From 35064811c0ac104acddd7777e00bfd9e054c2db6 Mon Sep 17 00:00:00 2001 From: hairball Date: Sat, 8 Feb 2014 03:21:02 +0000 Subject: Upgrade opus 1.0.2 -> 1.1 --- src/opus-1.0.2/celt/bands.c | 1302 ------------------------------------------- 1 file changed, 1302 deletions(-) delete mode 100644 src/opus-1.0.2/celt/bands.c (limited to 'src/opus-1.0.2/celt/bands.c') diff --git a/src/opus-1.0.2/celt/bands.c b/src/opus-1.0.2/celt/bands.c deleted file mode 100644 index 3be543c3..00000000 --- a/src/opus-1.0.2/celt/bands.c +++ /dev/null @@ -1,1302 +0,0 @@ -/* Copyright (c) 2007-2008 CSIRO - Copyright (c) 2007-2009 Xiph.Org Foundation - Copyright (c) 2008-2009 Gregory Maxwell - Written by Jean-Marc Valin and Gregory Maxwell */ -/* - Redistribution and use in source and binary forms, with or without - modification, are permitted provided that the following conditions - are met: - - - Redistributions of source code must retain the above copyright - notice, this list of conditions and the following disclaimer. - - - Redistributions in binary form must reproduce the above copyright - notice, this list of conditions and the following disclaimer in the - documentation and/or other materials provided with the distribution. - - THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER - OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, - EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, - PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR - PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF - LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING - NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS - SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. -*/ - -#ifdef HAVE_CONFIG_H -#include "config.h" -#endif - -#include -#include "bands.h" -#include "modes.h" -#include "vq.h" -#include "cwrs.h" -#include "stack_alloc.h" -#include "os_support.h" -#include "mathops.h" -#include "rate.h" - -opus_uint32 celt_lcg_rand(opus_uint32 seed) -{ - return 1664525 * seed + 1013904223; -} - -/* This is a cos() approximation designed to be bit-exact on any platform. Bit exactness - with this approximation is important because it has an impact on the bit allocation */ -static opus_int16 bitexact_cos(opus_int16 x) -{ - opus_int32 tmp; - opus_int16 x2; - tmp = (4096+((opus_int32)(x)*(x)))>>13; - celt_assert(tmp<=32767); - x2 = tmp; - x2 = (32767-x2) + FRAC_MUL16(x2, (-7651 + FRAC_MUL16(x2, (8277 + FRAC_MUL16(-626, x2))))); - celt_assert(x2<=32766); - return 1+x2; -} - -static int bitexact_log2tan(int isin,int icos) -{ - int lc; - int ls; - lc=EC_ILOG(icos); - ls=EC_ILOG(isin); - icos<<=15-lc; - isin<<=15-ls; - return (ls-lc)*(1<<11) - +FRAC_MUL16(isin, FRAC_MUL16(isin, -2597) + 7932) - -FRAC_MUL16(icos, FRAC_MUL16(icos, -2597) + 7932); -} - -#ifdef FIXED_POINT -/* Compute the amplitude (sqrt energy) in each of the bands */ -void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int M) -{ - int i, c, N; - const opus_int16 *eBands = m->eBands; - N = M*m->shortMdctSize; - c=0; do { - for (i=0;i 