diff --git a/doc/filters.texi b/doc/filters.texi index 83c48fe367..566bb94f97 100644 --- a/doc/filters.texi +++ b/doc/filters.texi @@ -3041,6 +3041,20 @@ Pass error signal estimated samples. Default value is @var{o}. @end table + +@item precision +Set which precision to use when processing samples. + +@table @option +@item auto +Auto pick internal sample format depending on other filters. + +@item float +Always use single-floating point precision sample format. + +@item double +Always use double-floating point precision sample format. +@end table @end table @section arnndn diff --git a/libavfilter/af_arls.c b/libavfilter/af_arls.c index 4f2eddffc4..85e4f92425 100644 --- a/libavfilter/af_arls.c +++ b/libavfilter/af_arls.c @@ -24,6 +24,7 @@ #include "audio.h" #include "avfilter.h" +#include "formats.h" #include "filters.h" #include "internal.h" @@ -43,6 +44,7 @@ typedef struct AudioRLSContext { float lambda; float delta; int output_mode; + int precision; int kernel_size; AVFrame *offset; @@ -54,6 +56,8 @@ typedef struct AudioRLSContext { AVFrame *frame[2]; + int (*filter_channels)(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs); + AVFloatDSPContext *fdsp; } AudioRLSContext; @@ -71,117 +75,32 @@ static const AVOption arls_options[] = { { "o", "output", 0, AV_OPT_TYPE_CONST, {.i64=OUT_MODE}, 0, 0, AT, "mode" }, { "n", "noise", 0, AV_OPT_TYPE_CONST, {.i64=NOISE_MODE}, 0, 0, AT, "mode" }, { "e", "error", 0, AV_OPT_TYPE_CONST, {.i64=ERROR_MODE}, 0, 0, AT, "mode" }, + { "precision", "set processing precision", OFFSET(precision), AV_OPT_TYPE_INT, {.i64=0}, 0, 2, A, "precision" }, + { "auto", "set auto processing precision", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, A, "precision" }, + { "float", "set single-floating point processing precision", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, A, "precision" }, + { "double","set double-floating point processing precision", 0, AV_OPT_TYPE_CONST, {.i64=2}, 0, 0, A, "precision" }, { NULL } }; AVFILTER_DEFINE_CLASS(arls); -static float fir_sample(AudioRLSContext *s, float sample, float *delay, - float *coeffs, float *tmp, int *offset) -{ - const int order = s->order; - float output; - - delay[*offset] = sample; - - memcpy(tmp, coeffs + order - *offset, order * sizeof(float)); - - output = s->fdsp->scalarproduct_float(delay, tmp, s->kernel_size); - - if (--(*offset) < 0) - *offset = order - 1; - - return output; -} - -static float process_sample(AudioRLSContext *s, float input, float desired, int ch) -{ - float *coeffs = (float *)s->coeffs->extended_data[ch]; - float *delay = (float *)s->delay->extended_data[ch]; - float *gains = (float *)s->gains->extended_data[ch]; - float *tmp = (float *)s->tmp->extended_data[ch]; - float *u = (float *)s->u->extended_data[ch]; - float *p = (float *)s->p->extended_data[ch]; - float *dp = (float *)s->dp->extended_data[ch]; - int *offsetp = (int *)s->offset->extended_data[ch]; - const int kernel_size = s->kernel_size; - const int order = s->order; - const float lambda = s->lambda; - int offset = *offsetp; - float g = lambda; - float output, e; - - delay[offset + order] = input; - - output = fir_sample(s, input, delay, coeffs, tmp, offsetp); - e = desired - output; - - for (int i = 0, pos = offset; i < order; i++, pos++) { - const int ikernel_size = i * kernel_size; - - u[i] = 0.f; - for (int k = 0, pos = offset; k < order; k++, pos++) - u[i] += p[ikernel_size + k] * delay[pos]; - - g += u[i] * delay[pos]; - } - - g = 1.f / g; - - for (int i = 0; i < order; i++) { - const int ikernel_size = i * kernel_size; - - gains[i] = u[i] * g; - coeffs[i] = coeffs[order + i] = coeffs[i] + gains[i] * e; - tmp[i] = 0.f; - for (int k = 0, pos = offset; k < order; k++, pos++) - tmp[i] += p[ikernel_size + k] * delay[pos]; - } - - for (int i = 0; i < order; i++) { - const int ikernel_size = i * kernel_size; - - for (int k = 0; k < order; k++) - dp[ikernel_size + k] = gains[i] * tmp[k]; - } - - for (int i = 0; i < order; i++) { - const int ikernel_size = i * kernel_size; - - for (int k = 0; k < order; k++) - p[ikernel_size + k] = (p[ikernel_size + k] - (dp[ikernel_size + k] + dp[kernel_size * k + i]) * 0.5f) * lambda; - } - - switch (s->output_mode) { - case IN_MODE: output = input; break; - case DESIRED_MODE: output = desired; break; - case OUT_MODE: output = desired - output; break; - case NOISE_MODE: output = input - output; break; - case ERROR_MODE: break; - } - return output; -} - -static int process_channels(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs) +static int query_formats(AVFilterContext *ctx) { AudioRLSContext *s = ctx->priv; - AVFrame *out = arg; - const int start = (out->ch_layout.nb_channels * jobnr) / nb_jobs; - const int end = (out->ch_layout.nb_channels * (jobnr+1)) / nb_jobs; + static const enum AVSampleFormat sample_fmts[3][3] = { + { AV_SAMPLE_FMT_FLTP, AV_SAMPLE_FMT_DBLP, AV_SAMPLE_FMT_NONE }, + { AV_SAMPLE_FMT_FLTP, AV_SAMPLE_FMT_NONE }, + { AV_SAMPLE_FMT_DBLP, AV_SAMPLE_FMT_NONE }, + }; + int ret; - for (int c = start; c < end; c++) { - const float *input = (const float *)s->frame[0]->extended_data[c]; - const float *desired = (const float *)s->frame[1]->extended_data[c]; - float *output = (float *)out->extended_data[c]; + if ((ret = ff_set_common_all_channel_counts(ctx)) < 0) + return ret; - for (int n = 0; n < out->nb_samples; n++) { - output[n] = process_sample(s, input[n], desired[n], c); - if (ctx->is_disabled) - output[n] = input[n]; - } - } + if ((ret = ff_set_common_formats_from_list(ctx, sample_fmts[s->precision])) < 0) + return ret; - return 0; + return ff_set_common_all_samplerates(ctx); } static int activate(AVFilterContext *ctx) @@ -216,7 +135,7 @@ static int activate(AVFilterContext *ctx) return AVERROR(ENOMEM); } - ff_filter_execute(ctx, process_channels, out, NULL, + ff_filter_execute(ctx, s->filter_channels, out, NULL, FFMIN(ctx->outputs[0]->ch_layout.nb_channels, ff_filter_get_nb_threads(ctx))); out->pts = s->frame[0]->pts; @@ -249,6 +168,13 @@ static int activate(AVFilterContext *ctx) return 0; } +#define DEPTH 32 +#include "arls_template.c" + +#undef DEPTH +#define DEPTH 64 +#include "arls_template.c" + static int config_output(AVFilterLink *outlink) { AVFilterContext *ctx = outlink->src; @@ -283,11 +209,27 @@ static int config_output(AVFilterLink *outlink) dst[0] = s->kernel_size - 1; } - for (int ch = 0; ch < s->p->ch_layout.nb_channels; ch++) { - float *dst = (float *)s->p->extended_data[ch]; + switch (outlink->format) { + case AV_SAMPLE_FMT_DBLP: + for (int ch = 0; ch < s->p->ch_layout.nb_channels; ch++) { + double *dst = (double *)s->p->extended_data[ch]; - for (int i = 0; i < s->kernel_size; i++) - dst[i * s->kernel_size + i] = s->delta; + for (int i = 0; i < s->kernel_size; i++) + dst[i * s->kernel_size + i] = s->delta; + } + + s->filter_channels = filter_channels_double; + break; + case AV_SAMPLE_FMT_FLTP: + for (int ch = 0; ch < s->p->ch_layout.nb_channels; ch++) { + float *dst = (float *)s->p->extended_data[ch]; + + for (int i = 0; i < s->kernel_size; i++) + dst[i * s->kernel_size + i] = s->delta; + } + + s->filter_channels = filter_channels_float; + break; } return 0; @@ -348,7 +290,7 @@ const AVFilter ff_af_arls = { .activate = activate, FILTER_INPUTS(inputs), FILTER_OUTPUTS(outputs), - FILTER_SINGLE_SAMPLEFMT(AV_SAMPLE_FMT_FLTP), + FILTER_QUERY_FUNC(query_formats), .