xref: /third_party/ffmpeg/libavcodec/sbcdsp.c (revision cabdff1a)
1/*
2 * Bluetooth low-complexity, subband codec (SBC)
3 *
4 * Copyright (C) 2017  Aurelien Jacobs <aurel@gnuage.org>
5 * Copyright (C) 2012-2013  Intel Corporation
6 * Copyright (C) 2008-2010  Nokia Corporation
7 * Copyright (C) 2004-2010  Marcel Holtmann <marcel@holtmann.org>
8 * Copyright (C) 2004-2005  Henryk Ploetz <henryk@ploetzli.ch>
9 * Copyright (C) 2005-2006  Brad Midgley <bmidgley@xmission.com>
10 *
11 * This file is part of FFmpeg.
12 *
13 * FFmpeg is free software; you can redistribute it and/or
14 * modify it under the terms of the GNU Lesser General Public
15 * License as published by the Free Software Foundation; either
16 * version 2.1 of the License, or (at your option) any later version.
17 *
18 * FFmpeg is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
21 * Lesser General Public License for more details.
22 *
23 * You should have received a copy of the GNU Lesser General Public
24 * License along with FFmpeg; if not, write to the Free Software
25 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 */
27
28/**
29 * @file
30 * SBC basic "building bricks"
31 */
32
33#include <stdint.h>
34#include <limits.h>
35#include <string.h>
36#include "libavutil/common.h"
37#include "libavutil/intmath.h"
38#include "libavutil/intreadwrite.h"
39#include "sbc.h"
40#include "sbcdsp.h"
41#include "sbcdsp_data.h"
42
43/*
44 * A reference C code of analysis filter with SIMD-friendly tables
45 * reordering and code layout. This code can be used to develop platform
46 * specific SIMD optimizations. Also it may be used as some kind of test
47 * for compiler autovectorization capabilities (who knows, if the compiler
48 * is very good at this stuff, hand optimized assembly may be not strictly
49 * needed for some platform).
50 *
51 * Note: It is also possible to make a simple variant of analysis filter,
52 * which needs only a single constants table without taking care about
53 * even/odd cases. This simple variant of filter can be implemented without
54 * input data permutation. The only thing that would be lost is the
55 * possibility to use pairwise SIMD multiplications. But for some simple
56 * CPU cores without SIMD extensions it can be useful. If anybody is
57 * interested in implementing such variant of a filter, sourcecode from
58 * bluez versions 4.26/4.27 can be used as a reference and the history of
59 * the changes in git repository done around that time may be worth checking.
60 */
61
62static av_always_inline void sbc_analyze_simd(const int16_t *in, int32_t *out,
63                                              const int16_t *consts,
64                                              unsigned subbands)
65{
66    int32_t t1[8];
67    int16_t t2[8];
68    int i, j, hop = 0;
69
70    /* rounding coefficient */
71    for (i = 0; i < subbands; i++)
72        t1[i] = 1 << (SBC_PROTO_FIXED_SCALE - 1);
73
74    /* low pass polyphase filter */
75    for (hop = 0; hop < 10*subbands; hop += 2*subbands)
76        for (i = 0; i < 2*subbands; i++)
77            t1[i >> 1] += in[hop + i] * consts[hop + i];
78
79    /* scaling */
80    for (i = 0; i < subbands; i++)
81        t2[i] = t1[i] >> SBC_PROTO_FIXED_SCALE;
82
83    memset(t1, 0, sizeof(t1));
84
85    /* do the cos transform */
86    for (i = 0; i < subbands/2; i++)
87        for (j = 0; j < 2*subbands; j++)
88            t1[j>>1] += t2[i * 2 + (j&1)] * consts[10*subbands + i*2*subbands + j];
89
90    for (i = 0; i < subbands; i++)
91        out[i] = t1[i] >> (SBC_COS_TABLE_FIXED_SCALE - SCALE_OUT_BITS);
92}
93
94static void sbc_analyze_4_simd(const int16_t *in, int32_t *out,
95                               const int16_t *consts)
96{
97    sbc_analyze_simd(in, out, consts, 4);
98}
99
100static void sbc_analyze_8_simd(const int16_t *in, int32_t *out,
101                               const int16_t *consts)
