2 * Copyright (c) 2012 Andrew D'Addesio
3 * Copyright (c) 2013-2014 Mozilla Corporation
4 * Copyright (c) 2017 Rostislav Pehlivanov <atomnuker@gmail.com>
6 * This file is part of FFmpeg.
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 #define CELT_PVQ_U(n, k) (ff_celt_pvq_u_row[FFMIN(n, k)][FFMAX(n, k)])
27 #define CELT_PVQ_V(n, k) (CELT_PVQ_U(n, k) + CELT_PVQ_U(n, (k) + 1))
29 static inline int16_t celt_cos(int16_t x)
31 x = (MUL16(x, x) + 4096) >> 13;
32 x = (32767-x) + ROUND_MUL16(x, (-7651 + ROUND_MUL16(x, (8277 + ROUND_MUL16(-626, x)))));
36 static inline int celt_log2tan(int isin, int icos)
43 return (ls << 11) - (lc << 11) +
44 ROUND_MUL16(isin, ROUND_MUL16(isin, -2597) + 7932) -
45 ROUND_MUL16(icos, ROUND_MUL16(icos, -2597) + 7932);
48 static inline int celt_bits2pulses(const uint8_t *cache, int bits)
50 // TODO: Find the size of cache and make it into an array in the parameters list
56 for (i = 0; i < 6; i++) {
57 int center = (low + high + 1) >> 1;
58 if (cache[center] >= bits)
64 return (bits - (low == 0 ? -1 : cache[low]) <= cache[high] - bits) ? low : high;
67 static inline int celt_pulses2bits(const uint8_t *cache, int pulses)
69 // TODO: Find the size of cache and make it into an array in the parameters list
70 return (pulses == 0) ? 0 : cache[pulses] + 1;
73 static inline void celt_normalize_residual(const int * av_restrict iy, float * av_restrict X,
77 for (i = 0; i < N; i++)
81 static void celt_exp_rotation_impl(float *X, uint32_t len, uint32_t stride,
88 for (i = 0; i < len - stride; i++) {
92 Xptr[stride] = c * x2 + s * x1;
93 *Xptr++ = c * x1 - s * x2;
96 Xptr = &X[len - 2 * stride - 1];
97 for (i = len - 2 * stride - 1; i >= 0; i--) {
101 Xptr[stride] = c * x2 + s * x1;
102 *Xptr-- = c * x1 - s * x2;
106 static inline void celt_exp_rotation(float *X, uint32_t len,
107 uint32_t stride, uint32_t K,
108 enum CeltSpread spread, const int encode)
110 uint32_t stride2 = 0;
115 if (2*K >= len || spread == CELT_SPREAD_NONE)
118 gain = (float)len / (len + (20 - 5*spread) * K);
119 theta = M_PI * gain * gain / 4;
124 if (len >= stride << 3) {
126 /* This is just a simple (equivalent) way of computing sqrt(len/stride) with rounding.
