lavfi: add atempo filter
authorPavel Koshevoy <pkoshevoy@gmail.com>
Sun, 17 Jun 2012 10:35:57 +0000 (12:35 +0200)
committerStefano Sabatini <stefasab@gmail.com>
Sun, 17 Jun 2012 10:50:35 +0000 (12:50 +0200)
Add atempo audio filter for adjusting audio tempo without affecting
pitch. This filter implements WSOLA algorithm with fast cross
correlation calculation in frequency domain.

Signed-off-by: Pavel Koshevoy <pavel@homestead.aragog.com>
Signed-off-by: Stefano Sabatini <stefasab@gmail.com>
Changelog
MAINTAINERS
configure
doc/filters.texi
libavfilter/Makefile
libavfilter/af_atempo.c [new file with mode: 0644]
libavfilter/allfilters.c
libavfilter/version.h

index facd0e9..729cbe5 100644 (file)
--- a/Changelog
+++ b/Changelog
@@ -7,6 +7,7 @@ version next:
 - Indeo Audio decoder
 - channelsplit audio filter
 - setnsamples audio filter
+- atempo filter
 
 
 version 0.11:
index ecee2c3..8d71dca 100644 (file)
@@ -275,6 +275,7 @@ Video filters:
   graphdump.c                           Nicolas George
   af_amerge.c                           Nicolas George
   af_astreamsync.c                      Nicolas George
+  af_atempo.c                           Pavel Koshevoy
   af_pan.c                              Nicolas George
   vsrc_mandelbrot.c                     Michael Niedermayer
   vf_yadif.c                            Michael Niedermayer
index ce5c188..82eafc6 100755 (executable)
--- a/configure
+++ b/configure
@@ -1702,6 +1702,7 @@ amovie_filter_deps="avcodec avformat"
 aresample_filter_deps="swresample"
 ass_filter_deps="libass"
 asyncts_filter_deps="avresample"
+atempo_filter_deps="avcodec"
 blackframe_filter_deps="gpl"
 boxblur_filter_deps="gpl"
 colormatrix_filter_deps="gpl"
index f2767cb..a354e97 100644 (file)
@@ -406,6 +406,24 @@ amovie=file.ogg [a] ; amovie=file.mp3 [b] ;
 [a2] [b2] amerge
 @end example
 
+@section atempo
+
+Adjust audio tempo.
+
+The filter accepts exactly one parameter, the audio tempo. If not
+specified then the filter will assume nominal 1.0 tempo. Tempo must
+be in the [0.5, 2.0] range.
+
+For example, to slow down audio to 80% tempo:
+@example
+atempo=0.8
+@end example
+
+For example, to speed up audio to 125% tempo:
+@example
+atempo=1.25
+@end example
+
 @section earwax
 
 Make audio easier to listen to on headphones.
index fee7641..72bb640 100644 (file)
@@ -9,6 +9,7 @@ FFLIBS-$(CONFIG_SCALE_FILTER)   += swscale
 FFLIBS-$(CONFIG_ACONVERT_FILTER)             += swresample
 FFLIBS-$(CONFIG_AMOVIE_FILTER)               += avformat avcodec
 FFLIBS-$(CONFIG_ARESAMPLE_FILTER)            += swresample
+FFLIBS-$(CONFIG_ATEMPO_FILTER)               += avcodec
 FFLIBS-$(CONFIG_MOVIE_FILTER)                += avformat avcodec
 FFLIBS-$(CONFIG_PAN_FILTER)                  += swresample
 FFLIBS-$(CONFIG_REMOVELOGO_FILTER)           += avformat avcodec
@@ -56,6 +57,7 @@ OBJS-$(CONFIG_ASHOWINFO_FILTER)              += af_ashowinfo.o
 OBJS-$(CONFIG_ASPLIT_FILTER)                 += split.o
 OBJS-$(CONFIG_ASTREAMSYNC_FILTER)            += af_astreamsync.o
 OBJS-$(CONFIG_ASYNCTS_FILTER)                += af_asyncts.o
+OBJS-$(CONFIG_ATEMPO_FILTER)                 += af_atempo.o
 OBJS-$(CONFIG_CHANNELSPLIT_FILTER)           += af_channelsplit.o
 OBJS-$(CONFIG_EARWAX_FILTER)                 += af_earwax.o
 OBJS-$(CONFIG_PAN_FILTER)                    += af_pan.o
diff --git a/libavfilter/af_atempo.c b/libavfilter/af_atempo.c
new file mode 100644 (file)
index 0000000..1ccc3f5
--- /dev/null
@@ -0,0 +1,1160 @@
+/*
+ * Copyright (c) 2012 Pavel Koshevoy <pkoshevoy at gmail dot com>
+ *
+ * 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
+ */
+
+/**
+ * @file
+ * tempo scaling audio filter -- an implementation of WSOLA algorithm
+ *
+ * Based on MIT licensed yaeAudioTempoFilter.h and yaeAudioFragment.h
+ * from Apprentice Video player by Pavel Koshevoy.
+ * https://sourceforge.net/projects/apprenticevideo/
+ *
+ * An explanation of SOLA algorithm is available at
+ * http://www.surina.net/article/time-and-pitch-scaling.html
+ *
+ * WSOLA is very similar to SOLA, only one major difference exists between
+ * these algorithms.  SOLA shifts audio fragments along the output stream,
+ * where as WSOLA shifts audio fragments along the input stream.
+ *
+ * The advantage of WSOLA algorithm is that the overlap region size is
+ * always the same, therefore the blending function is constant and
+ * can be precomputed.