102{
103    sbc_analyze_simd(in, out, consts, 8);
104}
105
106static inline void sbc_analyze_4b_4s_simd(SBCDSPContext *s,
107                                          int16_t *x, int32_t *out, int out_stride)
108{
109    /* Analyze blocks */
110    s->sbc_analyze_4(x + 12, out, ff_sbcdsp_analysis_consts_fixed4_simd_odd);
111    out += out_stride;
112    s->sbc_analyze_4(x + 8, out, ff_sbcdsp_analysis_consts_fixed4_simd_even);
113    out += out_stride;
114    s->sbc_analyze_4(x + 4, out, ff_sbcdsp_analysis_consts_fixed4_simd_odd);
115    out += out_stride;
116    s->sbc_analyze_4(x + 0, out, ff_sbcdsp_analysis_consts_fixed4_simd_even);
117}
118
119static inline void sbc_analyze_4b_8s_simd(SBCDSPContext *s,
120                                          int16_t *x, int32_t *out, int out_stride)
121{
122    /* Analyze blocks */
123    s->sbc_analyze_8(x + 24, out, ff_sbcdsp_analysis_consts_fixed8_simd_odd);
124    out += out_stride;
125    s->sbc_analyze_8(x + 16, out, ff_sbcdsp_analysis_consts_fixed8_simd_even);
126    out += out_stride;
127    s->sbc_analyze_8(x + 8, out, ff_sbcdsp_analysis_consts_fixed8_simd_odd);
128    out += out_stride;
129    s->sbc_analyze_8(x + 0, out, ff_sbcdsp_analysis_consts_fixed8_simd_even);
130}
131
132static inline void sbc_analyze_1b_8s_simd_even(SBCDSPContext *s,
133                                               int16_t *x, int32_t *out,
134                                               int out_stride);
135
136static inline void sbc_analyze_1b_8s_simd_odd(SBCDSPContext *s,
137                                              int16_t *x, int32_t *out,
138                                              int out_stride)
139{
140    s->sbc_analyze_8(x, out, ff_sbcdsp_analysis_consts_fixed8_simd_odd);
141    s->sbc_analyze_8s = sbc_analyze_1b_8s_simd_even;
142}
143
144static inline void sbc_analyze_1b_8s_simd_even(SBCDSPContext *s,
145                                               int16_t *x, int32_t *out,
146                                               int out_stride)
147{
148    s->sbc_analyze_8(x, out, ff_sbcdsp_analysis_consts_fixed8_simd_even);
149    s->sbc_analyze_8s = sbc_analyze_1b_8s_simd_odd;
150}
151
152/*
153 * Input data processing functions. The data is endian converted if needed,
154 * channels are deintrleaved and audio samples are reordered for use in
155 * SIMD-friendly analysis filter function. The results are put into "X"
156 * array, getting appended to the previous data (or it is better to say
157 * prepended, as the buffer is filled from top to bottom). Old data is
158 * discarded when neededed, but availability of (10 * nrof_subbands)
159 * contiguous samples is always guaranteed for the input to the analysis
160 * filter. This is achieved by copying a sufficient part of old data
161 * to the top of the buffer on buffer wraparound.
162 */
163
164static int sbc_enc_process_input_4s(int position, const uint8_t *pcm,
165                                    int16_t X[2][SBC_X_BUFFER_SIZE],
166                                    int nsamples, int nchannels)
167{
168    int c;
169
170    /* handle X buffer wraparound */
171    if (position < nsamples) {
172        for (c = 0; c < nchannels; c++)
173            memcpy(&X[c][SBC_X_BUFFER_SIZE - 40], &X[c][position],
174                            36 * sizeof(int16_t));
175        position = SBC_X_BUFFER_SIZE - 40;
176    }
177
178    /* copy/permutate audio samples */
179    for (; nsamples >= 8; nsamples -= 8, pcm += 16 * nchannels) {
180        position -= 8;
181        for (c = 0; c < nchannels; c++) {
182            int16_t *x = &X[c][position];
183            x[0] = AV_RN16(pcm + 14*nchannels + 2*c);
184            x[1] = AV_RN16(pcm +  6*nchannels + 2*c);
185            x[2] = AV_RN16(pcm + 12*nchannels + 2*c);
186            x[3] = AV_RN16(pcm +  8*nchannels + 2*c);
