Initially I used blockIdx.x in my code but I want to remove that and instead have a global value and use that in my block instead of blockidx.x. Since my code is too large and it hangs when I run it with large input sizes I thought think can help. I increment the counter atomically but when I run the code it hangs. Can anyone take a look at my code and see if I'm doing something wrong?
__device__ int counter = 0;
__global__ void kernel(int * ginput, int * goutput)
{
const int tid = threadIdx.x;
const int id = threadIdx.x + blockIdx.x * blockDim.x;
in myval = ginput[id];
if (tid == 0) {
atomicAdd(&counter, 1);
}
__syncthreads();
if (counter == 0) {
goutput[tid] = ...;
}
if (counter > 0) {
...
}
}
If I use blockIdx.x instead of counter in my code it works but I just want to replace it with the counter
If you want the counter to replace your usage of blockIdx.x (i.e. you want each block to have a unique value that it reads from counter), then something like this should work:
__device__ int counter = 0;
__global__ void kernel(int * ginput, int * goutput)
{
const int tid = threadIdx.x;
const int id = threadIdx.x + blockIdx.x * blockDim.x;
__shared__ int my_block_id;
if (tid == 0) {
my_block_id = atomicAdd(&counter, 1);
}
__syncthreads();
if (my_block_id == 0) {
goutput[tid] = ...;
}
if (my_block_id > 0) {
...
}
}
Your method is going to be troublesome, because if you do something like this:
if (counter > 5) ....
you are potentially reading a new updated value of counter from global memory, and any number of blocks may have updated that value, so the behavior will be unpredictable.