[Opus] Update to v1.2.1

third_party/opus/src/celt/bands.c

 /* 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
@@ skipping 47 matching lines @@
    return i;
 }
 
 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_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 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 LM)
+void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM, int arch)
 {
    int i, c, N;
    const opus_int16 *eBands = m->eBands;
+   (void)arch;
    N = m->shortMdctSize<<LM;
    c=0; do {
       for (i=0;i<end;i++)
       {
          int j;
          opus_val32 maxval=0;
          opus_val32 sum = 0;
 
          maxval = celt_maxabs32(&X[c*N+(eBands[i]<<LM)], (eBands[i+1]-eBands[i])<<LM);
          if (maxval > 0)
@@ skipping 39 matching lines @@
          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 (++j<M*eBands[i+1]);
       } while (++i<end);
    } while (++c<C);
 }
 
 #else /* 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 LM)
+void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM, int arch)
 {
    int i, c, N;
    const opus_int16 *eBands = m->eBands;
    N = m->shortMdctSize<<LM;
    c=0; do {
       for (i=0;i<end;i++)
       {
          opus_val32 sum;
-         sum = 1e-27f + celt_inner_prod_c(&X[c*N+(eBands[i]<<LM)], &X[c*N+(eBands[i]<<LM)], (eBands[i+1]-eBands[i])<<LM);
+         sum = 1e-27f + celt_inner_prod(&X[c*N+(eBands[i]<<LM)], &X[c*N+(eBands[i]<<LM)], (eBands[i+1]-eBands[i])<<LM, arch);
          bandE[i+c*m->nbEBands] = celt_sqrt(sum);
          /*printf ("%f ", bandE[i+c*m->nbEBands]);*/
       }
    } while (++c<C);
    /*printf ("\n");*/
 }
 
 /* Normalise each band such that the energy is one. */
 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)
 {
@@ skipping 39 matching lines @@
    for (i=start;i<end;i++)
    {
       int j, band_end;
       opus_val16 g;
       opus_val16 lg;
 #ifdef FIXED_POINT
       int shift;
 #endif
       j=M*eBands[i];
       band_end = M*eBands[i+1];
-      lg = ADD16(bandLogE[i], SHL16((opus_val16)eMeans[i],6));
+      lg = SATURATE16(ADD32(bandLogE[i], SHL32((opus_val32)eMeans[i],6)));
 #ifndef FIXED_POINT
       g = celt_exp2(MIN32(32.f, lg));
 #else
       /* Handle the integer part of the log energy */
       shift = 16-(lg>>DB_SHIFT);
       if (shift>31)
       {
          shift=0;
          g=0;
       } else {
          /* Handle the fractional part. */
          g = celt_exp2_frac(lg&((1<<DB_SHIFT)-1));
       }
       /* Handle extreme gains with negative shift. */
       if (shift<0)
       {
-         /* For shift < -2 we'd be likely to overflow, so we're capping
-               the gain here. This shouldn't happen unless the bitstream is
-               already corrupted. */
-         if (shift < -2)
+         /* For shift <= -2 and g > 16384 we'd be likely to overflow, so we're
+            capping the gain here, which is equivalent to a cap of 18 on lg.
+            This shouldn't trigger unless the bitstream is already corrupted. */
+         if (shift <= -2)
          {
-            g = 32767;
+            g = 16384;
             shift = -2;
          }
          do {
             *f++ = SHL32(MULT16_16(*x++, g), -shift);
          } while (++j<band_end);
       } else
 #endif
          /* Be careful of the fixed-point "else" just above when changing this code */
          do {
             *f++ = SHR32(MULT16_16(*x++, g), shift);
@@ skipping 93 matching lines @@
                renormalize = 1;
             }
          }
          /* We just added some energy, so we need to renormalise */
          if (renormalize)
             renormalise_vector(X, N0<<LM, Q15ONE, arch);
       } while (++c<C);
    }
 }
 