127 It's basically incrementing long as (stride2+0.5)^2 < len/stride. */
128 while ((stride2 * stride2 + stride2) * stride + (stride >> 2) < len)
132 /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
133 extract_collapse_mask().*/
135 for (i = 0; i < stride; i++) {
137 celt_exp_rotation_impl(X + i * len, len, 1, c, -s);
139 celt_exp_rotation_impl(X + i * len, len, stride2, s, -c);
142 celt_exp_rotation_impl(X + i * len, len, stride2, s, c);
143 celt_exp_rotation_impl(X + i * len, len, 1, c, s);
148 static inline uint32_t celt_extract_collapse_mask(const int *iy, uint32_t N, uint32_t B)
150 uint32_t collapse_mask;
157 /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
161 for (i = 0; i < B; i++)
162 for (j = 0; j < N0; j++)
163 collapse_mask |= (iy[i*N0+j]!=0)<<i;
164 return collapse_mask;
167 static inline void celt_stereo_merge(float *X, float *Y, float mid, int N)
170 float xp = 0, side = 0;
175 /* Compute the norm of X+Y and X-Y as |X|^2 + |Y|^2 +/- sum(xy) */
176 for (i = 0; i < N; i++) {
181 /* Compensating for the mid normalization */
184 E[0] = mid2 * mid2 + side - 2 * xp;
185 E[1] = mid2 * mid2 + side + 2 * xp;
186 if (E[0] < 6e-4f || E[1] < 6e-4f) {
187 for (i = 0; i < N; i++)
193 gain[0] = 1.0f / sqrtf(t);
195 gain[1] = 1.0f / sqrtf(t);
197 for (i = 0; i < N; i++) {
199 /* Apply mid scaling (side is already scaled) */
200 value[0] = mid * X[i];
202 X[i] = gain[0] * (value[0] - value[1]);
203 Y[i] = gain[1] * (value[0] + value[1]);
207 static void celt_interleave_hadamard(float *tmp, float *X, int N0,
208 int stride, int hadamard)
214 const uint8_t *ordery = ff_celt_hadamard_ordery + stride - 2;
215 for (i = 0; i < stride; i++)
216 for (j = 0; j < N0; j++)
217 tmp[j*stride+i] = X[ordery[i]*N0+j];
219 for (i = 0; i < stride; i++)
220 for (j = 0; j < N0; j++)
221 tmp[j*stride+i] = X[i*N0+j];
224 for (i = 0; i < N; i++)
228 static void celt_deinterleave_hadamard(float *tmp, float *X, int N0,
229 int stride, int hadamard)
235 const uint8_t *ordery = ff_celt_hadamard_ordery + stride - 2;
236 for (i = 0; i < stride; i++)
237 for (j = 0; j < N0; j++)
238 tmp[ordery[i]*N0+j] = X[j*stride+i];
240 for (i = 0; i < stride; i++)
241 for (j = 0; j < N0; j++)
242 tmp[i*N0+j] = X[j*stride+i];
245 for (i = 0; i < N; i++)
249 static void celt_haar1(float *X, int N0, int stride)
253 for (i = 0; i < stride; i++) {
254 for (j = 0; j < N0; j++) {
255 float x0 = X[stride * (2 * j + 0) + i];
256 float x1 = X[stride * (2 * j + 1) + i];
257 X[stride * (2 * j + 0) + i] = (x0 + x1) * M_SQRT1_2;
258 X[stride * (2 * j + 1) + i] = (x0 - x1) * M_SQRT1_2;
263 static inline int celt_compute_qn(int N, int b, int offset, int pulse_cap,
268 if (dualstereo && N == 2)
271 /* The upper limit ensures that in a stereo split with itheta==16384, we'll
272 * always have enough bits left over to code at least one pulse in the
273 * side; otherwise it would collapse, since it doesn't get folded. */
274 qb = FFMIN3(b - pulse_cap - (4 << 3), (b + N2 * offset) / N2, 8 << 3);
275 qn = (qb < (1 << 3 >> 1)) ? 1 : ((ff_celt_qn_exp2[qb & 0x7] >> (14 - (qb >> 3))) + 1) >> 1 << 1;
279 /* Convert the quantized vector to an index */
280 static inline uint32_t celt_icwrsi(uint32_t N, const int *y)
282 int i, idx = 0, sum = 0;
283 for (i = N - 1; i >= 0; i--) {
284 const uint32_t i_s = CELT_PVQ_U(N - i, sum + FFABS(y[i]) + 1);
285 idx += CELT_PVQ_U(N - i, sum) + (y[i] < 0)*i_s;
291 // this code was adapted from libopus
292 static inline uint64_t celt_cwrsi(uint32_t N, uint32_t K, uint32_t i, int *y)
302 /*Lots of pulses case:*/
304 const uint32_t *row = ff_celt_pvq_u_row[N];
306 /* Are the pulses in this dimension negative? */
311 /*Count how many pulses were placed in this dimension.*/
317 p = ff_celt_pvq_u_row[--K][N];
320 for (p = row[K]; p > i; p = row[K])
324 val = (k0 - K + s) ^ s;
327 } else { /*Lots of dimensions case:*/
328 /*Are there any pulses in this dimension at all?*/
329 p = ff_celt_pvq_u_row[K ][N];
330 q = ff_celt_pvq_u_row[K + 1][N];
332 if (p <= i && i < q) {
336 /*Are the pulses in this dimension negative?*/
340 /*Count how many pulses were placed in this dimension.*/
342 do p = ff_celt_pvq_u_row[--K][N];
346 val = (k0 - K + s) ^ s;
364 val = (k0 - K + s) ^ s;
377 static inline void celt_encode_pulses(OpusRangeCoder *rc, int *y, uint32_t N, uint32_t K)
379 ff_opus_rc_enc_uint(rc, celt_icwrsi(N, y), CELT_PVQ_V(N, K));
382 static inline float celt_decode_pulses(OpusRangeCoder *rc, int *y, uint32_t N, uint32_t K)
384 const uint32_t idx = ff_opus_rc_dec_uint(rc, CELT_PVQ_V(N, K));
385 return celt_cwrsi(N, K, idx, y);
389 * Faster than libopus's search, operates entirely in the signed domain.
