Del (_@_.com) on 2/11/12 wrote:
---------------------------
>Anon (no@thanks.com) on 2/9/12 wrote:
>---------------------------
>>Don't know about AMP and OpenACC, but when discussing potential implementations
>>of an app that would benefit from massive parallelism, the developers with experience
>>that I talked to unanimously said to use CUDA unless you absolutely had to run on
>>non-NVidia HW. They went so far as to suggest developing/debugging in CUDA and then
>>porting to OpenCL if necessary, just because the developer experience is so much better.
>
>Hm, I have heard another story from developers. They are telling me that one or
>the other is not a big deal. I might get a biased view though, since I am very reluctant
>to sponsoring anything CUDA-based. I happen to agree with David, and my life is
>to short to invest in a dying code base. Hence, it may be that developers are telling
>me what I want to hear, and I haven't really started looking closely at the codes
>yet (actually I have only looked at coding with CUDA so far, and it looked OK).

Well, my anecdotal evidence is likely just as biased - these were a bunch of HPC developers with ready access to clusters of machines with NV GPUs attached - i.e. developers in an environment where "specifying NV as part of the solution", as Michael S put it, is either very acceptable, or already done. So their "CUDA is so much easier to use" sentiments might partly be due to having much more experience versus using OpenCL.

For what it's worth, the thing I'll be working on has no such luxuries. It won't ever run on Windows, should be able to run on a multi-core x86 without a GPU, and isn't likely to get very far with management if some kind of commercial compiler tech is required. I'm not really seeing an alternative to OpenCL at the moment. ISPC looks really interesting, but has no GPU backends at the moment, and it's not clear if it ever will.