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, uint32_t K, 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;
288 av_assert0(sum == K);
292 // this code was adapted from libopus
293 static inline uint64_t celt_cwrsi(uint32_t N, uint32_t K, uint32_t i, int *y)
303 /*Lots of pulses case:*/
305 const uint32_t *row = ff_celt_pvq_u_row[N];
307 /* Are the pulses in this dimension negative? */
312 /*Count how many pulses were placed in this dimension.*/
318 p = ff_celt_pvq_u_row[--K][N];
321 for (p = row[K]; p > i; p = row[K])
325 val = (k0 - K + s) ^ s;
328 } else { /*Lots of dimensions case:*/
329 /*Are there any pulses in this dimension at all?*/
330 p = ff_celt_pvq_u_row[K ][N];
331 q = ff_celt_pvq_u_row[K + 1][N];
333 if (p <= i && i < q) {
337 /*Are the pulses in this dimension negative?*/
341 /*Count how many pulses were placed in this dimension.*/
343 do p = ff_celt_pvq_u_row[--K][N];
347 val = (k0 - K + s) ^ s;
365 val = (k0 - K + s) ^ s;
378 static inline void celt_encode_pulses(OpusRangeCoder *rc, int *y, uint32_t N, uint32_t K)
380 ff_opus_rc_enc_uint(rc, celt_icwrsi(N, K, y), CELT_PVQ_V(N, K));
383 static inline float celt_decode_pulses(OpusRangeCoder *rc, int *y, uint32_t N, uint32_t K)
385 const uint32_t idx = ff_opus_rc_dec_uint(rc, CELT_PVQ_V(N, K));
386 return celt_cwrsi(N, K, idx, y);
390 * Faster than libopus's search, operates entirely in the signed domain.
391 * Slightly worse/better depending on N, K and the input vector.
393 static void celt_pvq_search(float *X, int *y, int K, int N)
396 float res = 0.0f, y_norm = 0.0f, xy_norm = 0.0f;
398 for (i = 0; i < N; i++)
403 for (i = 0; i < N; i++) {
404 y[i] = lrintf(res*X[i]);
406 xy_norm += y[i]*X[i];
411 int max_idx = 0, phase = FFSIGN(K);
412 float max_den = 1.0f, max_num = 0.0f;
415 for (i = 0; i < N; i++) {
416 float xy_new = xy_norm + 1*phase*FFABS(X[i]);
417 float y_new = y_norm + 2*phase*FFABS(y[i]);
418 xy_new = xy_new * xy_new;
419 /* FIXME: the y[i] check makes the search slightly worse at Ks below 5 */
420 if (y[i] && (max_den*xy_new) > (y_new*max_num)) {
429 phase *= FFSIGN(X[max_idx]);
430 xy_norm += 1*phase*X[max_idx];
431 y_norm += 2*phase*y[max_idx];
436 static uint32_t celt_alg_quant(OpusRangeCoder *rc, float *X, uint32_t N, uint32_t K,
437 enum CeltSpread spread, uint32_t blocks, float gain)
441 celt_exp_rotation(X, N, blocks, K, spread, 1);
442 celt_pvq_search(X, y, K, N);
443 celt_encode_pulses(rc, y, N, K);
444 return celt_extract_collapse_mask(y, N, blocks);
447 /** Decode pulse vector and combine the result with the pitch vector to produce
448 the final normalised signal in the current band. */
449 static uint32_t celt_alg_unquant(OpusRangeCoder *rc, float *X, uint32_t N, uint32_t K,
450 enum CeltSpread spread, uint32_t blocks, float gain)
454 gain /= sqrtf(celt_decode_pulses(rc, y, N, K));
455 celt_normalize_residual(y, X, N, gain);
