2 * Copyright (c) 2012 Justin Ruggles <justin.ruggles@gmail.com>
4 * Triangular with Noise Shaping is based on opusfile.
5 * Copyright (c) 1994-2012 by the Xiph.Org Foundation and contributors
7 * This file is part of Libav.
9 * Libav is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * Libav is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with Libav; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 * Dithered Audio Sample Quantization
28 * Converts from dbl, flt, or s32 to s16 using dithering.
34 #include "libavutil/common.h"
35 #include "libavutil/lfg.h"
36 #include "libavutil/mem.h"
37 #include "libavutil/samplefmt.h"
38 #include "audio_convert.h"
42 typedef struct DitherState {
53 struct DitherContext {
54 DitherDSPContext ddsp;
55 enum AVResampleDitherMethod method;
57 ChannelMapInfo *ch_map_info;
59 int mute_dither_threshold; // threshold for disabling dither
60 int mute_reset_threshold; // threshold for resetting noise shaping
61 const float *ns_coef_b; // noise shaping coeffs
62 const float *ns_coef_a; // noise shaping coeffs
65 DitherState *state; // dither states for each channel
67 AudioData *flt_data; // input data in fltp
68 AudioData *s16_data; // dithered output in s16p
69 AudioConvert *ac_in; // converter for input to fltp
70 AudioConvert *ac_out; // converter for s16p to s16 (if needed)
72 void (*quantize)(int16_t *dst, const float *src, float *dither, int len);
76 /* mute threshold, in seconds */
77 #define MUTE_THRESHOLD_SEC 0.000333
79 /* scale factor for 16-bit output.
80 The signal is attenuated slightly to avoid clipping */
81 #define S16_SCALE 32753.0f
83 /* scale to convert lfg from INT_MIN/INT_MAX to -0.5/0.5 */
84 #define LFG_SCALE (1.0f / (2.0f * INT32_MAX))
86 /* noise shaping coefficients */
88 static const float ns_48_coef_b[4] = {
89 2.2374f, -0.7339f, -0.1251f, -0.6033f
92 static const float ns_48_coef_a[4] = {
93 0.9030f, 0.0116f, -0.5853f, -0.2571f
96 static const float ns_44_coef_b[4] = {
97 2.2061f, -0.4707f, -0.2534f, -0.6213f
100 static const float ns_44_coef_a[4] = {
101 1.0587f, 0.0676f, -0.6054f, -0.2738f
104 static void dither_int_to_float_rectangular_c(float *dst, int *src, int len)
107 for (i = 0; i < len; i++)
108 dst[i] = src[i] * LFG_SCALE;
111 static void dither_int_to_float_triangular_c(float *dst, int *src0, int len)
114 int *src1 = src0 + len;
116 for (i = 0; i < len; i++) {
117 float r = src0[i] * LFG_SCALE;
118 r += src1[i] * LFG_SCALE;
123 static void quantize_c(int16_t *dst, const float *src, float *dither, int len)
126 for (i = 0; i < len; i++)
127 dst[i] = av_clip_int16(lrintf(src[i] * S16_SCALE + dither[i]));
130 #define SQRT_1_6 0.40824829046386301723f
132 static void dither_highpass_filter(float *src, int len)
136 /* filter is from libswresample in FFmpeg */
137 for (i = 0; i < len - 2; i++)
138 src[i] = (-src[i] + 2 * src[i + 1] - src[i + 2]) * SQRT_1_6;
141 static int generate_dither_noise(DitherContext *c, DitherState *state,
145 int nb_samples = FFALIGN(min_samples, 16) + 16;
146 int buf_samples = nb_samples *
147 (c->method == AV_RESAMPLE_DITHER_RECTANGULAR ? 1 : 2);
148 unsigned int *noise_buf_ui;
150 av_freep(&state->noise_buf);
151 state->noise_buf_size = state->noise_buf_ptr = 0;
153 state->noise_buf = av_malloc(buf_samples * sizeof(*state->noise_buf));
154 if (!state->noise_buf)
155 return AVERROR(ENOMEM);
156 state->noise_buf_size = FFALIGN(min_samples, 16);
