cory-master/BabelStream
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[CUDA] If using managed memory, use device pointer for host reduction

Browse Source
This commit is contained in:
Tom Deakin 2016-12-19 05:08:19 -07:00
parent 62860284b2
commit 8d66a27131
1 changed files with 11 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
CUDAStream.cu
View File

@ -245,12 +245,23 @@ T CUDAStream<T>::dot()
dot_kernel<<<DOT_NUM_BLOCKS, TBSIZE, sizeof(T)*TBSIZE>>>(d_a, d_b, d_sum, array_size);
check_error();
#if defined(MANAGED) || defined(PAGEFAULT)
cudaDeviceSynchronize();
check_error();
#else
cudaMemcpy(sums, d_sum, DOT_NUM_BLOCKS*sizeof(T), cudaMemcpyDeviceToHost);
check_error();
#endif
T sum = 0.0;
for (int i = 0; i < DOT_NUM_BLOCKS; i++)
{
#if defined(MANAGED) || defined(PAGEFAULT)
sum += d_sum[i];
#else
sum += sums[i];
#endif
}
return sum;
}