456 celt_exp_rotation(X, N, blocks, K, spread, 0);
457 return celt_extract_collapse_mask(y, N, blocks);
460 uint32_t ff_celt_decode_band(CeltFrame *f, OpusRangeCoder *rc, const int band,
461 float *X, float *Y, int N, int b, uint32_t blocks,
462 float *lowband, int duration, float *lowband_out, int level,
463 float gain, float *lowband_scratch, int fill)
465 const uint8_t *cache;
466 int dualstereo, split;
467 int imid = 0, iside = 0;
475 float mid = 0, side = 0;
476 int longblocks = (B0 == 1);
479 N_B0 = N_B = N / blocks;
480 split = dualstereo = (Y != NULL);
483 /* special case for one sample */
486 for (i = 0; i <= dualstereo; i++) {
488 if (f->remaining2 >= 1<<3) {
489 sign = ff_opus_rc_get_raw(rc, 1);
490 f->remaining2 -= 1 << 3;
493 x[0] = sign ? -1.0f : 1.0f;
497 lowband_out[0] = X[0];
501 if (!dualstereo && level == 0) {
502 int tf_change = f->tf_change[band];
505 recombine = tf_change;
506 /* Band recombining to increase frequency resolution */
509 (recombine || ((N_B & 1) == 0 && tf_change < 0) || B0 > 1)) {
511 for (j = 0; j < N; j++)
512 lowband_scratch[j] = lowband[j];
513 lowband = lowband_scratch;
516 for (k = 0; k < recombine; k++) {
518 celt_haar1(lowband, N >> k, 1 << k);
519 fill = ff_celt_bit_interleave[fill & 0xF] | ff_celt_bit_interleave[fill >> 4] << 2;
521 blocks >>= recombine;
524 /* Increasing the time resolution */
525 while ((N_B & 1) == 0 && tf_change < 0) {
527 celt_haar1(lowband, N_B, blocks);
528 fill |= fill << blocks;
537 /* Reorganize the samples in time order instead of frequency order */
538 if (B0 > 1 && lowband)
539 celt_deinterleave_hadamard(f->scratch, lowband, N_B >> recombine,
540 B0 << recombine, longblocks);
543 /* If we need 1.5 more bit than we can produce, split the band in two. */
544 cache = ff_celt_cache_bits +
545 ff_celt_cache_index[(duration + 1) * CELT_MAX_BANDS + band];
546 if (!dualstereo && duration >= 0 && b > cache[cache[0]] + 12 && N > 2) {
552 fill = (fill & 1) | (fill << 1);
553 blocks = (blocks + 1) >> 1;
559 int mbits, sbits, delta;
566 /* Decide on the resolution to give to the split parameter theta */
567 pulse_cap = ff_celt_log_freq_range[band] + duration * 8;
568 offset = (pulse_cap >> 1) - (dualstereo && N == 2 ? CELT_QTHETA_OFFSET_TWOPHASE :
570 qn = (dualstereo && band >= f->intensity_stereo) ? 1 :
571 celt_compute_qn(N, b, offset, pulse_cap, dualstereo);
572 tell = opus_rc_tell_frac(rc);
574 /* Entropy coding of the angle. We use a uniform pdf for the
575 time split, a step for stereo, and a triangular one for the rest. */
576 if (dualstereo && N > 2)
577 itheta = ff_opus_rc_dec_uint_step(rc, qn/2);
578 else if (dualstereo || B0 > 1)
579 itheta = ff_opus_rc_dec_uint(rc, qn+1);
581 itheta = ff_opus_rc_dec_uint_tri(rc, qn);
582 itheta = itheta * 16384 / qn;
583 /* NOTE: Renormalising X and Y *may* help fixed-point a bit at very high rate.
