(L) [2013/01/07] [D-POWER] [OpenACC] Wayback!>> OpenACC is a new open parallel programming standard designed to enable the millions of scientific and technical Fortran and C programmers to easily take advantage of the transformative power of heterogeneous CPU/GPU computing systems.
Anyone has tried openacc? It seems interesting and very easy to use, but I'd like to know if it can deliver the same performance as hand-written cuda code.
(L) [2013/01/08] [spectral] [OpenACC] Wayback!I don't know OpenACC but you have to know that nothing can deliver the speed of hand tuned CUDA code...
Simply because it is so complex to acheive the "best" performance that we can only do this based on tuning experience.
For sure, it can help to improve the speed but don't expect some kind of magic.
You have to know that by example, even CUDA is faster than OpenCL because the compiler contains more optimizations [SMILEY ;-)]
(L) [2013/01/08] [Dade] [OpenACC] Wayback!>> spectral wrote:I don't know OpenACC but you have to know that nothing can deliver the speed of hand tuned CUDA code...
Something I have just finished to read: [LINK http://gpuscience.com/articles/programming-cuda-and-opencl-a-case-study-using-modern-c-libraries/]
>> spectral wrote:You have to know that by example, even CUDA is faster than OpenCL because the compiler contains more optimizations
CUDA isn't particularly fast if you don't own a NVIDIA GPU [SMILEY ;)]
More seriously, I have some doubt OpenACC/C++ AMP/Bold/etc. are particularly useful in our field: performance is just too important and they seem fast only when applied to simple and/or specific kind of problems. I think it is the same reason why we have bags full of tricks and we live with 32 bit float instead of using double.