2 * Copyright (c) 2019 Paul B Mahol
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 #include "libavutil/avassert.h"
24 #include "libavutil/audio_fifo.h"
25 #include "libavutil/avstring.h"
26 #include "libavutil/opt.h"
31 #include "af_anlmdndsp.h"
33 #define WEIGHT_LUT_NBITS 20
34 #define WEIGHT_LUT_SIZE (1<<WEIGHT_LUT_NBITS)
36 #define SQR(x) ((x) * (x))
38 typedef struct AudioNLMeansContext {
47 float pdiff_lut_scale;
48 float weight_lut[WEIGHT_LUT_SIZE];
64 AudioNLMDNDSPContext dsp;
65 } AudioNLMeansContext;
74 #define OFFSET(x) offsetof(AudioNLMeansContext, x)
75 #define AF AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
76 #define AFT AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_RUNTIME_PARAM
78 static const AVOption anlmdn_options[] = {
79 { "s", "set denoising strength", OFFSET(a), AV_OPT_TYPE_FLOAT, {.dbl=0.00001},0.00001, 10, AFT },
80 { "p", "set patch duration", OFFSET(pd), AV_OPT_TYPE_DURATION, {.i64=2000}, 1000, 100000, AF },
81 { "r", "set research duration", OFFSET(rd), AV_OPT_TYPE_DURATION, {.i64=6000}, 2000, 300000, AF },
82 { "o", "set output mode", OFFSET(om), AV_OPT_TYPE_INT, {.i64=OUT_MODE}, 0, NB_MODES-1, AFT, "mode" },
83 { "i", "input", 0, AV_OPT_TYPE_CONST, {.i64=IN_MODE}, 0, 0, AFT, "mode" },
84 { "o", "output", 0, AV_OPT_TYPE_CONST, {.i64=OUT_MODE}, 0, 0, AFT, "mode" },
85 { "n", "noise", 0, AV_OPT_TYPE_CONST, {.i64=NOISE_MODE},0, 0, AFT, "mode" },
86 { "m", "set smooth factor", OFFSET(m), AV_OPT_TYPE_FLOAT, {.dbl=11.}, 1, 15, AFT },
90 AVFILTER_DEFINE_CLASS(anlmdn);
92 static int query_formats(AVFilterContext *ctx)
94 AVFilterFormats *formats = NULL;
95 AVFilterChannelLayouts *layouts = NULL;
96 static const enum AVSampleFormat sample_fmts[] = {
102 formats = ff_make_format_list(sample_fmts);
104 return AVERROR(ENOMEM);
105 ret = ff_set_common_formats(ctx, formats);
109 layouts = ff_all_channel_counts();
111 return AVERROR(ENOMEM);
113 ret = ff_set_common_channel_layouts(ctx, layouts);
117 formats = ff_all_samplerates();
118 return ff_set_common_samplerates(ctx, formats);
121 static float compute_distance_ssd_c(const float *f1, const float *f2, ptrdiff_t K)
125 for (int k = -K; k <= K; k++)
126 distance += SQR(f1[k] - f2[k]);
131 static void compute_cache_c(float *cache, const float *f,
132 ptrdiff_t S, ptrdiff_t K,
133 ptrdiff_t i, ptrdiff_t jj)
137 for (int j = jj; j < jj + S; j++, v++)
138 cache[v] += -SQR(f[i - K - 1] - f[j - K - 1]) + SQR(f[i + K] - f[j + K]);
141 void ff_anlmdn_init(AudioNLMDNDSPContext *dsp)
143 dsp->compute_distance_ssd = compute_distance_ssd_c;
144 dsp->compute_cache = compute_cache_c;
147 ff_anlmdn_init_x86(dsp);
150 static int config_output(AVFilterLink *outlink)
152 AVFilterContext *ctx = outlink->src;
153 AudioNLMeansContext *s = ctx->priv;
156 s->K = av_rescale(s->pd, outlink->sample_rate, AV_TIME_BASE);
157 s->S = av_rescale(s->rd, outlink->sample_rate, AV_TIME_BASE);
160 s->pts = AV_NOPTS_VALUE;
162 s->N = s->H + (s->K + s->S) * 2;
164 av_log(ctx, AV_LOG_DEBUG, "K:%d S:%d H:%d N:%d\n", s->K, s->S, s->H, s->N);
166 av_frame_free(&s->in);
167 av_frame_free(&s->cache);
168 s->in = ff_get_audio_buffer(outlink, s->N);
170 return AVERROR(ENOMEM);
172 s->cache = ff_get_audio_buffer(outlink, s->S * 2);
174 return AVERROR(ENOMEM);
176 s->fifo = av_audio_fifo_alloc(outlink->format, outlink->channels, s->N);
178 return AVERROR(ENOMEM);
180 ret = av_audio_fifo_write(s->fifo, (void **)s->in->extended_data, s->K + s->S);
184 s->pdiff_lut_scale = 1.f / s->m * WEIGHT_LUT_SIZE;
185 for (int i = 0; i < WEIGHT_LUT_SIZE; i++) {
