| 1 | /* |
| 2 | * Copyright (c) 2012 Justin Ruggles <justin.ruggles@gmail.com> |
| 3 | * |
| 4 | * Triangular with Noise Shaping is based on opusfile. |
| 5 | * Copyright (c) 1994-2012 by the Xiph.Org Foundation and contributors |
| 6 | * |
| 7 | * This file is part of FFmpeg. |
| 8 | * |
| 9 | * FFmpeg 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. |
| 13 | * |
| 14 | * FFmpeg 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. |
| 18 | * |
| 19 | * You should have received a copy of the GNU Lesser General Public |
| 20 | * License along with FFmpeg; if not, write to the Free Software |
| 21 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 22 | */ |
| 23 | |
| 24 | /** |
| 25 | * @file |
| 26 | * Dithered Audio Sample Quantization |
| 27 | * |
| 28 | * Converts from dbl, flt, or s32 to s16 using dithering. |
| 29 | */ |
| 30 | |
| 31 | #include <math.h> |
| 32 | #include <stdint.h> |
| 33 | |
| 34 | #include "libavutil/attributes.h" |
| 35 | #include "libavutil/common.h" |
| 36 | #include "libavutil/lfg.h" |
| 37 | #include "libavutil/mem.h" |
| 38 | #include "libavutil/samplefmt.h" |
| 39 | #include "audio_convert.h" |
| 40 | #include "dither.h" |
| 41 | #include "internal.h" |
| 42 | |
| 43 | typedef struct DitherState { |
| 44 | int mute; |
| 45 | unsigned int seed; |
| 46 | AVLFG lfg; |
| 47 | float *noise_buf; |
| 48 | int noise_buf_size; |
| 49 | int noise_buf_ptr; |
| 50 | float dither_a[4]; |
| 51 | float dither_b[4]; |
| 52 | } DitherState; |
| 53 | |
| 54 | struct DitherContext { |
| 55 | DitherDSPContext ddsp; |
| 56 | enum AVResampleDitherMethod method; |
| 57 | int apply_map; |
| 58 | ChannelMapInfo *ch_map_info; |
| 59 | |
| 60 | int mute_dither_threshold; // threshold for disabling dither |
| 61 | int mute_reset_threshold; // threshold for resetting noise shaping |
| 62 | const float *ns_coef_b; // noise shaping coeffs |
| 63 | const float *ns_coef_a; // noise shaping coeffs |
| 64 | |
| 65 | int channels; |
| 66 | DitherState *state; // dither states for each channel |
| 67 | |
| 68 | AudioData *flt_data; // input data in fltp |
| 69 | AudioData *s16_data; // dithered output in s16p |
| 70 | AudioConvert *ac_in; // converter for input to fltp |
| 71 | AudioConvert *ac_out; // converter for s16p to s16 (if needed) |
| 72 | |
| 73 | void (*quantize)(int16_t *dst, const float *src, float *dither, int len); |
| 74 | int samples_align; |
| 75 | }; |
| 76 | |
| 77 | /* mute threshold, in seconds */ |
| 78 | #define MUTE_THRESHOLD_SEC 0.000333 |
| 79 | |
| 80 | /* scale factor for 16-bit output. |
| 81 | The signal is attenuated slightly to avoid clipping */ |
| 82 | #define S16_SCALE 32753.0f |
| 83 | |
| 84 | /* scale to convert lfg from INT_MIN/INT_MAX to -0.5/0.5 */ |
| 85 | #define LFG_SCALE (1.0f / (2.0f * INT32_MAX)) |
| 86 | |
| 87 | /* noise shaping coefficients */ |
| 88 | |
| 89 | static const float ns_48_coef_b[4] = { |
| 90 | 2.2374f, -0.7339f, -0.1251f, -0.6033f |
| 91 | }; |
| 92 | |
| 93 | static const float ns_48_coef_a[4] = { |
