path: root/libavcodec/opus_pvq.c (plain)
blob: ce93c4731d659e62ae6fd19b8ae32f852dbe6697

1	/*
2	* Copyright (c) 2012 Andrew D'Addesio
3	* Copyright (c) 2013-2014 Mozilla Corporation
4	* Copyright (c) 2017 Rostislav Pehlivanov <atomnuker@gmail.com>
5	*
6	* This file is part of FFmpeg.
7	*
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.
12	*
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.
17	*
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
21	*/
22
23	#include "opustab.h"
24	#include "opus_pvq.h"
25
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))
28
29	static inline int16_t celt_cos(int16_t x)
30	{
31	x = (MUL16(x, x) + 4096) >> 13;
32	x = (32767-x) + ROUND_MUL16(x, (-7651 + ROUND_MUL16(x, (8277 + ROUND_MUL16(-626, x)))));
33	return 1+x;
34	}
35
36	static inline int celt_log2tan(int isin, int icos)
37	{
38	int lc, ls;
39	lc = opus_ilog(icos);
40	ls = opus_ilog(isin);
41	icos <<= 15 - lc;
42	isin <<= 15 - ls;
43	return (ls << 11) - (lc << 11) +
44	ROUND_MUL16(isin, ROUND_MUL16(isin, -2597) + 7932) -
45	ROUND_MUL16(icos, ROUND_MUL16(icos, -2597) + 7932);
46	}
47
48	static inline int celt_bits2pulses(const uint8_t *cache, int bits)
49	{
50	// TODO: Find the size of cache and make it into an array in the parameters list
51	int i, low = 0, high;
52
53	high = cache[0];
54	bits--;
55
56	for (i = 0; i < 6; i++) {
57	int center = (low + high + 1) >> 1;
58	if (cache[center] >= bits)
59	high = center;
60	else
61	low = center;
62	}
63
64	return (bits - (low == 0 ? -1 : cache[low]) <= cache[high] - bits) ? low : high;
65	}
66
67	static inline int celt_pulses2bits(const uint8_t *cache, int pulses)
68	{
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;
71	}
72
73	static inline void celt_normalize_residual(const int * av_restrict iy, float * av_restrict X,
74	int N, float g)
75	{
76	int i;
77	for (i = 0; i < N; i++)
78	X[i] = g * iy[i];
79	}
80
81	static void celt_exp_rotation_impl(float *X, uint32_t len, uint32_t stride,
82	float c, float s)
83	{
84	float *Xptr;
85	int i;
86
87	Xptr = X;
88	for (i = 0; i < len - stride; i++) {
89	float x1, x2;
90	x1 = Xptr[0];
91	x2 = Xptr[stride];
92	Xptr[stride] = c * x2 + s * x1;
93	*Xptr++ = c * x1 - s * x2;
94	}
95
96	Xptr = &X[len - 2 * stride - 1];
97	for (i = len - 2 * stride - 1; i >= 0; i--) {
98	float x1, x2;
99	x1 = Xptr[0];
100	x2 = Xptr[stride];
101	Xptr[stride] = c * x2 + s * x1;
102	*Xptr-- = c * x1 - s * x2;
103	}
104	}
105
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)
109	{
110	uint32_t stride2 = 0;
111	float c, s;
112	float gain, theta;
113	int i;
114
115	if (2*K >= len || spread == CELT_SPREAD_NONE)
116	return;
117
118	gain = (float)len / (len + (20 - 5*spread) * K);
119	theta = M_PI * gain * gain / 4;
120
121	c = cosf(theta);
122	s = sinf(theta);
123
124	if (len >= stride << 3) {
125	stride2 = 1;
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)
129	stride2++;
130	}
131
132	/*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
133	extract_collapse_mask().*/
134	len /= stride;
135	for (i = 0; i < stride; i++) {
136	if (encode) {
137	celt_exp_rotation_impl(X + i * len, len, 1, c, -s);
138	if (stride2)
139	celt_exp_rotation_impl(X + i * len, len, stride2, s, -c);
140	} else {
141	if (stride2)
142	celt_exp_rotation_impl(X + i * len, len, stride2, s, c);
143	celt_exp_rotation_impl(X + i * len, len, 1, c, s);
144	}
145	}
146	}
147
148	static inline uint32_t celt_extract_collapse_mask(const int *iy, uint32_t N, uint32_t B)
149	{
150	uint32_t collapse_mask;
151	int N0;
152	int i, j;
153
154	if (B <= 1)
155	return 1;
156
157	/*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
158	exp_rotation().*/
159	N0 = N/B;
160	collapse_mask = 0;
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;
165	}
166
167	static inline void celt_stereo_merge(float *X, float *Y, float mid, int N)
168	{
169	int i;
170	float xp = 0, side = 0;
171	float E[2];
172	float mid2;
173	float t, gain[2];
174
175	/* Compute the norm of X+Y and X-Y as |X|^2 + |Y|^2 +/- sum(xy) */
176	for (i = 0; i < N; i++) {
177	xp += X[i] * Y[i];