584 Let's do that at higher complexity */
585 } else if (dualstereo) {
586 inv = (b > 2 << 3 && f->remaining2 > 2 << 3) ? ff_opus_rc_dec_log(rc, 2) : 0;
589 qalloc = opus_rc_tell_frac(rc) - tell;
596 fill = av_mod_uintp2(fill, blocks);
598 } else if (itheta == 16384) {
601 fill &= ((1 << blocks) - 1) << blocks;
604 imid = celt_cos(itheta);
605 iside = celt_cos(16384-itheta);
606 /* This is the mid vs side allocation that minimizes squared error
608 delta = ROUND_MUL16((N - 1) << 7, celt_log2tan(iside, imid));
611 mid = imid / 32768.0f;
612 side = iside / 32768.0f;
614 /* This is a special case for N=2 that only works for stereo and takes
615 advantage of the fact that mid and side are orthogonal to encode
616 the side with just one bit. */
617 if (N == 2 && dualstereo) {
623 /* Only need one bit for the side */
624 sbits = (itheta != 0 && itheta != 16384) ? 1 << 3 : 0;
627 f->remaining2 -= qalloc+sbits;
632 sign = ff_opus_rc_get_raw(rc, 1);
634 /* We use orig_fill here because we want to fold the side, but if
635 itheta==16384, we'll have cleared the low bits of fill. */
636 cm = ff_celt_decode_band(f, rc, band, x2, NULL, N, mbits, blocks,
637 lowband, duration, lowband_out, level, gain,
638 lowband_scratch, orig_fill);
639 /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse),
640 and there's no need to worry about mixing with the other channel. */
641 y2[0] = -sign * x2[1];
642 y2[1] = sign * x2[0];
654 /* "Normal" split code */
655 float *next_lowband2 = NULL;
656 float *next_lowband_out1 = NULL;
660 /* Give more bits to low-energy MDCTs than they would
661 * otherwise deserve */
662 if (B0 > 1 && !dualstereo && (itheta & 0x3fff)) {
664 /* Rough approximation for pre-echo masking */
665 delta -= delta >> (4 - duration);
667 /* Corresponds to a forward-masking slope of
668 * 1.5 dB per 10 ms */
669 delta = FFMIN(0, delta + (N << 3 >> (5 - duration)));
671 mbits = av_clip((b - delta) / 2, 0, b);
673 f->remaining2 -= qalloc;
675 if (lowband && !dualstereo)
676 next_lowband2 = lowband + N; /* >32-bit split case */
678 /* Only stereo needs to pass on lowband_out.
679 * Otherwise, it's handled at the end */
681 next_lowband_out1 = lowband_out;
683 next_level = level + 1;
685 rebalance = f->remaining2;
686 if (mbits >= sbits) {
687 /* In stereo mode, we do not apply a scaling to the mid
688 * because we need the normalized mid for folding later */
689 cm = ff_celt_decode_band(f, rc, band, X, NULL, N, mbits, blocks,
690 lowband, duration, next_lowband_out1,
691 next_level, dualstereo ? 1.0f : (gain * mid),
692 lowband_scratch, fill);
694 rebalance = mbits - (rebalance - f->remaining2);
695 if (rebalance > 3 << 3 && itheta != 0)
696 sbits += rebalance - (3 << 3);
698 /* For a stereo split, the high bits of fill are always zero,
699 * so no folding will be done to the side. */
700 cm |= ff_celt_decode_band(f, rc, band, Y, NULL, N, sbits, blocks,
701 next_lowband2, duration, NULL,
702 next_level, gain * side, NULL,
703 fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
705 /* For a stereo split, the high bits of fill are always zero,
706 * so no folding will be done to the side. */
707 cm = ff_celt_decode_band(f, rc, band, Y, NULL, N, sbits, blocks,
708 next_lowband2, duration, NULL,
709 next_level, gain * side, NULL,
710 fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
712 rebalance = sbits - (rebalance - f->remaining2);
713 if (rebalance > 3 << 3 && itheta != 16384)
714 mbits += rebalance - (3 << 3);