187            x[4] = AV_RN16(pcm +  0*nchannels + 2*c);
188            x[5] = AV_RN16(pcm +  4*nchannels + 2*c);
189            x[6] = AV_RN16(pcm +  2*nchannels + 2*c);
190            x[7] = AV_RN16(pcm + 10*nchannels + 2*c);
191        }
192    }
193
194    return position;
195}
196
197static int sbc_enc_process_input_8s(int position, const uint8_t *pcm,
198                                    int16_t X[2][SBC_X_BUFFER_SIZE],
199                                    int nsamples, int nchannels)
200{
201    int c;
202
203    /* handle X buffer wraparound */
204    if (position < nsamples) {
205        for (c = 0; c < nchannels; c++)
206            memcpy(&X[c][SBC_X_BUFFER_SIZE - 72], &X[c][position],
207                            72 * sizeof(int16_t));
208        position = SBC_X_BUFFER_SIZE - 72;
209    }
210
211    if (position % 16 == 8) {
212        position -= 8;
213        nsamples -= 8;
214        for (c = 0; c < nchannels; c++) {
215            int16_t *x = &X[c][position];
216            x[0] = AV_RN16(pcm + 14*nchannels + 2*c);
217            x[2] = AV_RN16(pcm + 12*nchannels + 2*c);
218            x[3] = AV_RN16(pcm +  0*nchannels + 2*c);
219            x[4] = AV_RN16(pcm + 10*nchannels + 2*c);
220            x[5] = AV_RN16(pcm +  2*nchannels + 2*c);
221            x[6] = AV_RN16(pcm +  8*nchannels + 2*c);
222            x[7] = AV_RN16(pcm +  4*nchannels + 2*c);
223            x[8] = AV_RN16(pcm +  6*nchannels + 2*c);
224        }
225        pcm += 16 * nchannels;
226    }
227
228    /* copy/permutate audio samples */
229    for (; nsamples >= 16; nsamples -= 16, pcm += 32 * nchannels) {
230        position -= 16;
231        for (c = 0; c < nchannels; c++) {
232            int16_t *x = &X[c][position];
233            x[0]  = AV_RN16(pcm + 30*nchannels + 2*c);
234            x[1]  = AV_RN16(pcm + 14*nchannels + 2*c);
235            x[2]  = AV_RN16(pcm + 28*nchannels + 2*c);
236            x[3]  = AV_RN16(pcm + 16*nchannels + 2*c);
237            x[4]  = AV_RN16(pcm + 26*nchannels + 2*c);
238            x[5]  = AV_RN16(pcm + 18*nchannels + 2*c);
239            x[6]  = AV_RN16(pcm + 24*nchannels + 2*c);
240            x[7]  = AV_RN16(pcm + 20*nchannels + 2*c);
241            x[8]  = AV_RN16(pcm + 22*nchannels + 2*c);
242            x[9]  = AV_RN16(pcm +  6*nchannels + 2*c);
243            x[10] = AV_RN16(pcm + 12*nchannels + 2*c);
244            x[11] = AV_RN16(pcm +  0*nchannels + 2*c);
245            x[12] = AV_RN16(pcm + 10*nchannels + 2*c);
246            x[13] = AV_RN16(pcm +  2*nchannels + 2*c);
247            x[14] = AV_RN16(pcm +  8*nchannels + 2*c);
248            x[15] = AV_RN16(pcm +  4*nchannels + 2*c);
249        }
250    }
251
252    if (nsamples == 8) {
253        position -= 8;
254        for (c = 0; c < nchannels; c++) {
255            int16_t *x = &X[c][position];
256            x[-7] = AV_RN16(pcm + 14*nchannels + 2*c);
257            x[1]  = AV_RN16(pcm +  6*nchannels + 2*c);
258            x[2]  = AV_RN16(pcm + 12*nchannels + 2*c);
259            x[3]  = AV_RN16(pcm +  0*nchannels + 2*c);
260            x[4]  = AV_RN16(pcm + 10*nchannels + 2*c);
261            x[5]  = AV_RN16(pcm +  2*nchannels + 2*c);
262            x[6]  = AV_RN16(pcm +  8*nchannels + 2*c);
263            x[7]  = AV_RN16(pcm +  4*nchannels + 2*c);
264        }
265    }
266
267    return position;
268}
269
270static void sbc_calc_scalefactors(int32_t sb_sample_f[16][2][8],
271                                  uint32_t scale_factor[2][8],
272                                  int blocks, int channels, int subbands)
273{
274    int ch, sb, blk;
275    for (ch = 0; ch < channels; ch++) {
276        for (sb = 0; sb < subbands; sb++) {
277            uint32_t x = 1 << SCALE_OUT_BITS;
278            for (blk = 0; blk < blocks; blk++) {
279                int32_t tmp = FFABS(sb_sample_f[blk][ch][sb]);
280                if (tmp != 0)