+ */
+
+#include <float.h>
+#include "libavcodec/avfft.h"
+#include "libavutil/avassert.h"
+#include "libavutil/avstring.h"
+#include "libavutil/eval.h"
+#include "libavutil/opt.h"
+#include "libavutil/samplefmt.h"
+#include "avfilter.h"
+#include "audio.h"
+#include "internal.h"
+
+/**
+ * A fragment of audio waveform
+ */
+typedef struct {
+    // index of the first sample of this fragment in the overall waveform;
+    // 0: input sample position
+    // 1: output sample position
+    int64_t position[2];
+
+    // original packed multi-channel samples:
+    uint8_t *data;
+
+    // number of samples in this fragment:
+    int nsamples;
+
+    // rDFT transform of the down-mixed mono fragment, used for
+    // fast waveform alignment via correlation in frequency domain:
+    FFTSample *xdat;
+} AudioFragment;
+
+/**
+ * Filter state machine states
+ */
+typedef enum {
+    YAE_LOAD_FRAGMENT,
+    YAE_ADJUST_POSITION,
+    YAE_RELOAD_FRAGMENT,
+    YAE_OUTPUT_OVERLAP_ADD,
+    YAE_FLUSH_OUTPUT,
+} FilterState;
+
+/**
+ * Filter state machine
+ */
+typedef struct {
+    // ring-buffer of input samples, necessary because some times
+    // input fragment position may be adjusted backwards:
+    uint8_t *buffer;
+
+    // ring-buffer maximum capacity, expressed in sample rate time base:
+    int ring;
+
+    // ring-buffer house keeping:
+    int size;
+    int head;
+    int tail;
+
+    // 0: input sample position corresponding to the ring buffer tail
+    // 1: output sample position
+    int64_t position[2];
+
+    // sample format:
+    enum AVSampleFormat format;
+
+    // number of channels:
+    int channels;
+
+    // row of bytes to skip from one sample to next, across multple channels;
+    // stride = (number-of-channels * bits-per-sample-per-channel) / 8
+    int stride;
+
+    // fragment window size, power-of-two integer:
+    int window;
+
+    // Hann window coefficients, for feathering
+    // (blending) the overlapping fragment region:
+    float *hann;
+
+    // tempo scaling factor:
+    double tempo;
+
+    // cumulative alignment drift:
+    int drift;
+
+    // current/previous fragment ring-buffer:
+    AudioFragment frag[2];
+
+    // current fragment index:
+    uint64_t nfrag;
+
+    // current state:
+    FilterState state;
+
+    // for fast correlation calculation in frequency domain:
+    RDFTContext *real_to_complex;
+    RDFTContext *complex_to_real;
+    FFTSample *correlation;
+
+    // for managing AVFilterPad.request_frame and AVFilterPad.filter_samples
+    int request_fulfilled;
+    AVFilterBufferRef *dst_buffer;
+    uint8_t *dst;
+    uint8_t *dst_end;
+    uint64_t nsamples_in;
+    uint64_t nsamples_out;
+} ATempoContext;
+
+/**
+ * Reset filter to initial state, do not deallocate existing local buffers.
+ */
+static void yae_clear(ATempoContext *atempo)
+{
+    atempo->size = 0;
+    atempo->head = 0;
+    atempo->tail = 0;
+
+    atempo->drift = 0;
+    atempo->nfrag = 0;
+    atempo->state = YAE_LOAD_FRAGMENT;
+
+    atempo->position[0] = 0;
+    atempo->position[1] = 0;
+
+    atempo->frag[0].position[0] = 0;
+    atempo->frag[0].position[1] = 0;
+    atempo->frag[0].nsamples    = 0;
+
+    atempo->frag[1].position[0] = 0;
+    atempo->frag[1].position[1] = 0;
+    atempo->frag[1].nsamples    = 0;
+
+    // shift left position of 1st fragment by half a window
+    // so that no re-normalization would be required for
+    // the left half of the 1st fragment:
+    atempo->frag[0].position[0] = -(int64_t)(atempo->window / 2);
+    atempo->frag[0].position[1] = -(int64_t)(atempo->window / 2);
+
+    avfilter_unref_bufferp(&atempo->dst_buffer);
+    atempo->dst     = NULL;
+    atempo->dst_end = NULL;
+
+    atempo->request_fulfilled = 0;
+    atempo->nsamples_in       = 0;
+    atempo->nsamples_out      = 0;
+}
+
+/**
+ * Reset filter to initial state and deallocate all buffers.
+ */
+static void yae_release_buffers(ATempoContext *atempo)
+{
+    yae_clear(atempo);
+
+    av_freep(&atempo->frag[0].data);
+    av_freep(&atempo->frag[1].data);
+    av_freep(&atempo->frag[0].xdat);
+    av_freep(&atempo->frag[1].xdat);
+
+    av_freep(&atempo->buffer);
+    av_freep(&atempo->hann);
+    av_freep(&atempo->correlation);
+
+    av_rdft_end(atempo->real_to_complex);
+    atempo->real_to_complex = NULL;
+
+    av_rdft_end(atempo->complex_to_real);
+    atempo->complex_to_real = NULL;
+}
+
+#define REALLOC_OR_FAIL(field, field_size)                      \
+    do {                                                        \
+        void * new_field = av_realloc(field, (field_size));     \
+        if (!new_field) {                                       \
+            yae_release_buffers(atempo);                        \
+            return AVERROR(ENOMEM);                             \
+        }                                                       \
+        field = new_field;                                      \
+    } while (0)
+
+/**
+ * Prepare filter for processing audio data of given format,
+ * sample rate and number of channels.
