cudanvidia

How do I use atomicMax on floating-point values in CUDA?

I have used atomicMax() to find the maximum value in the CUDA kernel:

__global__ void global_max(float* values, float* gl_max)
{
    int i=threadIdx.x + blockDim.x * blockIdx.x;
    float val=values[i];

    atomicMax(gl_max, val);
}

It is throwing the following error:

error: no instance of overloaded function "atomicMax" matches the argument list

The argument types are: (float *, float).

Solution

  • The short answer is that you can't. As you can see from the atomic function documentation, only integer arguments are supported for atomicMax and 64 bit integer arguments are only supported on compute capability 3.5 devices.