+/* Compute the weights to use for optimizing normalized distortion across
+   channels. We use the amplitude to weight square distortion, which means
+   that we use the square root of the value we would have been using if we
+   wanted to minimize the MSE in the non-normalized domain. This roughly
+   corresponds to some quick-and-dirty perceptual experiments I ran to
+   measure inter-aural masking (there doesn't seem to be any published data
+   on the topic). */
+static void compute_channel_weights(celt_ener Ex, celt_ener Ey, opus_val16 w[2])
+{
+   celt_ener minE;
+#if FIXED_POINT
+   int shift;
+#endif
+   minE = MIN32(Ex, Ey);
+   /* Adjustment to make the weights a bit more conservative. */
+   Ex = ADD32(Ex, minE/3);
+   Ey = ADD32(Ey, minE/3);
+#if FIXED_POINT
+   shift = celt_ilog2(EPSILON+MAX32(Ex, Ey))-14;
+#endif
+   w[0] = VSHR32(Ex, shift);
+   w[1] = VSHR32(Ey, shift);
+}
+
 static void intensity_stereo(const CELTMode *m, celt_norm * OPUS_RESTRICT X, const celt_norm * OPUS_RESTRICT Y, const celt_ener *bandE, int bandID, int N)
 {
    int i = bandID;
    int j;
    opus_val16 a1, a2;
    opus_val16 left, right;
    opus_val16 norm;
 #ifdef FIXED_POINT
    int shift = celt_zlog2(MAX32(bandE[i], bandE[i+m->nbEBands]))-13;
 #endif
@@ skipping 267 matching lines @@
    } else {
       qn = exp2_table8[qb&0x7]>>(14-(qb>>BITRES));
       qn = (qn+1)>>1<<1;
    }
    celt_assert(qn <= 256);
    return qn;
 }
 
 struct band_ctx {
    int encode;
+   int resynth;
    const CELTMode *m;
    int i;
    int intensity;
    int spread;
    int tf_change;
    ec_ctx *ec;
    opus_int32 remaining_bits;
    const celt_ener *bandE;
    opus_uint32 seed;
    int arch;
+   int theta_round;
+   int disable_inv;
+   int avoid_split_noise;
 };
 
 struct split_ctx {
    int inv;
    int imid;
    int iside;
    int delta;
    int itheta;
    int qalloc;
 };
@@ skipping 37 matching lines @@
       /* 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, ctx->arch);
    }
    tell = ec_tell_frac(ec);
    if (qn!=1)
    {
       if (encode)
-         itheta = (itheta*(opus_int32)qn+8192)>>14;
-
+      {
+         if (!stereo || ctx->theta_round == 0)
+         {
+            itheta = (itheta*(opus_int32)qn+8192)>>14;
+            if (!stereo && ctx->avoid_split_noise && itheta > 0 && itheta < qn)
+            {
+               /* Check if the selected value of theta will cause the bit allocation
+                  to inject noise on one side. If so, make sure the energy of that side
+                  is zero. */
+               int unquantized = celt_udiv((opus_int32)itheta*16384, qn);
+               imid = bitexact_cos((opus_int16)unquantized);
+               iside = bitexact_cos((opus_int16)(16384-unquantized));
+               delta = FRAC_MUL16((N-1)<<7,bitexact_log2tan(iside,imid));
+               if (delta > *b)
+                  itheta = qn;
+               else if (delta < -*b)
+                  itheta = 0;
+            }
+         } else {
+            int down;
+            /* Bias quantization towards itheta=0 and itheta=16384. */
+            int bias = itheta > 8192 ? 32767/qn : -32767/qn;
+            down = IMIN(qn-1, IMAX(0, (itheta*(opus_int32)qn + bias)>>14));
+            if (ctx->theta_round < 0)
+               itheta = down;
+            else
+               itheta = down+1;
+         }
+      }
       /* 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)
@@ skipping 57 matching lines @@
          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;
+         inv = itheta > 8192 && !ctx->disable_inv;
          if (inv)
          {
             int j;
             for (j=0;j<N;j++)
                Y[j] = -Y[j];
          }
          intensity_stereo(m, X, Y, bandE, i, N);
       }
       if (*b>2<<BITRES && ctx->remaining_bits > 2<<BITRES)
       {
          if (encode)
             ec_enc_bit_logp(ec, inv, 2);
          else
             inv = ec_dec_bit_logp(ec, 2);
       } else
          inv = 0;
+      /* inv flag override to avoid problems with downmixing. */
+      if (ctx->disable_inv)
+         inv = 0;
       itheta = 0;
    }
    qalloc = ec_tell_frac(ec) - tell;
    *b -= qalloc;
 