+ */
+static int yae_reset(ATempoContext *atempo,
+                     enum AVSampleFormat format,
+                     int sample_rate,
+                     int channels)
+{
+    const int sample_size = av_get_bytes_per_sample(format);
+    uint32_t nlevels  = 0;
+    uint32_t pot;
+    int i;
+
+    atempo->format   = format;
+    atempo->channels = channels;
+    atempo->stride   = sample_size * channels;
+
+    // pick a segment window size:
+    atempo->window = sample_rate / 24;
+
+    // adjust window size to be a power-of-two integer:
+    nlevels = av_log2(atempo->window);
+    pot = 1 << nlevels;
+    av_assert0(pot <= atempo->window);
+
+    if (pot < atempo->window) {
+        atempo->window = pot * 2;
+        nlevels++;
+    }
+
+    // initialize audio fragment buffers:
+    REALLOC_OR_FAIL(atempo->frag[0].data, atempo->window * atempo->stride);
+    REALLOC_OR_FAIL(atempo->frag[1].data, atempo->window * atempo->stride);
+    REALLOC_OR_FAIL(atempo->frag[0].xdat, atempo->window * sizeof(FFTComplex));
+    REALLOC_OR_FAIL(atempo->frag[1].xdat, atempo->window * sizeof(FFTComplex));
+
+    // initialize rDFT contexts:
+    av_rdft_end(atempo->real_to_complex);
+    atempo->real_to_complex = NULL;
+
+    av_rdft_end(atempo->complex_to_real);
+    atempo->complex_to_real = NULL;
+
+    atempo->real_to_complex = av_rdft_init(nlevels + 1, DFT_R2C);
+    if (!atempo->real_to_complex) {
+        yae_release_buffers(atempo);
+        return AVERROR(ENOMEM);
+    }
+
+    atempo->complex_to_real = av_rdft_init(nlevels + 1, IDFT_C2R);
+    if (!atempo->complex_to_real) {
+        yae_release_buffers(atempo);
+        return AVERROR(ENOMEM);
+    }
+
+    REALLOC_OR_FAIL(atempo->correlation, atempo->window * sizeof(FFTComplex));
+
+    atempo->ring = atempo->window * 3;
+    REALLOC_OR_FAIL(atempo->buffer, atempo->ring * atempo->stride);
+
+    // initialize the Hann window function:
+    REALLOC_OR_FAIL(atempo->hann, atempo->window * sizeof(float));
+
+    for (i = 0; i < atempo->window; i++) {
+        double t = (double)i / (double)(atempo->window - 1);
+        double h = 0.5 * (1.0 - cos(2.0 * M_PI * t));
+        atempo->hann[i] = (float)h;
+    }
+
+    yae_clear(atempo);
+    return 0;
+}
+
+static int yae_set_tempo(AVFilterContext *ctx, const char *arg_tempo)
+{
+    ATempoContext *atempo = ctx->priv;
+    char   *tail = NULL;
+    double tempo = av_strtod(arg_tempo, &tail);
+
+    if (tail && *tail) {
+        av_log(ctx, AV_LOG_ERROR, "Invalid tempo value '%s'\n", arg_tempo);
+        return AVERROR(EINVAL);
+    }
+
+    if (tempo < 0.5 || tempo > 2.0) {
+        av_log(ctx, AV_LOG_ERROR, "Tempo value %f exceeds [0.5, 2.0] range\n",
+               tempo);
+        return AVERROR(EINVAL);
+    }
+
+    atempo->tempo = tempo;
+    return 0;
+}
+
+inline static AudioFragment *yae_curr_frag(ATempoContext *atempo)
+{
+    return &atempo->frag[atempo->nfrag % 2];
+}
+
+inline static AudioFragment *yae_prev_frag(ATempoContext *atempo)
+{
+    return &atempo->frag[(atempo->nfrag + 1) % 2];
+}
+
+/**
+ * A helper macro for initializing complex data buffer with scalar data
+ * of a given type.
+ */
+#define yae_init_xdat(scalar_type, scalar_max)                          \
+    do {                                                                \
+        const uint8_t *src_end = src +                                  \
+            frag->nsamples * atempo->channels * sizeof(scalar_type);    \
+                                                                        \
+        FFTSample *xdat = frag->xdat;                                   \
+        scalar_type tmp;                                                \
+                                                                        \
+        if (atempo->channels == 1) {                                    \
+            for (; src < src_end; xdat++) {                             \
+                tmp = *(const scalar_type *)src;                        \
+                src += sizeof(scalar_type);                             \
+                                                                        \
+                *xdat = (FFTSample)tmp;                                 \
+            }                                                           \
+        } else {                                                        \
+            FFTSample s, max, ti, si;                                   \
+            int i;                                                      \
+                                                                        \
+            for (; src < src_end; xdat++) {                             \
+                tmp = *(const scalar_type *)src;                        \
+                src += sizeof(scalar_type);                             \
+                                                                        \
+                max = (FFTSample)tmp;                                   \
+                s = FFMIN((FFTSample)scalar_max,                        \
+                          (FFTSample)fabsf(max));                       \
+                                                                        \
+                for (i = 1; i < atempo->channels; i++) {                \
+                    tmp = *(const scalar_type *)src;                    \
+                    src += sizeof(scalar_type);                         \
+                                                                        \
+                    ti = (FFTSample)tmp;                                \
+                    si = FFMIN((FFTSample)scalar_max,                   \
+                               (FFTSample)fabsf(ti));                   \
+                                                                        \
+                    if (s < si) {                                       \
+                        s   = si;                                       \
+                        max = ti;                                       \
+                    }                                                   \
+                }                                                       \
+                                                                        \
+                *xdat = max;                                            \
+            }                                                           \
+        }                                                               \
+    } while (0)
+
+/**
+ * Initialize complex data buffer of a given audio fragment
+ * with down-mixed mono data of appropriate scalar type.
