I have a GTX260 that has double precision support, and have been using CUDA with double precision support for a while now, however, the Open CL conformant driver doesn’t set the cl_khr_fp64 string in the device extensions. Does anyone know when a double precision variant of the NVIDIA OpenCL driver will be released?
No double precision support?! It would be useless for many if not most scientific computation. I have read through nVidia’s OpenCL documents. My impression is that double precision is supported on GPUs with 1.3 computing capability. So I just convinced my boss to buy a $6000 workstation with Tesla C1060. What should I do now?
Use CUDA while this gets sorted out?
I don’t have GPU programming experience. So I want to start with OpenCL which is cross platform. I don’t want to rewrite my code several times.
Did nVidia’s OpenCL team intentionally mislead the readers when they wrote the documents or do they expect a new release with double precision support very soon?
Double precision is an optional extension. The OpenCL guide from NVidia indicates this so I don’t see how this is misleading anyone. I agree that it is important (and it is very important to me) but I think you may need to read the development document again.
I was referring to NVIDIA’s documents (programming guide, etc). When you read them,
you will get the impression that the double precision is included, although it will be converted
to single precision during computation on low computing capability devices. I suspect that
they most copy the CUDA documents without much thinking.