| 94 | 0.9030f, 0.0116f, -0.5853f, -0.2571f |
| 95 | }; |
| 96 | |
| 97 | static const float ns_44_coef_b[4] = { |
| 98 | 2.2061f, -0.4707f, -0.2534f, -0.6213f |
| 99 | }; |
| 100 | |
| 101 | static const float ns_44_coef_a[4] = { |
| 102 | 1.0587f, 0.0676f, -0.6054f, -0.2738f |
| 103 | }; |
| 104 | |
| 105 | static void dither_int_to_float_rectangular_c(float *dst, int *src, int len) |
| 106 | { |
| 107 | int i; |
| 108 | for (i = 0; i < len; i++) |
| 109 | dst[i] = src[i] * LFG_SCALE; |
| 110 | } |
| 111 | |
| 112 | static void dither_int_to_float_triangular_c(float *dst, int *src0, int len) |
| 113 | { |
| 114 | int i; |
| 115 | int *src1 = src0 + len; |
| 116 | |
| 117 | for (i = 0; i < len; i++) { |
| 118 | float r = src0[i] * LFG_SCALE; |
| 119 | r += src1[i] * LFG_SCALE; |
| 120 | dst[i] = r; |
| 121 | } |
| 122 | } |
| 123 | |
| 124 | static void quantize_c(int16_t *dst, const float *src, float *dither, int len) |
| 125 | { |
| 126 | int i; |
| 127 | for (i = 0; i < len; i++) |
| 128 | dst[i] = av_clip_int16(lrintf(src[i] * S16_SCALE + dither[i])); |
| 129 | } |
| 130 | |
| 131 | #define SQRT_1_6 0.40824829046386301723f |
| 132 | |
| 133 | static void dither_highpass_filter(float *src, int len) |
| 134 | { |
| 135 | int i; |
| 136 | |
| 137 | /* filter is from libswresample in FFmpeg */ |
| 138 | for (i = 0; i < len - 2; i++) |
| 139 | src[i] = (-src[i] + 2 * src[i + 1] - src[i + 2]) * SQRT_1_6; |
| 140 | } |
| 141 | |
| 142 | static int generate_dither_noise(DitherContext *c, DitherState *state, |
| 143 | int min_samples) |
| 144 | { |
| 145 | int i; |
| 146 | int nb_samples = FFALIGN(min_samples, 16) + 16; |
| 147 | int buf_samples = nb_samples * |
| 148 | (c->method == AV_RESAMPLE_DITHER_RECTANGULAR ? 1 : 2); |
| 149 | unsigned int *noise_buf_ui; |
| 150 | |
| 151 | av_freep(&state->noise_buf); |
| 152 | state->noise_buf_size = state->noise_buf_ptr = 0; |
| 153 | |
| 154 | state->noise_buf = av_malloc(buf_samples * sizeof(*state->noise_buf)); |
| 155 | if (!state->noise_buf) |
| 156 | return AVERROR(ENOMEM); |
| 157 | state->noise_buf_size = FFALIGN(min_samples, 16); |
| 158 | noise_buf_ui = (unsigned int *)state->noise_buf; |
| 159 | |
| 160 | av_lfg_init(&state->lfg, state->seed); |
| 161 | for (i = 0; i < buf_samples; i++) |
| 162 | noise_buf_ui[i] = av_lfg_get(&state->lfg); |
| 163 | |
| 164 | c->ddsp.dither_int_to_float(state->noise_buf, noise_buf_ui, nb_samples); |
| 165 | |
| 166 | if (c->method == AV_RESAMPLE_DITHER_TRIANGULAR_HP) |
| 167 | dither_highpass_filter(state->noise_buf, nb_samples); |
| 168 | |
| 169 | return 0; |
| 170 | } |
| 171 | |
| 172 | static void quantize_triangular_ns(DitherContext *c, DitherState *state, |
| 173 | int16_t *dst, const float *src, |
| 174 | int nb_samples) |
| 175 | { |
| 176 | int i, j; |
| 177 | float *dither = &state->noise_buf[state->noise_buf_ptr]; |
| 178 | |
| 179 | if (state->mute > c->mute_reset_threshold) |
| 180 | memset(state->dither_a, 0, sizeof(state->dither_a)); |
| 181 | |
| 182 | for (i = 0; i < nb_samples; i++) { |