+ */
+static void yae_downmix(ATempoContext *atempo, AudioFragment *frag)
+{
+    // shortcuts:
+    const uint8_t *src = frag->data;
+
+    // init complex data buffer used for FFT and Correlation:
+    memset(frag->xdat, 0, sizeof(FFTComplex) * atempo->window);
+
+    if (atempo->format == AV_SAMPLE_FMT_U8) {
+        yae_init_xdat(uint8_t, 127);
+    } else if (atempo->format == AV_SAMPLE_FMT_S16) {
+        yae_init_xdat(int16_t, 32767);
+    } else if (atempo->format == AV_SAMPLE_FMT_S32) {
+        yae_init_xdat(int, 2147483647);
+    } else if (atempo->format == AV_SAMPLE_FMT_FLT) {
+        yae_init_xdat(float, 1);
+    } else if (atempo->format == AV_SAMPLE_FMT_DBL) {
+        yae_init_xdat(double, 1);
+    }
+}
+
+/**
+ * Populate the internal data buffer on as-needed basis.
+ *
+ * @return
+ *   0 if requested data was already available or was successfully loaded,
+ *   AVERROR(EAGAIN) if more input data is required.
+ */
+static int yae_load_data(ATempoContext *atempo,
+                         const uint8_t **src_ref,
+                         const uint8_t *src_end,
+                         int64_t stop_here)
+{
+    // shortcut:
+    const uint8_t *src = *src_ref;
+    const int read_size = stop_here - atempo->position[0];
+
+    if (stop_here <= atempo->position[0]) {
+        return 0;
+    }
+
+    // samples are not expected to be skipped:
+    av_assert0(read_size <= atempo->ring);
+
+    while (atempo->position[0] < stop_here && src < src_end) {
+        int src_samples = (src_end - src) / atempo->stride;
+
+        // load data piece-wise, in order to avoid complicating the logic:
+        int nsamples = FFMIN(read_size, src_samples);
+        int na;
+        int nb;
+
+        nsamples = FFMIN(nsamples, atempo->ring);
+        na = FFMIN(nsamples, atempo->ring - atempo->tail);
+        nb = FFMIN(nsamples - na, atempo->ring);
+
+        if (na) {
+            uint8_t *a = atempo->buffer + atempo->tail * atempo->stride;
+            memcpy(a, src, na * atempo->stride);
+
+            src += na * atempo->stride;
+            atempo->position[0] += na;
+
+            atempo->size = FFMIN(atempo->size + na, atempo->ring);
+            atempo->tail = (atempo->tail + na) % atempo->ring;
+            atempo->head =
+                atempo->size < atempo->ring ?
+                atempo->tail - atempo->size :
+                atempo->tail;
+        }
+
+        if (nb) {
+            uint8_t *b = atempo->buffer;
+            memcpy(b, src, nb * atempo->stride);
+
+            src += nb * atempo->stride;
+            atempo->position[0] += nb;
+
+            atempo->size = FFMIN(atempo->size + nb, atempo->ring);
+            atempo->tail = (atempo->tail + nb) % atempo->ring;
+            atempo->head =
+                atempo->size < atempo->ring ?
+                atempo->tail - atempo->size :
+                atempo->tail;
+        }
+    }
+
+    // pass back the updated source buffer pointer:
+    *src_ref = src;
+
+    // sanity check:
+    av_assert0(atempo->position[0] <= stop_here);
+
+    return atempo->position[0] == stop_here ? 0 : AVERROR(EAGAIN);
+}
+
+/**
+ * Populate current audio fragment data buffer.
+ *
+ * @return
+ *   0 when the fragment is ready,
+ *   AVERROR(EAGAIN) if more input data is required.
+ */
+static int yae_load_frag(ATempoContext *atempo,
+                         const uint8_t **src_ref,
+                         const uint8_t *src_end)
+{
+    // shortcuts:
+    AudioFragment *frag = yae_curr_frag(atempo);
+    uint8_t *dst;
+    int64_t missing, start, zeros;
+    uint32_t nsamples;
+    const uint8_t *a, *b;
+    int i0, i1, n0, n1, na, nb;
+
+    int64_t stop_here = frag->position[0] + atempo->window;
+    if (src_ref && yae_load_data(atempo, src_ref, src_end, stop_here) != 0) {
+        return AVERROR(EAGAIN);
+    }
+
+    // calculate the number of samples we don't have:
+    missing =
+        stop_here > atempo->position[0] ?
+        stop_here - atempo->position[0] : 0;
+
+    nsamples =
+        missing < (int64_t)atempo->window ?
+        (uint32_t)(atempo->window - missing) : 0;
+
+    // setup the output buffer:
+    frag->nsamples = nsamples;
+    dst = frag->data;
+
+    start = atempo->position[0] - atempo->size;
+    zeros = 0;
+
+    if (frag->position[0] < start) {
+        // what we don't have we substitute with zeros:
+        zeros = FFMIN(start - frag->position[0], (int64_t)nsamples);
+        av_assert0(zeros != nsamples);
+
+        memset(dst, 0, zeros * atempo->stride);
+        dst += zeros * atempo->stride;
+    }
+
+    if (zeros == nsamples) {
+        return 0;
+    }
+
+    // get the remaining data from the ring buffer:
+    na = (atempo->head < atempo->tail ?
