[Haskell-cafe] RE: [Haskell] ANNOUNCE: GPipe-1.0.0: A functional graphics API for programmable GPUs

[Tobias Bexelius]

Hi Petr,
 
Your question is completely justified! Yes, I beleive GPipe is a good foundation for such GPGPU-computations (general purpose GPU-programming), giving you easy access to data parallelism.
 
A way of doing this is to start with one or more equally sized textures that is your input (1 component depth textures perserves most floating point precision) . Then you rasterize a quad that covers the entire FrameBuffer, and for each fragment point sample these input textures and perform some computation on the samples. The resulting framebuffer can then be converted to a normal list in main memory, or converted to another texture to be used as input in another pass. 
 
To gather data, e.g. sum up all values in a texture, you could make a fragment program that uses a half-sized quad in which you sample two parts of the texture which you add and return in a half-sized texture. If you repeat this log2 texturesize times, you end up with one single value that is the sum of all values.
 
All texture loading and retrieving functions in GPipe are in IO, but safe to wrap in unsafePerformIO as long as you guarantee that the Ptr's are safe. So I think its easy to modularize GPGPU computations in GPipe and for instance create something like:
> simpleGPUmap :: (Fragment Float -> Fragment Float) -> [Float] -> [Float]
 
So, why dont you have a go on it? It might just turn out to be a pretty useful library... ;)
 
 
Cheers
/Tobias
 
 
 
> Date: Wed, 7 Oct 2009 16:47:15 +0200
> From: deb at pudlak.name
> To: tobias_bexelius at hotmail.com
> CC: haskell-cafe at haskell.org
> Subject: Re: [Haskell] ANNOUNCE: GPipe-1.0.0: A functional graphics API for programmable GPUs
> 
> Hi Tobias,
> 
> (I'm completely new to GPU programming, so my question may be completely
> stupid or unrelated. Please be patient :-).)
> 
> Some time ago I needed to perform some large-scale computations
> (searching for first-order logic models) and a friend told me that GPUs
> can be used to perform many simple computations in parallel. Could GPipe
> be used for such a task? I.e. to program some non-graphical,
> parallelized algorithm, which could be run on a GPU cluster?
> 
> Thanks for your answer,
> 
> Petr
> 
> On Sun, Oct 04, 2009 at 08:32:56PM +0200, Tobias Bexelius wrote:
> > I'm proud to announce the first release of GPipe-1.0.0: A functional graphics
> > API for programmable GPUs.
> > 
> > GPipe models the entire graphics pipeline in a purely functional, immutable
> > and typesafe way. It is built on top of the programmable pipeline (i.e.
> > non-fixed function) of OpenGL 2.1 and uses features such as vertex buffer
> > objects (VBO's), texture objects and GLSL shader code synthetisation to create
> > fast graphics programs. Buffers, textures and shaders are cached internally to
> > ensure fast framerate, and GPipe is also capable of managing multiple windows
> > and contexts. By creating your own instances of GPipes classes, it's possible
> > to use additional datatypes on the GPU.
> > 
> > You'll need full OpenGL 2.1 support, including GLSL 1.20 to use GPipe. Thanks
> > to OpenGLRaw, you may still build GPipe programs on machines lacking this
> > support.
> > 
> > The package, including full documentation, can be found at:
> > http://hackage.haskell.org/package/GPipe-1.0.0
> > 
> > Of course, you may also install it with:
> > cabal install gpipe
> > 
> > 
> > Cheers!
> > Tobias Bexelius
> > 
> > ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
> > kolla in resten av Windows LiveT. Inte bara e-post - Windows LiveT är mycket
> > mer än din inkorg. Mer än bara meddelanden







Windows Live: Håll dina vänner uppdaterade om vad du gör online.
 		 	   		  
_________________________________________________________________
Hitta kärleken nu i vår!
http://dejting.se.msn.com/channel/index.aspx?trackingid=1002952
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.haskell.org/pipermail/haskell-cafe/attachments/20091007/abaf6d9a/attachment.html