178	side += Y[i] * Y[i];
179	}
180
181	/* Compensating for the mid normalization */
182	xp *= mid;
183	mid2 = mid;
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++)
188	Y[i] = X[i];
189	return;
190	}
191
192	t = E[0];
193	gain[0] = 1.0f / sqrtf(t);
194	t = E[1];
195	gain[1] = 1.0f / sqrtf(t);
196
197	for (i = 0; i < N; i++) {
198	float value[2];
199	/* Apply mid scaling (side is already scaled) */
200	value[0] = mid * X[i];
201	value[1] = Y[i];
202	X[i] = gain[0] * (value[0] - value[1]);
203	Y[i] = gain[1] * (value[0] + value[1]);
204	}
205	}
206
207	static void celt_interleave_hadamard(float *tmp, float *X, int N0,
208	int stride, int hadamard)
209	{
210	int i, j;
211	int N = N0*stride;
212
213	if (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];
218	} else {
219	for (i = 0; i < stride; i++)
220	for (j = 0; j < N0; j++)
221	tmp[j*stride+i] = X[i*N0+j];
222	}
223
224	for (i = 0; i < N; i++)
225	X[i] = tmp[i];
226	}
227
228	static void celt_deinterleave_hadamard(float *tmp, float *X, int N0,
229	int stride, int hadamard)
230	{
231	int i, j;
232	int N = N0*stride;
233
234	if (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];
239	} else {
240	for (i = 0; i < stride; i++)
241	for (j = 0; j < N0; j++)
242	tmp[i*N0+j] = X[j*stride+i];
243	}
244
245	for (i = 0; i < N; i++)
246	X[i] = tmp[i];
247	}
248
249	static void celt_haar1(float *X, int N0, int stride)
250	{
251	int i, j;
252	N0 >>= 1;
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;
259	}
260	}
261	}
262
263	static inline int celt_compute_qn(int N, int b, int offset, int pulse_cap,
264	int dualstereo)
265	{
266	int qn, qb;
267	int N2 = 2 * N - 1;
268	if (dualstereo && N == 2)
269	N2--;
270
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;
276	return qn;
277	}
278
279	/* Convert the quantized vector to an index */
280	static inline uint32_t celt_icwrsi(uint32_t N, uint32_t K, const int *y)
281	{
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;
286	sum += FFABS(y[i]);
287	}
288	return idx;
289	}
290
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)
293	{
294	uint64_t norm = 0;
295	uint32_t p;
296	int s, val;
297	int k0;
298
299	while (N > 2) {
300	uint32_t q;
301
302	/*Lots of pulses case:*/
303	if (K >= N) {
304	const uint32_t *row = ff_celt_pvq_u_row[N];
305
306	/* Are the pulses in this dimension negative? */
307	p = row[K + 1];
308	s = -(i >= p);
309	i -= p & s;
310
311	/*Count how many pulses were placed in this dimension.*/
312	k0 = K;
313	q = row[N];
314	if (q > i) {
315	K = N;
316	do {
317	p = ff_celt_pvq_u_row[--K][N];
318	} while (p > i);
319	} else
320	for (p = row[K]; p > i; p = row[K])
321	K--;
322
323	i -= p;
324	val = (k0 - K + s) ^ s;
325	norm += val * val;
326	*y++ = val;
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];
331
332	if (p <= i && i < q) {
333	i -= p;
334	*y++ = 0;
335	} else {
336	/*Are the pulses in this dimension negative?*/
337	s = -(i >= q);
338	i -= q & s;
339
340	/*Count how many pulses were placed in this dimension.*/
341	k0 = K;
342	do p = ff_celt_pvq_u_row[--K][N];
343	while (p > i);
344
345	i -= p;
346	val = (k0 - K + s) ^ s;
347	norm += val * val;
348	*y++ = val;
349	}
350	}
351	N--;
352	}
353
354	/* N == 2 */
355	p = 2 * K + 1;
356	s = -(i >= p);
357	i -= p & s;
358	k0 = K;
359	K = (i + 1) / 2;
360
361	if (K)
362	i -= 2 * K - 1;
363
364	val = (k0 - K + s) ^ s;
365	norm += val * val;
366	*y++ = val;
367
368	/* N==1 */
369	s = -i;
370	val = (K + s) ^ s;
371	norm += val * val;
372	*y = val;
373
374	return norm;
375	}
376
377	static inline void celt_encode_pulses(OpusRangeCoder *rc, int *y, uint32_t N, uint32_t K)
378	{
379	ff_opus_rc_enc_uint(rc, celt_icwrsi(N, K, y), CELT_PVQ_V(N, K));
380	}
381
382	static inline float celt_decode_pulses(OpusRangeCoder *rc, int *y, uint32_t N, uint32_t K)
383	{
384	const uint32_t idx = ff_opus_rc_dec_uint(rc, CELT_PVQ_V(N, K));
385	return celt_cwrsi(N, K, idx, y);
386	}
387
388	/*
389	* Faster than libopus's search, operates entirely in the signed domain.
390	* Slightly worse/better depending on N, K and the input vector.
