math-api | Put LaTeX Math equation on everywhere as img tag | Math library

CUDA Appendix Suggests that the value provided by the API are not accurate. There is a difference of, 1/2 ulps in their value. (when compared between host and device).

However, for all practical purposes, these values are accurate.

Anyway if you want to perform symbolic operations with high-level precision, using float precision makes it inaccurate.

As a helpful comment pointed out, I was overthinking the problem. The following works, and uses a simple bool to int conversion:

The kernel will compile if you cast the argument to the type indicated in the CUDA math API documentation:

Presumably FP rounding mode.

The exact result is usually between two representable float values, and the rounding mode determines which of the two is picked. So the difference between 2 rounding modes will be 1 unit in the last place (of the significand aka mantissa), or 0 if they round the same way.

The result you are obtaining is correct.

The intrinsic you are using implements double precision division with a specific IEEE 754-2008 rounding mode for the unit in the last place (ULP) of the significand. This controls what happens when a result cannot be exactly represented in the selected format. In this case you have selected round up, which means the last digit of the significand produced in the division result is rounded up (toward +∞). In your case all rounding modes should produce the same result because the result can be exactly represented in IEEE 754 binary64 format (it is a round power of 2).

Please read everything here before writing any more floating point code.

Ideas?

Switch to CUDA 9.2

Your code compiles without error on CUDA 9.2, but throws the errors you indicate on CUDA 9.1. If you have CUDA 9.1 installed, then the documentation for it is already installed on your machine. On a typical linux install, it will be located in /usr/local/cuda-9.1/doc. If you look at /usr/local/cuda-9.1/doc/pdf/CUDA_Math_API.pdf you will see that the corresponding functions are only marked __device__, so this change was indeed made between CUDA 9.1 and CUDA 9.2

Going via floating point gives you a much more precise result, slightly lower instruction count on most architectures, and potentially a higher throughput: