2 * Copyright (c) 2008-2009 Rob Sykes <robs@users.sourceforge.net>
3 * Copyright (c) 2017 Paul B Mahol
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 #include "libavutil/avassert.h"
23 #include "libavutil/opt.h"
25 #include "libavcodec/avfft.h"
31 typedef struct SincContext {
34 int sample_rate, nb_samples;
35 float att, beta, phase, Fc0, Fc1, tbw0, tbw1;
43 RDFTContext *rdft, *irdft;
46 static int request_frame(AVFilterLink *outlink)
48 AVFilterContext *ctx = outlink->src;
49 SincContext *s = ctx->priv;
50 const float *coeffs = s->coeffs;
51 AVFrame *frame = NULL;
54 nb_samples = FFMIN(s->nb_samples, s->n - s->pts);
58 if (!(frame = ff_get_audio_buffer(outlink, nb_samples)))
59 return AVERROR(ENOMEM);
61 memcpy(frame->data[0], coeffs + s->pts, nb_samples * sizeof(float));
66 return ff_filter_frame(outlink, frame);
69 static int query_formats(AVFilterContext *ctx)
71 SincContext *s = ctx->priv;
72 static const int64_t chlayouts[] = { AV_CH_LAYOUT_MONO, -1 };
73 int sample_rates[] = { s->sample_rate, -1 };
74 static const enum AVSampleFormat sample_fmts[] = { AV_SAMPLE_FMT_FLT,
76 AVFilterFormats *formats;
77 AVFilterChannelLayouts *layouts;
80 formats = ff_make_format_list(sample_fmts);
82 return AVERROR(ENOMEM);
83 ret = ff_set_common_formats (ctx, formats);
87 layouts = ff_make_format64_list(chlayouts);
89 return AVERROR(ENOMEM);
90 ret = ff_set_common_channel_layouts(ctx, layouts);
94 formats = ff_make_format_list(sample_rates);
96 return AVERROR(ENOMEM);
97 return ff_set_common_samplerates(ctx, formats);
100 static float bessel_I_0(float x)
102 float term = 1, sum = 1, last_sum, x2 = x / 2;
109 sum += term *= y * y;
110 } while (sum != last_sum);
115 static float *make_lpf(int num_taps, float Fc, float beta, float rho,
116 float scale, int dc_norm)
118 int i, m = num_taps - 1;
119 float *h = av_calloc(num_taps, sizeof(*h)), sum = 0;
120 float mult = scale / bessel_I_0(beta), mult1 = 1.f / (.5f * m + rho);
122 av_assert0(Fc >= 0 && Fc <= 1);
124 for (i = 0; i <= m / 2; i++) {
125 float z = i - .5f * m, x = z * M_PI, y = z * mult1;
126 h[i] = x ? sinf(Fc * x) / x : Fc;
127 sum += h[i] *= bessel_I_0(beta * sqrtf(1.f - y * y)) * mult;
134 for (i = 0; dc_norm && i < num_taps; i++)
140 static float kaiser_beta(float att, float tr_bw)
143 static const float coefs[][4] = {
144 {-6.784957e-10, 1.02856e-05, 0.1087556, -0.8988365 + .001},
145 {-6.897885e-10, 1.027433e-05, 0.10876, -0.8994658 + .002},
146 {-1.000683e-09, 1.030092e-05, 0.1087677, -0.9007898 + .003},
147 {-3.654474e-10, 1.040631e-05, 0.1087085, -0.8977766 + .006},
148 {8.106988e-09, 6.983091e-06, 0.1091387, -0.9172048 + .015},
149 {9.519571e-09, 7.272678e-06, 0.1090068, -0.9140768 + .025},
150 {-5.626821e-09, 1.342186e-05, 0.1083999, -0.9065452 + .05},
151 {-9.965946e-08, 5.073548e-05, 0.1040967, -0.7672778 + .085},
