On Thu, Mar 16, 2017 at 1:01 PM, wm4 <nfx...@googlemail.com> wrote:
> On Thu, 16 Mar 2017 12:43:10 +0700
> Muhammad Faiz <mfc...@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 <mfc...@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.
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
