[gmx-developers] gromacs on GPU
vvishal at stanford.edu
Fri May 13 21:18:22 CEST 2005
The single most annoying thing about GPUs is that they can not do
indirected writes. i.e. you can not do something of the form y[i] = x
where i is a random (i.e. non streaming) index.
Algorithmically speaking you can do this:
for i = 0 to n
f[i] = some_function( inputs )
You can not do this:
for i = 0 to n
j = some_array[i]
f[j] = some_function( inputs )
Another thing you can not really do in GPUs is true for-loops. All your
loops have to be unrollable at compile time.
Not sure if that was too vague, do email me if you would like to discuss
this further off the list. :)
On Fri, 13 May 2005 zrotsch at gmail.com wrote:
> Sorry if this is not the most right forum to post my question here,
> but since I am newbie, I will try to proceed here.
> To put it philosophically, developing computer program procedures
> involves mainly development of syntactic (formal symbol manipulation)
> engine and only partly with regard to specific hardware (since it is
> represented on system's level as formal code with regard to device
> Can someone formulate in the very basic way why it is so hard to
> develop a mathematical engine for running GROMACS on GPU (for a group
> of science oriented people)?
> Thank you.
> gmx-developers mailing list
> gmx-developers at gromacs.org
> Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-developers-request at gromacs.org.
More information about the gromacs.org_gmx-developers