| 183 | float err = 0; |
| 184 | float sample = src[i] * S16_SCALE; |
| 185 | |
| 186 | for (j = 0; j < 4; j++) { |
| 187 | err += c->ns_coef_b[j] * state->dither_b[j] - |
| 188 | c->ns_coef_a[j] * state->dither_a[j]; |
| 189 | } |
| 190 | for (j = 3; j > 0; j--) { |
| 191 | state->dither_a[j] = state->dither_a[j - 1]; |
| 192 | state->dither_b[j] = state->dither_b[j - 1]; |
| 193 | } |
| 194 | state->dither_a[0] = err; |
| 195 | sample -= err; |
| 196 | |
| 197 | if (state->mute > c->mute_dither_threshold) { |
| 198 | dst[i] = av_clip_int16(lrintf(sample)); |
| 199 | state->dither_b[0] = 0; |
| 200 | } else { |
| 201 | dst[i] = av_clip_int16(lrintf(sample + dither[i])); |
| 202 | state->dither_b[0] = av_clipf(dst[i] - sample, -1.5f, 1.5f); |
| 203 | } |
| 204 | |
| 205 | state->mute++; |
| 206 | if (src[i]) |
| 207 | state->mute = 0; |
| 208 | } |
| 209 | } |
| 210 | |
| 211 | static int convert_samples(DitherContext *c, int16_t **dst, float * const *src, |
| 212 | int channels, int nb_samples) |
| 213 | { |
| 214 | int ch, ret; |
| 215 | int aligned_samples = FFALIGN(nb_samples, 16); |
| 216 | |
| 217 | for (ch = 0; ch < channels; ch++) { |
| 218 | DitherState *state = &c->state[ch]; |
| 219 | |
| 220 | if (state->noise_buf_size < aligned_samples) { |
| 221 | ret = generate_dither_noise(c, state, nb_samples); |
| 222 | if (ret < 0) |
| 223 | return ret; |
| 224 | } else if (state->noise_buf_size - state->noise_buf_ptr < aligned_samples) { |
| 225 | state->noise_buf_ptr = 0; |
| 226 | } |
| 227 | |
| 228 | if (c->method == AV_RESAMPLE_DITHER_TRIANGULAR_NS) { |
| 229 | quantize_triangular_ns(c, state, dst[ch], src[ch], nb_samples); |
| 230 | } else { |
| 231 | c->quantize(dst[ch], src[ch], |
| 232 | &state->noise_buf[state->noise_buf_ptr], |
| 233 | FFALIGN(nb_samples, c->samples_align)); |
| 234 | } |
| 235 | |
| 236 | state->noise_buf_ptr += aligned_samples; |
| 237 | } |
| 238 | |
| 239 | return 0; |
| 240 | } |
| 241 | |
| 242 | int ff_convert_dither(DitherContext *c, AudioData *dst, AudioData *src) |
| 243 | { |
| 244 | int ret; |
| 245 | AudioData *flt_data; |
| 246 | |
| 247 | /* output directly to dst if it is planar */ |
| 248 | if (dst->sample_fmt == AV_SAMPLE_FMT_S16P) |
| 249 | c->s16_data = dst; |
| 250 | else { |
| 251 | /* make sure s16_data is large enough for the output */ |
| 252 | ret = ff_audio_data_realloc(c->s16_data, src->nb_samples); |
| 253 | if (ret < 0) |
| 254 | return ret; |
| 255 | } |
| 256 | |
| 257 | if (src->sample_fmt != AV_SAMPLE_FMT_FLTP || c->apply_map) { |
| 258 | /* make sure flt_data is large enough for the input */ |
| 259 | ret = ff_audio_data_realloc(c->flt_data, src->nb_samples); |
| 260 | if (ret < 0) |
| 261 | return ret; |
| 262 | flt_data = c->flt_data; |
| 263 | } |
| 264 | |
| 265 | if (src->sample_fmt != AV_SAMPLE_FMT_FLTP) { |
| 266 | /* convert input samples to fltp and scale to s16 range */ |
| 267 | ret = ff_audio_convert(c->ac_in, flt_data, src); |
| 268 | if (ret < 0) |
| 269 | return ret; |
| 270 | } else if (c->apply_map) { |
