Calling MKL BLAS routines via OpenCL kernels on Xeon Phi natively

[Gaurav_K_Intel]

Dear All,

In brief - Is it possible to call MKL BLAS routines via a OpenCL kernel which is executing natively on a Xeon Phi card? 

Detail - There is a CFD code which I am helping with running on Phi natively. It uses DGEMM for interpolation on the mesh. The code is also ported to OpenCL for AMD GPUs. For some reasons, we need to run this code (OpenCL version of it) natively on the Phi card and invoke the relevant MKL blas calls via the kernel.Is it possible to do this in the newly released OpenCL SDK?

Thanks and Regards,

GK 

[Dmitry_K_Intel]

As for now MKL routines cannot be used by OpenCL applications on MIC. It *is* possible to use both offload compiler and OpenCL for MIC but:

1. this will result in device oversubscription

2. data sharing will be possible only by transfering it from offloaded region to host and then back to device using OpenCL buffers and vise versa.

We are looking right now for possible solutions of this problem.