157 noise_buf_ui = (unsigned int *)state->noise_buf;
159 av_lfg_init(&state->lfg, state->seed);
160 for (i = 0; i < buf_samples; i++)
161 noise_buf_ui[i] = av_lfg_get(&state->lfg);
163 c->ddsp.dither_int_to_float(state->noise_buf, noise_buf_ui, nb_samples);
165 if (c->method == AV_RESAMPLE_DITHER_TRIANGULAR_HP)
166 dither_highpass_filter(state->noise_buf, nb_samples);
171 static void quantize_triangular_ns(DitherContext *c, DitherState *state,
172 int16_t *dst, const float *src,
176 float *dither = &state->noise_buf[state->noise_buf_ptr];
178 if (state->mute > c->mute_reset_threshold)
179 memset(state->dither_a, 0, sizeof(state->dither_a));
181 for (i = 0; i < nb_samples; i++) {
183 float sample = src[i] * S16_SCALE;
185 for (j = 0; j < 4; j++) {
186 err += c->ns_coef_b[j] * state->dither_b[j] -
187 c->ns_coef_a[j] * state->dither_a[j];
189 for (j = 3; j > 0; j--) {
190 state->dither_a[j] = state->dither_a[j - 1];
191 state->dither_b[j] = state->dither_b[j - 1];
193 state->dither_a[0] = err;
196 if (state->mute > c->mute_dither_threshold) {
197 dst[i] = av_clip_int16(lrintf(sample));
198 state->dither_b[0] = 0;
200 dst[i] = av_clip_int16(lrintf(sample + dither[i]));
201 state->dither_b[0] = av_clipf(dst[i] - sample, -1.5f, 1.5f);
210 static int convert_samples(DitherContext *c, int16_t **dst, float * const *src,
211 int channels, int nb_samples)
214 int aligned_samples = FFALIGN(nb_samples, 16);
216 for (ch = 0; ch < channels; ch++) {
217 DitherState *state = &c->state[ch];
219 if (state->noise_buf_size < aligned_samples) {
220 ret = generate_dither_noise(c, state, nb_samples);
223 } else if (state->noise_buf_size - state->noise_buf_ptr < aligned_samples) {
224 state->noise_buf_ptr = 0;
227 if (c->method == AV_RESAMPLE_DITHER_TRIANGULAR_NS) {
228 quantize_triangular_ns(c, state, dst[ch], src[ch], nb_samples);
230 c->quantize(dst[ch], src[ch],
231 &state->noise_buf[state->noise_buf_ptr],
232 FFALIGN(nb_samples, c->samples_align));
235 state->noise_buf_ptr += aligned_samples;
241 int ff_convert_dither(DitherContext *c, AudioData *dst, AudioData *src)
246 /* output directly to dst if it is planar */
247 if (dst->sample_fmt == AV_SAMPLE_FMT_S16P)
250 /* make sure s16_data is large enough for the output */
251 ret = ff_audio_data_realloc(c->s16_data, src->nb_samples);
256 if (src->sample_fmt != AV_SAMPLE_FMT_FLTP || c->apply_map) {
257 /* make sure flt_data is large enough for the input */
258 ret = ff_audio_data_realloc(c->flt_data, src->nb_samples);
261 flt_data = c->flt_data;
264 if (src->sample_fmt != AV_SAMPLE_FMT_FLTP) {
265 /* convert input samples to fltp and scale to s16 range */
266 ret = ff_audio_convert(c->ac_in, flt_data, src);
269 } else if (c->apply_map) {
270 ret = ff_audio_data_copy(flt_data, src, c->ch_map_info);
277 /* check alignment and padding constraints */
278 if (c->method != AV_RESAMPLE_DITHER_TRIANGULAR_NS) {
279 int ptr_align = FFMIN(flt_data->ptr_align, c->s16_data->ptr_align);
280 int samples_align = FFMIN(flt_data->samples_align, c->s16_data->samples_align);
281 int aligned_len = FFALIGN(src->nb_samples, c->ddsp.samples_align);
283 if (!(ptr_align % c->ddsp.ptr_align) && samples_align >= aligned_len) {
284 c->quantize = c->ddsp.quantize;
285 c->samples_align = c->ddsp.samples_align;
287 c->quantize = quantize_c;
288 c->samples_align = 1;
292 ret = convert_samples(c, (int16_t **)c->s16_data->data,
293 (float * const *)flt_data->data, src->channels,