    if (itheta == 0)
    {
       imid = 32767;
       iside = 0;
       *fill &= (1<<B)-1;
@@ skipping 15 matching lines @@
    sctx->inv = inv;
    sctx->imid = imid;
    sctx->iside = iside;
    sctx->delta = delta;
    sctx->itheta = itheta;
    sctx->qalloc = qalloc;
 }
 static unsigned quant_band_n1(struct band_ctx *ctx, celt_norm *X, celt_norm *Y, int b,
       celt_norm *lowband_out)
 {
-#ifdef RESYNTH
-   int resynth = 1;
-#else
-   int resynth = !ctx->encode;
-#endif
    int c;
    int stereo;
    celt_norm *x = X;
    int encode;
    ec_ctx *ec;
 
    encode = ctx->encode;
    ec = ctx->ec;
 
    stereo = Y != NULL;
    c=0; do {
       int sign=0;
       if (ctx->remaining_bits>=1<<BITRES)
       {
          if (encode)
          {
             sign = x[0]<0;
             ec_enc_bits(ec, sign, 1);
          } else {
             sign = ec_dec_bits(ec, 1);
          }
          ctx->remaining_bits -= 1<<BITRES;
          b-=1<<BITRES;
       }
-      if (resynth)
+      if (ctx->resynth)
          x[0] = sign ? -NORM_SCALING : NORM_SCALING;
       x = Y;
    } while (++c<1+stereo);
    if (lowband_out)
       lowband_out[0] = SHR16(X[0],4);
    return 1;
 }
 
 /* This function is responsible for encoding and decoding a mono partition.
    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_partition(struct band_ctx *ctx, celt_norm *X,
       int N, int b, int B, celt_norm *lowband,
       int LM,
       opus_val16 gain, int fill)
 {
    const unsigned char *cache;
    int q;
    int curr_bits;
    int imid=0, iside=0;
    int B0=B;
    opus_val16 mid=0, side=0;
    unsigned cm=0;
-#ifdef RESYNTH
-   int resynth = 1;
-#else
-   int resynth = !ctx->encode;
-#endif
    celt_norm *Y=NULL;
    int encode;
    const CELTMode *m;
    int i;
    int spread;
    ec_ctx *ec;
 
    encode = ctx->encode;
    m = ctx->m;
    i = ctx->i;
@@ skipping 11 matching lines @@
       celt_norm *next_lowband2=NULL;
       opus_int32 rebalance;
 
       N >>= 1;
       Y = X+N;
       LM -= 1;
       if (B==1)
          fill = (fill&1)|(fill<<1);
       B = (B+1)>>1;
 