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL | AVFILTER_FLAG_SLICE_THREADS, .process_command = ff_filter_process_command, diff --git a/libavfilter/arls_template.c b/libavfilter/arls_template.c new file mode 100644 index 0000000000..e5b91d0a85 --- /dev/null +++ b/libavfilter/arls_template.c @@ -0,0 +1,158 @@ +/* + * This file is part of FFmpeg. + * + * FFmpeg is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * FFmpeg is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with FFmpeg; if not, write to the Free Software + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +#undef ONE +#undef ftype +#undef SAMPLE_FORMAT +#if DEPTH == 32 +#define SAMPLE_FORMAT float +#define ftype float +#define ONE 1.f +#else +#define SAMPLE_FORMAT double +#define ftype double +#define ONE 1.0 +#endif + +#define fn3(a,b) a##_##b +#define fn2(a,b) fn3(a,b) +#define fn(a) fn2(a, SAMPLE_FORMAT) + +#if DEPTH == 64 +static double scalarproduct_double(const double *v1, const double *v2, int len) +{ + double p = 0.0; + + for (int i = 0; i < len; i++) + p += v1[i] * v2[i]; + + return p; +} +#endif + +static ftype fn(fir_sample)(AudioRLSContext *s, ftype sample, ftype *delay, + ftype *coeffs, ftype *tmp, int *offset) +{ + const int order = s->order; + ftype output; + + delay[*offset] = sample; + + memcpy(tmp, coeffs + order - *offset, order * sizeof(ftype)); + +#if DEPTH == 32 + output = s->fdsp->scalarproduct_float(delay, tmp, s->kernel_size); +#else + output = scalarproduct_double(delay, tmp, s->kernel_size); +#endif + + if (--(*offset) < 0) + *offset = order - 1; + + return output; +} + +static ftype fn(process_sample)(AudioRLSContext *s, ftype input, ftype desired, int ch) +{ + ftype *coeffs = (ftype *)s->coeffs->extended_data[ch]; + ftype *delay = (ftype *)s->delay->extended_data[ch]; + ftype *gains = (ftype *)s->gains->extended_data[ch]; + ftype *tmp = (ftype *)s->tmp->extended_data[ch]; + ftype *u = (ftype *)s->u->extended_data[ch]; + ftype *p = (ftype *)s->p->extended_data[ch]; + ftype *dp = (ftype *)s->dp->extended_data[ch]; + int *offsetp = (int *)s->offset->extended_data[ch]; + const int kernel_size = s->kernel_size; + const int order = s->order; + const ftype lambda = s->lambda; + int offset = *offsetp; + ftype g = lambda; + ftype output, e; + + delay[offset + order] = input; + + output = fn(fir_sample)(s, input, delay, coeffs, tmp, offsetp); + e = desired - output; + + for (int i = 0, pos = offset; i < order; i++, pos++) { + const int ikernel_size = i * kernel_size; + + u[i] = 0.f; + for (int k = 0, pos = offset; k < order; k++, pos++) + u[i] += p[ikernel_size + k] * delay[pos]; + + g += u[i] * delay[pos]; + } + + g = 1.f / g; + + for (int i = 0; i < order; i++) { + const int ikernel_size = i * kernel_size; + + gains[i] = u[i] * g; + coeffs[i] = coeffs[order + i] = coeffs[i] + gains[i] * e; + tmp[i] = 0.f; + for (int k = 0, pos = offset; k < order; k++, pos++) + tmp[i] += p[ikernel_size + k] * delay[pos]; + } + + for (int i = 0; i < order; i++) { + const int ikernel_size = i * kernel_size; + + for (int k = 0; k < order; k++) + dp[ikernel_size + k] = gains[i] * tmp[k]; + } + + for (int i = 0; i < order; i++) { + const int ikernel_size = i * kernel_size; + + for (int k = 0; k < order; k++) + p[ikernel_size + k] = (p[ikernel_size + k] - (dp[ikernel_size + k] + dp[kernel_size * k + i]) * 0.5f) * lambda; + } + + switch (s->output_mode) { + case IN_MODE: output = input; break; + case DESIRED_MODE: output = desired; break; + case OUT_MODE: output = desired - output; break; + case NOISE_MODE: output = input - output; break; + case ERROR_MODE: break; + } + return output; +} + +static int fn(filter_channels)(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs) +{ + AudioRLSContext *s = ctx->priv; + AVFrame *out = arg; + const int start = (out->ch_layout.nb_channels * jobnr) / nb_jobs; + const int end = (out->ch_layout.nb_channels * (jobnr+1)) / nb_jobs; + + for (int c = start; c < end; c++) { + const ftype *input = (const ftype *)s->frame[0]->extended_data[c]; + const ftype *desired = (const ftype *)s->frame[1]->extended_data[c]; + ftype *output = (ftype *)out->extended_data[c]; + + for (int n = 0; n < out->nb_samples; n++) { + output[n] = fn(process_sample)(s, input[n], desired[n], c); + if (ctx->is_disabled) + output[n] = input[n]; + } + } + + return 0; +}