Hi Jonny, Jonathan Hunt <[email protected]> writes: > Am I correct in my understanding of the OpenCL spec that an > out-of-order queue can be emulated by creating a new queue for each > kernel? Obviously, it is up to the underlying implementation as to > whether it executes them out of order but it is certainly allowable to > execute kernels from two different queues simultaneously and/or > out-of-order?
This is true in theory but often not in practice. Creating a command queue can be an expensive operation. At one point, creating a stream (CUDA's equivalent of a queue) in CUDA required that the entire GPU finish whatever it was doing. Not sure to what extent that is still true. > Tangentially related, I couldn't find any mailing lists about OpenCL? > Is stackoverflow the best place for discussions/questions about OpenCL > (i.e. not pyopencl specific) - any other recommended forums? Brian tackled this question, I'd just like to add that my experience with getting a response out of the Khronos forums has been mixed. As of today, it seems like Nvidia's forums are also back up. For better or worse (and likely as a move to decrease OpenCL's visibility) they seem to have merged the CL and CUDA forums: https://devtalk.nvidia.com/default/board/57/cuda-programming-and-performance/ Andreas _______________________________________________ PyOpenCL mailing list [email protected] http://lists.tiker.net/listinfo/pyopencl
