Re: [eigen] Merging NVCC support branch

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

On Tue, Nov 5, 2013 at 4:51 PM, Marek Otahal <markotahal@xxxxxxxxx> wrote:

Hi,

On Tue, Nov 5, 2013 at 4:37 PM, Gael Guennebaud <gael.guennebaud@xxxxxxxxx> wrote:

OpenCL is C only, so no chance to call Eigen from it.

Not sure I understand it, but you'd be calling openCL from Eigen, right? There's support for C++11 in oCL : http://simpleopencl.blogspot.cz/2013/06/tutorial-simple-start-with-opencl-and-c.html

The current port works the other-way round: it allows to call Eigen *from* cuda. This is very useful for all applications working with many tiny matrices.

Deporting linear algebra operations for very large dense matrices on the GPU is already possible through Magma: https://github.com/bravegag/eigen-magma

Then, the missing bit is to deport coefficient-wise operations for which we have to generate kernels on the fly. Once the nvcc port is complete, this is just a matter of replacing for loops of our evaluator by a CUDA kernel launch with templated kernel taking as argument the _expression_. Of course, the data has to be stored/moved to the device memory.

gael