186 float w = -i / s->pdiff_lut_scale;
188 s->weight_lut[i] = expf(w);
191 ff_anlmdn_init(&s->dsp);
196 static int filter_channel(AVFilterContext *ctx, void *arg, int ch, int nb_jobs)
198 AudioNLMeansContext *s = ctx->priv;
202 const int om = s->om;
203 const float *f = (const float *)(s->in->extended_data[ch]) + K;
204 float *cache = (float *)s->cache->extended_data[ch];
205 const float sw = (65536.f / (4 * K + 2)) / sqrtf(s->a);
206 float *dst = (float *)out->extended_data[ch] + s->offset;
207 const float smooth = s->m;
209 for (int i = S; i < s->H + S; i++) {
210 float P = 0.f, Q = 0.f;
214 for (int j = i - S; j <= i + S; j++) {
217 cache[v++] = s->dsp.compute_distance_ssd(f + i, f + j, K);
220 s->dsp.compute_cache(cache, f, S, K, i, i - S);
221 s->dsp.compute_cache(cache + S, f, S, K, i, i + 1);
224 for (int j = 0; j < 2 * S && !ctx->is_disabled; j++) {
225 const float distance = cache[j];
226 unsigned weight_lut_idx;
229 if (distance < 0.f) {
236 weight_lut_idx = w * s->pdiff_lut_scale;
237 av_assert2(weight_lut_idx < WEIGHT_LUT_SIZE);
238 w = s->weight_lut[weight_lut_idx];
239 P += w * f[i - S + j + (j >= S)];
247 case IN_MODE: dst[i - S] = f[i]; break;
248 case OUT_MODE: dst[i - S] = P / Q; break;
249 case NOISE_MODE: dst[i - S] = f[i] - (P / Q); break;
256 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
258 AVFilterContext *ctx = inlink->dst;
259 AVFilterLink *outlink = ctx->outputs[0];
260 AudioNLMeansContext *s = ctx->priv;
262 int available, wanted, ret;
264 if (s->pts == AV_NOPTS_VALUE)
267 ret = av_audio_fifo_write(s->fifo, (void **)in->extended_data,
272 available = av_audio_fifo_size(s->fifo);
273 wanted = (available / s->H) * s->H;
275 if (wanted >= s->H && available >= s->N) {
276 out = ff_get_audio_buffer(outlink, wanted);
278 return AVERROR(ENOMEM);
281 while (available >= s->N) {
282 ret = av_audio_fifo_peek(s->fifo, (void **)s->in->extended_data, s->N);
286 ctx->internal->execute(ctx, filter_channel, out, NULL, inlink->channels);
288 av_audio_fifo_drain(s->fifo, s->H);
296 out->nb_samples = s->offset;
297 if (s->eof_left >= 0) {
298 out->nb_samples = FFMIN(s->eof_left, s->offset);
299 s->eof_left -= out->nb_samples;
301 s->pts += av_rescale_q(s->offset, (AVRational){1, outlink->sample_rate}, outlink->time_base);
303 return ff_filter_frame(outlink, out);
309 static int request_frame(AVFilterLink *outlink)
311 AVFilterContext *ctx = outlink->src;
312 AudioNLMeansContext *s = ctx->priv;
315 ret = ff_request_frame(ctx->inputs[0]);
317 if (ret == AVERROR_EOF && s->eof_left != 0) {
321 s->eof_left = av_audio_fifo_size(s->fifo) - (s->S + s->K);
322 if (s->eof_left <= 0)
324 in = ff_get_audio_buffer(outlink, s->H);
326 return AVERROR(ENOMEM);
328 return filter_frame(ctx->inputs[0], in);
334 static av_cold void uninit(AVFilterContext *ctx)
336 AudioNLMeansContext *s = ctx->priv;
338 av_audio_fifo_free(s->fifo);
339 av_frame_free(&s->in);
340 av_frame_free(&s->cache);
343 static const AVFilterPad inputs[] = {
346 .type = AVMEDIA_TYPE_AUDIO,
347 .filter_frame = filter_frame,
352 static const AVFilterPad outputs[] = {
355 .type = AVMEDIA_TYPE_AUDIO,
356 .config_props = config_output,
357 .request_frame = request_frame,
362 AVFilter ff_af_anlmdn = {
364 .description = NULL_IF_CONFIG_SMALL("Reduce broadband noise from stream using Non-Local Means."),
365 .query_formats = query_formats,
366 .priv_size = sizeof(AudioNLMeansContext),
367 .priv_class = &anlmdn_class,
371 .process_command = ff_filter_process_command,
372 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL |
373 AVFILTER_FLAG_SLICE_THREADS,