281                    x |= tmp - 1;
282            }
283            scale_factor[ch][sb] = (31 - SCALE_OUT_BITS) - ff_clz(x);
284        }
285    }
286}
287
288static int sbc_calc_scalefactors_j(int32_t sb_sample_f[16][2][8],
289                                   uint32_t scale_factor[2][8],
290                                   int blocks, int subbands)
291{
292    int blk, joint = 0;
293    int32_t tmp0, tmp1;
294    uint32_t x, y;
295
296    /* last subband does not use joint stereo */
297    int sb = subbands - 1;
298    x = 1 << SCALE_OUT_BITS;
299    y = 1 << SCALE_OUT_BITS;
300    for (blk = 0; blk < blocks; blk++) {
301        tmp0 = FFABS(sb_sample_f[blk][0][sb]);
302        tmp1 = FFABS(sb_sample_f[blk][1][sb]);
303        if (tmp0 != 0)
304            x |= tmp0 - 1;
305        if (tmp1 != 0)
306            y |= tmp1 - 1;
307    }
308    scale_factor[0][sb] = (31 - SCALE_OUT_BITS) - ff_clz(x);
309    scale_factor[1][sb] = (31 - SCALE_OUT_BITS) - ff_clz(y);
310
311    /* the rest of subbands can use joint stereo */
312    while (--sb >= 0) {
313        int32_t sb_sample_j[16][2];
314        x = 1 << SCALE_OUT_BITS;
315        y = 1 << SCALE_OUT_BITS;
316        for (blk = 0; blk < blocks; blk++) {
317            tmp0 = sb_sample_f[blk][0][sb];
318            tmp1 = sb_sample_f[blk][1][sb];
319            sb_sample_j[blk][0] = (tmp0 >> 1) + (tmp1 >> 1);
320            sb_sample_j[blk][1] = (tmp0 >> 1) - (tmp1 >> 1);
321            tmp0 = FFABS(tmp0);
322            tmp1 = FFABS(tmp1);
323            if (tmp0 != 0)
324                x |= tmp0 - 1;
325            if (tmp1 != 0)
326                y |= tmp1 - 1;
327        }
328        scale_factor[0][sb] = (31 - SCALE_OUT_BITS) -
329            ff_clz(x);
330        scale_factor[1][sb] = (31 - SCALE_OUT_BITS) -
331            ff_clz(y);
332        x = 1 << SCALE_OUT_BITS;
333        y = 1 << SCALE_OUT_BITS;
334        for (blk = 0; blk < blocks; blk++) {
335            tmp0 = FFABS(sb_sample_j[blk][0]);
336            tmp1 = FFABS(sb_sample_j[blk][1]);
337            if (tmp0 != 0)
338                x |= tmp0 - 1;
339            if (tmp1 != 0)
340                y |= tmp1 - 1;
341        }
342        x = (31 - SCALE_OUT_BITS) - ff_clz(x);
343        y = (31 - SCALE_OUT_BITS) - ff_clz(y);
344
345        /* decide whether to use joint stereo for this subband */
346        if ((scale_factor[0][sb] + scale_factor[1][sb]) > x + y) {
347            joint |= 1 << (subbands - 1 - sb);
348            scale_factor[0][sb] = x;
349            scale_factor[1][sb] = y;
350            for (blk = 0; blk < blocks; blk++) {
351                sb_sample_f[blk][0][sb] = sb_sample_j[blk][0];
352                sb_sample_f[blk][1][sb] = sb_sample_j[blk][1];
353            }
354        }
355    }
356
357    /* bitmask with the information about subbands using joint stereo */
358    return joint;
359}
360
361/*
362 * Detect CPU features and setup function pointers
363 */
364av_cold void ff_sbcdsp_init(SBCDSPContext *s)
365{
366    /* Default implementation for analyze functions */
367    s->sbc_analyze_4 = sbc_analyze_4_simd;
368    s->sbc_analyze_8 = sbc_analyze_8_simd;
369    s->sbc_analyze_4s = sbc_analyze_4b_4s_simd;
370    if (s->increment == 1)
371        s->sbc_analyze_8s = sbc_analyze_1b_8s_simd_odd;
372    else
373        s->sbc_analyze_8s = sbc_analyze_4b_8s_simd;
374
375    /* Default implementation for input reordering / deinterleaving */
376    s->sbc_enc_process_input_4s = sbc_enc_process_input_4s;
377    s->sbc_enc_process_input_8s = sbc_enc_process_input_8s;
378
379    /* Default implementation for scale factors calculation */
380    s->sbc_calc_scalefactors = sbc_calc_scalefactors;
381    s->sbc_calc_scalefactors_j = sbc_calc_scalefactors_j;
382
383#if ARCH_ARM
384    ff_sbcdsp_init_arm(s);
385#elif ARCH_X86
386    ff_sbcdsp_init_x86(s);
387#endif
388}
389