原子地址 - CUDA function

Question

I apply atomicAdd function to add 10 in each array component The results are Not identical to my expection. Could you tell me why the value of list[1] is 12, while I expect 11=1+10. Total threads are 5. The initial array values are

slist[0]=1
slist[1]=2
slist[2]=3
slist[3]=4
slist[4]=5

the results are

list[0]= 1, list[0]= 1
list[0]= 1, list[1]= 12
list[0]= 1, list[2]= 13
list[0]= 1, list[3]= 14
list[0]= 1, list[4]= 15

__global__ void RunAtomicAdd(int* slist, int* val)
{
    int id = threadIdx.x;
    slist[0] = atomicAdd((slist +id), 10);
    printf("list[0]= %d, list[%d]= %d \n", slist[0], id, slist[id]);
}

Answer 1

Note that atomicAdd does not return the updated value, instead it returns the old value: cuda atomicAdd example fails to yield correct output

So all of your outputs are expected. In slist[0] , even if you update the value with atomicAdd , you immediately overwrite it with the output of atomicAdd , the old value. This does not happen with the rest of the id , except they do indeed store 1 in slist[0] , all of them.

You may want to have a new array to store the result of atomicAdd .