Hi,

I'm going to write an application with heavy use of matrix calculation. Actually the size of the matrices is small, i.e. 4x4, but I have to do *a lot* of matrix-per-matrix calculation. I'm planning to use CUDA to speedup the process. On the other hand I want to have an adapter between my app and the basic technology so I'm looking for something that can work even without CUDA. Eigen library seems great but I don't understand, reading the docs, what is the current status of CUDA support in Eigen. Is it full supported? If not, what is "backended" whit CUDA and what not? I need to use a function like cublasSgemmBatched() of cuBLAS to do parallel processing of the matrix calculation, is it possible via some high level method of Eigen? As last option I could write my own class and copy to/from gpu the data via a raw data pointer. The method data() of the class matrix should work for my goal, isn't it?

Thanks.

Mark

Currently Eigen do not call CUDA by itself, basically we only added the required __DEVICE__ token at the right place (plus some other adjustments) so that one can use Eigen *from* a CUDA kernel. So you can write a generic function and call it from CUDA or from the CPU. A typical example is our respective unit test in test/cuda_*

ggael wrote:Currently Eigen do not call CUDA by itself, basically we only added the required __DEVICE__ token at the right place (plus some other adjustments) so that one can use Eigen *from* a CUDA kernel. So you can write a generic function and call it from CUDA or from the CPU. A typical example is our respective unit test in test/cuda_*

I didn't find test for cuda in the latest release, it's only present in dev branch, so I guess it's not ready for production use. Can you confirm that I can go writing my own methods using the data() method of the matrix class?

Indeed, CUDA support is only in the devel branch. If you only want to access to the internal storage of an Eigen::Matrix , then using mat.data() is safe.