I have read up on floating precision and there might be an issue with rounding modes.
https://developer.nvidia.com/content/precision-performance-floating-point-and-ieee-754-compliance-nvidia-gpus Of course this is rounding for float/double register results, not the round() function, but I can imagine the same circuitry being re-used. In a nutshell, there are 4 different rounding modes: - rn: round to nearest, ties to even - rz: round towards zero - ru: round towards +infty - rd: round towards -infty My round(88.9f) == 88.0 problem could occur when "rd" is active for some reason. And now my question: Is the rounding mode set somewhere in pyopencl? Kind regards, Kai On 14.04.2014 02:04, Andreas Kloeckner wrote: > Kai Ruhl <[email protected]> writes: > >> thanks for the quick response -- I also tested it on a Titan Black >> which gives correct results. It is *just* the Ti780. I use the x-swat >> nvidia drivers 334 for Ubuntu 14.04. >> >> In the meantime, I used the AMD SDK lib to compile pyopencl with >> OCL1.2, still the same error. > > As in, AMD OpenCL gives you 88.0 as well? > > Andreas > _______________________________________________ PyOpenCL mailing list [email protected] http://lists.tiker.net/listinfo/pyopencl
