Accéder au contenu.
Menu Sympa

starpu-devel - [Starpu-devel] initialised SCRATCH buffer

Objet : Developers list for StarPU

Archives de la liste

[Starpu-devel] initialised SCRATCH buffer


Chronologique Discussions 
  • From: Xavier Lacoste <xavier.lacoste@inria.fr>
  • To: starpu-devel@lists.gforge.inria.fr
  • Subject: [Starpu-devel] initialised SCRATCH buffer
  • Date: Thu, 05 Jan 2012 10:34:46 +0100
  • List-archive: <http://lists.gforge.inria.fr/pipermail/starpu-devel>
  • List-id: "Developers list. For discussion of new features, code changes, etc." <starpu-devel.lists.gforge.inria.fr>

Hello,

I would like to know if it was possible to initialize a SCRATCH buffer using the a function of the codelets parameters ?

I have a CUDA codelet which as to perform a GEMM with one one matrix and a sub-bloc of this column and use the product to update an other matrix in several sub-blocs.
So I have one GEMM and several AXPY (one AXPY per column of each sub-bloc under the bloc used for the product).
I understood that the number of CUDA calls explain the poor perfomances of my CUDA codelet.

Thus I wrote a CUDA kernel for sparse GEMM based on MAGMA Tesla CUDA kernel.
I had to the MAGMA kernel 2 arrays containing the indexes of each bloc of the product and the corresponding indexes where I have to write in the destination buffer.

My problem is now I want to use this kernel with StarPU in PaStiX but I don't want to allocate one buffer by task.
I would like StarPU to build the array using a function of the codelet arguments at the same time he moves data to the executing device.

Am I clear in my explanation ?
Is this the good way to do ?
Is this already available in StarPU ?
If not, can this feature be added to StarPU ?

Thanks,

Xavier Lacoste.





Archives gérées par MHonArc 2.6.19+.

Haut de le page