+          atempo->tail - atempo->head :
+          atempo->ring - atempo->head);
+
+    nb = atempo->head < atempo->tail ? 0 : atempo->tail;
+
+    // sanity check:
+    av_assert0(nsamples <= zeros + na + nb);
+
+    a = atempo->buffer + atempo->head * atempo->stride;
+    b = atempo->buffer;
+
+    i0 = frag->position[0] + zeros - start;
+    i1 = i0 < na ? 0 : i0 - na;
+
+    n0 = i0 < na ? FFMIN(na - i0, (int)(nsamples - zeros)) : 0;
+    n1 = nsamples - zeros - n0;
+
+    if (n0) {
+        memcpy(dst, a + i0 * atempo->stride, n0 * atempo->stride);
+        dst += n0 * atempo->stride;
+    }
+
+    if (n1) {
+        memcpy(dst, b + i1 * atempo->stride, n1 * atempo->stride);
+        dst += n1 * atempo->stride;
+    }
+
+    return 0;
+}
+
+/**
+ * Prepare for loading next audio fragment.
+ */
+static void yae_advance_to_next_frag(ATempoContext *atempo)
+{
+    const double fragment_step = atempo->tempo * (double)(atempo->window / 2);
+
+    const AudioFragment *prev;
+    AudioFragment       *frag;
+
+    atempo->nfrag++;
+    prev = yae_prev_frag(atempo);
+    frag = yae_curr_frag(atempo);
+
+    frag->position[0] = prev->position[0] + (int64_t)fragment_step;
+    frag->position[1] = prev->position[1] + atempo->window / 2;
+    frag->nsamples    = 0;
+}
+
+/**
+ * Calculate cross-correlation via rDFT.
+ *
+ * Multiply two vectors of complex numbers (result of real_to_complex rDFT)
+ * and transform back via complex_to_real rDFT.
+ */
+static void yae_xcorr_via_rdft(FFTSample *xcorr,
+                               RDFTContext *complex_to_real,
+                               const FFTComplex *xa,
+                               const FFTComplex *xb,
+                               const int window)
+{
+    FFTComplex *xc = (FFTComplex *)xcorr;
+    int i;
+
+    // NOTE: first element requires special care -- Given Y = rDFT(X),
+    // Im(Y[0]) and Im(Y[N/2]) are always zero, therefore av_rdft_calc
+    // stores Re(Y[N/2]) in place of Im(Y[0]).
+
+    xc->re = xa->re * xb->re;
+    xc->im = xa->im * xb->im;
+    xa++;
+    xb++;
+    xc++;
+
+    for (i = 1; i < window; i++, xa++, xb++, xc++) {
+        xc->re = (xa->re * xb->re + xa->im * xb->im);
+        xc->im = (xa->im * xb->re - xa->re * xb->im);
+    }
+
+    // apply inverse rDFT:
+    av_rdft_calc(complex_to_real, xcorr);
+}
+
+/**
+ * Calculate alignment offset for given fragment
+ * relative to the previous fragment.
+ *
+ * @return alignment offset of current fragment relative to previous.
+ */
+static int yae_align(AudioFragment *frag,
+                     const AudioFragment *prev,
+                     const int window,
+                     const int delta_max,
+                     const int drift,
+                     FFTSample *correlation,
+                     RDFTContext *complex_to_real)
+{
+    int       best_offset = -drift;
+    FFTSample best_metric = -FLT_MAX;
+    FFTSample *xcorr;
+
+    int i0;
+    int i1;
+    int i;
+
+    yae_xcorr_via_rdft(correlation,
+                       complex_to_real,
+                       (const FFTComplex *)prev->xdat,
+                       (const FFTComplex *)frag->xdat,
+                       window);
+
+    // identify search window boundaries:
+    i0 = FFMAX(window / 2 - delta_max - drift, 0);
+    i0 = FFMIN(i0, window);
+
+    i1 = FFMIN(window / 2 + delta_max - drift, window - window / 16);
+    i1 = FFMAX(i1, 0);
+
+    // identify cross-correlation peaks within search window:
+    xcorr = correlation + i0;
+
+    for (i = i0; i < i1; i++, xcorr++) {
+        FFTSample metric = *xcorr;
+
+        // normalize:
+        FFTSample drifti = (FFTSample)(drift + i);
+        metric *= drifti;
+
+        if (metric > best_metric) {
+            best_metric = metric;
+            best_offset = i - window / 2;
+        }
+    }
+
+    return best_offset;
+}
+
+/**
+ * Adjust current fragment position for better alignment
+ * with previous fragment.
+ *
+ * @return alignment correction.
+ */
+static int yae_adjust_position(ATempoContext *atempo)
+{
+    const AudioFragment *prev = yae_prev_frag(atempo);
+    AudioFragment       *frag = yae_curr_frag(atempo);
+
+    const int delta_max  = atempo->window / 2;
+    const int correction = yae_align(frag,
+                                     prev,
+                                     atempo->window,
+                                     delta_max,
+                                     atempo->drift,
+                                     atempo->correlation,
+                                     atempo->complex_to_real);
+
+    if (correction) {
+        // adjust fragment position:
+        frag->position[0] -= correction;
+
+        // clear so that the fragment can be reloaded:
+        frag->nsamples = 0;
+
+        // update cumulative correction drift counter:
+        atempo->drift += correction;
+    }
+
+    return correction;
+}
+
+/**
+ * A helper macro for blending the overlap region of previous
+ * and current audio fragment.
