| Message ID | 20170316054310.22561-1-mfcc64@gmail.com |
|---|---|
| State | Accepted |
| Commit | c52638cca255737eb060dcdedf5be4414e622e82 |
| Headers | show |
On Thu, 16 Mar 2017 12:43:10 +0700 Muhammad Faiz <mfcc64@gmail.com> wrote: > use fltp when doing s32 -> s32 resampling > because s32p has no simd optimization > > benchmark: > old 17.913s > new 7.584s (use fma3) > > Signed-off-by: Muhammad Faiz <mfcc64@gmail.com> > --- > libswresample/swresample.c | 2 ++ > 1 file changed, 2 insertions(+) > > diff --git a/libswresample/swresample.c b/libswresample/swresample.c > index f2e6600..74c96dc 100644 > --- a/libswresample/swresample.c > +++ b/libswresample/swresample.c > @@ -223,6 +223,8 @@ av_cold int swr_init(struct SwrContext *s){ > }else if( av_get_planar_sample_fmt(s-> in_sample_fmt) == AV_SAMPLE_FMT_S32P > && av_get_planar_sample_fmt(s->out_sample_fmt) == AV_SAMPLE_FMT_S32P > && !s->rematrix > + && s->out_sample_rate == s->in_sample_rate > + && !(s->flags & SWR_FLAG_RESAMPLE) > && s->engine != SWR_ENGINE_SOXR){ > s->int_sample_fmt= AV_SAMPLE_FMT_S32P; > }else if(av_get_bytes_per_sample(s->in_sample_fmt) <= 4){ Wouldn't going over float lose bits? I guess most times, s32 is actually 24 bit padded (since we have no 24 bit format). But it could be true 32 bit as well (with all bits used). Or does it use double, or am I missing something else?
On Thu, Mar 16, 2017 at 1:01 PM, wm4 <nfxjfg@googlemail.com> wrote: > On Thu, 16 Mar 2017 12:43:10 +0700 > Muhammad Faiz <mfcc64@gmail.com> wrote: > >> use fltp when doing s32 -> s32 resampling >> because s32p has no simd optimization >> >> benchmark: >> old 17.913s >> new 7.584s (use fma3) >> >> Signed-off-by: Muhammad Faiz <mfcc64@gmail.com> >> --- >> libswresample/swresample.c | 2 ++ >> 1 file changed, 2 insertions(+) >> >> diff --git a/libswresample/swresample.c b/libswresample/swresample.c >> index f2e6600..74c96dc 100644 >> --- a/libswresample/swresample.c >> +++ b/libswresample/swresample.c >> @@ -223,6 +223,8 @@ av_cold int swr_init(struct SwrContext *s){ >> }else if( av_get_planar_sample_fmt(s-> in_sample_fmt) == AV_SAMPLE_FMT_S32P >> && av_get_planar_sample_fmt(s->out_sample_fmt) == AV_SAMPLE_FMT_S32P >> && !s->rematrix >> + && s->out_sample_rate == s->in_sample_rate >> + && !(s->flags & SWR_FLAG_RESAMPLE) >> && s->engine != SWR_ENGINE_SOXR){ >> s->int_sample_fmt= AV_SAMPLE_FMT_S32P; >> }else if(av_get_bytes_per_sample(s->in_sample_fmt) <= 4){ > > Wouldn't going over float lose bits? Yes, of course. But when resampling with default opt, aliasing happens at about -90dB which is only 15 bits. Keeping 32 bits has no point at all. > I guess most times, s32 is > actually 24 bit padded (since we have no 24 bit format). But it could > be true 32 bit as well (with all bits used). Or does it use double, or > am I missing something else? Using double is overkill, i think.