391	*/
392	static void celt_pvq_search(float *X, int *y, int K, int N)
393	{
394	int i;
395	float res = 0.0f, y_norm = 0.0f, xy_norm = 0.0f;
396
397	for (i = 0; i < N; i++)
398	res += FFABS(X[i]);
399
400	res = K/(res + FLT_EPSILON);
401
402	for (i = 0; i < N; i++) {
403	y[i] = lrintf(res*X[i]);
404	y_norm += y[i]*y[i];
405	xy_norm += y[i]*X[i];
406	K -= FFABS(y[i]);
407	}
408
409	while (K) {
410	int max_idx = 0, phase = FFSIGN(K);
411	float max_den = 1.0f, max_num = 0.0f;
412	y_norm += 1.0f;
413
414	for (i = 0; i < N; i++) {
415	/* If the sum has been overshot and the best place has 0 pulses allocated
416	* to it, attempting to decrease it further will actually increase the
417	* sum. Prevent this by disregarding any 0 positions when decrementing. */
418	const int ca = 1 ^ ((y[i] == 0) & (phase < 0));
419	float xy_new = xy_norm + 1*phase*FFABS(X[i]);
420	float y_new = y_norm + 2*phase*FFABS(y[i]);
421	xy_new = xy_new * xy_new;
422	if (ca && (max_den*xy_new) > (y_new*max_num)) {
423	max_den = y_new;
424	max_num = xy_new;
425	max_idx = i;
426	}
427	}
428
429	K -= phase;
430
431	phase *= FFSIGN(X[max_idx]);
432	xy_norm += 1*phase*X[max_idx];
433	y_norm += 2*phase*y[max_idx];
434	y[max_idx] += phase;
435	}
436	}
437
438	static uint32_t celt_alg_quant(OpusRangeCoder *rc, float *X, uint32_t N, uint32_t K,
439	enum CeltSpread spread, uint32_t blocks, float gain)
440	{
441	int y[176];
442
443	celt_exp_rotation(X, N, blocks, K, spread, 1);
444	celt_pvq_search(X, y, K, N);
445	celt_encode_pulses(rc, y, N, K);
446	return celt_extract_collapse_mask(y, N, blocks);
447	}
448
449	/** Decode pulse vector and combine the result with the pitch vector to produce
450	the final normalised signal in the current band. */
451	static uint32_t celt_alg_unquant(OpusRangeCoder *rc, float *X, uint32_t N, uint32_t K,
452	enum CeltSpread spread, uint32_t blocks, float gain)
453	{
454	int y[176];
455
456	gain /= sqrtf(celt_decode_pulses(rc, y, N, K));
457	celt_normalize_residual(y, X, N, gain);
458	celt_exp_rotation(X, N, blocks, K, spread, 0);
459	return celt_extract_collapse_mask(y, N, blocks);
460	}
461
462	uint32_t ff_celt_decode_band(CeltFrame *f, OpusRangeCoder *rc, const int band,
463	float *X, float *Y, int N, int b, uint32_t blocks,
464	float *lowband, int duration, float *lowband_out, int level,
465	float gain, float *lowband_scratch, int fill)
466	{
467	const uint8_t *cache;
468	int dualstereo, split;
469	int imid = 0, iside = 0;
470	uint32_t N0 = N;
471	int N_B;
472	int N_B0;
473	int B0 = blocks;
474	int time_divide = 0;
475	int recombine = 0;
476	int inv = 0;
477	float mid = 0, side = 0;
478	int longblocks = (B0 == 1);
479	uint32_t cm = 0;
480
481	N_B0 = N_B = N / blocks;
482	split = dualstereo = (Y != NULL);
483
484	if (N == 1) {
485	/* special case for one sample */
486	int i;
487	float *x = X;
488	for (i = 0; i <= dualstereo; i++) {
489	int sign = 0;
490	if (f->remaining2 >= 1<<3) {
491	sign = ff_opus_rc_get_raw(rc, 1);
492	f->remaining2 -= 1 << 3;
493	b -= 1 << 3;
494	}
495	x[0] = sign ? -1.0f : 1.0f;
496	x = Y;
497	}
498	if (lowband_out)
499	lowband_out[0] = X[0];
500	return 1;
501	}
502
503	if (!dualstereo && level == 0) {
504	int tf_change = f->tf_change[band];
505	int k;
506	if (tf_change > 0)
507	recombine = tf_change;
508	/* Band recombining to increase frequency resolution */
509
510	if (lowband &&
511	(recombine || ((N_B & 1) == 0 && tf_change < 0) || B0 > 1)) {
512	int j;
513	for (j = 0; j < N; j++)
514	lowband_scratch[j] = lowband[j];
515	lowband = lowband_scratch;
516	}
517
518	for (k = 0; k < recombine; k++) {
519	if (lowband)
520	celt_haar1(lowband, N >> k, 1 << k);
521	fill = ff_celt_bit_interleave[fill & 0xF] | ff_celt_bit_interleave[fill >> 4] << 2;
522	}
523	blocks >>= recombine;
524	N_B <<= recombine;
525
526	/* Increasing the time resolution */
527	while ((N_B & 1) == 0 && tf_change < 0) {
528	if (lowband)
529	celt_haar1(lowband, N_B, blocks);
530	fill |= fill << blocks;
531	blocks <<= 1;
532	N_B >>= 1;
533	time_divide++;
534	tf_change++;
535	}
536	B0 = blocks;
537	N_B0 = N_B;
538
539	/* Reorganize the samples in time order instead of frequency order */
540	if (B0 > 1 && lowband)
541	celt_deinterleave_hadamard(f->scratch, lowband, N_B >> recombine,
542	B0 << recombine, longblocks);
543	}
544