390 * Slightly worse/better depending on N, K and the input vector.
392 static void celt_pvq_search(float *X, int *y, int K, int N)
395 float res = 0.0f, y_norm = 0.0f, xy_norm = 0.0f;
397 for (i = 0; i < N; i++)
402 for (i = 0; i < N; i++) {
403 y[i] = lrintf(res*X[i]);
405 xy_norm += y[i]*X[i];
410 int max_idx = 0, phase = FFSIGN(K);
411 float max_den = 1.0f, max_num = 0.0f;
414 for (i = 0; i < N; i++) {
415 float xy_new = xy_norm + 1*phase*FFABS(X[i]);
416 float y_new = y_norm + 2*phase*FFABS(y[i]);
417 xy_new = xy_new * xy_new;
418 if ((max_den*xy_new) > (y_new*max_num)) {
427 phase *= FFSIGN(X[max_idx]);
428 xy_norm += 1*phase*X[max_idx];
429 y_norm += 2*phase*y[max_idx];
434 static uint32_t celt_alg_quant(OpusRangeCoder *rc, float *X, uint32_t N, uint32_t K,
435 enum CeltSpread spread, uint32_t blocks, float gain)
439 celt_exp_rotation(X, N, blocks, K, spread, 1);
440 celt_pvq_search(X, y, K, N);
441 celt_encode_pulses(rc, y, N, K);
442 return celt_extract_collapse_mask(y, N, blocks);
445 /** Decode pulse vector and combine the result with the pitch vector to produce
446 the final normalised signal in the current band. */
447 static uint32_t celt_alg_unquant(OpusRangeCoder *rc, float *X, uint32_t N, uint32_t K,
448 enum CeltSpread spread, uint32_t blocks, float gain)
452 gain /= sqrtf(celt_decode_pulses(rc, y, N, K));
453 celt_normalize_residual(y, X, N, gain);
454 celt_exp_rotation(X, N, blocks, K, spread, 0);
455 return celt_extract_collapse_mask(y, N, blocks);
458 uint32_t ff_celt_decode_band(CeltFrame *f, OpusRangeCoder *rc, const int band,
459 float *X, float *Y, int N, int b, uint32_t blocks,
460 float *lowband, int duration, float *lowband_out, int level,
461 float gain, float *lowband_scratch, int fill)
463 const uint8_t *cache;
464 int dualstereo, split;
465 int imid = 0, iside = 0;
473 float mid = 0, side = 0;
474 int longblocks = (B0 == 1);
477 N_B0 = N_B = N / blocks;
478 split = dualstereo = (Y != NULL);
481 /* special case for one sample */
484 for (i = 0; i <= dualstereo; i++) {
486 if (f->remaining2 >= 1<<3) {
487 sign = ff_opus_rc_get_raw(rc, 1);
488 f->remaining2 -= 1 << 3;
491 x[0] = sign ? -1.0f : 1.0f;
495 lowband_out[0] = X[0];
499 if (!dualstereo && level == 0) {
500 int tf_change = f->tf_change[band];
503 recombine = tf_change;
504 /* Band recombining to increase frequency resolution */
507 (recombine || ((N_B & 1) == 0 && tf_change < 0) || B0 > 1)) {
509 for (j = 0; j < N; j++)
510 lowband_scratch[j] = lowband[j];
511 lowband = lowband_scratch;
514 for (k = 0; k < recombine; k++) {
516 celt_haar1(lowband, N >> k, 1 << k);
517 fill = ff_celt_bit_interleave[fill & 0xF] | ff_celt_bit_interleave[fill >> 4] << 2;
519 blocks >>= recombine;
522 /* Increasing the time resolution */
523 while ((N_B & 1) == 0 && tf_change < 0) {
525 celt_haar1(lowband, N_B, blocks);
526 fill |= fill << blocks;
535 /* Reorganize the samples in time order instead of frequency order */
536 if (B0 > 1 && lowband)
537 celt_deinterleave_hadamard(f->scratch, lowband, N_B >> recombine,
