I would like to ask, regarding the following program, can each d_sigma successfully read the old value of d_A before the atomic operation updates it? Or will there be random and uncertain results?
int i = blockIdx.x * blockDim.x + threadIdx.x;
float d_sigma[i]=d_A[i];
atomicAdd(&d_A[i], 10.0f);
atomicAdd(&d_A[i-1], 10.0f);
atomicAdd(&d_A[i+1], 10.0f);
I think there is a best practice, not to mix non-atomic and atomic accesses.
Not sure, how true this is any longer.
Please also state, whether d_A is global memory (probably, because of the naming) or e.g. shared memory.
cuda provides no guarantees about the order of thread execution, except those that you explicitly write in code. You haven’t included any here.
Since each thread, therefore, is writing to 3 separate locations, two of which are also written and read by adjacent threads, there is no reason to assume any ordering of things. You cannot say whether a given location d_A[i] will be read before any atomic updates, or which atomic updates will be applied, when it is read.
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.