29 Oct
2010
29 Oct
'10
6:18 a.m.
On 29/10/10 08:07, Manjunath Kudlur wrote:
Wow Jolly Jumper ! This is news. We were looking at OpenCL for a while as the TMP on CUDA was bleak but this is interesting !
Glad to be of help. A higher level programming model for the GPU, say, a CUDA "backend" for NT2, should be very cool.
We're deep in this. Our current experiment on OCL are OKish in term of performance but require some fine tuning on the CL JIT part to avoid recompilation. CUDA could be better but NVIDIA centric. I guess we'll have to do like we did for SIMD and offer multiple back-ends.