OPUS_RESTRICT norm, * OPUS_RESTRICT norm2; |
| 1180 | VARDECL(celt_norm, _norm); |
| 1181 | VARDECL(celt_norm, lowband_scratch); |
| 1182 | int B; |
| 1183 | int M; |
| 1184 | int lowband_offset; |
| 1185 | int update_lowband = 1; |
| 1186 | int C = Y_ != NULL ? 2 : 1; |
| 1187 | #ifdef RESYNTH |
| 1188 | int resynth = 1; |
| 1189 | #else |
| 1190 | int resynth = !encode; |
| 1191 | #endif |
| 1192 | SAVE_STACK; |
| 1193 | |
| 1194 | M = 1<<LM; |
| 1195 | B = shortBlocks ? M : 1; |
| 1196 | ALLOC(_norm, C*M*eBands[m->nbEBands], celt_norm); |
| 1197 | ALLOC(lowband_scratch, M*(eBands[m->nbEBands]-eBands[m->nbEBands-1]), celt_norm); |
| 1198 | norm = _norm; |
| 1199 | norm2 = norm + M*eBands[m->nbEBands]; |
| 1200 | |
| 1201 | lowband_offset = 0; |
| 1202 | for (i=start;i<end;i++) |
| 1203 | { |
| 1204 | opus_int32 tell; |
| 1205 | int b; |
| 1206 | int N; |
| 1207 | opus_int32 curr_balance; |
| 1208 | int effective_lowband=-1; |
| 1209 | celt_norm * OPUS_RESTRICT X, * OPUS_RESTRICT Y; |
| 1210 | int tf_change=0; |
| 1211 | unsigned x_cm; |
| 1212 | unsigned y_cm; |
| 1213 | |
| 1214 | X = X_+M*eBands[i]; |
| 1215 | if (Y_!=NULL) |
| 1216 | Y = Y_+M*eBands[i]; |
| 1217 | else |
| 1218 | Y = NULL; |
| 1219 | N = M*eBands[i+1]-M*eBands[i]; |
| 1220 | tell = ec_tell_frac(ec); |
| 1221 | |
| 1222 | /* Compute how many bits we want to allocate to this band */ |
| 1223 | if (i != start) |
| 1224 | balance -= tell; |
| 1225 | remaining_bits = total_bits-tell-1; |
| 1226 | if (i <= codedBands-1) |
| 1227 | { |
| 1228 | curr_balance = balance / IMIN(3, codedBands-i); |
| 1229 | b = IMAX(0, IMIN(16383, IMIN(remaining_bits+1,pulses[i]+curr_balance))); |
| 1230 | } else { |
| 1231 | b = 0; |
| 1232 | } |
| 1233 | |
| 1234 | if (resynth && M*eBands[i]-N >= M*eBands[start] && (update_lowband || lowband_offset==0)) |
| 1235 | lowband_offset = i; |
| 1236 | |
| 1237 | tf_change = tf_res[i]; |
| 1238 | if (i>=m->effEBands) |
| 1239 | { |
| 1240 | X=norm; |
| 1241 | if (Y_!=NULL) |
| 1242 | Y = norm; |
| 1243 | } |
| 1244 | |
| 1245 | /* Get a conservative estimate of the collapse_mask's for the bands we're |
| 1246 | going to be folding from. */ |
| 1247 | if (lowband_offset != 0 && (spread!=SPREAD_AGGRESSIVE || B>1 || tf_change<0)) |
| 1248 | { |
| 1249 | int fold_start; |
| 1250 | int fold_end; |
| 1251 | int fold_i; |
| 1252 | /* This ensures we never repeat spectral content within one band */ |
| 1253 | effective_lowband = IMAX(M*eBands[start], M*eBands[lowband_offset]-N); |
| 1254 | fold_start = lowband_offset; |
| 1255 | while(M*eBands[--fold_start] > effective_lowband); |
| 1256 | fold_end = lowband_offset-1; |
| 1257 | while(M*eBands[++fold_end] < effective_lowband+N); |
| 1258 | x_cm = y_cm = 0; |
| 1259 | fold_i = fold_start; do { |
| 1260 | x_cm |= collapse_masks[fold_i*C+0]; |
| 1261 | y_cm |= collapse_masks[fold_i*C+C-1]; |
| 1262 | } while (++fold_i<fold_end); |
| 1263 | } |
| 1264 | /* Otherwise, we'll be using the LCG to fold, so all blocks will (almost |
| 1265 | always) be non-zero.*/ |
| 1266 | else |
| 1267 | x_cm = y_cm = (1<<B)-1; |
| 1268 | |
| 1269 | if (dual_stereo && i==intensity) |
| 1270 | { |
| 1271 | int j; |
| 1272 | |
| 1273 | /* Switch off dual stereo to do intensity */ |
| 1274 | dual_stereo = 0; |
| 1275 | if (resynth) |
| 1276 | for (j=M*eBands[start];j<M*eBands[i];j++) |
| 1277 | norm[j] = HALF32(norm[j]+norm2[j]); |
| 1278 | } |
| 1279 | if (dual_stereo) |
| 1280 | { |
| 1281 | x_cm = quant_band(encode, m, i, X, NULL, N, b/2, spread, B, intensity, tf_change, |
| 1282 | effective_lowband != -1 ? norm+effective_lowband : NULL, ec, &remaining_bits, LM, |
| 1283 | norm+M*eBands[i], bandE, 0, seed, Q15ONE, lowband_scratch, x_cm); |
| 1284 | y_cm = quant_band(encode, m, i, Y, NULL, N, b/2, spread, B, intensity, tf_change, |
| 1285 | effective_lowband != -1 ? norm2+effective_lowband : NULL, ec, &remaining_bits, LM, |
| 1286 | norm2+M*eBands[i], bandE, 0, seed, Q15ONE, lowband_scratch, y_cm); |
| 1287 | } else { |
| 1288 | x_cm = quant_band(encode, m, i, X, Y, N, b, spread, B, intensity, tf_change, |
| 1289 | effective_lowband != -1 ? norm+effective_lowband : NULL, ec, &remaining_bits, LM, |
| 1290 | norm+M*eBands[i], bandE, 0, seed, Q15ONE, lowband_scratch, x_cm|y_cm); |
| 1291 | y_cm = x_cm; |
| 1292 | } |
| 1293 | collapse_masks[i*C+0] = (unsigned char)x_cm; |
| 1294 | collapse_masks[i*C+C-1] = (unsigned char)y_cm; |
| 1295 | balance += pulses[i] + tell; |
| 1296 | |
| 1297 | /* Update the folding position only as long as we have 1 bit/sample depth */ |
| 1298 | update_lowband = b>(N<<BITRES); |
| 1299 | } |
| 1300 | RESTORE_STACK; |
| 1301 | } |
| 1302 | |