CODEWITHSUNDEEP
cudaoverloadingfast-math
tommsch
tommsch

Reputation: 688

__host__ __device__ functions calling overloaded functions

I do not understand whether there is function overloading in Cuda or not. I want to explain my problem on the following two functions, which I want to be able to use both on the GPU and the CPU, and I don't care about precision:

__host__ __device__
float myabs( float v ) {
    return abs( v + 1 ); //I want the floating point absolute value
}

__host__ __device__
float mycos( float v ) {
    return 2.f*cos( v );
}
  1. Which function of abs, resp. cos should I call, and why?
    • std::abs / abs / fabs / fabsf / anythingelse
    • std::cos / cos / cosf / __cosf / anythingelse

(Since __cosf is a Cuda-intrinsic and std::abs/std::cos are not available in Cuda, I assume I have to use preprocessor directives inside my functions for those choices.)

  1. Which headers should I include?

  2. Does the answer to the first two questions depend on whether I compile with fast-math flags (e.g. -ffast-math).

If this important for the answer, I am compiling with nvcc 10.2 under Ubuntu 18.04.4., but I am rather interrested in a platform independent answer.

Upvotes: 0

Views: 569

Answers (1)

talonmies
talonmies

Reputation: 72342

  1. Which function of abs, resp. cos should I call, and why?

If you are using floating point arguments, then conventionally you would use fabs and cosf. Those are the standard CUDA Math API implementations (and they correspond to the names of equivalent C standard library functions).

  1. Which headers should I include?

Conventionally you should include either math.h or cmath

  1. Does the answer to the first two questions depend on whether I compile with fast-math flags (e.g. -ffast-math).

No. Neither of those functions will be substituted for fast intrinsics by fast math.

Upvotes: 1

Related Questions