Boost logo

Boost :

Subject: Re: [boost] [gsoc] Interest check for 3d geometry proposal
From: joel falcou (joel.falcou_at_[hidden])
Date: 2009-03-27 00:16:01

Le Ven 27 mars 2009 03:00, Fernando Cacciola a écrit :
> To
measure the convenience of using the GPU against the slow down of the
> bus bottleneck depends on what you offload to the GPU.

The OP want to offoad 4 element vector operation ont he GPU, one at a
time ...
It doesn't requir emuch more analysis to see it'll fail.

> Stating that *G*raphics algorithms, among all things,
wouldn't gain much
> from using the *Graphics*PU can't make sense
IMO. Otherwise, there
> wouldn't be any reason to have a GPU to
begin with (or GPUs would still be
> just dumb renderers as they
used to be in the distant past)
> FWIW you can google for the term
GPGPU to see how the field of computing
> in the GPU is

We've been digging GPGPU since two years now so we
start to knwo when it fails and when it does not. Small amount of data is
usually a bad start, small amount of dtata + few operations is road to
disaster. Hence my remark concerning the OP idea.

FYI, some
graphic algorithm horribly fail in term of efficiency on GPU. I don't
consider havign a speed-up of 6 or 8 with a GPU for n optical flow
algortihm a success. BEst GPGPU alorithm efficiency are achieved with
algorithm that mimick the operation density of 3D rendering (database
search, N-bod like applciation etc). Offloading a few 1x4 vector multiply
won't cut it

Boost list run by bdawes at, gregod at, cpdaniel at, john at