+ */
+#define yae_blend(scalar_type)                                          \
+    do {                                                                \
+        const scalar_type *aaa = (const scalar_type *)a;                \
+        const scalar_type *bbb = (const scalar_type *)b;                \
+                                                                        \
+        scalar_type *out     = (scalar_type *)dst;                      \
+        scalar_type *out_end = (scalar_type *)dst_end;                  \
+        int64_t i;                                                      \
+                                                                        \
+        for (i = 0; i < overlap && out < out_end;                       \
+             i++, atempo->position[1]++, wa++, wb++) {                  \
+            float w0 = *wa;                                             \
+            float w1 = *wb;                                             \
+            int j;                                                      \
+                                                                        \
+            for (j = 0; j < atempo->channels;                           \
+                 j++, aaa++, bbb++, out++) {                            \
+                float t0 = (float)*aaa;                                 \
+                float t1 = (float)*bbb;                                 \
+                                                                        \
+                *out =                                                  \
+                    frag->position[0] + i < 0 ?                         \
+                    *aaa :                                              \
+                    (scalar_type)(t0 * w0 + t1 * w1);                   \
+            }                                                           \
+        }                                                               \
+        dst = (uint8_t *)out;                                           \
+    } while (0)
+
+/**
+ * Blend the overlap region of previous and current audio fragment
+ * and output the results to the given destination buffer.
+ *
+ * @return
+ *   0 if the overlap region was completely stored in the dst buffer,
+ *   AVERROR(EAGAIN) if more destination buffer space is required.
+ */
+static int yae_overlap_add(ATempoContext *atempo,
+                           uint8_t **dst_ref,
+                           uint8_t *dst_end)
+{
+    // shortcuts:
+    const AudioFragment *prev = yae_prev_frag(atempo);
+    const AudioFragment *frag = yae_curr_frag(atempo);
+
+    const int64_t start_here = FFMAX(atempo->position[1],
+                                     frag->position[1]);
+
+    const int64_t stop_here = FFMIN(prev->position[1] + prev->nsamples,
+                                    frag->position[1] + frag->nsamples);
+
+    const int64_t overlap = stop_here - start_here;
+
+    const int64_t ia = start_here - prev->position[1];
+    const int64_t ib = start_here - frag->position[1];
+
+    const float *wa = atempo->hann + ia;
+    const float *wb = atempo->hann + ib;
+
+    const uint8_t *a = prev->data + ia * atempo->stride;
+    const uint8_t *b = frag->data + ib * atempo->stride;
+
+    uint8_t *dst = *dst_ref;
+
+    av_assert0(start_here <= stop_here &&
+               frag->position[1] <= start_here &&
+               overlap <= frag->nsamples);
+
+    if (atempo->format == AV_SAMPLE_FMT_U8) {
+        yae_blend(uint8_t);
+    } else if (atempo->format == AV_SAMPLE_FMT_S16) {
+        yae_blend(int16_t);
+    } else if (atempo->format == AV_SAMPLE_FMT_S32) {
+        yae_blend(int);
+    } else if (atempo->format == AV_SAMPLE_FMT_FLT) {
+        yae_blend(float);
+    } else if (atempo->format == AV_SAMPLE_FMT_DBL) {
+        yae_blend(double);
+    }
+
+    // pass-back the updated destination buffer pointer:
+    *dst_ref = dst;
+
+    return atempo->position[1] == stop_here ? 0 : AVERROR(EAGAIN);
+}
+
+/**
+ * Feed as much data to the filter as it is able to consume
+ * and receive as much processed data in the destination buffer
+ * as it is able to produce or store.
+ */
+static void
+yae_apply(ATempoContext *atempo,
+          const uint8_t **src_ref,
+          const uint8_t *src_end,
+          uint8_t **dst_ref,
+          uint8_t *dst_end)
+{
+    while (1) {
+        if (atempo->state == YAE_LOAD_FRAGMENT) {
+            // load additional data for the current fragment:
+            if (yae_load_frag(atempo, src_ref, src_end) != 0) {
+                break;
+            }
+
+            // down-mix to mono:
+            yae_downmix(atempo, yae_curr_frag(atempo));
+
+            // apply rDFT:
+            av_rdft_calc(atempo->real_to_complex, yae_curr_frag(atempo)->xdat);
+
+            // must load the second fragment before alignment can start:
+            if (!atempo->nfrag) {
+                yae_advance_to_next_frag(atempo);
+                continue;
+            }
+
+            atempo->state = YAE_ADJUST_POSITION;
+        }
+
+        if (atempo->state == YAE_ADJUST_POSITION) {
+            // adjust position for better alignment:
+            if (yae_adjust_position(atempo)) {
+                // reload the fragment at the corrected position, so that the
+                // Hann window blending would not require normalization:
+                atempo->state = YAE_RELOAD_FRAGMENT;
+            } else {
+                atempo->state = YAE_OUTPUT_OVERLAP_ADD;
+            }
+        }
+
+        if (atempo->state == YAE_RELOAD_FRAGMENT) {
+            // load additional data if necessary due to position adjustment:
+            if (yae_load_frag(atempo, src_ref, src_end) != 0) {
+                break;
+            }
+
+            // down-mix to mono:
+            yae_downmix(atempo, yae_curr_frag(atempo));
+
+            // apply rDFT:
+            av_rdft_calc(atempo->real_to_complex, yae_curr_frag(atempo)->xdat);
+
+            atempo->state = YAE_OUTPUT_OVERLAP_ADD;
+        }
+
+        if (atempo->state == YAE_OUTPUT_OVERLAP_ADD) {
+            // overlap-add and output the result:
+            if (yae_overlap_add(atempo, dst_ref, dst_end) != 0) {
+                break;
+            }
+
+            // advance to the next fragment, repeat:
+            yae_advance_to_next_frag(atempo);
+            atempo->state = YAE_LOAD_FRAGMENT;
+        }
+    }
+}
+
+/**
+ * Flush any buffered data from the filter.
+ *
+ * @return
+ *   0 if all data was completely stored in the dst buffer,
+ *   AVERROR(EAGAIN) if more destination buffer space is required.