152 {1.604808e-07, -5.856462e-05, 0.1185998, -1.34824 + .1},
153 {-1.511964e-07, 6.363034e-05, 0.1064627, -0.9876665 + .18},
155 float realm = logf(tr_bw / .0005f) / logf(2.f);
156 float const *c0 = coefs[av_clip((int)realm, 0, FF_ARRAY_ELEMS(coefs) - 1)];
157 float const *c1 = coefs[av_clip(1 + (int)realm, 0, FF_ARRAY_ELEMS(coefs) - 1)];
158 float b0 = ((c0[0] * att + c0[1]) * att + c0[2]) * att + c0[3];
159 float b1 = ((c1[0] * att + c1[1]) * att + c1[2]) * att + c1[3];
161 return b0 + (b1 - b0) * (realm - (int)realm);
164 return .1102f * (att - 8.7f);
166 return .58417f * powf(att - 20.96f, .4f) + .07886f * (att - 20.96f);
170 static void kaiser_params(float att, float Fc, float tr_bw, float *beta, int *num_taps)
172 *beta = *beta < 0.f ? kaiser_beta(att, tr_bw * .5f / Fc): *beta;
173 att = att < 60.f ? (att - 7.95f) / (2.285f * M_PI * 2.f) :
174 ((.0007528358f-1.577737e-05 * *beta) * *beta + 0.6248022f) * *beta + .06186902f;
175 *num_taps = !*num_taps ? ceilf(att/tr_bw + 1) : *num_taps;
178 static float *lpf(float Fn, float Fc, float tbw, int *num_taps, float att, float *beta, int round)
182 if ((Fc /= Fn) <= 0.f || Fc >= 1.f) {
187 att = att ? att : 120.f;
189 kaiser_params(att, Fc, (tbw ? tbw / Fn : .05f) * .5f, beta, num_taps);
193 *num_taps = av_clip(n, 11, 32767);
195 *num_taps = 1 + 2 * (int)((int)((*num_taps / 2) * Fc + .5f) / Fc + .5f);
198 return make_lpf(*num_taps |= 1, Fc, *beta, 0.f, 1.f, 0);
201 static void invert(float *h, int n)
203 for (int i = 0; i < n; i++)
209 #define PACK(h, n) h[1] = h[n]
210 #define UNPACK(h, n) h[n] = h[1], h[n + 1] = h[1] = 0;
211 #define SQR(a) ((a) * (a))
213 static float safe_log(float x)
221 static int fir_to_phase(SincContext *s, float **h, int *len, int *post_len, float phase)
223 float *pi_wraps, *work, phase1 = (phase > 50.f ? 100.f - phase : phase) / 50.f;
224 int i, work_len, begin, end, imp_peak = 0, peak = 0;
225 float imp_sum = 0, peak_imp_sum = 0;
226 float prev_angle2 = 0, cum_2pi = 0, prev_angle1 = 0, cum_1pi = 0;
228 for (i = *len, work_len = 2 * 2 * 8; i > 1; work_len <<= 1, i >>= 1);
230 /* The first part is for work (+2 for (UN)PACK), the latter for pi_wraps. */
231 work = av_calloc((work_len + 2) + (work_len / 2 + 1), sizeof(float));
233 return AVERROR(ENOMEM);
234 pi_wraps = &work[work_len + 2];
236 memcpy(work, *h, *len * sizeof(*work));
238 av_rdft_end(s->rdft);
239 av_rdft_end(s->irdft);
240 s->rdft = s->irdft = NULL;
241 s->rdft = av_rdft_init(av_log2(work_len), DFT_R2C);
242 s->irdft = av_rdft_init(av_log2(work_len), IDFT_C2R);
243 if (!s->rdft || !s->irdft) {
245 return AVERROR(ENOMEM);
248 av_rdft_calc(s->rdft, work); /* Cepstral: */
249 UNPACK(work, work_len);
251 for (i = 0; i <= work_len; i += 2) {
252 float angle = atan2f(work[i + 1], work[i]);
253 float detect = 2 * M_PI;