716 /* In stereo mode, we do not apply a scaling to the mid because
717 * we need the normalized mid for folding later */
718 cm |= ff_celt_decode_band(f, rc, band, X, NULL, N, mbits, blocks,
719 lowband, duration, next_lowband_out1,
720 next_level, dualstereo ? 1.0f : (gain * mid),
721 lowband_scratch, fill);
725 /* This is the basic no-split case */
726 uint32_t q = celt_bits2pulses(cache, b);
727 uint32_t curr_bits = celt_pulses2bits(cache, q);
728 f->remaining2 -= curr_bits;
730 /* Ensures we can never bust the budget */
731 while (f->remaining2 < 0 && q > 0) {
732 f->remaining2 += curr_bits;
733 curr_bits = celt_pulses2bits(cache, --q);
734 f->remaining2 -= curr_bits;
738 /* Finally do the actual quantization */
739 cm = celt_alg_unquant(rc, X, N, (q < 8) ? q : (8 + (q & 7)) << ((q >> 3) - 1),
740 f->spread, blocks, gain);
742 /* If there's no pulse, fill the band anyway */
744 uint32_t cm_mask = (1 << blocks) - 1;
747 for (j = 0; j < N; j++)
752 for (j = 0; j < N; j++)
753 X[j] = (((int32_t)celt_rng(f)) >> 20);
756 /* Folded spectrum */
757 for (j = 0; j < N; j++) {
758 /* About 48 dB below the "normal" folding level */
759 X[j] = lowband[j] + (((celt_rng(f)) & 0x8000) ? 1.0f / 256 : -1.0f / 256);
763 celt_renormalize_vector(X, N, gain);
768 /* This code is used by the decoder and by the resynthesis-enabled encoder */
772 celt_stereo_merge(X, Y, mid, N);
774 for (j = 0; j < N; j++)
777 } else if (level == 0) {
780 /* Undo the sample reorganization going from time order to frequency order */
782 celt_interleave_hadamard(f->scratch, X, N_B>>recombine,
783 B0<<recombine, longblocks);
785 /* Undo time-freq changes that we did earlier */
788 for (k = 0; k < time_divide; k++) {
792 celt_haar1(X, N_B, blocks);
795 for (k = 0; k < recombine; k++) {
796 cm = ff_celt_bit_deinterleave[cm];
797 celt_haar1(X, N0>>k, 1<<k);
799 blocks <<= recombine;
801 /* Scale output for later folding */
805 for (j = 0; j < N0; j++)
806 lowband_out[j] = n * X[j];
808 cm = av_mod_uintp2(cm, blocks);
814 /* This has to be, AND MUST BE done by the psychoacoustic system, this has a very
815 * big impact on the entire quantization and especially huge on transients */
816 static int celt_calc_theta(const float *X, const float *Y, int coupling, int N)
819 float e[2] = { 0.0f, 0.0f };
820 for (j = 0; j < N; j++) {
821 if (coupling) { /* Coupling case */
822 e[0] += (X[j] + Y[j])*(X[j] + Y[j]);
823 e[1] += (X[j] - Y[j])*(X[j] - Y[j]);
829 return lrintf(32768.0f*atan2f(sqrtf(e[1]), sqrtf(e[0]))/M_PI);
832 static void celt_stereo_is_decouple(float *X, float *Y, float e_l, float e_r, int N)
835 const float energy_n = 1.0f/(sqrtf(e_l*e_l + e_r*e_r) + FLT_EPSILON);
838 for (i = 0; i < N; i++)
839 X[i] = e_l*X[i] + e_r*Y[i];
842 static void celt_stereo_ms_decouple(float *X, float *Y, int N)
845 const float decouple_norm = 1.0f/sqrtf(2.0f);
846 for (i = 0; i < N; i++) {
847 const float Xret = X[i];
848 X[i] = (X[i] + Y[i])*decouple_norm;
849 Y[i] = (Y[i] - Xret)*decouple_norm;
853 uint32_t ff_celt_encode_band(CeltFrame *f, OpusRangeCoder *rc, const int band,
854 float *X, float *Y, int N, int b, uint32_t blocks,
855 float *lowband, int duration, float *lowband_out, int level,
856 float gain, float *lowband_scratch, int fill)
858 const uint8_t *cache;
859 int dualstereo, split;
860 int imid = 0, iside = 0;
862 int N_B = N / blocks;
868 float mid = 0, side = 0;
869 int longblocks = (B0 == 1);
872 split = dualstereo = (Y != NULL);