0) - { - int shift = celt_ilog2(maxval)-10; - j=M*eBands[i]; do { - sum = MAC16_16(sum, EXTRACT16(VSHR32(X[j+c*N],shift)), - EXTRACT16(VSHR32(X[j+c*N],shift))); - } while (++jnbEBands] = EPSILON+VSHR32(EXTEND32(celt_sqrt(sum)),-shift); - } else { - bandE[i+c*m->nbEBands] = EPSILON; - } - /*printf ("%f ", bandE[i+c*m->nbEBands]);*/ - } - } while (++ceBands; - N = M*m->shortMdctSize; - c=0; do { - i=0; do { - opus_val16 g; - int j,shift; - opus_val16 E; - shift = celt_zlog2(bandE[i+c*m->nbEBands])-13; - E = VSHR32(bandE[i+c*m->nbEBands], shift); - g = EXTRACT16(celt_rcp(SHL32(E,3))); - j=M*eBands[i]; do { - X[j+c*N] = MULT16_16_Q15(VSHR32(freq[j+c*N],shift-1),g); - } while (++jeBands; - N = M*m->shortMdctSize; - c=0; do { - for (i=0;inbEBands] = celt_sqrt(sum); - /*printf ("%f ", bandE[i+c*m->nbEBands]);*/ - } - } while (++ceBands; - N = M*m->shortMdctSize; - c=0; do { - for (i=0;inbEBands]); - for (j=M*eBands[i];jeBands; - N = M*m->shortMdctSize; - celt_assert2(C<=2, "denormalise_bands() not implemented for >2 channels"); - c=0; do { - celt_sig * OPUS_RESTRICT f; - const celt_norm * OPUS_RESTRICT x; - f = freq+c*N; - x = X+c*N; - for (i=0;inbEBands],1); - j=M*eBands[i]; - band_end = M*eBands[i+1]; - do { - *f++ = SHL32(MULT16_32_Q15(*x, g),2); - x++; - } while (++jeBands[i+1]-m->eBands[i]; - /* depth in 1/8 bits */ - depth = (1+pulses[i])/((m->eBands[i+1]-m->eBands[i])<>1; - t = SHL32(t, (7-shift)<<1); - sqrt_1 = celt_rsqrt_norm(t); - } -#else - thresh = .5f*celt_exp2(-.125f*depth); - sqrt_1 = celt_rsqrt(N0<nbEBands+i]; - prev2 = prev2logE[c*m->nbEBands+i]; - if (C==1) - { - prev1 = MAX16(prev1,prev1logE[m->nbEBands+i]); - prev2 = MAX16(prev2,prev2logE[m->nbEBands+i]); - } - Ediff = EXTEND32(logE[c*m->nbEBands+i])-EXTEND32(MIN16(prev1,prev2)); - Ediff = MAX32(0, Ediff); - -#ifdef FIXED_POINT - if (Ediff < 16384) - { - opus_val32 r32 = SHR32(celt_exp2(-EXTRACT16(Ediff)),1); - r = 2*MIN16(16383,r32); - } else { - r = 0; - } - if (LM==3) - r = MULT16_16_Q14(23170, MIN32(23169, r)); - r = SHR16(MIN16(thresh, r),1); - r = SHR32(MULT16_16_Q15(sqrt_1, r),shift); -#else - /* r needs to be multiplied by 2 or 2*sqrt(2) depending on LM because - short blocks don't have the same energy as long */ - r = 2.f*celt_exp2(-Ediff); - if (LM==3) - r *= 1.41421356f; - r = MIN16(thresh, r); - r = r*sqrt_1; -#endif - X = X_+c*size+(m->eBands[i]<nbEBands]))-13; -#endif - left = VSHR32(bandE[i],shift); - right = VSHR32(bandE[i+m->nbEBands],shift); - norm = EPSILON + celt_sqrt(EPSILON+MULT16_16(left,left)+MULT16_16(right,right)); - a1 = DIV32_16(SHL32(EXTEND32(left),14),norm); - a2 = DIV32_16(SHL32(EXTEND32(right),14),norm); - for (j=0;j>1; - kr = celt_ilog2(Er)>>1; -#endif - t = VSHR32(El, (kl-7)<<1); - lgain = celt_rsqrt_norm(t); - t = VSHR32(Er, (kr-7)<<1); - rgain = celt_rsqrt_norm(t); - -#ifdef FIXED_POINT - if (kl < 7) - kl = 7; - if (kr < 7) - kr = 7; -#endif - - for (j=0;jeBands; - int decision; - int hf_sum=0; - - celt_assert(end>0); - - N0 = M*m->shortMdctSize; - - if (M*(eBands[end]-eBands[end-1]) <= 