Hi Karan,
It's normal for quant_all_bands() to take most of the CPU since it also
does most of the decoding work. I don't have any particular trick to
give you that I haven't already applied in the code. My only general
advice is that you should try git master and make sure that the ARMv5E
optimizations are enabled.
Cheers,
Jean-Marc
On 01/08/14 12:52 AM, karan.tandel wrote:> Hi,
> I am working on performance tuning of opus library on ARMv5E
> platform. Our marvell soc also have wmmx simd coprocessor for speed
> optimization. Now, I have profiled quant_all_bands() function and it
> seems its taking almost 51% time of whole frame decoding. I have tried
> to find out the hot spots parts in this function and it seems
> quant_all_partition() function is taking too much time. Now I m trying
> to locate the main hotspot parts in this function but as this function
> is very much complex and its recursively called to splitting the bands I
> am unable to detect the hotspot part in this function. So can you guys
> give me some insights on optimizing any part of this function? I have
> tried applying wmmx simd on alg_unquant() function but its not showing
> any speed optimization. I am stuck in this crucial parts. Any help,
> insights on optimizing this function will be such a great help to me.
> Thanks in advance.
>
> Thanks,
> Karan
>
> _______________________________________________
> opus mailing list
> opus at xiph.org
> http://lists.xiph.org/mailman/listinfo/opus
>