538 B0 << recombine, longblocks);
541 /* If we need 1.5 more bit than we can produce, split the band in two. */
542 cache = ff_celt_cache_bits +
543 ff_celt_cache_index[(duration + 1) * CELT_MAX_BANDS + band];
544 if (!dualstereo && duration >= 0 && b > cache[cache[0]] + 12 && N > 2) {
550 fill = (fill & 1) | (fill << 1);
551 blocks = (blocks + 1) >> 1;
557 int mbits, sbits, delta;
564 /* Decide on the resolution to give to the split parameter theta */
565 pulse_cap = ff_celt_log_freq_range[band] + duration * 8;
566 offset = (pulse_cap >> 1) - (dualstereo && N == 2 ? CELT_QTHETA_OFFSET_TWOPHASE :
568 qn = (dualstereo && band >= f->intensity_stereo) ? 1 :
569 celt_compute_qn(N, b, offset, pulse_cap, dualstereo);
570 tell = opus_rc_tell_frac(rc);
572 /* Entropy coding of the angle. We use a uniform pdf for the
573 time split, a step for stereo, and a triangular one for the rest. */
574 if (dualstereo && N > 2)
575 itheta = ff_opus_rc_dec_uint_step(rc, qn/2);
576 else if (dualstereo || B0 > 1)
577 itheta = ff_opus_rc_dec_uint(rc, qn+1);
579 itheta = ff_opus_rc_dec_uint_tri(rc, qn);
580 itheta = itheta * 16384 / qn;
581 /* NOTE: Renormalising X and Y *may* help fixed-point a bit at very high rate.
582 Let's do that at higher complexity */
583 } else if (dualstereo) {
584 inv = (b > 2 << 3 && f->remaining2 > 2 << 3) ? ff_opus_rc_dec_log(rc, 2) : 0;
587 qalloc = opus_rc_tell_frac(rc) - tell;
594 fill = av_mod_uintp2(fill, blocks);
596 } else if (itheta == 16384) {
599 fill &= ((1 << blocks) - 1) << blocks;
602 imid = celt_cos(itheta);
603 iside = celt_cos(16384-itheta);
604 /* This is the mid vs side allocation that minimizes squared error
606 delta = ROUND_MUL16((N - 1) << 7, celt_log2tan(iside, imid));
609 mid = imid / 32768.0f;
610 side = iside / 32768.0f;
612 /* This is a special case for N=2 that only works for stereo and takes
613 advantage of the fact that mid and side are orthogonal to encode
614 the side with just one bit. */
615 if (N == 2 && dualstereo) {
621 /* Only need one bit for the side */
622 sbits = (itheta != 0 && itheta != 16384) ? 1 << 3 : 0;
625 f->remaining2 -= qalloc+sbits;
630 sign = ff_opus_rc_get_raw(rc, 1);
632 /* We use orig_fill here because we want to fold the side, but if
633 itheta==16384, we'll have cleared the low bits of fill. */
634 cm = ff_celt_decode_band(f, rc, band, x2, NULL, N, mbits, blocks,
635 lowband, duration, lowband_out, level, gain,
636 lowband_scratch, orig_fill);
637 /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse),
638 and there's no need to worry about mixing with the other channel. */
639 y2[0] = -sign * x2[1];
640 y2[1] = sign * x2[0];
652 /* "Normal" split code */
653 float *next_lowband2 = NULL;
654 float *next_lowband_out1 = NULL;
658 /* Give more bits to low-energy MDCTs than they would
659 * otherwise deserve */
660 if (B0 > 1 && !dualstereo && (itheta & 0x3fff)) {
662 /* Rough approximation for pre-echo masking */
663 delta -= delta >> (4 - duration);
665 /* Corresponds to a forward-masking slope of