545	/* If we need 1.5 more bit than we can produce, split the band in two. */
546	cache = ff_celt_cache_bits +
547	ff_celt_cache_index[(duration + 1) * CELT_MAX_BANDS + band];
548	if (!dualstereo && duration >= 0 && b > cache[cache[0]] + 12 && N > 2) {
549	N >>= 1;
550	Y = X + N;
551	split = 1;
552	duration -= 1;
553	if (blocks == 1)
554	fill = (fill & 1) | (fill << 1);
555	blocks = (blocks + 1) >> 1;
556	}
557
558	if (split) {
559	int qn;
560	int itheta = 0;
561	int mbits, sbits, delta;
562	int qalloc;
563	int pulse_cap;
564	int offset;
565	int orig_fill;
566	int tell;
567
568	/* Decide on the resolution to give to the split parameter theta */
569	pulse_cap = ff_celt_log_freq_range[band] + duration * 8;
570	offset = (pulse_cap >> 1) - (dualstereo && N == 2 ? CELT_QTHETA_OFFSET_TWOPHASE :
571	CELT_QTHETA_OFFSET);
572	qn = (dualstereo && band >= f->intensity_stereo) ? 1 :
573	celt_compute_qn(N, b, offset, pulse_cap, dualstereo);
574	tell = opus_rc_tell_frac(rc);
575	if (qn != 1) {
576	/* Entropy coding of the angle. We use a uniform pdf for the
577	time split, a step for stereo, and a triangular one for the rest. */
578	if (dualstereo && N > 2)
579	itheta = ff_opus_rc_dec_uint_step(rc, qn/2);
580	else if (dualstereo || B0 > 1)
581	itheta = ff_opus_rc_dec_uint(rc, qn+1);
582	else
583	itheta = ff_opus_rc_dec_uint_tri(rc, qn);
584	itheta = itheta * 16384 / qn;
585	/* NOTE: Renormalising X and Y *may* help fixed-point a bit at very high rate.
586	Let's do that at higher complexity */
587	} else if (dualstereo) {
588	inv = (b > 2 << 3 && f->remaining2 > 2 << 3) ? ff_opus_rc_dec_log(rc, 2) : 0;
589	itheta = 0;
590	}
591	qalloc = opus_rc_tell_frac(rc) - tell;
592	b -= qalloc;
593
594	orig_fill = fill;
595	if (itheta == 0) {
596	imid = 32767;
597	iside = 0;
598	fill = av_mod_uintp2(fill, blocks);
599	delta = -16384;
600	} else if (itheta == 16384) {
601	imid = 0;
602	iside = 32767;
603	fill &= ((1 << blocks) - 1) << blocks;
604	delta = 16384;
605	} else {
606	imid = celt_cos(itheta);
607	iside = celt_cos(16384-itheta);
608	/* This is the mid vs side allocation that minimizes squared error
609	in that band. */
610	delta = ROUND_MUL16((N - 1) << 7, celt_log2tan(iside, imid));
611	}
612
613	mid = imid / 32768.0f;
614	side = iside / 32768.0f;
615
616	/* This is a special case for N=2 that only works for stereo and takes
617	advantage of the fact that mid and side are orthogonal to encode
618	the side with just one bit. */
619	if (N == 2 && dualstereo) {
620	int c;
621	int sign = 0;
622	float tmp;
623	float *x2, *y2;
624	mbits = b;
625	/* Only need one bit for the side */
626	sbits = (itheta != 0 && itheta != 16384) ? 1 << 3 : 0;
627	mbits -= sbits;
628	c = (itheta > 8192);
629	f->remaining2 -= qalloc+sbits;
630
631	x2 = c ? Y : X;
632	y2 = c ? X : Y;
633	if (sbits)
634	sign = ff_opus_rc_get_raw(rc, 1);
635	sign = 1 - 2 * sign;
636	/* We use orig_fill here because we want to fold the side, but if
637	itheta==16384, we'll have cleared the low bits of fill. */
638	cm = ff_celt_decode_band(f, rc, band, x2, NULL, N, mbits, blocks,
639	lowband, duration, lowband_out, level, gain,
640	lowband_scratch, orig_fill);
641	/* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse),
642	and there's no need to worry about mixing with the other channel. */
643	y2[0] = -sign * x2[1];
644	y2[1] = sign * x2[0];
645	X[0] *= mid;
646	X[1] *= mid;
647	Y[0] *= side;
648	Y[1] *= side;
649	tmp = X[0];
650	X[0] = tmp - Y[0];
651	Y[0] = tmp + Y[0];
652	tmp = X[1];
653	X[1] = tmp - Y[1];
654	Y[1] = tmp + Y[1];
655	} else {
656	/* "Normal" split code */
657	float *next_lowband2 = NULL;
658	float *next_lowband_out1 = NULL;
659	int next_level = 0;
660	int rebalance;
661
662	/* Give more bits to low-energy MDCTs than they would
663	* otherwise deserve */
664	if (B0 > 1 && !dualstereo && (itheta & 0x3fff)) {
665	if (itheta > 8192)
666	/* Rough approximation for pre-echo masking */
667	delta -= delta >> (4 - duration);
668	else
669	/* Corresponds to a forward-masking slope of
670	* 1.5 dB per 10 ms */
671	delta = FFMIN(0, delta + (N << 3 >> (5 - duration)));
672	}
673	mbits = av_clip((b - delta) / 2, 0, b);
674	sbits = b - mbits;
675	f->remaining2 -= qalloc;
676
677	if (lowband && !dualstereo)
678	next_lowband2 = lowband + N; /* >32-bit split case */
679
680	/* Only stereo needs to pass on lowband_out.