254 float delta = angle - prev_angle2;
255 float adjust = detect * ((delta < -detect * .7f) - (delta > detect * .7f));
261 delta = angle - prev_angle1;
262 adjust = detect * ((delta < -detect * .7f) - (delta > detect * .7f));
264 cum_1pi += fabsf(adjust); /* fabs for when 2pi and 1pi have combined */
265 pi_wraps[i >> 1] = cum_1pi;
267 work[i] = safe_log(sqrtf(SQR(work[i]) + SQR(work[i + 1])));
271 PACK(work, work_len);
272 av_rdft_calc(s->irdft, work);
274 for (i = 0; i < work_len; i++)
275 work[i] *= 2.f / work_len;
277 for (i = 1; i < work_len / 2; i++) { /* Window to reject acausal components */
279 work[i + work_len / 2] = 0;
281 av_rdft_calc(s->rdft, work);
283 for (i = 2; i < work_len; i += 2) /* Interpolate between linear & min phase */
284 work[i + 1] = phase1 * i / work_len * pi_wraps[work_len >> 1] + (1 - phase1) * (work[i + 1] + pi_wraps[i >> 1]) - pi_wraps[i >> 1];
286 work[0] = exp(work[0]);
287 work[1] = exp(work[1]);
288 for (i = 2; i < work_len; i += 2) {
289 float x = expf(work[i]);
291 work[i ] = x * cosf(work[i + 1]);
292 work[i + 1] = x * sinf(work[i + 1]);
295 av_rdft_calc(s->irdft, work);
296 for (i = 0; i < work_len; i++)
297 work[i] *= 2.f / work_len;
300 for (i = 0; i <= (int) (pi_wraps[work_len >> 1] / M_PI + .5f); i++) {
302 if (fabs(imp_sum) > fabs(peak_imp_sum)) {
303 peak_imp_sum = imp_sum;
306 if (work[i] > work[imp_peak]) /* For debug check only */
310 while (peak && fabsf(work[peak - 1]) > fabsf(work[peak]) && (work[peak - 1] * work[peak] > 0)) {
316 } else if (phase1 == 1) {
317 begin = peak - *len / 2;
319 begin = (.997f - (2 - phase1) * .22f) * *len + .5f;
320 end = (.997f + (0 - phase1) * .22f) * *len + .5f;
321 begin = peak - (begin & ~3);
322 end = peak + 1 + ((end + 3) & ~3);
324 *h = av_realloc_f(*h, *len, sizeof(**h));
327 return AVERROR(ENOMEM);
331 for (i = 0; i < *len; i++) {
332 (*h)[i] = work[(begin + (phase > 50.f ? *len - 1 - i : i) + work_len) & (work_len - 1)];
334 *post_len = phase > 50 ? peak - begin : begin + *len - (peak + 1);
336 av_log(s, AV_LOG_DEBUG, "%d nPI=%g peak-sum@%i=%g (val@%i=%g); len=%i post=%i (%g%%)\n",
337 work_len, pi_wraps[work_len >> 1] / M_PI, peak, peak_imp_sum, imp_peak,
338 work[imp_peak], *len, *post_len, 100.f - 100.f * *post_len / (*len - 1));
345 static int config_output(AVFilterLink *outlink)
347 AVFilterContext *ctx = outlink->src;
348 SincContext *s = ctx->priv;
349 float Fn = s->sample_rate * .5f;
351 int i, n, post_peak, longer;
353 outlink->sample_rate = s->sample_rate;
356 if (s->Fc0 >= Fn || s->Fc1 >= Fn) {
357 av_log(ctx, AV_LOG_ERROR,
358 "filter frequency must be less than %d/2.\n", s->sample_rate);
359 return AVERROR(EINVAL);
362 h[0] = lpf(Fn, s->Fc0, s->tbw0, &s->num_taps[0], s->att, &s->beta, s->round);
363 h[1] = lpf(Fn, s->Fc1, s->tbw1, &s->num_taps[1], s->att, &s->beta, s->round);