+ */
+static int yae_flush(ATempoContext *atempo,
+                     uint8_t **dst_ref,
+                     uint8_t *dst_end)
+{
+    AudioFragment *frag = yae_curr_frag(atempo);
+    int64_t overlap_end;
+    int64_t start_here;
+    int64_t stop_here;
+    int64_t offset;
+
+    const uint8_t *src;
+    uint8_t *dst;
+
+    int src_size;
+    int dst_size;
+    int nbytes;
+
+    atempo->state = YAE_FLUSH_OUTPUT;
+
+    if (atempo->position[0] == frag->position[0] + frag->nsamples &&
+        atempo->position[1] == frag->position[1] + frag->nsamples) {
+        // the current fragment is already flushed:
+        return 0;
+    }
+
+    if (frag->position[0] + frag->nsamples < atempo->position[0]) {
+        // finish loading the current (possibly partial) fragment:
+        yae_load_frag(atempo, NULL, NULL);
+
+        if (atempo->nfrag) {
+            // down-mix to mono:
+            yae_downmix(atempo, frag);
+
+            // apply rDFT:
+            av_rdft_calc(atempo->real_to_complex, frag->xdat);
+
+            // align current fragment to previous fragment:
+            if (yae_adjust_position(atempo)) {
+                // reload the current fragment due to adjusted position:
+                yae_load_frag(atempo, NULL, NULL);
+            }
+        }
+    }
+
+    // flush the overlap region:
+    overlap_end = frag->position[1] + FFMIN(atempo->window / 2,
+                                            frag->nsamples);
+
+    while (atempo->position[1] < overlap_end) {
+        if (yae_overlap_add(atempo, dst_ref, dst_end) != 0) {
+            return AVERROR(EAGAIN);
+        }
+    }
+
+    // flush the remaininder of the current fragment:
+    start_here = FFMAX(atempo->position[1], overlap_end);
+    stop_here  = frag->position[1] + frag->nsamples;
+    offset     = start_here - frag->position[1];
+    av_assert0(start_here <= stop_here && frag->position[1] <= start_here);
+
+    src = frag->data + offset * atempo->stride;
+    dst = (uint8_t *)*dst_ref;
+
+    src_size = (int)(stop_here - start_here) * atempo->stride;
+    dst_size = dst_end - dst;
+    nbytes = FFMIN(src_size, dst_size);
+
+    memcpy(dst, src, nbytes);
+    dst += nbytes;
+
+    atempo->position[1] += (nbytes / atempo->stride);
+
+    // pass-back the updated destination buffer pointer:
+    *dst_ref = (uint8_t *)dst;
+
+    return atempo->position[1] == stop_here ? 0 : AVERROR(EAGAIN);
+}
+
+static av_cold int init(AVFilterContext *ctx, const char *args, void *opaque)
+{
+    ATempoContext *atempo = ctx->priv;
+
+    // NOTE: this assumes that the caller has memset ctx->priv to 0:
+    atempo->format = AV_SAMPLE_FMT_NONE;
+    atempo->tempo  = 1.0;
+    atempo->state  = YAE_LOAD_FRAGMENT;
+
+    return args ? yae_set_tempo(ctx, args) : 0;
+}
+
+static av_cold void uninit(AVFilterContext *ctx)
+{
+    ATempoContext *atempo = ctx->priv;
+    yae_release_buffers(atempo);
+}
+
+static int query_formats(AVFilterContext *ctx)
+{
+    AVFilterChannelLayouts *layouts = NULL;
+    AVFilterFormats        *formats = NULL;
+
+    // WSOLA necessitates an internal sliding window ring buffer
+    // for incoming audio stream.
+    //
+    // Planar sample formats are too cumbersome to store in a ring buffer,
+    // therefore planar sample formats are not supported.
+    //
+    enum AVSampleFormat sample_fmts[] = {
+        AV_SAMPLE_FMT_U8,
+        AV_SAMPLE_FMT_S16,
+        AV_SAMPLE_FMT_S32,
+        AV_SAMPLE_FMT_FLT,
+        AV_SAMPLE_FMT_DBL,
+        AV_SAMPLE_FMT_NONE
+    };
+
+    layouts = ff_all_channel_layouts();
+    if (!layouts) {
+        return AVERROR(ENOMEM);
+    }
+    ff_set_common_channel_layouts(ctx, layouts);
+
+    formats = ff_make_format_list(sample_fmts);
+    if (!formats) {
+        return AVERROR(ENOMEM);
+    }
+    ff_set_common_formats(ctx, formats);
+
+    formats = ff_all_samplerates();
+    if (!formats) {
+        return AVERROR(ENOMEM);
+    }
+    ff_set_common_samplerates(ctx, formats);
+
+    return 0;
+}
+
+static int config_props(AVFilterLink *inlink)
+{
+    AVFilterContext  *ctx = inlink->dst;
+    ATempoContext *atempo = ctx->priv;
+
+    enum AVSampleFormat format = inlink->format;
+    int sample_rate = (int)inlink->sample_rate;
+    int channels = av_get_channel_layout_nb_channels(inlink->channel_layout);
+
+    return yae_reset(atempo, format, sample_rate, channels);
+}
+
+static void push_samples(ATempoContext *atempo,
+                         AVFilterLink *outlink,
+                         int n_out)
+{
+    atempo->dst_buffer->audio->sample_rate = outlink->sample_rate;
+    atempo->dst_buffer->audio->nb_samples  = n_out;
+
+    // adjust the PTS:
+    atempo->dst_buffer->pts =
+        av_rescale_q(atempo->nsamples_out,
+                     (AVRational){ 1, outlink->sample_rate },