681	* Otherwise, it's handled at the end */
682	if (dualstereo)
683	next_lowband_out1 = lowband_out;
684	else
685	next_level = level + 1;
686
687	rebalance = f->remaining2;
688	if (mbits >= sbits) {
689	/* In stereo mode, we do not apply a scaling to the mid
690	* because we need the normalized mid for folding later */
691	cm = ff_celt_decode_band(f, rc, band, X, NULL, N, mbits, blocks,
692	lowband, duration, next_lowband_out1,
693	next_level, dualstereo ? 1.0f : (gain * mid),
694	lowband_scratch, fill);
695
696	rebalance = mbits - (rebalance - f->remaining2);
697	if (rebalance > 3 << 3 && itheta != 0)
698	sbits += rebalance - (3 << 3);
699
700	/* For a stereo split, the high bits of fill are always zero,
701	* so no folding will be done to the side. */
702	cm |= ff_celt_decode_band(f, rc, band, Y, NULL, N, sbits, blocks,
703	next_lowband2, duration, NULL,
704	next_level, gain * side, NULL,
705	fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
706	} else {
707	/* For a stereo split, the high bits of fill are always zero,
708	* so no folding will be done to the side. */
709	cm = ff_celt_decode_band(f, rc, band, Y, NULL, N, sbits, blocks,
710	next_lowband2, duration, NULL,
711	next_level, gain * side, NULL,
712	fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
713
714	rebalance = sbits - (rebalance - f->remaining2);
715	if (rebalance > 3 << 3 && itheta != 16384)
716	mbits += rebalance - (3 << 3);
717
718	/* In stereo mode, we do not apply a scaling to the mid because
719	* we need the normalized mid for folding later */
720	cm |= ff_celt_decode_band(f, rc, band, X, NULL, N, mbits, blocks,
721	lowband, duration, next_lowband_out1,
722	next_level, dualstereo ? 1.0f : (gain * mid),
723	lowband_scratch, fill);
724	}
725	}
726	} else {
727	/* This is the basic no-split case */
728	uint32_t q = celt_bits2pulses(cache, b);
729	uint32_t curr_bits = celt_pulses2bits(cache, q);
730	f->remaining2 -= curr_bits;
731
732	/* Ensures we can never bust the budget */
733	while (f->remaining2 < 0 && q > 0) {
734	f->remaining2 += curr_bits;
735	curr_bits = celt_pulses2bits(cache, --q);
736	f->remaining2 -= curr_bits;
737	}
738
739	if (q != 0) {
740	/* Finally do the actual quantization */
741	cm = celt_alg_unquant(rc, X, N, (q < 8) ? q : (8 + (q & 7)) << ((q >> 3) - 1),
742	f->spread, blocks, gain);
743	} else {
744	/* If there's no pulse, fill the band anyway */
745	int j;
746	uint32_t cm_mask = (1 << blocks) - 1;
747	fill &= cm_mask;
748	if (!fill) {
749	for (j = 0; j < N; j++)
750	X[j] = 0.0f;
751	} else {
752	if (!lowband) {
753	/* Noise */
754	for (j = 0; j < N; j++)
755	X[j] = (((int32_t)celt_rng(f)) >> 20);
756	cm = cm_mask;
757	} else {
758	/* Folded spectrum */
759	for (j = 0; j < N; j++) {
760	/* About 48 dB below the "normal" folding level */
761	X[j] = lowband[j] + (((celt_rng(f)) & 0x8000) ? 1.0f / 256 : -1.0f / 256);
762	}
763	cm = fill;
764	}
765	celt_renormalize_vector(X, N, gain);
766	}
767	}
768	}
769
770	/* This code is used by the decoder and by the resynthesis-enabled encoder */
771	if (dualstereo) {
772	int j;
773	if (N != 2)
774	celt_stereo_merge(X, Y, mid, N);
775	if (inv) {
776	for (j = 0; j < N; j++)
777	Y[j] *= -1;
778	}
779	} else if (level == 0) {
780	int k;
781
782	/* Undo the sample reorganization going from time order to frequency order */
783	if (B0 > 1)
784	celt_interleave_hadamard(f->scratch, X, N_B>>recombine,
785	B0<<recombine, longblocks);
786
787	/* Undo time-freq changes that we did earlier */
788	N_B = N_B0;
789	blocks = B0;
790	for (k = 0; k < time_divide; k++) {
791	blocks >>= 1;
792	N_B <<= 1;
793	cm |= cm >> blocks;
794	celt_haar1(X, N_B, blocks);
795	}
796