| 271 | ret = ff_audio_data_copy(flt_data, src, c->ch_map_info); |
| 272 | if (ret < 0) |
| 273 | return ret; |
| 274 | } else { |
| 275 | flt_data = src; |
| 276 | } |
| 277 | |
| 278 | /* check alignment and padding constraints */ |
| 279 | if (c->method != AV_RESAMPLE_DITHER_TRIANGULAR_NS) { |
| 280 | int ptr_align = FFMIN(flt_data->ptr_align, c->s16_data->ptr_align); |
| 281 | int samples_align = FFMIN(flt_data->samples_align, c->s16_data->samples_align); |
| 282 | int aligned_len = FFALIGN(src->nb_samples, c->ddsp.samples_align); |
| 283 | |
| 284 | if (!(ptr_align % c->ddsp.ptr_align) && samples_align >= aligned_len) { |
| 285 | c->quantize = c->ddsp.quantize; |
| 286 | c->samples_align = c->ddsp.samples_align; |
| 287 | } else { |
| 288 | c->quantize = quantize_c; |
| 289 | c->samples_align = 1; |
| 290 | } |
| 291 | } |
| 292 | |
| 293 | ret = convert_samples(c, (int16_t **)c->s16_data->data, |
| 294 | (float * const *)flt_data->data, src->channels, |
| 295 | src->nb_samples); |
| 296 | if (ret < 0) |
| 297 | return ret; |
| 298 | |
| 299 | c->s16_data->nb_samples = src->nb_samples; |
| 300 | |
| 301 | /* interleave output to dst if needed */ |
| 302 | if (dst->sample_fmt == AV_SAMPLE_FMT_S16) { |
| 303 | ret = ff_audio_convert(c->ac_out, dst, c->s16_data); |
| 304 | if (ret < 0) |
| 305 | return ret; |
| 306 | } else |
| 307 | c->s16_data = NULL; |
| 308 | |
| 309 | return 0; |
| 310 | } |
| 311 | |
| 312 | void ff_dither_free(DitherContext **cp) |
| 313 | { |
| 314 | DitherContext *c = *cp; |
| 315 | int ch; |
| 316 | |
| 317 | if (!c) |
| 318 | return; |
| 319 | ff_audio_data_free(&c->flt_data); |
| 320 | ff_audio_data_free(&c->s16_data); |
| 321 | ff_audio_convert_free(&c->ac_in); |
| 322 | ff_audio_convert_free(&c->ac_out); |
| 323 | for (ch = 0; ch < c->channels; ch++) |
| 324 | av_free(c->state[ch].noise_buf); |
| 325 | av_free(c->state); |
| 326 | av_freep(cp); |
| 327 | } |
| 328 | |
| 329 | static av_cold void dither_init(DitherDSPContext *ddsp, |
| 330 | enum AVResampleDitherMethod method) |
| 331 | { |
| 332 | ddsp->quantize = quantize_c; |
| 333 | ddsp->ptr_align = 1; |
| 334 | ddsp->samples_align = 1; |
| 335 | |
| 336 | if (method == AV_RESAMPLE_DITHER_RECTANGULAR) |
| 337 | ddsp->dither_int_to_float = dither_int_to_float_rectangular_c; |
| 338 | else |
| 339 | ddsp->dither_int_to_float = dither_int_to_float_triangular_c; |
| 340 | |
| 341 | if (ARCH_X86) |
| 342 | ff_dither_init_x86(ddsp, method); |
| 343 | } |
| 344 | |
| 345 | DitherContext *ff_dither_alloc(AVAudioResampleContext *avr, |
| 346 | enum AVSampleFormat out_fmt, |
| 347 | enum AVSampleFormat in_fmt, |
| 348 | int channels, int sample_rate, int apply_map) |
| 349 | { |
| 350 | AVLFG seed_gen; |
| 351 | DitherContext *c; |
| 352 | int ch; |
| 353 | |
| 354 | if (av_get_packed_sample_fmt(out_fmt) != AV_SAMPLE_FMT_S16 || |
| 355 | av_get_bytes_per_sample(in_fmt) <= 2) { |
| 356 | av_log(avr, AV_LOG_ERROR, "dithering %s to %s is not supported\n", |
| 357 | av_get_sample_fmt_name(in_fmt), av_get_sample_fmt_name(out_fmt)); |