666 * 1.5 dB per 10 ms */
667 delta = FFMIN(0, delta + (N << 3 >> (5 - duration)));
669 mbits = av_clip((b - delta) / 2, 0, b);
671 f->remaining2 -= qalloc;
673 if (lowband && !dualstereo)
674 next_lowband2 = lowband + N; /* >32-bit split case */
676 /* Only stereo needs to pass on lowband_out.
677 * Otherwise, it's handled at the end */
679 next_lowband_out1 = lowband_out;
681 next_level = level + 1;
683 rebalance = f->remaining2;
684 if (mbits >= sbits) {
685 /* In stereo mode, we do not apply a scaling to the mid
686 * because we need the normalized mid for folding later */
687 cm = ff_celt_decode_band(f, rc, band, X, NULL, N, mbits, blocks,
688 lowband, duration, next_lowband_out1,
689 next_level, dualstereo ? 1.0f : (gain * mid),
690 lowband_scratch, fill);
692 rebalance = mbits - (rebalance - f->remaining2);
693 if (rebalance > 3 << 3 && itheta != 0)
694 sbits += rebalance - (3 << 3);
696 /* For a stereo split, the high bits of fill are always zero,
697 * so no folding will be done to the side. */
698 cm |= ff_celt_decode_band(f, rc, band, Y, NULL, N, sbits, blocks,
699 next_lowband2, duration, NULL,
700 next_level, gain * side, NULL,
701 fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
703 /* For a stereo split, the high bits of fill are always zero,
704 * so no folding will be done to the side. */
705 cm = ff_celt_decode_band(f, rc, band, Y, NULL, N, sbits, blocks,
706 next_lowband2, duration, NULL,
707 next_level, gain * side, NULL,
708 fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
710 rebalance = sbits - (rebalance - f->remaining2);
711 if (rebalance > 3 << 3 && itheta != 16384)
712 mbits += rebalance - (3 << 3);
714 /* In stereo mode, we do not apply a scaling to the mid because
715 * we need the normalized mid for folding later */
716 cm |= ff_celt_decode_band(f, rc, band, X, NULL, N, mbits, blocks,
717 lowband, duration, next_lowband_out1,
718 next_level, dualstereo ? 1.0f : (gain * mid),
719 lowband_scratch, fill);
723 /* This is the basic no-split case */
724 uint32_t q = celt_bits2pulses(cache, b);
725 uint32_t curr_bits = celt_pulses2bits(cache, q);
726 f->remaining2 -= curr_bits;
728 /* Ensures we can never bust the budget */
729 while (f->remaining2 < 0 && q > 0) {
730 f->remaining2 += curr_bits;
731 curr_bits = celt_pulses2bits(cache, --q);
732 f->remaining2 -= curr_bits;
736 /* Finally do the actual quantization */
737 cm = celt_alg_unquant(rc, X, N, (q < 8) ? q : (8 + (q & 7)) << ((q >> 3) - 1),
738 f->spread, blocks, gain);
740 /* If there's no pulse, fill the band anyway */
742 uint32_t cm_mask = (1 << blocks) - 1;
745 for (j = 0; j < N; j++)
750 for (j = 0; j < N; j++)
751 X[j] = (((int32_t)celt_rng(f)) >> 20);
754 /* Folded spectrum */
755 for (j = 0; j < N; j++) {
756 /* About 48 dB below the "normal" folding level */
757 X[j] = lowband[j] + (((celt_rng(f)) & 0x8000) ? 1.0f / 256 : -1.0f / 256);
761 celt_renormalize_vector(X, N, gain);
766 /* This code is used by the decoder and by the resynthesis-enabled encoder */