-      compute_theta(ctx, &sctx, X, Y, N, &b, B, B0,
-            LM, 0, &fill);
+      compute_theta(ctx, &sctx, X, Y, N, &b, B, B0, LM, 0, &fill);
       imid = sctx.imid;
       iside = sctx.iside;
       delta = sctx.delta;
       itheta = sctx.itheta;
       qalloc = sctx.qalloc;
 #ifdef FIXED_POINT
       mid = imid;
       side = iside;
 #else
       mid = (1.f/32768)*imid;
@@ skipping 13 matching lines @@
       mbits = IMAX(0, IMIN(b, (b-delta)/2));
       sbits = b-mbits;
       ctx->remaining_bits -= qalloc;
 
       if (lowband)
          next_lowband2 = lowband+N; /* >32-bit split case */
 
       rebalance = ctx->remaining_bits;
       if (mbits >= sbits)
       {
-         cm = quant_partition(ctx, X, N, mbits, B,
-               lowband, LM,
+         cm = quant_partition(ctx, X, N, mbits, B, lowband, LM,
                MULT16_16_P15(gain,mid), fill);
          rebalance = mbits - (rebalance-ctx->remaining_bits);
          if (rebalance > 3<<BITRES && itheta!=0)
             sbits += rebalance - (3<<BITRES);
-         cm |= quant_partition(ctx, Y, N, sbits, B,
-               next_lowband2, LM,
+         cm |= quant_partition(ctx, Y, N, sbits, B, next_lowband2, LM,
                MULT16_16_P15(gain,side), fill>>B)<<(B0>>1);
       } else {
-         cm = quant_partition(ctx, Y, N, sbits, B,
-               next_lowband2, LM,
+         cm = quant_partition(ctx, Y, N, sbits, B, next_lowband2, LM,
                MULT16_16_P15(gain,side), fill>>B)<<(B0>>1);
          rebalance = sbits - (rebalance-ctx->remaining_bits);
          if (rebalance > 3<<BITRES && itheta!=16384)
             mbits += rebalance - (3<<BITRES);
-         cm |= quant_partition(ctx, X, N, mbits, B,
-               lowband, LM,
+         cm |= quant_partition(ctx, X, N, mbits, B, lowband, LM,
                MULT16_16_P15(gain,mid), fill);
       }
    } else {
       /* This is the basic no-split case */
       q = bits2pulses(m, i, LM, b);
       curr_bits = pulses2bits(m, i, LM, q);
       ctx->remaining_bits -= curr_bits;
 
       /* Ensures we can never bust the budget */
       while (ctx->remaining_bits < 0 && q > 0)
       {
          ctx->remaining_bits += curr_bits;
          q--;
          curr_bits = pulses2bits(m, i, LM, q);
          ctx->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
-                 );
+            cm = alg_quant(X, N, K, spread, B, ec, gain, ctx->resynth, ctx->arch);
          } 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)
+         if (ctx->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<<B)-1;
             fill &= cm_mask;
             if (!fill)
             {
                OPUS_CLEAR(X, N);
             } else {
@@ skipping 36 matching lines @@
       opus_val16 gain, celt_norm *lowband_scratch, int fill)
 {
    int N0=N;
    int N_B=N;
    int N_B0;
    int B0=B;
    int time_divide=0;
    int recombine=0;
    int longBlocks;
    unsigned cm=0;
-#ifdef RESYNTH
-   int resynth = 1;
-#else
-   int resynth = !ctx->encode;
-#endif
    int k;
    int encode;
    int tf_change;
 
    encode = ctx->encode;
    tf_change = ctx->tf_change;
 
    longBlocks = B0==1;
 
    N_B = celt_udiv(N_B, B);
@@ skipping 46 matching lines @@
 
    /* Reorganize the samples in time order instead of frequency order */
    if (B0>1)
    {
       if (encode)
          deinterleave_hadamard(X, N_B>>recombine, B0<<recombine, longBlocks);
       if (lowband)
          deinterleave_hadamard(lowband, N_B>>recombine, B0<<recombine, longBlocks);
    }
 
-   cm = quant_partition(ctx, X, N, b, B, lowband,
-         LM, gain, fill);
+   cm = quant_partition(ctx, X, N, b, B, lowband, LM, gain, fill);
 
    /* This code is used by the decoder and by the resynthesis-enabled encoder */
-   if (resynth)
+   if (ctx->resynth)
    {
       /* Undo the sample reorganization going from time order to frequency order */
       if (B0>1)
          interleave_hadamard(X, N_B>>recombine, B0<<recombine, longBlocks);
 
       /* Undo time-freq changes that we did earlier */
       N_B = N_B0;
       B = B0;
       for (k=0;k<time_divide;k++)
       {
@@ skipping 32 matching lines @@
 /* This function is responsible for encoding and decoding a band for the stereo case. */
 static unsigned quant_band_stereo(struct band_ctx *ctx, celt_norm *X, celt_norm *Y,
       int N, int b, int B, celt_norm *lowband,
       int LM, celt_norm *lowband_out,
       celt_norm *lowband_scratch, int fill)
 {
    int imid=0, iside=0;
    int inv = 0;
    opus_val16 mid=0, side=0;
    unsigned cm=0;
-#ifdef RESYNTH
-   int resynth = 1;
-#else
-   int resynth = !ctx->encode;
-#endif
    int mbits, sbits, delta;
    int itheta;
    int qalloc;
    struct split_ctx sctx;
    int orig_fill;
    int encode;
    ec_ctx *ec;
 
    encode = ctx->encode;
    ec = ctx->ec;
 
    /* Special case for one sample */
    if (N==1)
    {
       return quant_band_n1(ctx, X, Y, b, lowband_out);
    }
 
    orig_fill = fill;
 
-   compute_theta(ctx, &sctx, X, Y, N, &b, B, B,
-         LM, 1, &fill);
+   compute_theta(ctx, &sctx, X, Y, N, &b, B, B, LM, 1, &fill);
    inv = sctx.inv;
    imid = sctx.imid;
    iside = sctx.iside;
    delta = sctx.delta;
    itheta = sctx.itheta;
    qalloc = sctx.qalloc;
 #ifdef FIXED_POINT
    mid = imid;
    side = iside;
 #else
@@ skipping 27 matching lines @@
             /* 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(ctx, x2, N, mbits, B, lowband,
-            LM, lowband_out, Q15ONE, lowband_scratch, orig_fill);
+      cm = quant_band(ctx, x2, N, mbits, B, lowband, LM, lowband_out, Q15ONE,
+            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)
+      if (ctx->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 */
       opus_int32 rebalance;
 
       mbits = IMAX(0, IMIN(b, (b-delta)/2));
       sbits = b-mbits;
       ctx->remaining_bits -= qalloc;
 
       rebalance = ctx->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(ctx, X, N, mbits, B,
-               lowband, LM, lowband_out,
-               Q15ONE, lowband_scratch, fill);
+         cm = quant_band(ctx, X, N, mbits, B, lowband, LM, lowband_out, Q15ONE,
+               lowband_scratch, fill);
          rebalance = mbits - (rebalance-ctx->remaining_bits);
          if (rebalance > 3<<BITRES && itheta!=0)
             sbits += rebalance - (3<<BITRES);
 
          /* For a stereo split, the high bits of fill are always zero, so no
             folding will be done to the side. */
-         cm |= quant_band(ctx, Y, N, sbits, B,
-               NULL, LM, NULL,
-               side, NULL, fill>>B);
+         cm |= quant_band(ctx, Y, N, sbits, B, NULL, LM, NULL, side, NULL, fill>>B);
       } 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(ctx, Y, N, sbits, B,
-               NULL, LM, NULL,
-               side, NULL, fill>>B);
+         cm = quant_band(ctx, Y, N, sbits, B, NULL, LM, NULL, side, NULL, fill>>B);
          rebalance = sbits - (rebalance-ctx->remaining_bits);
          if (rebalance > 3<<BITRES && itheta!=16384)
             mbits += rebalance - (3<<BITRES);
          /* In stereo mode, we do not apply a scaling to the mid because we need the normalized
             mid for folding later. */
-         cm |= quant_band(ctx, X, N, mbits, B,
-               lowband, LM, lowband_out,
-               Q15ONE, lowband_scratch, fill);
+         cm |= quant_band(ctx, X, N, mbits, B, lowband, LM, lowband_out, Q15ONE,
+               lowband_scratch, fill);
       }
    }
 
 
    /* This code is used by the decoder and by the resynthesis-enabled encoder */
-   if (resynth)
+   if (ctx->resynth)
    {
       if (N!=2)
          stereo_merge(X, Y, mid, N, ctx->arch);
       if (inv)
       {
          int j;
          for (j=0;j<N;j++)
             Y[j] = -Y[j];
       }
    }
    return cm;
 }
 
+static void special_hybrid_folding(const CELTMode *m, celt_norm *norm, celt_norm *norm2, int start, int M, int dual_stereo)
+{
+   int n1, n2;
+   const opus_int16 * OPUS_RESTRICT eBands = m->eBands;
+   n1 = M*(eBands[start+1]-eBands[start]);
+   n2 = M*(eBands[start+2]-eBands[start+1]);
+   /* Duplicate enough of the first band folding data to be able to fold the second band.
+      Copies no data for CELT-only mode. */
+   OPUS_COPY(&norm[n1], &norm[2*n1 - n2], n2-n1);
+   if (dual_stereo)
+      OPUS_COPY(&norm2[n1], &norm2[2*n1 - n2], n2-n1);
+}
 
 void quant_all_bands(int encode, const CELTMode *m, int start, int end,
       celt_norm *X_, celt_norm *Y_, unsigned char *collapse_masks,
       const celt_ener *bandE, int *pulses, int shortBlocks, int spread,
       int dual_stereo, int intensity, int *tf_res, opus_int32 total_bits,
       opus_int32 balance, ec_ctx *ec, int LM, int codedBands,
-      opus_uint32 *seed, int arch)
+      opus_uint32 *seed, int complexity, int arch, int disable_inv)
 {
    int i;
    opus_int32 remaining_bits;
    const opus_int16 * OPUS_RESTRICT eBands = m->eBands;
    celt_norm * OPUS_RESTRICT norm, * OPUS_RESTRICT norm2;
    VARDECL(celt_norm, _norm);
+   VARDECL(celt_norm, _lowband_scratch);
+   VARDECL(celt_norm, X_save);
+   VARDECL(celt_norm, Y_save);
+   VARDECL(celt_norm, X_save2);
+   VARDECL(celt_norm, Y_save2);
+   VARDECL(celt_norm, norm_save2);
+   int resynth_alloc;
    celt_norm *lowband_scratch;
    int B;
    int M;
    int lowband_offset;
    int update_lowband = 1;
    int C = Y_ != NULL ? 2 : 1;
    int norm_offset;
+   int theta_rdo = encode && Y_!=NULL && !dual_stereo && complexity>=8;
 #ifdef RESYNTH
    int resynth = 1;
 #else
-   int resynth = !encode;
+   int resynth = !encode || theta_rdo;
 #endif
    struct band_ctx ctx;
    SAVE_STACK;
 
    M = 1<<LM;
    B = shortBlocks ? M : 1;
    norm_offset = M*eBands[start];
    /* No need to allocate norm for the last band because we don't need an
       output in that band. */
    ALLOC(_norm, C*(M*eBands[m->nbEBands-1]-norm_offset), celt_norm);
    norm = _norm;
    norm2 = norm + M*eBands[m->nbEBands-1]-norm_offset;
-   /* We can use the last band as scratch space because we don't need that
-      scratch space for the last band. */
-   lowband_scratch = X_+M*eBands[m->nbEBands-1];
+
+   /* For decoding, we can use the last band as scratch space because we don't need that
+      scratch space for the last band and we don't care about the data there until we're
+      decoding the last band. */
+   if (encode && resynth)
+      resynth_alloc = M*(eBands[m->nbEBands]-eBands[m->nbEBands-1]);
+   else
+      resynth_alloc = ALLOC_NONE;
+   ALLOC(_lowband_scratch, resynth_alloc, celt_norm);
+   if (encode && resynth)
+      lowband_scratch = _lowband_scratch;
+   else
+      lowband_scratch = X_+M*eBands[m->nbEBands-1];
+   ALLOC(X_save, resynth_alloc, celt_norm);
+   ALLOC(Y_save, resynth_alloc, celt_norm);
+   ALLOC(X_save2, resynth_alloc, celt_norm);
+   ALLOC(Y_save2, resynth_alloc, celt_norm);
+   ALLOC(norm_save2, resynth_alloc, celt_norm);
 
    lowband_offset = 0;
    ctx.bandE = bandE;
    ctx.ec = ec;
    ctx.encode = encode;
    ctx.intensity = intensity;
    ctx.m = m;
    ctx.seed = *seed;
    ctx.spread = spread;
    ctx.arch = arch;
+   ctx.disable_inv = disable_inv;
+   ctx.resynth = resynth;
+   ctx.theta_round = 0;
+   /* Avoid injecting noise in the first band on transients. */
+   ctx.avoid_split_noise = B > 1;
    for (i=start;i<end;i++)
    {
       opus_int32 tell;
       int b;
       int N;
       opus_int32 curr_balance;
       int effective_lowband=-1;
       celt_norm * OPUS_RESTRICT X, * OPUS_RESTRICT Y;
       int tf_change=0;
       unsigned x_cm;
@@ skipping 17 matching lines @@
       remaining_bits = total_bits-tell-1;
       ctx.remaining_bits = remaining_bits;
       if (i <= codedBands-1)
       {
          curr_balance = celt_sudiv(balance, IMIN(3, codedBands-i));
          b = IMAX(0, IMIN(16383, IMIN(remaining_bits+1,pulses[i]+curr_balance)));
       } else {
          b = 0;
       }
 
+#ifdef ENABLE_UPDATE_DRAFT
+      if (resynth && (M*eBands[i]-N >= M*eBands[start] || i==start+1) && (update_lowband || lowband_offset==0))
+            lowband_offset = i;
+      if (i == start+1)
+         special_hybrid_folding(m, norm, norm2, start, M, dual_stereo);
+#else
       if (resynth && M*eBands[i]-N >= M*eBands[start] && (update_lowband || lowband_offset==0))
             lowband_offset = i;
+#endif
 
       tf_change = tf_res[i];
       ctx.tf_change = tf_change;
       if (i>=m->effEBands)
       {
          X=norm;
          if (Y_!=NULL)
             Y = norm;
          lowband_scratch = NULL;
       }
-      if (i==end-1)
+      if (last && !theta_rdo)
          lowband_scratch = NULL;
 
       /* 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(0, M*eBands[lowband_offset]-norm_offset-N);
          fold_start = lowband_offset;
          while(M*eBands[--fold_start] > effective_lowband+norm_offset);
          fold_end = lowband_offset-1;
+#ifdef ENABLE_UPDATE_DRAFT
+         while(++fold_end < i && M*eBands[fold_end] < effective_lowband+norm_offset+N);
+#else
          while(M*eBands[++fold_end] < effective_lowband+norm_offset+N);
+#endif
          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<fold_end);
       }
       /* Otherwise, we'll be using the LCG to fold, so all blocks will (almost
          always) be non-zero. */
       else
          x_cm = y_cm = (1<<B)-1;
@@ skipping 12 matching lines @@
       {
          x_cm = quant_band(&ctx, X, N, b/2, B,
                effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
                last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm);
          y_cm = quant_band(&ctx, Y, N, b/2, B,
                effective_lowband != -1 ? norm2+effective_lowband : NULL, LM,
                last?NULL:norm2+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, y_cm);
       } else {
          if (Y!=NULL)
          {
-            x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
-                  effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
-                        last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, x_cm|y_cm);
+            if (theta_rdo && i < intensity)
+            {
+               ec_ctx ec_save, ec_save2;
+               struct band_ctx ctx_save, ctx_save2;
+               opus_val32 dist0, dist1;
+               unsigned cm, cm2;
+               int nstart_bytes, nend_bytes, save_bytes;
+               unsigned char *bytes_buf;
+               unsigned char bytes_save[1275];
+               opus_val16 w[2];
+               compute_channel_weights(bandE[i], bandE[i+m->nbEBands], w);
+               /* Make a copy. */
+               cm = x_cm|y_cm;
+               ec_save = *ec;
+               ctx_save = ctx;
+               OPUS_COPY(X_save, X, N);
+               OPUS_COPY(Y_save, Y, N);
+               /* Encode and round down. */
+               ctx.theta_round = -1;
+               x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
+                     effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
+                     last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, cm);
+               dist0 = MULT16_32_Q15(w[0], celt_inner_prod(X_save, X, N, arch)) + MULT16_32_Q15(w[1], celt_inner_prod(Y_save, Y, N, arch));
+
+               /* Save first result. */
+               cm2 = x_cm;
+               ec_save2 = *ec;
+               ctx_save2 = ctx;
+               OPUS_COPY(X_save2, X, N);
+               OPUS_COPY(Y_save2, Y, N);
+               if (!last)
+                  OPUS_COPY(norm_save2, norm+M*eBands[i]-norm_offset, N);
+               nstart_bytes = ec_save.offs;
+               nend_bytes = ec_save.storage;
+               bytes_buf = ec_save.buf+nstart_bytes;
+               save_bytes = nend_bytes-nstart_bytes;
+               OPUS_COPY(bytes_save, bytes_buf, save_bytes);
+
+               /* Restore */
+               *ec = ec_save;
+               ctx = ctx_save;
+               OPUS_COPY(X, X_save, N);
+               OPUS_COPY(Y, Y_save, N);
+               if (i == start+1)
+                  special_hybrid_folding(m, norm, norm2, start, M, dual_stereo);
+               /* Encode and round up. */
+               ctx.theta_round = 1;
+               x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
+                     effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
+                     last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, cm);
+               dist1 = MULT16_32_Q15(w[0], celt_inner_prod(X_save, X, N, arch)) + MULT16_32_Q15(w[1], celt_inner_prod(Y_save, Y, N, arch));
+               if (dist0 >= dist1) {
+                  x_cm = cm2;
+                  *ec = ec_save2;
+                  ctx = ctx_save2;
+                  OPUS_COPY(X, X_save2, N);
+                  OPUS_COPY(Y, Y_save2, N);
+                  if (!last)
+                     OPUS_COPY(norm+M*eBands[i]-norm_offset, norm_save2, N);
+                  OPUS_COPY(bytes_buf, bytes_save, save_bytes);
+               }
+            } else {
+               ctx.theta_round = 0;
+               x_cm = quant_band_stereo(&ctx, X, Y, N, b, B,
+                     effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
+                     last?NULL:norm+M*eBands[i]-norm_offset, lowband_scratch, x_cm|y_cm);
+            }
          } else {
             x_cm = quant_band(&ctx, X, N, b, B,
                   effective_lowband != -1 ? norm+effective_lowband : NULL, LM,
-                        last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm|y_cm);
+                  last?NULL:norm+M*eBands[i]-norm_offset, Q15ONE, lowband_scratch, x_cm|y_cm);
          }
          y_cm = x_cm;
       }
       collapse_masks[i*C+0] = (unsigned char)x_cm;
       collapse_masks[i*C+C-1] = (unsigned char)y_cm;
       balance += pulses[i] + tell;
 
       /* Update the folding position only as long as we have 1 bit/sample depth. */
       update_lowband = b>(N<<BITRES);
+      /* We only need to avoid noise on a split for the first band. After that, we
+         have folding. */
+      ctx.avoid_split_noise = 0;
    }
    *seed = ctx.seed;
 
    RESTORE_STACK;
 }