The use of _syncthreads()

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello everyone,

Based on this post https://stackoverflow.com/questions/15240432/does-syncthreads-synchronize-all-threads-in-the-grid, I can understand that this function is used to synchronize threads inside a block.

In my code, I used the function and I’m not sure if it’s correct, I need some guidance, here’s my kernel that adds two vectors.

__global__ void
vectorAdd (const float *A, const float *B, float *C, int numElements)
{
    int nIter = 1000;

    for (int k = 0; k < nIter; ++k)
    {
         int i = blockDim.x * blockIdx.x + threadIdx.x;

          if (i < numElements)
          {
              C[i] = A[i] + B[i];
          }
          __syncthreads();
    }

}

PS: I’m using the loop inside the kernel, because I want it to run for longer for power consumption measurements.

2

syncthreads isn’t needed there at all

it serves no useful purpose

however your usage there is not illegal

3

Thank you for your quick response.

Yet , I saw impact on time, it increases when using it : especially for small kernel like this one.

So better take it off, because it costs regarding the little time small kernels take.