770 celt_stereo_merge(X, Y, mid, N);
772 for (j = 0; j < N; j++)
775 } else if (level == 0) {
778 /* Undo the sample reorganization going from time order to frequency order */
780 celt_interleave_hadamard(f->scratch, X, N_B>>recombine,
781 B0<<recombine, longblocks);
783 /* Undo time-freq changes that we did earlier */
786 for (k = 0; k < time_divide; k++) {
790 celt_haar1(X, N_B, blocks);
793 for (k = 0; k < recombine; k++) {
794 cm = ff_celt_bit_deinterleave[cm];
795 celt_haar1(X, N0>>k, 1<<k);
797 blocks <<= recombine;
799 /* Scale output for later folding */
803 for (j = 0; j < N0; j++)
804 lowband_out[j] = n * X[j];
806 cm = av_mod_uintp2(cm, blocks);
812 /* This has to be, AND MUST BE done by the psychoacoustic system, this has a very
813 * big impact on the entire quantization and especially huge on transients */
814 static int celt_calc_theta(const float *X, const float *Y, int coupling, int N)
817 float e[2] = { 0.0f, 0.0f };
818 for (j = 0; j < N; j++) {
819 if (coupling) { /* Coupling case */
820 e[0] += (X[j] + Y[j])*(X[j] + Y[j]);
821 e[1] += (X[j] - Y[j])*(X[j] - Y[j]);
827 return lrintf(32768.0f*atan2f(sqrtf(e[1]), sqrtf(e[0]))/M_PI);
830 static void celt_stereo_is_decouple(float *X, float *Y, float e_l, float e_r, int N)
833 const float energy_n = 1.0f/(sqrtf(e_l*e_l + e_r*e_r) + FLT_EPSILON);
836 for (i = 0; i < N; i++)
837 X[i] = e_l*X[i] + e_r*Y[i];
840 static void celt_stereo_ms_decouple(float *X, float *Y, int N)
843 const float decouple_norm = 1.0f/sqrtf(2.0f);
844 for (i = 0; i < N; i++) {
845 const float Xret = X[i];
846 X[i] = (X[i] + Y[i])*decouple_norm;
847 Y[i] = (Y[i] - Xret)*decouple_norm;
851 uint32_t ff_celt_encode_band(CeltFrame *f, OpusRangeCoder *rc, const int band,
852 float *X, float *Y, int N, int b, uint32_t blocks,
853 float *lowband, int duration, float *lowband_out, int level,
854 float gain, float *lowband_scratch, int fill)
856 const uint8_t *cache;
857 int dualstereo, split;
858 int imid = 0, iside = 0;
866 float mid = 0, side = 0;
867 int longblocks = (B0 == 1);
870 //N_B0 = N_B = N / blocks;
871 split = dualstereo = (Y != NULL);
874 /* special case for one sample - the decoder's output will be +- 1.0f!!! */
877 for (i = 0; i <= dualstereo; i++) {
878 if (f->remaining2 >= 1<<3) {
879 ff_opus_rc_put_raw(rc, x[0] < 0, 1);
880 f->remaining2 -= 1 << 3;
886 lowband_out[0] = X[0];
890 if (!dualstereo && level == 0) {
891 int tf_change = f->tf_change[band];
894 recombine = tf_change;
895 /* Band recombining to increase frequency resolution */
898 (recombine || ((N_B & 1) == 0 && tf_change < 0) || B0 > 1)) {
900 for (j = 0; j < N; j++)
901 lowband_scratch[j] = lowband[j];
902 lowband = lowband_scratch;
905 for (k = 0; k < recombine; k++) {
906 celt_haar1(X, N >> k, 1 << k);
907 fill = ff_celt_bit_interleave[fill & 0xF] | ff_celt_bit_interleave[fill >> 4] << 2;
909 blocks >>= recombine;
912 /* Increasing the time resolution */
913 while ((N_B & 1) == 0 && tf_change < 0) {
914 celt_haar1(X, N_B, blocks);
915 fill |= fill << blocks;