366 invert(h[0], s->num_taps[0]);
368 longer = s->num_taps[1] > s->num_taps[0];
369 n = s->num_taps[longer];
372 for (i = 0; i < s->num_taps[!longer]; i++)
373 h[longer][i + (n - s->num_taps[!longer]) / 2] += h[!longer][i];
376 invert(h[longer], n);
381 if (s->phase != 50.f) {
382 int ret = fir_to_phase(s, &h[longer], &n, &post_peak, s->phase);
389 s->n = 1 << (av_log2(n) + 1);
390 s->rdft_len = 1 << av_log2(n);
391 s->coeffs = av_calloc(s->n, sizeof(*s->coeffs));
393 return AVERROR(ENOMEM);
395 for (i = 0; i < n; i++)
396 s->coeffs[i] = h[longer][i];
399 av_rdft_end(s->rdft);
400 av_rdft_end(s->irdft);
401 s->rdft = s->irdft = NULL;
406 static av_cold void uninit(AVFilterContext *ctx)
408 SincContext *s = ctx->priv;
410 av_freep(&s->coeffs);
411 av_rdft_end(s->rdft);
412 av_rdft_end(s->irdft);
413 s->rdft = s->irdft = NULL;
416 static const AVFilterPad sinc_outputs[] = {
419 .type = AVMEDIA_TYPE_AUDIO,
420 .config_props = config_output,
421 .request_frame = request_frame,
426 #define AF AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
427 #define OFFSET(x) offsetof(SincContext, x)
429 static const AVOption sinc_options[] = {
430 { "sample_rate", "set sample rate", OFFSET(sample_rate), AV_OPT_TYPE_INT, {.i64=44100}, 1, INT_MAX, AF },
431 { "r", "set sample rate", OFFSET(sample_rate), AV_OPT_TYPE_INT, {.i64=44100}, 1, INT_MAX, AF },
432 { "nb_samples", "set the number of samples per requested frame", OFFSET(nb_samples), AV_OPT_TYPE_INT, {.i64=1024}, 1, INT_MAX, AF },
433 { "n", "set the number of samples per requested frame", OFFSET(nb_samples), AV_OPT_TYPE_INT, {.i64=1024}, 1, INT_MAX, AF },
434 { "hp", "set high-pass filter frequency", OFFSET(Fc0), AV_OPT_TYPE_FLOAT, {.dbl=0}, 0, INT_MAX, AF },
435 { "lp", "set low-pass filter frequency", OFFSET(Fc1), AV_OPT_TYPE_FLOAT, {.dbl=0}, 0, INT_MAX, AF },
436 { "phase", "set filter phase response", OFFSET(phase), AV_OPT_TYPE_FLOAT, {.dbl=50}, 0, 100, AF },
437 { "beta", "set kaiser window beta", OFFSET(beta), AV_OPT_TYPE_FLOAT, {.dbl=-1}, -1, 256, AF },
438 { "att", "set stop-band attenuation", OFFSET(att), AV_OPT_TYPE_FLOAT, {.dbl=120}, 40, 180, AF },
439 { "round", "enable rounding", OFFSET(round), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, AF },
440 { "hptaps", "set number of taps for high-pass filter", OFFSET(num_taps[0]), AV_OPT_TYPE_INT, {.i64=0}, 0, 32768, AF },
441 { "lptaps", "set number of taps for low-pass filter", OFFSET(num_taps[1]), AV_OPT_TYPE_INT, {.i64=0}, 0, 32768, AF },
445 AVFILTER_DEFINE_CLASS(sinc);
447 const AVFilter ff_asrc_sinc = {
449 .description = NULL_IF_CONFIG_SMALL("Generate a sinc kaiser-windowed low-pass, high-pass, band-pass, or band-reject FIR coefficients."),
450 .priv_size = sizeof(SincContext),
451 .priv_class = &sinc_class,
452 .query_formats = query_formats,
455 .outputs = sinc_outputs,