diff mbox

[FFmpeg-devel] avfilter/af_firequalizer: add fft2 option

Message ID 1478135713-16489-1-git-send-email-mfcc64@gmail.com
State Accepted
Commit b4e9252ae3d1094b3573299001684ae35d5a5d4f
Headers show

Commit Message

Muhammad Faiz Nov. 3, 2016, 1:15 a.m. UTC
2-channels convolution using complex fft
improves speed significantly

not sure if it should be enabled by default
so disable it by default

Signed-off-by: Muhammad Faiz <mfcc64@gmail.com>
---
 doc/filters.texi                |  4 +++
 libavfilter/af_firequalizer.c   | 80 ++++++++++++++++++++++++++++++++++++++++-
 tests/filtergraphs/firequalizer |  8 +++--
 3 files changed, 88 insertions(+), 4 deletions(-)

Comments

Muhammad Faiz Nov. 4, 2016, 2:54 a.m. UTC | #1
On 11/3/16, Muhammad Faiz <mfcc64@gmail.com> wrote:
> 2-channels convolution using complex fft
> improves speed significantly
>
> not sure if it should be enabled by default
> so disable it by default
>
> Signed-off-by: Muhammad Faiz <mfcc64@gmail.com>
> ---
>  doc/filters.texi                |  4 +++
>  libavfilter/af_firequalizer.c   | 80
> ++++++++++++++++++++++++++++++++++++++++-
>  tests/filtergraphs/firequalizer |  8 +++--
>  3 files changed, 88 insertions(+), 4 deletions(-)

applied

thanks
diff mbox

Patch

diff --git a/doc/filters.texi b/doc/filters.texi
index 867b932..5bced37 100644
--- a/doc/filters.texi
+++ b/doc/filters.texi
@@ -2585,6 +2585,10 @@  Set file for dumping, suitable for gnuplot.
 @item dumpscale
 Set scale for dumpfile. Acceptable values are same with scale option.
 Default is linlog.
+
+@item fft2
+Enable 2-channel convolution using complex FFT. This improves speed significantly.
+Default is disabled.
 @end table
 
 @subsection Examples
diff --git a/libavfilter/af_firequalizer.c b/libavfilter/af_firequalizer.c
index 4988717..5c6fd54 100644
--- a/libavfilter/af_firequalizer.c
+++ b/libavfilter/af_firequalizer.c
@@ -69,6 +69,7 @@  typedef struct {
     RDFTContext   *analysis_irdft;
     RDFTContext   *rdft;
     RDFTContext   *irdft;
+    FFTContext    *fft_ctx;
     int           analysis_rdft_len;
     int           rdft_len;
 
@@ -97,6 +98,7 @@  typedef struct {
     int           scale;
     char          *dumpfile;
     int           dumpscale;
+    int           fft2;
 
     int           nb_gain_entry;
     int           gain_entry_err;
@@ -132,6 +134,7 @@  static const AVOption firequalizer_options[] = {
         { "loglog", "logarithmic-freq logarithmic-gain", 0, AV_OPT_TYPE_CONST, { .i64 = SCALE_LOGLOG }, 0, 0, FLAGS, "scale" },
     { "dumpfile", "set dump file", OFFSET(dumpfile), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
     { "dumpscale", "set dump scale", OFFSET(dumpscale), AV_OPT_TYPE_INT, { .i64 = SCALE_LINLOG }, 0, NB_SCALE-1, FLAGS, "scale" },
+    { "fft2", "set 2-channels fft", OFFSET(fft2), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, FLAGS },
     { NULL }
 };
 
@@ -143,7 +146,9 @@  static void common_uninit(FIREqualizerContext *s)
     av_rdft_end(s->analysis_irdft);
     av_rdft_end(s->rdft);
     av_rdft_end(s->irdft);
+    av_fft_end(s->fft_ctx);
     s->analysis_rdft = s->analysis_irdft = s->rdft = s->irdft = NULL;
+    s->fft_ctx = NULL;
 
     av_freep(&s->analysis_buf);
     av_freep(&s->dump_buf);
@@ -230,6 +235,70 @@  static void fast_convolute(FIREqualizerContext *s, const float *kernel_buf, floa
     }
 }
 
