Is there a CUDA equivalent to std::numeric_limits?

Question

I would like to determine the maximum int value in a CUDA kernel. Unfortunately I can't find anything similar to std::numeric_limits for CUDA. Trying to use the ::std function results in a error:

error : calling a __host__ function("std::numeric_limits<int> ::max") from a __global__ function("xyz_kernel") is not allowed C:\cuda.cu(153) (col. 10)

Is there a way to determine the desired value from withing a kernel, or should I just pass it as a parameter?

Answer 1

It exists but it is not as generic as std::numeric_limits. See this page for the list.

For example, you can have NPP_MAX_32U but this is specific to 32-bit unsigned rather than to the int type, whose width is system-dependent.

Answer 2

I've written something like that:

CUDA-compatible version of <limits> (named limits.cuh)

Essentially this means adding __host__ __device__ to a bunch of functions, and putting all of the structures within a namespace other than std.

Notes:

1. Licensing. It's based on <limits> from libstdc++, so GPL with the standard library exception.
2. This has undergone some, but not extensive, testing, with device-side code (only).
3. There may be some macro clashes with the actual <limits> (I'm not sure).