1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libopus/celt/vq.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,415 @@
1.4 +/* Copyright (c) 2007-2008 CSIRO
1.5 + Copyright (c) 2007-2009 Xiph.Org Foundation
1.6 + Written by Jean-Marc Valin */
1.7 +/*
1.8 + Redistribution and use in source and binary forms, with or without
1.9 + modification, are permitted provided that the following conditions
1.10 + are met:
1.11 +
1.12 + - Redistributions of source code must retain the above copyright
1.13 + notice, this list of conditions and the following disclaimer.
1.14 +
1.15 + - Redistributions in binary form must reproduce the above copyright
1.16 + notice, this list of conditions and the following disclaimer in the
1.17 + documentation and/or other materials provided with the distribution.
1.18 +
1.19 + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
1.20 + ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
1.21 + LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
1.22 + A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
1.23 + OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
1.24 + EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
1.25 + PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
1.26 + PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
1.27 + LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
1.28 + NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
1.29 + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
1.30 +*/
1.31 +
1.32 +#ifdef HAVE_CONFIG_H
1.33 +#include "config.h"
1.34 +#endif
1.35 +
1.36 +#include "mathops.h"
1.37 +#include "cwrs.h"
1.38 +#include "vq.h"
1.39 +#include "arch.h"
1.40 +#include "os_support.h"
1.41 +#include "bands.h"
1.42 +#include "rate.h"
1.43 +
1.44 +static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_val16 s)
1.45 +{
1.46 + int i;
1.47 + celt_norm *Xptr;
1.48 + Xptr = X;
1.49 + for (i=0;i<len-stride;i++)
1.50 + {
1.51 + celt_norm x1, x2;
1.52 + x1 = Xptr[0];
1.53 + x2 = Xptr[stride];
1.54 + Xptr[stride] = EXTRACT16(SHR32(MULT16_16(c,x2) + MULT16_16(s,x1), 15));
1.55 + *Xptr++ = EXTRACT16(SHR32(MULT16_16(c,x1) - MULT16_16(s,x2), 15));
1.56 + }
1.57 + Xptr = &X[len-2*stride-1];
1.58 + for (i=len-2*stride-1;i>=0;i--)
1.59 + {
1.60 + celt_norm x1, x2;
1.61 + x1 = Xptr[0];
1.62 + x2 = Xptr[stride];
1.63 + Xptr[stride] = EXTRACT16(SHR32(MULT16_16(c,x2) + MULT16_16(s,x1), 15));
1.64 + *Xptr-- = EXTRACT16(SHR32(MULT16_16(c,x1) - MULT16_16(s,x2), 15));
1.65 + }
1.66 +}
1.67 +
1.68 +static void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread)
1.69 +{
1.70 + static const int SPREAD_FACTOR[3]={15,10,5};
1.71 + int i;
1.72 + opus_val16 c, s;
1.73 + opus_val16 gain, theta;
1.74 + int stride2=0;
1.75 + int factor;
1.76 +
1.77 + if (2*K>=len || spread==SPREAD_NONE)
1.78 + return;
1.79 + factor = SPREAD_FACTOR[spread-1];
1.80 +
1.81 + gain = celt_div((opus_val32)MULT16_16(Q15_ONE,len),(opus_val32)(len+factor*K));
1.82 + theta = HALF16(MULT16_16_Q15(gain,gain));
1.83 +
1.84 + c = celt_cos_norm(EXTEND32(theta));
1.85 + s = celt_cos_norm(EXTEND32(SUB16(Q15ONE,theta))); /* sin(theta) */
1.86 +
1.87 + if (len>=8*stride)
1.88 + {
1.89 + stride2 = 1;
1.90 + /* This is just a simple (equivalent) way of computing sqrt(len/stride) with rounding.
1.91 + It's basically incrementing long as (stride2+0.5)^2 < len/stride. */
1.92 + while ((stride2*stride2+stride2)*stride + (stride>>2) < len)
1.93 + stride2++;
1.94 + }
1.95 + /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
1.96 + extract_collapse_mask().*/
1.97 + len /= stride;
1.98 + for (i=0;i<stride;i++)
1.99 + {
1.100 + if (dir < 0)
1.101 + {
1.102 + if (stride2)
1.103 + exp_rotation1(X+i*len, len, stride2, s, c);
1.104 + exp_rotation1(X+i*len, len, 1, c, s);
1.105 + } else {
1.106 + exp_rotation1(X+i*len, len, 1, c, -s);
1.107 + if (stride2)
1.108 + exp_rotation1(X+i*len, len, stride2, s, -c);
1.109 + }
1.110 + }
1.111 +}
1.112 +
1.113 +/** Takes the pitch vector and the decoded residual vector, computes the gain
1.114 + that will give ||p+g*y||=1 and mixes the residual with the pitch. */
1.115 +static void normalise_residual(int * OPUS_RESTRICT iy, celt_norm * OPUS_RESTRICT X,
1.116 + int N, opus_val32 Ryy, opus_val16 gain)
1.117 +{
1.118 + int i;
1.119 +#ifdef FIXED_POINT
1.120 + int k;
1.121 +#endif
1.122 + opus_val32 t;
1.123 + opus_val16 g;
1.124 +
1.125 +#ifdef FIXED_POINT
1.126 + k = celt_ilog2(Ryy)>>1;
1.127 +#endif
1.128 + t = VSHR32(Ryy, 2*(k-7));
1.129 + g = MULT16_16_P15(celt_rsqrt_norm(t),gain);
1.130 +
1.131 + i=0;
1.132 + do
1.133 + X[i] = EXTRACT16(PSHR32(MULT16_16(g, iy[i]), k+1));
1.134 + while (++i < N);
1.135 +}
1.136 +
1.137 +static unsigned extract_collapse_mask(int *iy, int N, int B)
1.138 +{
1.139 + unsigned collapse_mask;
1.140 + int N0;
1.141 + int i;
1.142 + if (B<=1)
1.143 + return 1;
1.144 + /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
1.145 + exp_rotation().*/
1.146 + N0 = N/B;
1.147 + collapse_mask = 0;
1.148 + i=0; do {
1.149 + int j;
1.150 + j=0; do {
1.151 + collapse_mask |= (iy[i*N0+j]!=0)<<i;
1.152 + } while (++j<N0);
1.153 + } while (++i<B);
1.154 + return collapse_mask;
1.155 +}
1.156 +
1.157 +unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc
1.158 +#ifdef RESYNTH
1.159 + , opus_val16 gain
1.160 +#endif
1.161 + )
1.162 +{
1.163 + VARDECL(celt_norm, y);
1.164 + VARDECL(int, iy);
1.165 + VARDECL(opus_val16, signx);
1.166 + int i, j;
1.167 + opus_val16 s;
1.168 + int pulsesLeft;
1.169 + opus_val32 sum;
1.170 + opus_val32 xy;
1.171 + opus_val16 yy;
1.172 + unsigned collapse_mask;
1.173 + SAVE_STACK;
1.174 +
1.175 + celt_assert2(K>0, "alg_quant() needs at least one pulse");
1.176 + celt_assert2(N>1, "alg_quant() needs at least two dimensions");
1.177 +
1.178 + ALLOC(y, N, celt_norm);
1.179 + ALLOC(iy, N, int);
1.180 + ALLOC(signx, N, opus_val16);
1.181 +
1.182 + exp_rotation(X, N, 1, B, K, spread);
1.183 +
1.184 + /* Get rid of the sign */
1.185 + sum = 0;
1.186 + j=0; do {
1.187 + if (X[j]>0)
1.188 + signx[j]=1;
1.189 + else {
1.190 + signx[j]=-1;
1.191 + X[j]=-X[j];
1.192 + }
1.193 + iy[j] = 0;
1.194 + y[j] = 0;
1.195 + } while (++j<N);
1.196 +
1.197 + xy = yy = 0;
1.198 +
1.199 + pulsesLeft = K;
1.200 +
1.201 + /* Do a pre-search by projecting on the pyramid */
1.202 + if (K > (N>>1))
1.203 + {
1.204 + opus_val16 rcp;
1.205 + j=0; do {
1.206 + sum += X[j];
1.207 + } while (++j<N);
1.208 +
1.209 + /* If X is too small, just replace it with a pulse at 0 */
1.210 +#ifdef FIXED_POINT
1.211 + if (sum <= K)
1.212 +#else
1.213 + /* Prevents infinities and NaNs from causing too many pulses
1.214 + to be allocated. 64 is an approximation of infinity here. */
1.215 + if (!(sum > EPSILON && sum < 64))
1.216 +#endif
1.217 + {
1.218 + X[0] = QCONST16(1.f,14);
1.219 + j=1; do
1.220 + X[j]=0;
1.221 + while (++j<N);
1.222 + sum = QCONST16(1.f,14);
1.223 + }
1.224 + rcp = EXTRACT16(MULT16_32_Q16(K-1, celt_rcp(sum)));
1.225 + j=0; do {
1.226 +#ifdef FIXED_POINT
1.227 + /* It's really important to round *towards zero* here */
1.228 + iy[j] = MULT16_16_Q15(X[j],rcp);
1.229 +#else
1.230 + iy[j] = (int)floor(rcp*X[j]);
1.231 +#endif
1.232 + y[j] = (celt_norm)iy[j];
1.233 + yy = MAC16_16(yy, y[j],y[j]);
1.234 + xy = MAC16_16(xy, X[j],y[j]);
1.235 + y[j] *= 2;
1.236 + pulsesLeft -= iy[j];
1.237 + } while (++j<N);
1.238 + }
1.239 + celt_assert2(pulsesLeft>=1, "Allocated too many pulses in the quick pass");
1.240 +
1.241 + /* This should never happen, but just in case it does (e.g. on silence)
1.242 + we fill the first bin with pulses. */
1.243 +#ifdef FIXED_POINT_DEBUG
1.244 + celt_assert2(pulsesLeft<=N+3, "Not enough pulses in the quick pass");
1.245 +#endif
1.246 + if (pulsesLeft > N+3)
1.247 + {
1.248 + opus_val16 tmp = (opus_val16)pulsesLeft;
1.249 + yy = MAC16_16(yy, tmp, tmp);
1.250 + yy = MAC16_16(yy, tmp, y[0]);
1.251 + iy[0] += pulsesLeft;
1.252 + pulsesLeft=0;
1.253 + }
1.254 +
1.255 + s = 1;
1.256 + for (i=0;i<pulsesLeft;i++)
1.257 + {
1.258 + int best_id;
1.259 + opus_val32 best_num = -VERY_LARGE16;
1.260 + opus_val16 best_den = 0;
1.261 +#ifdef FIXED_POINT
1.262 + int rshift;
1.263 +#endif
1.264 +#ifdef FIXED_POINT
1.265 + rshift = 1+celt_ilog2(K-pulsesLeft+i+1);
1.266 +#endif
1.267 + best_id = 0;
1.268 + /* The squared magnitude term gets added anyway, so we might as well
1.269 + add it outside the loop */
1.270 + yy = ADD32(yy, 1);
1.271 + j=0;
1.272 + do {
1.273 + opus_val16 Rxy, Ryy;
1.274 + /* Temporary sums of the new pulse(s) */
1.275 + Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[j])),rshift));
1.276 + /* We're multiplying y[j] by two so we don't have to do it here */
1.277 + Ryy = ADD16(yy, y[j]);
1.278 +
1.279 + /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that
1.280 + Rxy is positive because the sign is pre-computed) */
1.281 + Rxy = MULT16_16_Q15(Rxy,Rxy);
1.282 + /* The idea is to check for num/den >= best_num/best_den, but that way
1.283 + we can do it without any division */
1.284 + /* OPT: Make sure to use conditional moves here */
1.285 + if (MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num))
1.286 + {
1.287 + best_den = Ryy;
1.288 + best_num = Rxy;
1.289 + best_id = j;
1.290 + }
1.291 + } while (++j<N);
1.292 +
1.293 + /* Updating the sums of the new pulse(s) */
1.294 + xy = ADD32(xy, EXTEND32(X[best_id]));
1.295 + /* We're multiplying y[j] by two so we don't have to do it here */
1.296 + yy = ADD16(yy, y[best_id]);
1.297 +
1.298 + /* Only now that we've made the final choice, update y/iy */
1.299 + /* Multiplying y[j] by 2 so we don't have to do it everywhere else */
1.300 + y[best_id] += 2*s;
1.301 + iy[best_id]++;
1.302 + }
1.303 +
1.304 + /* Put the original sign back */
1.305 + j=0;
1.306 + do {
1.307 + X[j] = MULT16_16(signx[j],X[j]);
1.308 + if (signx[j] < 0)
1.309 + iy[j] = -iy[j];
1.310 + } while (++j<N);
1.311 + encode_pulses(iy, N, K, enc);
1.312 +
1.313 +#ifdef RESYNTH
1.314 + normalise_residual(iy, X, N, yy, gain);
1.315 + exp_rotation(X, N, -1, B, K, spread);
1.316 +#endif
1.317 +
1.318 + collapse_mask = extract_collapse_mask(iy, N, B);
1.319 + RESTORE_STACK;
1.320 + return collapse_mask;
1.321 +}
1.322 +
1.323 +/** Decode pulse vector and combine the result with the pitch vector to produce
1.324 + the final normalised signal in the current band. */
1.325 +unsigned alg_unquant(celt_norm *X, int N, int K, int spread, int B,
1.326 + ec_dec *dec, opus_val16 gain)
1.327 +{
1.328 + int i;
1.329 + opus_val32 Ryy;
1.330 + unsigned collapse_mask;
1.331 + VARDECL(int, iy);
1.332 + SAVE_STACK;
1.333 +
1.334 + celt_assert2(K>0, "alg_unquant() needs at least one pulse");
1.335 + celt_assert2(N>1, "alg_unquant() needs at least two dimensions");
1.336 + ALLOC(iy, N, int);
1.337 + decode_pulses(iy, N, K, dec);
1.338 + Ryy = 0;
1.339 + i=0;
1.340 + do {
1.341 + Ryy = MAC16_16(Ryy, iy[i], iy[i]);
1.342 + } while (++i < N);
1.343 + normalise_residual(iy, X, N, Ryy, gain);
1.344 + exp_rotation(X, N, -1, B, K, spread);
1.345 + collapse_mask = extract_collapse_mask(iy, N, B);
1.346 + RESTORE_STACK;
1.347 + return collapse_mask;
1.348 +}
1.349 +
1.350 +void renormalise_vector(celt_norm *X, int N, opus_val16 gain)
1.351 +{
1.352 + int i;
1.353 +#ifdef FIXED_POINT
1.354 + int k;
1.355 +#endif
1.356 + opus_val32 E = EPSILON;
1.357 + opus_val16 g;
1.358 + opus_val32 t;
1.359 + celt_norm *xptr = X;
1.360 + for (i=0;i<N;i++)
1.361 + {
1.362 + E = MAC16_16(E, *xptr, *xptr);
1.363 + xptr++;
1.364 + }
1.365 +#ifdef FIXED_POINT
1.366 + k = celt_ilog2(E)>>1;
1.367 +#endif
1.368 + t = VSHR32(E, 2*(k-7));
1.369 + g = MULT16_16_P15(celt_rsqrt_norm(t),gain);
1.370 +
1.371 + xptr = X;
1.372 + for (i=0;i<N;i++)
1.373 + {
1.374 + *xptr = EXTRACT16(PSHR32(MULT16_16(g, *xptr), k+1));
1.375 + xptr++;
1.376 + }
1.377 + /*return celt_sqrt(E);*/
1.378 +}
1.379 +
1.380 +int stereo_itheta(celt_norm *X, celt_norm *Y, int stereo, int N)
1.381 +{
1.382 + int i;
1.383 + int itheta;
1.384 + opus_val16 mid, side;
1.385 + opus_val32 Emid, Eside;
1.386 +
1.387 + Emid = Eside = EPSILON;
1.388 + if (stereo)
1.389 + {
1.390 + for (i=0;i<N;i++)
1.391 + {
1.392 + celt_norm m, s;
1.393 + m = ADD16(SHR16(X[i],1),SHR16(Y[i],1));
1.394 + s = SUB16(SHR16(X[i],1),SHR16(Y[i],1));
1.395 + Emid = MAC16_16(Emid, m, m);
1.396 + Eside = MAC16_16(Eside, s, s);
1.397 + }
1.398 + } else {
1.399 + for (i=0;i<N;i++)
1.400 + {
1.401 + celt_norm m, s;
1.402 + m = X[i];
1.403 + s = Y[i];
1.404 + Emid = MAC16_16(Emid, m, m);
1.405 + Eside = MAC16_16(Eside, s, s);
1.406 + }
1.407 + }
1.408 + mid = celt_sqrt(Emid);
1.409 + side = celt_sqrt(Eside);
1.410 +#ifdef FIXED_POINT
1.411 + /* 0.63662 = 2/pi */
1.412 + itheta = MULT16_16_Q15(QCONST16(0.63662f,15),celt_atan2p(side, mid));
1.413 +#else
1.414 + itheta = (int)floor(.5f+16384*0.63662f*atan2(side,mid));
1.415 +#endif
1.416 +
1.417 + return itheta;
1.418 +}