797	for (k = 0; k < recombine; k++) {
798	cm = ff_celt_bit_deinterleave[cm];
799	celt_haar1(X, N0>>k, 1<<k);
800	}
801	blocks <<= recombine;
802
803	/* Scale output for later folding */
804	if (lowband_out) {
805	int j;
806	float n = sqrtf(N0);
807	for (j = 0; j < N0; j++)
808	lowband_out[j] = n * X[j];
809	}
810	cm = av_mod_uintp2(cm, blocks);
811	}
812
813	return cm;
814	}
815
816	/* This has to be, AND MUST BE done by the psychoacoustic system, this has a very
817	* big impact on the entire quantization and especially huge on transients */
818	static int celt_calc_theta(const float *X, const float *Y, int coupling, int N)
819	{
820	int j;
821	float e[2] = { 0.0f, 0.0f };
822	for (j = 0; j < N; j++) {
823	if (coupling) { /* Coupling case */
824	e[0] += (X[j] + Y[j])*(X[j] + Y[j]);
825	e[1] += (X[j] - Y[j])*(X[j] - Y[j]);
826	} else {
827	e[0] += X[j]*X[j];
828	e[1] += Y[j]*Y[j];
829	}
830	}
831	return lrintf(32768.0f*atan2f(sqrtf(e[1]), sqrtf(e[0]))/M_PI);
832	}
833
834	static void celt_stereo_is_decouple(float *X, float *Y, float e_l, float e_r, int N)
835	{
836	int i;
837	const float energy_n = 1.0f/(sqrtf(e_l*e_l + e_r*e_r) + FLT_EPSILON);
838	e_l *= energy_n;
839	e_r *= energy_n;
840	for (i = 0; i < N; i++)
841	X[i] = e_l*X[i] + e_r*Y[i];
842	}
843
844	static void celt_stereo_ms_decouple(float *X, float *Y, int N)
845	{
846	int i;
847	const float decouple_norm = 1.0f/sqrtf(2.0f);
848	for (i = 0; i < N; i++) {
849	const float Xret = X[i];
850	X[i] = (X[i] + Y[i])*decouple_norm;
851	Y[i] = (Y[i] - Xret)*decouple_norm;
852	}
853	}
854
855	uint32_t ff_celt_encode_band(CeltFrame *f, OpusRangeCoder *rc, const int band,
856	float *X, float *Y, int N, int b, uint32_t blocks,
857	float *lowband, int duration, float *lowband_out, int level,
858	float gain, float *lowband_scratch, int fill)
859	{
860	const uint8_t *cache;
861	int dualstereo, split;
862	int imid = 0, iside = 0;
863	//uint32_t N0 = N;
864	int N_B = N / blocks;
865	//int N_B0 = N_B;
866	int B0 = blocks;
867	int time_divide = 0;
868	int recombine = 0;
869	int inv = 0;
870	float mid = 0, side = 0;
871	int longblocks = (B0 == 1);
872	uint32_t cm = 0;
873
874	split = dualstereo = (Y != NULL);
875
876	if (N == 1) {
877	/* special case for one sample - the decoder's output will be +- 1.0f!!! */
878	int i;
879	float *x = X;
880	for (i = 0; i <= dualstereo; i++) {
881	if (f->remaining2 >= 1<<3) {
882	ff_opus_rc_put_raw(rc, x[0] < 0, 1);
883	f->remaining2 -= 1 << 3;
884	b -= 1 << 3;
885	}
886	x = Y;
887	}
888	if (lowband_out)
889	lowband_out[0] = X[0];
890	return 1;
891	}
892
893	if (!dualstereo && level == 0) {
894	int tf_change = f->tf_change[band];
895	int k;
896	if (tf_change > 0)
897	recombine = tf_change;
898	/* Band recombining to increase frequency resolution */
899
900	if (lowband &&
901	(recombine || ((N_B & 1) == 0 && tf_change < 0) || B0 > 1)) {
902	int j;
903	for (j = 0; j < N; j++)
904	lowband_scratch[j] = lowband[j];
905	lowband = lowband_scratch;
906	}
907
908	for (k = 0; k < recombine; k++) {
909	celt_haar1(X, N >> k, 1 << k);
910	fill = ff_celt_bit_interleave[fill & 0xF] | ff_celt_bit_interleave[fill >> 4] << 2;
911	}
912	blocks >>= recombine;
913	N_B <<= recombine;
914
915	/* Increasing the time resolution */
916	while ((N_B & 1) == 0 && tf_change < 0) {
917	celt_haar1(X, N_B, blocks);
918	fill |= fill << blocks;
919	blocks <<= 1;
920	N_B >>= 1;
921	time_divide++;
922	tf_change++;
923	}
924	B0 = blocks;
925	//N_B0 = N_B;
926
927	/* Reorganize the samples in time order instead of frequency order */
928	if (B0 > 1)
929	celt_deinterleave_hadamard(f->scratch, X, N_B >> recombine,
930	B0 << recombine, longblocks);
931	}
932
