How to prec-div=true to OpenCL compiler

In NVIDIA_OpenCL_ProgrammingGuide.pdf, page 41, for ulp error from x/y, it said “0 for compute capability ≥ 2 when compiled with -prec-div=true”. It is not clear to me how to pass the option -prec-dev=true to clBuildProgram. I got an error if I passed that option.

Does anyone know how to do it?


Best I could establish, there is no such option for the OpenCL compiler. I will follow up with the documentation team. Thanks for bringing this documentation issue to our attention.

[added later]

The compiler team informs me that for sm_2x targets, the OpenCL compiler defaults to single-precision division with IEEE rounding to-nearest-or-even, so there is no need to apply a compiler switch for this.

This is helpful. Thank you very much.