Unsupported operation while compiling atomicMin

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi everyone,

I am getting this error message “Error: unsupported operation” while trying to compile this piece of code :

int a = 20;
int b = 30;
atomicMin(&b, a-3);//the log trace points out to this line

The official doc states clearly that CUDA supports :

int atomicMin(int* address, int val);

I would like also to stress that I am using this code snippet within the body of a kernel. Is that the reason why I am getting this “Unsupported operation” error? If so how can I circumvent this problem (Id o need to calculate the min of two integers within the kernel since I am unable to make a host call for std::min from the device)?

Could anyone please point me to the error source here? Thank you.

2

An atomic operation must be done either on global or shared memory.

https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#atomic-functions

(read the first sentence)

This:

int b = 30;

is neither. It is in the local memory space. Atomics are unsupported on local memory.

You don’t need atomics to calculate things when using only a single thread. The nature of local memory usage is that it is visible only to a single thread. To find the minimum of two values in local memory associated with a particular thread, CUDA provides suitable functions for you.

It should be sufficient to do:

int a = 20;
int b = 30;
b = min(b, a-3);
3

@Robert_Crovella : Thank you!