diff options
| -rw-r--r-- | nerv/lib/matrix/generic/cukernel.cu | 5 |
1 files changed, 3 insertions, 2 deletions
diff --git a/nerv/lib/matrix/generic/cukernel.cu b/nerv/lib/matrix/generic/cukernel.cu index f996fdd..e1063af 100644 --- a/nerv/lib/matrix/generic/cukernel.cu +++ b/nerv/lib/matrix/generic/cukernel.cu @@ -231,8 +231,9 @@ __global__ void cudak_(update_select_rows)(MATRIX_ELEM *c, const MATRIX_ELEM *a, int i = blockIdx.y * blockDim.y + threadIdx.y; if (i >= nrow_a || j >= ncol_a) return; int i_c = lrintf(idx[i]); - //c[j + i_c * stride_c] = c[j + i_c * stride_c] * (1 - beta * alpha) + a[j + i * stride_a] * alpha; - atomicAdd_nvidia(&c[j + i_c * stride_c], c[j + i_c * stride_c] * (- beta * alpha) + a[j + i * stride_a] * alpha); + //critical: i_c could conflict among threads(same index in the idx array), so atomicAdd is used + //c[j + i_c * stride_c] = c[j + i_c * stride_c] * (1 - beta * alpha) + a[j + i * stride_a] * alpha; + atomicAdd_nvidia(c + j + i_c * stride_c, c[j + i_c * stride_c] * (- beta * alpha) + a[j + i * stride_a] * alpha); } __global__ void cudak_(expand_frm)(const MATRIX_ELEM *a, MATRIX_ELEM *b, |