+                     outlink->time_base);
+
+    ff_filter_samples(outlink, atempo->dst_buffer);
+    atempo->dst_buffer = NULL;
+    atempo->dst        = NULL;
+    atempo->dst_end    = NULL;
+
+    atempo->nsamples_out += n_out;
+}
+
+static void filter_samples(AVFilterLink *inlink,
+                           AVFilterBufferRef *src_buffer)
+{
+    AVFilterContext  *ctx = inlink->dst;
+    ATempoContext *atempo = ctx->priv;
+    AVFilterLink *outlink = ctx->outputs[0];
+
+    int n_in = src_buffer->audio->nb_samples;
+    int n_out = (int)(0.5 + ((double)n_in) / atempo->tempo);
+
+    const uint8_t *src = src_buffer->data[0];
+    const uint8_t *src_end = src + n_in * atempo->stride;
+
+    while (src < src_end) {
+        if (!atempo->dst_buffer) {
+            atempo->dst_buffer = ff_get_audio_buffer(outlink,
+                                                     AV_PERM_WRITE,
+                                                     n_out);
+            avfilter_copy_buffer_ref_props(atempo->dst_buffer, src_buffer);
+
+            atempo->dst = atempo->dst_buffer->data[0];
+            atempo->dst_end = atempo->dst + n_out * atempo->stride;
+        }
+
+        yae_apply(atempo, &src, src_end, &atempo->dst, atempo->dst_end);
+
+        if (atempo->dst == atempo->dst_end) {
+            push_samples(atempo, outlink, n_out);
+            atempo->request_fulfilled = 1;
+        }
+    }
+
+    atempo->nsamples_in += n_in;
+    avfilter_unref_bufferp(&src_buffer);
+}
+
+static int request_frame(AVFilterLink *outlink)
+{
+    AVFilterContext  *ctx = outlink->src;
+    ATempoContext *atempo = ctx->priv;
+    int ret;
+
+    atempo->request_fulfilled = 0;
+    do {
+        ret = avfilter_request_frame(ctx->inputs[0]);
+    }
+    while (!atempo->request_fulfilled && ret >= 0);
+
+    if (ret == AVERROR_EOF) {
+        // flush the filter:
+        int n_max = atempo->ring;
+        int n_out;
+        int err = AVERROR(EAGAIN);
+
+        while (err == AVERROR(EAGAIN)) {
+            if (!atempo->dst_buffer) {
+                atempo->dst_buffer = ff_get_audio_buffer(outlink,
+                                                         AV_PERM_WRITE,
+                                                         n_max);
+
+                atempo->dst = atempo->dst_buffer->data[0];
+                atempo->dst_end = atempo->dst + n_max * atempo->stride;
+            }
+
+            err = yae_flush(atempo, &atempo->dst, atempo->dst_end);
+
+            n_out = ((atempo->dst - atempo->dst_buffer->data[0]) /
+                     atempo->stride);
+
+            if (n_out) {
+                push_samples(atempo, outlink, n_out);
+            }
+        }
+
+        avfilter_unref_bufferp(&atempo->dst_buffer);
+        atempo->dst     = NULL;
+        atempo->dst_end = NULL;
+
+        return AVERROR_EOF;
+    }
+
+    return ret;
+}
+
+static int process_command(AVFilterContext *ctx,
+                           const char *cmd,
+                           const char *arg,
+                           char *res,
+                           int res_len,
+                           int flags)
+{
+    return !strcmp(cmd, "tempo") ? yae_set_tempo(ctx, arg) : AVERROR(ENOSYS);
+}
+
+AVFilter avfilter_af_atempo = {
+    .name            = "atempo",
+    .description     = NULL_IF_CONFIG_SMALL("Adjust audio tempo."),
+    .init            = init,
+    .uninit          = uninit,
+    .query_formats   = query_formats,
+    .process_command = process_command,
+    .priv_size       = sizeof(ATempoContext),
+
+    .inputs    = (const AVFilterPad[]) {
+        { .name            = "default",
+          .type            = AVMEDIA_TYPE_AUDIO,
+          .filter_samples  = filter_samples,
+          .config_props    = config_props,
+          .min_perms       = AV_PERM_READ, },
+        { .name = NULL}
+    },
+
+    .outputs   = (const AVFilterPad[]) {
+        { .name            = "default",
+          .request_frame   = request_frame,
+          .type            = AVMEDIA_TYPE_AUDIO, },
+        { .name = NULL}
+    },
+};
index 8257ee5..35065d5 100644 (file)
@@ -45,6 +45,7 @@ void avfilter_register_all(void)
     REGISTER_FILTER (ASPLIT,      asplit,      af);
     REGISTER_FILTER (ASTREAMSYNC, astreamsync, af);
     REGISTER_FILTER (ASYNCTS,     asyncts,     af);
+    REGISTER_FILTER (ATEMPO,      atempo,      af);
     REGISTER_FILTER (CHANNELSPLIT,channelsplit,af);
     REGISTER_FILTER (EARWAX,      earwax,      af);
     REGISTER_FILTER (PAN,         pan,         af);
index 861fe99..8b3265f 100644 (file)
@@ -29,7 +29,7 @@
 #include "libavutil/avutil.h"
 
 #define LIBAVFILTER_VERSION_MAJOR  2
-#define LIBAVFILTER_VERSION_MINOR 80
+#define LIBAVFILTER_VERSION_MINOR 81
 #define LIBAVFILTER_VERSION_MICRO 100
 
 #define LIBAVFILTER_VERSION_INT AV_VERSION_INT(LIBAVFILTER_VERSION_MAJOR, \