933	/* If we need 1.5 more bit than we can produce, split the band in two. */
934	cache = ff_celt_cache_bits +
935	ff_celt_cache_index[(duration + 1) * CELT_MAX_BANDS + band];
936	if (!dualstereo && duration >= 0 && b > cache[cache[0]] + 12 && N > 2) {
937	N >>= 1;
938	Y = X + N;
939	split = 1;
940	duration -= 1;
941	if (blocks == 1)
942	fill = (fill & 1) | (fill << 1);
943	blocks = (blocks + 1) >> 1;
944	}
945
946	if (split) {
947	int qn;
948	int itheta = celt_calc_theta(X, Y, dualstereo, N);
949	int mbits, sbits, delta;
950	int qalloc;
951	int pulse_cap;
952	int offset;
953	int orig_fill;
954	int tell;
955
956	/* Decide on the resolution to give to the split parameter theta */
957	pulse_cap = ff_celt_log_freq_range[band] + duration * 8;
958	offset = (pulse_cap >> 1) - (dualstereo && N == 2 ? CELT_QTHETA_OFFSET_TWOPHASE :
959	CELT_QTHETA_OFFSET);
960	qn = (dualstereo && band >= f->intensity_stereo) ? 1 :
961	celt_compute_qn(N, b, offset, pulse_cap, dualstereo);
962	tell = opus_rc_tell_frac(rc);
963
964	if (qn != 1) {
965
966	itheta = (itheta*qn + 8192) >> 14;
967
968	/* Entropy coding of the angle. We use a uniform pdf for the
969	* time split, a step for stereo, and a triangular one for the rest. */
970	if (dualstereo && N > 2)
971	ff_opus_rc_enc_uint_step(rc, itheta, qn / 2);
972	else if (dualstereo || B0 > 1)
973	ff_opus_rc_enc_uint(rc, itheta, qn + 1);
974	else
975	ff_opus_rc_enc_uint_tri(rc, itheta, qn);
976	itheta = itheta * 16384 / qn;
977
978	if (dualstereo) {
979	if (itheta == 0)
980	celt_stereo_is_decouple(X, Y, f->block[0].lin_energy[band], f->block[1].lin_energy[band], N);
981	else
982	celt_stereo_ms_decouple(X, Y, N);
983	}
984	} else if (dualstereo) {
985	inv = itheta > 8192;
986	if (inv)
987	{
988	int j;
989	for (j=0;j<N;j++)
990	Y[j] = -Y[j];
991	}
992	celt_stereo_is_decouple(X, Y, f->block[0].lin_energy[band], f->block[1].lin_energy[band], N);
993
994	if (b > 2 << 3 && f->remaining2 > 2 << 3) {
995	ff_opus_rc_enc_log(rc, inv, 2);
996	} else {
997	inv = 0;
998	}
999
1000	itheta = 0;
1001	}
1002	qalloc = opus_rc_tell_frac(rc) - tell;
1003	b -= qalloc;
1004
1005	orig_fill = fill;
1006	if (itheta == 0) {
1007	imid = 32767;
1008	iside = 0;
1009	fill = av_mod_uintp2(fill, blocks);
1010	delta = -16384;
1011	} else if (itheta == 16384) {
1012	imid = 0;
1013	iside = 32767;
1014	fill &= ((1 << blocks) - 1) << blocks;
1015	delta = 16384;
1016	} else {
1017	imid = celt_cos(itheta);
1018	iside = celt_cos(16384-itheta);
1019	/* This is the mid vs side allocation that minimizes squared error
1020	in that band. */
1021	delta = ROUND_MUL16((N - 1) << 7, celt_log2tan(iside, imid));
1022	}
1023
1024	mid = imid / 32768.0f;
1025	side = iside / 32768.0f;
1026
1027	/* This is a special case for N=2 that only works for stereo and takes
1028	advantage of the fact that mid and side are orthogonal to encode
1029	the side with just one bit. */
1030	if (N == 2 && dualstereo) {
1031	int c;
1032	int sign = 0;
1033	float tmp;
1034	float *x2, *y2;
1035	mbits = b;
1036	/* Only need one bit for the side */
1037	sbits = (itheta != 0 && itheta != 16384) ? 1 << 3 : 0;
1038	mbits -= sbits;
1039	c = (itheta > 8192);
1040	f->remaining2 -= qalloc+sbits;
1041
1042	x2 = c ? Y : X;
1043	y2 = c ? X : Y;
1044	if (sbits) {
1045	sign = x2[0]*y2[1] - x2[1]*y2[0] < 0;
1046	ff_opus_rc_put_raw(rc, sign, 1);
1047	}
1048	sign = 1 - 2 * sign;
1049	/* We use orig_fill here because we want to fold the side, but if
1050	itheta==16384, we'll have cleared the low bits of fill. */
1051	cm = ff_celt_encode_band(f, rc, band, x2, NULL, N, mbits, blocks,
1052	lowband, duration, lowband_out, level, gain,
1053	lowband_scratch, orig_fill);