On Thu, 16 Mar 2017 17:00:17 +0700 Muhammad Faiz <mfcc64@gmail.com> wrote: > On Thu, Mar 16, 2017 at 1:01 PM, wm4 <nfxjfg@googlemail.com> wrote: > > On Thu, 16 Mar 2017 12:43:10 +0700 > > Muhammad Faiz <mfcc64@gmail.com> wrote: > > > >> use fltp when doing s32 -> s32 resampling > >> because s32p has no simd optimization > >> > >> benchmark: > >> old 17.913s > >> new 7.584s (use fma3) > >> > >> Signed-off-by: Muhammad Faiz <mfcc64@gmail.com> > >> --- > >> libswresample/swresample.c | 2 ++ > >> 1 file changed, 2 insertions(+) > >> > >> diff --git a/libswresample/swresample.c b/libswresample/swresample.c > >> index f2e6600..74c96dc 100644 > >> --- a/libswresample/swresample.c > >> +++ b/libswresample/swresample.c > >> @@ -223,6 +223,8 @@ av_cold int swr_init(struct SwrContext *s){ > >> }else if( av_get_planar_sample_fmt(s-> in_sample_fmt) == AV_SAMPLE_FMT_S32P > >> && av_get_planar_sample_fmt(s->out_sample_fmt) == AV_SAMPLE_FMT_S32P > >> && !s->rematrix > >> + && s->out_sample_rate == s->in_sample_rate > >> + && !(s->flags & SWR_FLAG_RESAMPLE) > >> && s->engine != SWR_ENGINE_SOXR){ > >> s->int_sample_fmt= AV_SAMPLE_FMT_S32P; > >> }else if(av_get_bytes_per_sample(s->in_sample_fmt) <= 4){ > > > > Wouldn't going over float lose bits? > > Yes, of course. But when resampling with default opt, aliasing happens at > about -90dB which is only 15 bits. Keeping 32 bits has no point at all. > > > I guess most times, s32 is > > actually 24 bit padded (since we have no 24 bit format). But it could > > be true 32 bit as well (with all bits used). Or does it use double, or > > am I missing something else? > > Using double is overkill, i think. From my naive point of view this seems ok, then.
On Fri, Mar 17, 2017 at 1:14 AM, wm4 <nfxjfg@googlemail.com> wrote: > On Thu, 16 Mar 2017 17:00:17 +0700 > Muhammad Faiz <mfcc64@gmail.com> wrote: > >> On Thu, Mar 16, 2017 at 1:01 PM, wm4 <nfxjfg@googlemail.com> wrote: >> > On Thu, 16 Mar 2017 12:43:10 +0700 >> > Muhammad Faiz <mfcc64@gmail.com> wrote: >> > >> >> use fltp when doing s32 -> s32 resampling >> >> because s32p has no simd optimization >> >> >> >> benchmark: >> >> old 17.913s >> >> new 7.584s (use fma3) >> >> >> >> Signed-off-by: Muhammad Faiz <mfcc64@gmail.com> >> >> --- >> >> libswresample/swresample.c | 2 ++ >> >> 1 file changed, 2 insertions(+) >> >> >> >> diff --git a/libswresample/swresample.c b/libswresample/swresample.c >> >> index f2e6600..74c96dc 100644 >> >> --- a/libswresample/swresample.c >> >> +++ b/libswresample/swresample.c >> >> @@ -223,6 +223,8 @@ av_cold int swr_init(struct SwrContext *s){ >> >> }else if( av_get_planar_sample_fmt(s-> in_sample_fmt) == AV_SAMPLE_FMT_S32P >> >> && av_get_planar_sample_fmt(s->out_sample_fmt) == AV_SAMPLE_FMT_S32P >> >> && !s->rematrix >> >> + && s->out_sample_rate == s->in_sample_rate >> >> + && !(s->flags & SWR_FLAG_RESAMPLE) >> >> && s->engine != SWR_ENGINE_SOXR){ >> >> s->int_sample_fmt= AV_SAMPLE_FMT_S32P; >> >> }else if(av_get_bytes_per_sample(s->in_sample_fmt) <= 4){ >> > >> > Wouldn't going over float lose bits? >> >> Yes, of course. But when resampling with default opt, aliasing happens at >> about -90dB which is only 15 bits. Keeping 32 bits has no point at all. >> >> > I guess most times, s32 is >> > actually 24 bit padded (since we have no 24 bit format). But it could >> > be true 32 bit as well (with all bits used). Or does it use double, or >> > am I missing something else? >> >> Using double is overkill, i think. > > From my naive point of view this seems ok, then. Applied Thank's
diff --git a/libswresample/swresample.c b/libswresample/swresample.c index f2e6600..74c96dc 100644 --- a/libswresample/swresample.c +++ b/libswresample/swresample.c @@ -223,6 +223,8 @@ av_cold int swr_init(struct SwrContext *s){ }else if( av_get_planar_sample_fmt(s-> in_sample_fmt) == AV_SAMPLE_FMT_S32P && av_get_planar_sample_fmt(s->out_sample_fmt) == AV_SAMPLE_FMT_S32P && !s->rematrix + && s->out_sample_rate == s->in_sample_rate + && !(s->flags & SWR_FLAG_RESAMPLE) && s->engine != SWR_ENGINE_SOXR){ s->int_sample_fmt= AV_SAMPLE_FMT_S32P; }else if(av_get_bytes_per_sample(s->in_sample_fmt) <= 4){
use fltp when doing s32 -> s32 resampling because s32p has no simd optimization benchmark: old 17.913s new 7.584s (use fma3) Signed-off-by: Muhammad Faiz <mfcc64@gmail.com> --- libswresample/swresample.c | 2 ++ 1 file changed, 2 insertions(+)