875 /* special case for one sample - the decoder's output will be +- 1.0f!!! */
878 for (i = 0; i <= dualstereo; i++) {
879 if (f->remaining2 >= 1<<3) {
880 ff_opus_rc_put_raw(rc, x[0] < 0, 1);
881 f->remaining2 -= 1 << 3;
887 lowband_out[0] = X[0];
891 if (!dualstereo && level == 0) {
892 int tf_change = f->tf_change[band];
895 recombine = tf_change;
896 /* Band recombining to increase frequency resolution */
899 (recombine || ((N_B & 1) == 0 && tf_change < 0) || B0 > 1)) {
901 for (j = 0; j < N; j++)
902 lowband_scratch[j] = lowband[j];
903 lowband = lowband_scratch;
906 for (k = 0; k < recombine; k++) {
907 celt_haar1(X, N >> k, 1 << k);
908 fill = ff_celt_bit_interleave[fill & 0xF] | ff_celt_bit_interleave[fill >> 4] << 2;
910 blocks >>= recombine;
913 /* Increasing the time resolution */
914 while ((N_B & 1) == 0 && tf_change < 0) {
915 celt_haar1(X, N_B, blocks);
916 fill |= fill << blocks;
925 /* Reorganize the samples in time order instead of frequency order */
927 celt_deinterleave_hadamard(f->scratch, X, N_B >> recombine,
928 B0 << recombine, longblocks);
931 /* If we need 1.5 more bit than we can produce, split the band in two. */
932 cache = ff_celt_cache_bits +
933 ff_celt_cache_index[(duration + 1) * CELT_MAX_BANDS + band];
934 if (!dualstereo && duration >= 0 && b > cache[cache[0]] + 12 && N > 2) {
940 fill = (fill & 1) | (fill << 1);
941 blocks = (blocks + 1) >> 1;
946 int itheta = celt_calc_theta(X, Y, dualstereo, N);
947 int mbits, sbits, delta;
954 /* Decide on the resolution to give to the split parameter theta */
955 pulse_cap = ff_celt_log_freq_range[band] + duration * 8;
956 offset = (pulse_cap >> 1) - (dualstereo && N == 2 ? CELT_QTHETA_OFFSET_TWOPHASE :
958 qn = (dualstereo && band >= f->intensity_stereo) ? 1 :
959 celt_compute_qn(N, b, offset, pulse_cap, dualstereo);
960 tell = opus_rc_tell_frac(rc);
964 itheta = (itheta*qn + 8192) >> 14;
966 /* Entropy coding of the angle. We use a uniform pdf for the
967 * time split, a step for stereo, and a triangular one for the rest. */
968 if (dualstereo && N > 2)
969 ff_opus_rc_enc_uint_step(rc, itheta, qn / 2);
970 else if (dualstereo || B0 > 1)
971 ff_opus_rc_enc_uint(rc, itheta, qn + 1);
973 ff_opus_rc_enc_uint_tri(rc, itheta, qn);
974 itheta = itheta * 16384 / qn;
978 celt_stereo_is_decouple(X, Y, f->block[0].lin_energy[band], f->block[1].lin_energy[band], N);
980 celt_stereo_ms_decouple(X, Y, N);
982 } else if (dualstereo) {
990 celt_stereo_is_decouple(X, Y, f->block[0].lin_energy[band], f->block[1].lin_energy[band], N);
992 if (b > 2 << 3 && f->remaining2 > 2 << 3) {
993 ff_opus_rc_enc_log(rc, inv, 2);
1000 qalloc = opus_rc_tell_frac(rc) - tell;
1007 fill = av_mod_uintp2(fill, blocks);
1009 } else if (itheta == 16384) {
1012 fill &= ((1 << blocks) - 1) << blocks;
1015 imid = celt_cos(itheta);
1016 iside = celt_cos(16384-itheta);
1017 /* This is the mid vs side allocation that minimizes squared error
1019 delta = ROUND_MUL16((N - 1) << 7, celt_log2tan(iside, imid));
1022 mid = imid / 32768.0f;
1023 side = iside / 32768.0f;
1025 /* This is a special case for N=2 that only works for stereo and takes
1026 advantage of the fact that mid and side are orthogonal to encode
1027 the side with just one bit. */
1028 if (N == 2 && dualstereo) {
1034 /* Only need one bit for the side */
1035 sbits = (itheta != 0 && itheta != 16384) ? 1 << 3 : 0;
1037 c = (itheta > 8192);
1038 f->remaining2 -= qalloc+sbits;
1043 sign = x2[0]*y2[1] - x2[1]*y2[0] < 0;