1054	/* We don't split N=2 bands, so cm is either 1 or 0 (for a fold-collapse),
1055	and there's no need to worry about mixing with the other channel. */
1056	y2[0] = -sign * x2[1];
1057	y2[1] = sign * x2[0];
1058	X[0] *= mid;
1059	X[1] *= mid;
1060	Y[0] *= side;
1061	Y[1] *= side;
1062	tmp = X[0];
1063	X[0] = tmp - Y[0];
1064	Y[0] = tmp + Y[0];
1065	tmp = X[1];
1066	X[1] = tmp - Y[1];
1067	Y[1] = tmp + Y[1];
1068	} else {
1069	/* "Normal" split code */
1070	float *next_lowband2 = NULL;
1071	float *next_lowband_out1 = NULL;
1072	int next_level = 0;
1073	int rebalance;
1074
1075	/* Give more bits to low-energy MDCTs than they would
1076	* otherwise deserve */
1077	if (B0 > 1 && !dualstereo && (itheta & 0x3fff)) {
1078	if (itheta > 8192)
1079	/* Rough approximation for pre-echo masking */
1080	delta -= delta >> (4 - duration);
1081	else
1082	/* Corresponds to a forward-masking slope of
1083	* 1.5 dB per 10 ms */
1084	delta = FFMIN(0, delta + (N << 3 >> (5 - duration)));
1085	}
1086	mbits = av_clip((b - delta) / 2, 0, b);
1087	sbits = b - mbits;
1088	f->remaining2 -= qalloc;
1089
1090	if (lowband && !dualstereo)
1091	next_lowband2 = lowband + N; /* >32-bit split case */
1092
1093	/* Only stereo needs to pass on lowband_out.
1094	* Otherwise, it's handled at the end */
1095	if (dualstereo)
1096	next_lowband_out1 = lowband_out;
1097	else
1098	next_level = level + 1;
1099
1100	rebalance = f->remaining2;
1101	if (mbits >= sbits) {
1102	/* In stereo mode, we do not apply a scaling to the mid
1103	* because we need the normalized mid for folding later */
1104	cm = ff_celt_encode_band(f, rc, band, X, NULL, N, mbits, blocks,
1105	lowband, duration, next_lowband_out1,
1106	next_level, dualstereo ? 1.0f : (gain * mid),
1107	lowband_scratch, fill);
1108
1109	rebalance = mbits - (rebalance - f->remaining2);
1110	if (rebalance > 3 << 3 && itheta != 0)
1111	sbits += rebalance - (3 << 3);
1112
1113	/* For a stereo split, the high bits of fill are always zero,
1114	* so no folding will be done to the side. */
1115	cm |= ff_celt_encode_band(f, rc, band, Y, NULL, N, sbits, blocks,
1116	next_lowband2, duration, NULL,
1117	next_level, gain * side, NULL,
1118	fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
1119	} else {
1120	/* For a stereo split, the high bits of fill are always zero,
1121	* so no folding will be done to the side. */
1122	cm = ff_celt_encode_band(f, rc, band, Y, NULL, N, sbits, blocks,
1123	next_lowband2, duration, NULL,
1124	next_level, gain * side, NULL,
1125	fill >> blocks) << ((B0 >> 1) & (dualstereo - 1));
1126
1127	rebalance = sbits - (rebalance - f->remaining2);
1128	if (rebalance > 3 << 3 && itheta != 16384)
1129	mbits += rebalance - (3 << 3);
1130
1131	/* In stereo mode, we do not apply a scaling to the mid because
1132	* we need the normalized mid for folding later */
1133	cm |= ff_celt_encode_band(f, rc, band, X, NULL, N, mbits, blocks,
1134	lowband, duration, next_lowband_out1,
1135	next_level, dualstereo ? 1.0f : (gain * mid),
1136	lowband_scratch, fill);
1137	}
1138	}
1139	} else {
1140	/* This is the basic no-split case */
1141	uint32_t q = celt_bits2pulses(cache, b);
1142	uint32_t curr_bits = celt_pulses2bits(cache, q);
1143	f->remaining2 -= curr_bits;
1144
1145	/* Ensures we can never bust the budget */
1146	while (f->remaining2 < 0 && q > 0) {
1147	f->remaining2 += curr_bits;
1148	curr_bits = celt_pulses2bits(cache, --q);
1149	f->remaining2 -= curr_bits;
1150	}
1151
1152	if (q != 0) {
1153	/* Finally do the actual quantization */
1154	cm = celt_alg_quant(rc, X, N, (q < 8) ? q : (8 + (q & 7)) << ((q >> 3) - 1),
1155	f->spread, blocks, gain);
1156	}
1157	}
1158
1159	return cm;
1160	}
1161