8) - return SPREAD_NONE; - c=0; do { - for (i=0;im->nbEBands-4) - hf_sum += 32*(tcount[1]+tcount[0])/N; - tmp = (2*tcount[2] >= N) + (2*tcount[1] >= N) + (2*tcount[0] >= N); - sum += tmp*256; - nbBands++; - } - } while (++cnbEBands+end); - *hf_average = (*hf_average+hf_sum)>>1; - hf_sum = *hf_average; - if (*tapset_decision==2) - hf_sum += 4; - else if (*tapset_decision==0) - hf_sum -= 4; - if (hf_sum > 22) - *tapset_decision=2; - else if (hf_sum > 18) - *tapset_decision=1; - else - *tapset_decision=0; - } - /*printf("%d %d %d\n", hf_sum, *hf_average, *tapset_decision);*/ - celt_assert(nbBands>0); /*M*(eBands[end]-eBands[end-1]) <= 8 assures this*/ - sum /= nbBands; - /* Recursive averaging */ - sum = (sum+*average)>>1; - *average = sum; - /* Hysteresis */ - sum = (3*sum + (((3-last_decision)<<7) + 64) + 2)>>2; - if (sum < 80) - { - decision = SPREAD_AGGRESSIVE; - } else if (sum < 256) - { - decision = SPREAD_NORMAL; - } else if (sum < 384) - { - decision = SPREAD_LIGHT; - } else { - decision = SPREAD_NONE; - } -#ifdef FUZZING - decision = rand()&0x3; - *tapset_decision=rand()%3; -#endif - return decision; -} - -#ifdef MEASURE_NORM_MSE - -float MSE[30] = {0}; -int nbMSEBands = 0; -int MSECount[30] = {0}; - -void dump_norm_mse(void) -{ - int i; - for (i=0;inbEBands;i++) - { - int j; - int c; - float g; - if (bandE0[i]<10 || (C==2 && bandE0[i+m->nbEBands]<1)) - continue; - c=0; do { - g = bandE[i+c*m->nbEBands]/(1e-15+bandE0[i+c*m->nbEBands]); - for (j=M*m->eBands[i];jeBands[i+1];j++) - MSE[i] += (g*X[j+c*N]-X0[j+c*N])*(g*X[j+c*N]-X0[j+c*N]); - } while (++cnbEBands; -} - -#endif - -/* Indexing table for converting from natural Hadamard to ordery Hadamard - This is essentially a bit-reversed Gray, on top of which we've added - an inversion of the order because we want the DC at the end rather than - the beginning. The lines are for N=2, 4, 8, 16 */ -static const int ordery_table[] = { - 1, 0, - 3, 0, 2, 1, - 7, 0, 4, 3, 6, 1, 5, 2, - 15, 0, 8, 7, 12, 3, 11, 4, 14, 1, 9, 6, 13, 2, 10, 5, -}; - -static void deinterleave_hadamard(celt_norm *X, int N0, int stride, int hadamard) -{ - int i,j; - VARDECL(celt_norm, tmp); - int N; - SAVE_STACK; - N = N0*stride; - ALLOC(tmp, N, celt_norm); - celt_assert(stride>0); - if (hadamard) - { - const int *ordery = ordery_table+stride-2; - for (i=0;i>= 1; - for (i=0;i>1)) { - qn = 1; - } else { - qn = exp2_table8[qb&0x7]>>(14-(qb>>BITRES)); - qn = (qn+1)>>1<<1; - } - celt_assert(qn <= 256); - return qn; -} - -/* This function is responsible for encoding and decoding a band for both - the mono and stereo case. Even in the mono case, it can split the band - in two and transmit the energy difference with the two half-bands. It - can be called recursively so bands can end up being split in 8 parts. */ -static unsigned quant_band(int encode, const CELTMode *m, int i, celt_norm *X, celt_norm *Y, - int N, int b, int spread, int B, int intensity, int tf_change, celt_norm *lowband, ec_ctx *ec, - opus_int32 *remaining_bits, int LM, celt_norm *lowband_out, const celt_ener *bandE, int level, - opus_uint32 *seed, opus_val16 gain, celt_norm *lowband_scratch, int fill) -{ - const unsigned char *cache; - int q; - int curr_bits; - int stereo, split; - int imid=0, iside=0; - int N0=N; - int N_B=N; - int N_B0; - int B0=B; - int time_divide=0; - int recombine=0; - int inv = 0; - opus_val16 mid=0, side=0; - int longBlocks; - unsigned cm=0; -#ifdef RESYNTH - int resynth = 1; -#else - int resynth = !encode; -#endif - - longBlocks = B0==1; - - N_B /= B; - N_B0 = N_B; - - split = stereo = Y != NULL; - - /* Special case for one sample */ - if (N==1) - { - int c; - celt_norm *x = X; - c=0; do { - int sign=0; - if (*remaining_bits>=1<0) - recombine = tf_change; - /* Band recombining to increase frequency resolution */ - - if (lowband && (recombine || ((N_B&1) == 0 && tf_change<0) || B0>1)) - { - int j; - for (j=0;j>k, 1<>k, 1<>4]<<2; - } - B>>=recombine; - N_B<<=recombine; - - /* Increasing the time resolution */ - while ((N_B&1) == 0 && tf_change<0) - { - if (encode) - haar1(X, N_B, B); - if (lowband) - haar1(lowband, N_B, B); - fill |= fill<>= 1; - time_divide++; - tf_change++; - } - B0=B; - N_B0 = N_B; - - /* Reorganize the samples in time order instead of frequency order */ - if (B0>1) - { - if (encode) - deinterleave_hadamard(X, N_B>>recombine, B0<>recombine, B0<cache.bits + m->cache.index[(LM+1)*m->nbEBands+i]; - if (!stereo && LM != -1 && b > cache[cache[0]]+12 && N>2) - { - N >>= 1; - Y = X+N; - split = 1; - LM -= 1; - if (B==1) - fill = (fill&1)|(fill<<1); - B = (B+1)>>1; - } - - if (split) - { - int qn; - int itheta=0; - int mbits, sbits, delta; - int qalloc; - int pulse_cap; - int offset; - int orig_fill; - opus_int32 tell; - - /* Decide on the resolution to give to the split parameter theta */ - pulse_cap = m->logN[i]+LM*(1<>1) - (stereo&&N==2 ? QTHETA_OFFSET_TWOPHASE : QTHETA_OFFSET); - qn = compute_qn(N, b, offset, pulse_cap, stereo); - if (stereo && i>=intensity) - qn = 1; - if (encode) - { - /* theta is the atan() of the ratio between the (normalized) - side and mid. With just that parameter, we can re-scale both - mid and side because we know that 1) they have unit norm and - 2) they are orthogonal. */ - itheta = stereo_itheta(X, Y, stereo, N); - } - tell = ec_tell_frac(ec); - if (qn!=1) - { - if (encode) - itheta = (itheta*qn+8192)>>14; - - /* Entropy coding of the angle. We use a uniform pdf for the - time split, a step for stereo, and a triangular one for the rest. */ - if (stereo && N>2) - { - int p0 = 3; - int x = itheta; - int x0 = qn/2; - int ft = p0*(x0+1) + x0; - /* Use a probability of p0 up to itheta=8192 and then use 1 after */ - if (encode) - { - ec_encode(ec,x<=x0?p0*x:(x-1-x0)+(x0+1)*p0,x<=x0?p0*(x+1):(x-x0)+(x0+1)*p0,ft); - } else { - int fs; - fs=ec_decode(ec,ft); - if (fs<(x0+1)*p0) - x=fs/p0; - else - x=x0+1+(fs-(x0+1)*p0); - 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); - itheta = x; - } - } else if (B0>1 || stereo) { - /* Uniform pdf */ - if (encode) - ec_enc_uint(ec, itheta, qn+1); - else - itheta = ec_dec_uint(ec, qn+1); - } else { - int fs=1, ft; - ft = ((qn>>1)+1)*((qn>>1)+1); - if (encode) - { - int fl; - - fs = itheta <= (qn>>1) ? itheta + 1 : qn + 1 - itheta; - fl = itheta <= (qn>>1) ? itheta*(itheta + 1)>>1 : - ft - ((qn + 1 - itheta)*(qn + 2 - itheta)>>1); - - ec_encode(ec, fl, fl+fs, ft); - } else { - /* Triangular pdf */ - int fl=0; - int fm; - fm = ec_decode(ec, ft); - - if (fm < ((qn>>1)*((qn>>1) + 1)>>1)) - { - itheta = (isqrt32(8*(opus_uint32)fm + 1) - 1)>>1; - fs = itheta + 1; - fl = itheta*(itheta + 1)>>1; - } - else - { - itheta = (2*(qn + 1) - - isqrt32(8*(opus_uint32)(ft - fm - 1) + 1))>>1; - fs = qn + 1 - itheta; - fl = ft - ((qn + 1 - itheta)*(qn + 2 - itheta)>>1); - } - - ec_dec_update(ec, fl, fl+fs, ft); - } - } - itheta = (opus_int32)itheta*16384/qn; - if (encode && stereo) - { - if (itheta==0) - intensity_stereo(m, X, Y, bandE, i, N); - else - stereo_split(X, Y, N); - } - /* NOTE: Renormalising X and Y *may* help fixed-point a bit at very high rate. - Let's do that at higher complexity */ - } else if (stereo) { - if (encode) - { - inv = itheta > 8192; - if (inv) - { - int j; - for (j=0;j2< 2< 8192; - *remaining_bits -= qalloc+sbits; - - x2 = c ? Y : X; - y2 = c ? X : Y; - if (sbits) - { - if (encode) - { - /* Here we only need to encode a sign for the side */ - sign = x2[0]*y2[1] - x2[1]*y2[0] < 0; - ec_enc_bits(ec, sign, 1); - } else { - sign = ec_dec_bits(ec, 1); - } - } - sign = 1-2*sign; - /* We use orig_fill here because we want to fold the side, but if - itheta==16384, we'll have cleared the low bits of fill. */ - 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); - /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse), - and there's no need to worry about mixing with the other channel. */ - y2[0] = -sign*x2[1]; - y2[1] = sign*x2[0]; - if (resynth) - { - celt_norm tmp; - X[0] = MULT16_16_Q15(mid, X[0]); - X[1] = MULT16_16_Q15(mid, X[1]); - Y[0] = MULT16_16_Q15(side, Y[0]); - Y[1] = MULT16_16_Q15(side, Y[1]); - tmp = X[0]; - X[0] = SUB16(tmp,Y[0]); - Y[0] = ADD16(tmp,Y[0]); - tmp = X[1]; - X[1] = SUB16(tmp,Y[1]); - Y[1] = ADD16(tmp,Y[1]); - } - } else { - /* "Normal" split code */ - celt_norm *next_lowband2=NULL; - celt_norm *next_lowband_out1=NULL; - int next_level=0; - opus_int32 rebalance; - - /* Give more bits to low-energy MDCTs than they would otherwise deserve */ - if (B0>1 && !stereo && (itheta&0x3fff)) - { - if (itheta > 8192) - /* Rough approximation for pre-echo masking */ - delta -= delta>>(4-LM); - else - /* Corresponds to a forward-masking slope of 1.5 dB per 10 ms */ - delta = IMIN(0, delta + (N<>(5-LM))); - } - mbits = IMAX(0, IMIN(b, (b-delta)/2)); - sbits = b-mbits; - *remaining_bits -= qalloc; - - if (lowband && !stereo) - next_lowband2 = lowband+N; /* >32-bit split case */ - - /* Only stereo needs to pass on lowband_out. Otherwise, it's - handled at the end */ - if (stereo) - next_lowband_out1 = lowband_out; - else - next_level = level+1; - - rebalance = *remaining_bits; - if (mbits >= sbits) - { - /* In stereo mode, we do not apply a scaling to the mid because we need the normalized - mid for folding later */ - cm = quant_band(encode, m, i, X, NULL, N, mbits, spread, B, intensity, tf_change, - lowband, ec, remaining_bits, LM, next_lowband_out1, - NULL, next_level, seed, stereo ? Q15ONE : MULT16_16_P15(gain,mid), lowband_scratch, fill); - rebalance = mbits - (rebalance-*remaining_bits); - if (rebalance > 3<>B)<<((B0>>1)&(stereo-1)); - } else { - /* For a stereo split, the high bits of fill are always zero, so no - folding will be done to the side. */ - cm = quant_band(encode, m, i, Y, NULL, N, sbits, spread, B, intensity, tf_change, - next_lowband2, ec, remaining_bits, LM, NULL, - NULL, next_level, seed, MULT16_16_P15(gain,side), NULL, fill>>B)<<((B0>>1)&(stereo-1)); - rebalance = sbits - (rebalance-*remaining_bits); - if (rebalance > 3< 0) - { - *remaining_bits += curr_bits; - q--; - curr_bits = pulses2bits(m, i, LM, q); - *remaining_bits -= curr_bits; - } - - if (q!=0) - { - int K = get_pulses(q); - - /* Finally do the actual quantization */ - if (encode) - { - cm = alg_quant(X, N, K, spread, B, ec -#ifdef RESYNTH - , gain -#endif - ); - } else { - cm = alg_unquant(X, N, K, spread, B, ec, gain); - } - } else { - /* If there's no pulse, fill the band anyway */ - int j; - if (resynth) - { - unsigned cm_mask; - /*B can be as large as 16, so this shift might overflow an int on a - 16-bit platform; use a long to get defined behavior.*/ - cm_mask = (unsigned)(1UL<>20); - } - cm = cm_mask; - } else { - /* Folded spectrum */ - for (j=0;j1) - interleave_hadamard(X, N_B>>recombine, B0<>= 1; - N_B <<= 1; - cm |= cm>>B; - haar1(X, N_B, B); - } - - for (k=0;k>k, 1<eBands; - celt_norm * OPUS_RESTRICT norm, * OPUS_RESTRICT norm2; - VARDECL(celt_norm, _norm); - VARDECL(celt_norm, lowband_scratch); - int B; - int M; - int lowband_offset; - int update_lowband = 1; - int C = Y_ != NULL ? 2 : 1; -#ifdef RESYNTH - int resynth = 1; -#else - int resynth = !encode; -#endif - SAVE_STACK; - - M = 1<nbEBands], celt_norm); - ALLOC(lowband_scratch, M*(eBands[m->nbEBands]-eBands[m->nbEBands-1]), celt_norm); - norm = _norm; - norm2 = norm + M*eBands[m->nbEBands]; - - lowband_offset = 0; - for (i=start;i= M*eBands[start] && (update_lowband || lowband_offset==0)) - lowband_offset = i; - - tf_change = tf_res[i]; - if (i>=m->effEBands) - { - X=norm; - if (Y_!=NULL) - Y = norm; - } - - /* Get a conservative estimate of the collapse_mask's for the bands we're - going to be folding from. */ - if (lowband_offset != 0 && (spread!=SPREAD_AGGRESSIVE || B>1 || tf_change<0)) - { - int fold_start; - int fold_end; - int fold_i; - /* This ensures we never repeat spectral content within one band */ - effective_lowband = IMAX(M*eBands[start], M*eBands[lowband_offset]-N); - fold_start = lowband_offset; - while(M*eBands[--fold_start] > effective_lowband); - fold_end = lowband_offset-1; - while(M*eBands[++fold_end] < effective_lowband+N); - x_cm = y_cm = 0; - fold_i = fold_start; do { - x_cm |= collapse_masks[fold_i*C+0]; - y_cm |= collapse_masks[fold_i*C+C-1]; - } while (++fold_i(N<