+static void fast_convolute2(FIREqualizerContext *s, const float *kernel_buf, FFTComplex *conv_buf,
+                            OverlapIndex *idx, float *data0, float *data1, int nsamples)
+{
+    if (nsamples <= s->nsamples_max) {
+        FFTComplex *buf = conv_buf + idx->buf_idx * s->rdft_len;
+        FFTComplex *obuf = conv_buf + !idx->buf_idx * s->rdft_len + idx->overlap_idx;
+        int center = s->fir_len/2;
+        int k;
+        float tmp;
+
+        memset(buf, 0, center * sizeof(*buf));
+        for (k = 0; k < nsamples; k++) {
+            buf[center+k].re = data0[k];
+            buf[center+k].im = data1[k];
+        }
+        memset(buf + center + nsamples, 0, (s->rdft_len - nsamples - center) * sizeof(*buf));
+        av_fft_permute(s->fft_ctx, buf);
+        av_fft_calc(s->fft_ctx, buf);
+
+        /* swap re <-> im, do backward fft using forward fft_ctx */
+        /* normalize with 0.5f */
+        tmp = buf[0].re;
+        buf[0].re = 0.5f * kernel_buf[0] * buf[0].im;
+        buf[0].im = 0.5f * kernel_buf[0] * tmp;
+        for (k = 1; k < s->rdft_len/2; k++) {
+            int m = s->rdft_len - k;
+            tmp = buf[k].re;
+            buf[k].re = 0.5f * kernel_buf[k] * buf[k].im;
+            buf[k].im = 0.5f * kernel_buf[k] * tmp;
+            tmp = buf[m].re;
+            buf[m].re = 0.5f * kernel_buf[k] * buf[m].im;
+            buf[m].im = 0.5f * kernel_buf[k] * tmp;
+        }
+        tmp = buf[k].re;
+        buf[k].re = 0.5f * kernel_buf[k] * buf[k].im;
+        buf[k].im = 0.5f * kernel_buf[k] * tmp;
+
+        av_fft_permute(s->fft_ctx, buf);
+        av_fft_calc(s->fft_ctx, buf);
+
+        for (k = 0; k < s->rdft_len - idx->overlap_idx; k++) {
+            buf[k].re += obuf[k].re;
+            buf[k].im += obuf[k].im;
+        }
+
+        /* swapped re <-> im */
+        for (k = 0; k < nsamples; k++) {
+            data0[k] = buf[k].im;
+            data1[k] = buf[k].re;
+        }
+        idx->buf_idx = !idx->buf_idx;
+        idx->overlap_idx = nsamples;
+    } else {
+        while (nsamples > s->nsamples_max * 2) {
+            fast_convolute2(s, kernel_buf, conv_buf, idx, data0, data1, s->nsamples_max);
+            data0 += s->nsamples_max;
+            data1 += s->nsamples_max;
+            nsamples -= s->nsamples_max;
+        }
+        fast_convolute2(s, kernel_buf, conv_buf, idx, data0, data1, nsamples/2);
+        fast_convolute2(s, kernel_buf, conv_buf, idx, data0 + nsamples/2, data1 + nsamples/2, nsamples - nsamples/2);
+    }
+}
+
 static void dump_fir(AVFilterContext *ctx, FILE *fp, int ch)
 {
     FIREqualizerContext *s = ctx->priv;
@@ -598,6 +667,9 @@  static int config_input(AVFilterLink *inlink)
     if (!(s->rdft = av_rdft_init(rdft_bits, DFT_R2C)) || !(s->irdft = av_rdft_init(rdft_bits, IDFT_C2R)))
         return AVERROR(ENOMEM);
 
+    if (s->fft2 && !s->multi && inlink->channels > 1 && !(s->fft_ctx = av_fft_init(rdft_bits, 0)))
+        return AVERROR(ENOMEM);
+
     for ( ; rdft_bits <= RDFT_BITS_MAX; rdft_bits++) {
         s->analysis_rdft_len = 1 << rdft_bits;
         if (inlink->sample_rate <= s->accuracy * s->analysis_rdft_len)
@@ -640,7 +712,13 @@  static int filter_frame(AVFilterLink *inlink, AVFrame *frame)
     FIREqualizerContext *s = ctx->priv;
     int ch;
 
-    for (ch = 0; ch < inlink->channels; ch++) {
+    for (ch = 0; ch + 1 < inlink->channels && s->fft_ctx; ch += 2) {
+        fast_convolute2(s, s->kernel_buf, (FFTComplex *)(s->conv_buf + 2 * ch * s->rdft_len),
+                        s->conv_idx + ch, (float *) frame->extended_data[ch],
+                        (float *) frame->extended_data[ch+1], frame->nb_samples);
+    }
+
+    for ( ; ch < inlink->channels; ch++) {
         fast_convolute(s, s->kernel_buf + (s->multi ? ch * s->rdft_len : 0),
                        s->conv_buf + 2 * ch * s->rdft_len, s->conv_idx + ch,
                        (float *) frame->extended_data[ch], frame->nb_samples);
diff --git a/tests/filtergraphs/firequalizer b/tests/filtergraphs/firequalizer
index ee432e2..777dfdd 100644
--- a/tests/filtergraphs/firequalizer
+++ b/tests/filtergraphs/firequalizer
@@ -1,4 +1,5 @@ 
 firequalizer    =
+    fft2        = on:
     gain        = 'sin(0.001*f) - 1':
     delay       = 0.05,
 
@@ -11,14 +12,15 @@  firequalizer    =
     wfunc       = nuttall,
 
 firequalizer    =
+    fft2        = on:
+    gain_entry  = 'entry(1000, 0); entry(5000, 0.1); entry(10000, 0.2)',
+
+firequalizer    =
     gain        = 'if (ch, -0.3 * sin(0.001*f), -0.8 * sin(0.001*f)) - 1':
     delay       = 0.05:
     multi       = on,
 
 firequalizer    =
-    gain_entry  = 'entry(1000, 0); entry(5000, 0.1); entry(10000, 0.2)',
-
-firequalizer    =
     gain_entry  = 'entry(1000, 0.2); entry(5000, 0.1); entry(10000, 0)',
 
 volume          = 2.8dB,