924 /* Reorganize the samples in time order instead of frequency order */
926 celt_deinterleave_hadamard(f->scratch, X, N_B >> recombine,
927 B0 << recombine, longblocks);
930 /* If we need 1.5 more bit than we can produce, split the band in two. */
931 cache = ff_celt_cache_bits +
932 ff_celt_cache_index[(duration + 1) * CELT_MAX_BANDS + band];
933 if (!dualstereo && duration >= 0 && b > cache[cache[0]] + 12 && N > 2) {
939 fill = (fill & 1) | (fill << 1);
940 blocks = (blocks + 1) >> 1;
945 int itheta = celt_calc_theta(X, Y, dualstereo, N);
946 int mbits, sbits, delta;
953 /* Decide on the resolution to give to the split parameter theta */
954 pulse_cap = ff_celt_log_freq_range[band] + duration * 8;
955 offset = (pulse_cap >> 1) - (dualstereo && N == 2 ? CELT_QTHETA_OFFSET_TWOPHASE :
957 qn = (dualstereo && band >= f->intensity_stereo) ? 1 :
958 celt_compute_qn(N, b, offset, pulse_cap, dualstereo);
959 tell = opus_rc_tell_frac(rc);
963 itheta = (itheta*qn + 8192) >> 14;
965 /* Entropy coding of the angle. We use a uniform pdf for the
966 * time split, a step for stereo, and a triangular one for the rest. */
967 if (dualstereo && N > 2)
968 ff_opus_rc_enc_uint_step(rc, itheta, qn / 2);
969 else if (dualstereo || B0 > 1)
970 ff_opus_rc_enc_uint(rc, itheta, qn + 1);
972 ff_opus_rc_enc_uint_tri(rc, itheta, qn);
973 itheta = itheta * 16384 / qn;
977 celt_stereo_is_decouple(X, Y, f->block[0].lin_energy[band], f->block[1].lin_energy[band], N);
979 celt_stereo_ms_decouple(X, Y, N);
981 } else if (dualstereo) {
989 celt_stereo_is_decouple(X, Y, f->block[0].lin_energy[band], f->block[1].lin_energy[band], N);
991 if (b > 2 << 3 && f->remaining2 > 2 << 3) {
992 ff_opus_rc_enc_log(rc, inv, 2);
999 qalloc = opus_rc_tell_frac(rc) - tell;
1006 fill = av_mod_uintp2(fill, blocks);
1008 } else if (itheta == 16384) {
1011 fill &= ((1 << blocks) - 1) << blocks;
1014 imid = celt_cos(itheta);
1015 iside = celt_cos(16384-itheta);
1016 /* This is the mid vs side allocation that minimizes squared error
1018 delta = ROUND_MUL16((N - 1) << 7, celt_log2tan(iside, imid));
1021 mid = imid / 32768.0f;
1022 side = iside / 32768.0f;
1024 /* This is a special case for N=2 that only works for stereo and takes
1025 advantage of the fact that mid and side are orthogonal to encode
1026 the side with just one bit. */
1027 if (N == 2 && dualstereo) {
1033 /* Only need one bit for the side */
1034 sbits = (itheta != 0 && itheta != 16384) ? 1 << 3 : 0;
1036 c = (itheta > 8192);
1037 f->remaining2 -= qalloc+sbits;
1042 sign = x2[0]*y2[1] - x2[1]*y2[0] < 0;
1043 ff_opus_rc_put_raw(rc, sign, 1);
1045 sign = 1 - 2 * sign;
1046 /* We use orig_fill here because we want to fold the side, but if
1047 itheta==16384, we'll have cleared the low bits of fill. */
1048 cm = ff_celt_encode_band(f, rc, band, x2, NULL, N, mbits, blocks,
1049 lowband, duration, lowband_out, level, gain,
1050 lowband_scratch, orig_fill);
1051 /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse),
1052 and there's no need to worry about mixing with the other channel. */