1044 ff_opus_rc_put_raw(rc, sign, 1);
1046 sign = 1 - 2 * sign;
1047 /* We use orig_fill here because we want to fold the side, but if
1048 itheta==16384, we'll have cleared the low bits of fill. */
1049 cm = ff_celt_encode_band(f, rc, band, x2, NULL, N, mbits, blocks,
1050 lowband, duration, lowband_out, level, gain,
1051 lowband_scratch, orig_fill);
1052 /* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse),
1053 and there's no need to worry about mixing with the other channel. */
1054 y2[0] = -sign * x2[1];
1055 y2[1] = sign * x2[0];
1067 /* "Normal" split code */
1068 float *next_lowband2 = NULL;
1069 float *next_lowband_out1 = NULL;
1073 /* Give more bits to low-energy MDCTs than they would
1074 * otherwise deserve */
1075 if (B0 > 1 && !dualstereo && (itheta & 0x3fff)) {
1077 /* Rough approximation for pre-echo masking */
1078 delta -= delta >> (4 - duration);
1080 /* Corresponds to a forward-masking slope of
1081 * 1.5 dB per 10 ms */
1082 delta = FFMIN(0, delta + (N << 3 >> (5 - duration)));
1084 mbits = av_clip((b - delta) / 2, 0, b);
1086 f->remaining2 -= qalloc;
1088 if (lowband && !dualstereo)
1089 next_lowband2 = lowband + N; /* >32-bit split case */
1091 /* Only stereo needs to pass on lowband_out.
1092 * Otherwise, it's handled at the end */
1094 next_lowband_out1 = lowband_out;
1096 next_level = level + 1;
1098 rebalance = f->remaining2;
1099 if (mbits >= sbits) {
1100 /* In stereo mode, we do not apply a scaling to the mid
1101 * because we need the normalized mid for folding later */
1102 cm = ff_celt_encode_band(f, rc, band, X, NULL, N, mbits, blocks,
1103 lowband, duration, next_lowband_out1,
1104 next_level, dualstereo ? 1.0f : (gain * mid),
1105 lowband_scratch, fill);
1107 rebalance = mbits - (rebalance - f->remaining2);
1108 if (rebalance > 3 << 3 && itheta != 0)
1109 sbits += rebalance - (3 << 3);
1111 /* For a stereo split, the high bits of fill are always zero,
1112 * so no folding will be done to the side. */
1113 cm |= ff_celt_encode_band(f, rc, band, Y, NULL, N, sbits, blocks,
1114 next_lowband2, duration, NULL,
1115 next_level, gain * side, NULL,
1116 fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
1118 /* For a stereo split, the high bits of fill are always zero,
1119 * so no folding will be done to the side. */
1120 cm = ff_celt_encode_band(f, rc, band, Y, NULL, N, sbits, blocks,
1121 next_lowband2, duration, NULL,
1122 next_level, gain * side, NULL,
1123 fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
1125 rebalance = sbits - (rebalance - f->remaining2);
1126 if (rebalance > 3 << 3 && itheta != 16384)
1127 mbits += rebalance - (3 << 3);
1129 /* In stereo mode, we do not apply a scaling to the mid because
1130 * we need the normalized mid for folding later */
1131 cm |= ff_celt_encode_band(f, rc, band, X, NULL, N, mbits, blocks,
1132 lowband, duration, next_lowband_out1,
1133 next_level, dualstereo ? 1.0f : (gain * mid),
1134 lowband_scratch, fill);
1138 /* This is the basic no-split case */
1139 uint32_t q = celt_bits2pulses(cache, b);
1140 uint32_t curr_bits = celt_pulses2bits(cache, q);
1141 f->remaining2 -= curr_bits;
1143 /* Ensures we can never bust the budget */
1144 while (f->remaining2 < 0 && q > 0) {
1145 f->remaining2 += curr_bits;
1146 curr_bits = celt_pulses2bits(cache, --q);
1147 f->remaining2 -= curr_bits;
1151 /* Finally do the actual quantization */
1152 cm = celt_alg_quant(rc, X, N, (q < 8) ? q : (8 + (q & 7)) << ((q >> 3) - 1),
1153 f->spread, blocks, gain);