Objet : Developers list for StarPU
Archives de la liste
- From: Samuel Thibault <samuel.thibault@ens-lyon.org>
- To: Starpu-devel@lists.gforge.inria.fr
- Subject: [Starpu-devel] poll: Should performance models depend on matrix ld?
- Date: Wed, 19 Nov 2014 12:01:25 +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,
While working on the MPI layer, we have been thinking about the
footprint used for performance models: for matrices for instance, it
takes into account nx and ny, and notably distinguishes e.g. 128x16
vs 16x128, but it does not take into account the ld value. Perhaps
it should since I guess a kernel will not have the same efficiency
when used on a contiguous tile, or on a tile within a bigger matrix?
That would however mean changing the footprint of all matrices &
such performance models, and thus people would have to either rerun
performance measurements, or fix the footprint value in the existing
performance model files, so we'd probably only do this in the trunk.
So what do people think about it: should performance models depend on
e.g. matrix ld?
Samuel
- [Starpu-devel] poll: Should performance models depend on matrix ld?, Samuel Thibault, 19/11/2014
Archives gérées par MHonArc 2.6.19+.