298 c->s16_data->nb_samples = src->nb_samples;
300 /* interleave output to dst if needed */
301 if (dst->sample_fmt == AV_SAMPLE_FMT_S16) {
302 ret = ff_audio_convert(c->ac_out, dst, c->s16_data);
311 void ff_dither_free(DitherContext **cp)
313 DitherContext *c = *cp;
318 ff_audio_data_free(&c->flt_data);
319 ff_audio_data_free(&c->s16_data);
320 ff_audio_convert_free(&c->ac_in);
321 ff_audio_convert_free(&c->ac_out);
322 for (ch = 0; ch < c->channels; ch++)
323 av_free(c->state[ch].noise_buf);
328 static void dither_init(DitherDSPContext *ddsp,
329 enum AVResampleDitherMethod method)
331 ddsp->quantize = quantize_c;
333 ddsp->samples_align = 1;
335 if (method == AV_RESAMPLE_DITHER_RECTANGULAR)
336 ddsp->dither_int_to_float = dither_int_to_float_rectangular_c;
338 ddsp->dither_int_to_float = dither_int_to_float_triangular_c;
341 ff_dither_init_x86(ddsp, method);
344 DitherContext *ff_dither_alloc(AVAudioResampleContext *avr,
345 enum AVSampleFormat out_fmt,
346 enum AVSampleFormat in_fmt,
347 int channels, int sample_rate, int apply_map)
353 if (av_get_packed_sample_fmt(out_fmt) != AV_SAMPLE_FMT_S16 ||
354 av_get_bytes_per_sample(in_fmt) <= 2) {
355 av_log(avr, AV_LOG_ERROR, "dithering %s to %s is not supported\n",
356 av_get_sample_fmt_name(in_fmt), av_get_sample_fmt_name(out_fmt));
360 c = av_mallocz(sizeof(*c));
364 c->apply_map = apply_map;
366 c->ch_map_info = &avr->ch_map_info;
368 if (avr->dither_method == AV_RESAMPLE_DITHER_TRIANGULAR_NS &&
369 sample_rate != 48000 && sample_rate != 44100) {
370 av_log(avr, AV_LOG_WARNING, "sample rate must be 48000 or 44100 Hz "
371 "for triangular_ns dither. using triangular_hp instead.\n");
372 avr->dither_method = AV_RESAMPLE_DITHER_TRIANGULAR_HP;
374 c->method = avr->dither_method;
375 dither_init(&c->ddsp, c->method);
377 if (c->method == AV_RESAMPLE_DITHER_TRIANGULAR_NS) {
378 if (sample_rate == 48000) {
379 c->ns_coef_b = ns_48_coef_b;
380 c->ns_coef_a = ns_48_coef_a;
382 c->ns_coef_b = ns_44_coef_b;
383 c->ns_coef_a = ns_44_coef_a;
387 /* Either s16 or s16p output format is allowed, but s16p is used
388 internally, so we need to use a temp buffer and interleave if the output
390 if (out_fmt != AV_SAMPLE_FMT_S16P) {
391 c->s16_data = ff_audio_data_alloc(channels, 1024, AV_SAMPLE_FMT_S16P,
392 "dither s16 buffer");
396 c->ac_out = ff_audio_convert_alloc(avr, out_fmt, AV_SAMPLE_FMT_S16P,
397 channels, sample_rate, 0);
402 if (in_fmt != AV_SAMPLE_FMT_FLTP || c->apply_map) {
403 c->flt_data = ff_audio_data_alloc(channels, 1024, AV_SAMPLE_FMT_FLTP,
404 "dither flt buffer");
408 if (in_fmt != AV_SAMPLE_FMT_FLTP) {
409 c->ac_in = ff_audio_convert_alloc(avr, AV_SAMPLE_FMT_FLTP, in_fmt,
410 channels, sample_rate, c->apply_map);
415 c->state = av_mallocz(channels * sizeof(*c->state));
418 c->channels = channels;
420 /* calculate thresholds for turning off dithering during periods of
421 silence to avoid replacing digital silence with quiet dither noise */
422 c->mute_dither_threshold = lrintf(sample_rate * MUTE_THRESHOLD_SEC);
423 c->mute_reset_threshold = c->mute_dither_threshold * 4;
425 /* initialize dither states */
426 av_lfg_init(&seed_gen, 0xC0FFEE);
427 for (ch = 0; ch < channels; ch++) {
428 DitherState *state = &c->state[ch];
429 state->mute = c->mute_reset_threshold + 1;
430 state->seed = av_lfg_get(&seed_gen);
431 generate_dither_noise(c, state, FFMAX(32768, sample_rate / 2));