[x264-devel] Re: Speed-up method

[Guillaume POIRIER]

Hi,

On 3/3/07, Tomas Carnecky <tom at dbservice.com> wrote:
> Guillaume POIRIER wrote:
> > x264 spends the vast majority of its time doing integer computations,
> > not floats.
> >
> > Last time I looked at it, integer performance on GPU isn't that great
> > compared to its float performance.
> > I'd like to point out that communications between CPU and GPU are
> > expensive, and that there's no shared memory between them.
>
> PCI-E has a very big bandwidth, even for GPU->CPU transport (maximum
> with my 7800 is ~1150 MB/s).

But latency is very poor (compared to a HT link (AMD) or shared L2/FSB
communication(intel) ). That can easily kill all speed-ups you gained
from the GPU. You then need to let the GPU do a lot of work without
having the data going back and forth between the CPU and GPU, which
requires serious threading effort (it's not garanteed that x264's
threading model maps too well on heterogenous archs like CPU+GPU).

To put it in a nutshell, GPU power is certainly not a free ride...
free lunch is over!

;-)

Guillaume

-- 
This is the x264-devel mailing-list
To unsubscribe, go to: http://developers.videolan.org/lists.html