[pulseaudio-discuss] [PATCH] resampler: Resample first followed by remapping if have more out channels than in channels
Peter Meerwald
pmeerw at pmeerw.net
Thu Feb 7 11:14:06 PST 2013
From: Peter Meerwald <p.meerwald at bct-electronic.com>
The patch intends to reduce computational load when resampling AND remapping. The PA
resampler performs the following steps:
sample format conversion -> remapping -> resampling -> sample format conversion
In case the number of output channels is higher than the number of input channels, the
resampler has to be run more often than necessary. E.g. in case of mono to 4-channel remapping,
the resampler runs on 4 channels separately.
To ímprove this, the PA resampler pipeline is made adaptive:
if out-channels <= in-channels:
sample format conversion -> remapping -> resampling -> sample format conversion
if out-channels > in-channels:
sample format conversion -> resampling -> remapping -> sample format conversion
Signed-off-by: Peter Meerwald <p.meerwald at bct-electronic.com>
---
src/pulsecore/resampler.c | 44 +++++++++++++++++++++++++++++++-------------
1 file changed, 31 insertions(+), 13 deletions(-)
diff --git a/src/pulsecore/resampler.c b/src/pulsecore/resampler.c
index 649294d..50aaf4f 100644
--- a/src/pulsecore/resampler.c
+++ b/src/pulsecore/resampler.c
@@ -67,6 +67,7 @@ struct pa_resampler {
bool remap_buf_contains_leftover_data;
pa_sample_format_t work_format;
+ uint8_t work_channels;
pa_convert_func_t to_work_format_func;
pa_convert_func_t from_work_format_func;
@@ -328,6 +329,16 @@ pa_resampler* pa_resampler_new(
}
}
+ if (r->o_ss.channels <= r->i_ss.channels)
+ r->work_channels = r->o_ss.channels;
+ else
+ r->work_channels = r->i_ss.channels;
+
+ pa_log_debug("Resampler:\n rate %d -> %d (method %s),\n format %s -> %s (intermediate %s),\n channels %d %d (resampling %d)",
+ a->rate, b->rate, pa_resample_method_to_string(r->method),
+ pa_sample_format_to_string(a->format), pa_sample_format_to_string(b->format), pa_sample_format_to_string(r->work_format),
+ a->channels, b->channels, r->work_channels);
+
/* initialize implementation */
if (init_table[method](r) < 0)
goto fail;
@@ -1140,10 +1151,10 @@ static pa_memchunk *resample(pa_resampler *r, pa_memchunk *input) {
return input;
in_n_samples = (unsigned) (input->length / r->w_sz);
- in_n_frames = (unsigned) (in_n_samples / r->o_ss.channels);
+ in_n_frames = (unsigned) (in_n_samples / r->work_channels);
out_n_frames = ((in_n_frames*r->o_ss.rate)/r->i_ss.rate)+EXTRA_FRAMES;
- out_n_samples = out_n_frames * r->o_ss.channels;
+ out_n_samples = out_n_frames * r->work_channels;
r->resample_buf.index = 0;
r->resample_buf.length = r->w_sz * out_n_samples;
@@ -1157,7 +1168,7 @@ static pa_memchunk *resample(pa_resampler *r, pa_memchunk *input) {
}
r->impl_resample(r, input, in_n_frames, &r->resample_buf, &out_n_frames);
- r->resample_buf.length = out_n_frames * r->w_sz * r->o_ss.channels;
+ r->resample_buf.length = out_n_frames * r->w_sz * r->work_channels;
return &r->resample_buf;
}
@@ -1210,8 +1221,15 @@ void pa_resampler_run(pa_resampler *r, const pa_memchunk *in, pa_memchunk *out)
buf = (pa_memchunk*) in;
buf = convert_to_work_format(r, buf);
- buf = remap_channels(r, buf);
- buf = resample(r, buf);
+ /* Try to save resampling effort: if we have more output channels than
+ * input channels, do resampling first, then remapping. */
+ if (r->o_ss.channels <= r->i_ss.channels) {
+ buf = remap_channels(r, buf);
+ buf = resample(r, buf);
+ } else {
+ buf = resample(r, buf);
+ buf = remap_channels(r, buf);
+ }
if (buf->length) {
buf = convert_from_work_format(r, buf);
@@ -1276,8 +1294,8 @@ static void libsamplerate_resample(pa_resampler *r, const pa_memchunk *input, un
pa_assert_se(src_process(r->src.state, &data) == 0);
if (data.input_frames_used < in_n_frames) {
- void *leftover_data = data.data_in + data.input_frames_used * r->o_ss.channels;
- size_t leftover_length = (in_n_frames - data.input_frames_used) * sizeof(float) * r->o_ss.channels;
+ void *leftover_data = data.data_in + data.input_frames_used * r->work_channels;
+ size_t leftover_length = (in_n_frames - data.input_frames_used) * sizeof(float) * r->work_channels;
save_leftover(r, leftover_data, leftover_length);
}
@@ -1413,7 +1431,7 @@ static int speex_init(pa_resampler *r) {
pa_log_info("Choosing speex quality setting %i.", q);
- if (!(r->speex.state = speex_resampler_init(r->o_ss.channels, r->i_ss.rate, r->o_ss.rate, q, &err)))
+ if (!(r->speex.state = speex_resampler_init(r->work_channels, r->i_ss.rate, r->o_ss.rate, q, &err)))
return -1;
return 0;
@@ -1432,7 +1450,7 @@ static void trivial_resample(pa_resampler *r, const pa_memchunk *input, unsigned
pa_assert(output);
pa_assert(out_n_frames);
- fz = r->w_sz * r->o_ss.channels;
+ fz = r->w_sz * r->work_channels;
src = pa_memblock_acquire_chunk(input);
dst = pa_memblock_acquire_chunk(output);
@@ -1616,7 +1634,7 @@ static void ffmpeg_resample(pa_resampler *r, const pa_memchunk *input, unsigned
pa_assert(output);
pa_assert(out_n_frames);
- for (c = 0; c < r->o_ss.channels; c++) {
+ for (c = 0; c < r->work_channels; c++) {
unsigned u;
pa_memblock *b, *w;
int16_t *p, *t, *k, *q, *s;
@@ -1631,7 +1649,7 @@ static void ffmpeg_resample(pa_resampler *r, const pa_memchunk *input, unsigned
k = p;
for (u = 0; u < in_n_frames; u++) {
*k = *t;
- t += r->o_ss.channels;
+ t += r->work_channels;
k ++;
}
pa_memblock_release(input->memblock);
@@ -1645,7 +1663,7 @@ static void ffmpeg_resample(pa_resampler *r, const pa_memchunk *input, unsigned
q, p,
&consumed_frames,
(int) in_n_frames, (int) *out_n_frames,
- c >= (unsigned) (r->o_ss.channels-1));
+ c >= (unsigned) (r->work_channels-1));
pa_memblock_release(b);
pa_memblock_unref(b);
@@ -1659,7 +1677,7 @@ static void ffmpeg_resample(pa_resampler *r, const pa_memchunk *input, unsigned
for (u = 0; u < used_frames; u++) {
*s = *q;
q++;
- s += r->o_ss.channels;
+ s += r->work_channels;
}
pa_memblock_release(output->memblock);
pa_memblock_release(w);
--
1.8.1.1
More information about the pulseaudio-discuss
mailing list