1053 y2[0] = -sign * x2[1];
1054 y2[1] = sign * x2[0];
1066 /* "Normal" split code */
1067 float *next_lowband2 = NULL;
1068 float *next_lowband_out1 = NULL;
1072 /* Give more bits to low-energy MDCTs than they would
1073 * otherwise deserve */
1074 if (B0 > 1 && !dualstereo && (itheta & 0x3fff)) {
1076 /* Rough approximation for pre-echo masking */
1077 delta -= delta >> (4 - duration);
1079 /* Corresponds to a forward-masking slope of
1080 * 1.5 dB per 10 ms */
1081 delta = FFMIN(0, delta + (N << 3 >> (5 - duration)));
1083 mbits = av_clip((b - delta) / 2, 0, b);
1085 f->remaining2 -= qalloc;
1087 if (lowband && !dualstereo)
1088 next_lowband2 = lowband + N; /* >32-bit split case */
1090 /* Only stereo needs to pass on lowband_out.
1091 * Otherwise, it's handled at the end */
1093 next_lowband_out1 = lowband_out;
1095 next_level = level + 1;
1097 rebalance = f->remaining2;
1098 if (mbits >= sbits) {
1099 /* In stereo mode, we do not apply a scaling to the mid
1100 * because we need the normalized mid for folding later */
1101 cm = ff_celt_encode_band(f, rc, band, X, NULL, N, mbits, blocks,
1102 lowband, duration, next_lowband_out1,
1103 next_level, dualstereo ? 1.0f : (gain * mid),
1104 lowband_scratch, fill);
1106 rebalance = mbits - (rebalance - f->remaining2);
1107 if (rebalance > 3 << 3 && itheta != 0)
1108 sbits += rebalance - (3 << 3);
1110 /* For a stereo split, the high bits of fill are always zero,
1111 * so no folding will be done to the side. */
1112 cm |= ff_celt_encode_band(f, rc, band, Y, NULL, N, sbits, blocks,
1113 next_lowband2, duration, NULL,
1114 next_level, gain * side, NULL,
1115 fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
1117 /* For a stereo split, the high bits of fill are always zero,
1118 * so no folding will be done to the side. */
1119 cm = ff_celt_encode_band(f, rc, band, Y, NULL, N, sbits, blocks,
1120 next_lowband2, duration, NULL,
1121 next_level, gain * side, NULL,
1122 fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
1124 rebalance = sbits - (rebalance - f->remaining2);
1125 if (rebalance > 3 << 3 && itheta != 16384)
1126 mbits += rebalance - (3 << 3);
1128 /* In stereo mode, we do not apply a scaling to the mid because
1129 * we need the normalized mid for folding later */
1130 cm |= ff_celt_encode_band(f, rc, band, X, NULL, N, mbits, blocks,
1131 lowband, duration, next_lowband_out1,
1132 next_level, dualstereo ? 1.0f : (gain * mid),
1133 lowband_scratch, fill);
1137 /* This is the basic no-split case */
1138 uint32_t q = celt_bits2pulses(cache, b);
1139 uint32_t curr_bits = celt_pulses2bits(cache, q);
1140 f->remaining2 -= curr_bits;
1142 /* Ensures we can never bust the budget */
1143 while (f->remaining2 < 0 && q > 0) {
1144 f->remaining2 += curr_bits;
1145 curr_bits = celt_pulses2bits(cache, --q);
1146 f->remaining2 -= curr_bits;
1150 /* Finally do the actual quantization */
1151 cm = celt_alg_quant(rc, X, N, (q < 8) ? q : (8 + (q & 7)) << ((q >> 3) - 1),
1152 f->spread, blocks, gain);