| 358 | return NULL; |
| 359 | } |
| 360 | |
| 361 | c = av_mallocz(sizeof(*c)); |
| 362 | if (!c) |
| 363 | return NULL; |
| 364 | |
| 365 | c->apply_map = apply_map; |
| 366 | if (apply_map) |
| 367 | c->ch_map_info = &avr->ch_map_info; |
| 368 | |
| 369 | if (avr->dither_method == AV_RESAMPLE_DITHER_TRIANGULAR_NS && |
| 370 | sample_rate != 48000 && sample_rate != 44100) { |
| 371 | av_log(avr, AV_LOG_WARNING, "sample rate must be 48000 or 44100 Hz " |
| 372 | "for triangular_ns dither. using triangular_hp instead.\n"); |
| 373 | avr->dither_method = AV_RESAMPLE_DITHER_TRIANGULAR_HP; |
| 374 | } |
| 375 | c->method = avr->dither_method; |
| 376 | dither_init(&c->ddsp, c->method); |
| 377 | |
| 378 | if (c->method == AV_RESAMPLE_DITHER_TRIANGULAR_NS) { |
| 379 | if (sample_rate == 48000) { |
| 380 | c->ns_coef_b = ns_48_coef_b; |
| 381 | c->ns_coef_a = ns_48_coef_a; |
| 382 | } else { |
| 383 | c->ns_coef_b = ns_44_coef_b; |
| 384 | c->ns_coef_a = ns_44_coef_a; |
| 385 | } |
| 386 | } |
| 387 | |
| 388 | /* Either s16 or s16p output format is allowed, but s16p is used |
| 389 | internally, so we need to use a temp buffer and interleave if the output |
| 390 | format is s16 */ |
| 391 | if (out_fmt != AV_SAMPLE_FMT_S16P) { |
| 392 | c->s16_data = ff_audio_data_alloc(channels, 1024, AV_SAMPLE_FMT_S16P, |
| 393 | "dither s16 buffer"); |
| 394 | if (!c->s16_data) |
| 395 | goto fail; |
| 396 | |
| 397 | c->ac_out = ff_audio_convert_alloc(avr, out_fmt, AV_SAMPLE_FMT_S16P, |
| 398 | channels, sample_rate, 0); |
| 399 | if (!c->ac_out) |
| 400 | goto fail; |
| 401 | } |
| 402 | |
| 403 | if (in_fmt != AV_SAMPLE_FMT_FLTP || c->apply_map) { |
| 404 | c->flt_data = ff_audio_data_alloc(channels, 1024, AV_SAMPLE_FMT_FLTP, |
| 405 | "dither flt buffer"); |
| 406 | if (!c->flt_data) |
| 407 | goto fail; |
| 408 | } |
| 409 | if (in_fmt != AV_SAMPLE_FMT_FLTP) { |
| 410 | c->ac_in = ff_audio_convert_alloc(avr, AV_SAMPLE_FMT_FLTP, in_fmt, |
| 411 | channels, sample_rate, c->apply_map); |
| 412 | if (!c->ac_in) |
| 413 | goto fail; |
| 414 | } |
| 415 | |
| 416 | c->state = av_mallocz(channels * sizeof(*c->state)); |
| 417 | if (!c->state) |
| 418 | goto fail; |
| 419 | c->channels = channels; |
| 420 | |
| 421 | /* calculate thresholds for turning off dithering during periods of |
| 422 | silence to avoid replacing digital silence with quiet dither noise */ |
| 423 | c->mute_dither_threshold = lrintf(sample_rate * MUTE_THRESHOLD_SEC); |
| 424 | c->mute_reset_threshold = c->mute_dither_threshold * 4; |
| 425 | |
| 426 | /* initialize dither states */ |
| 427 | av_lfg_init(&seed_gen, 0xC0FFEE); |
| 428 | for (ch = 0; ch < channels; ch++) { |
| 429 | DitherState *state = &c->state[ch]; |
| 430 | state->mute = c->mute_reset_threshold + 1; |
| 431 | state->seed = av_lfg_get(&seed_gen); |
| 432 | generate_dither_noise(c, state, FFMAX(32768, sample_rate / 2)); |
| 433 | } |
| 434 | |
| 435 | return c; |
| 436 | |
| 437 | fail: |
| 438 | ff_dither_free(&c); |
| 439 | return NULL; |
| 440 | } |