diff options
author | Determinant <[email protected]> | 2016-03-15 15:46:05 +0800 |
---|---|---|
committer | Determinant <[email protected]> | 2016-03-15 15:46:05 +0800 |
commit | 07fc1e2794027d44c255e1062c4491346b101a08 (patch) | |
tree | 8e7217b9c5e9570b94af5aaad3f94d1a37cfe40b /nerv/lib/matrix/generic/cukernel.cu | |
parent | a5a4d2735b595fc9fadc9c7e91198786d3c0e078 (diff) | |
parent | e15307f071813e2eb56f7f83229b91141961325a (diff) |
Merge branch 'master' of github.com:liuq901/nerv into liuq901-master
Diffstat (limited to 'nerv/lib/matrix/generic/cukernel.cu')
-rw-r--r-- | nerv/lib/matrix/generic/cukernel.cu | 22 |
1 files changed, 22 insertions, 0 deletions
diff --git a/nerv/lib/matrix/generic/cukernel.cu b/nerv/lib/matrix/generic/cukernel.cu index fc630ad..cf9d213 100644 --- a/nerv/lib/matrix/generic/cukernel.cu +++ b/nerv/lib/matrix/generic/cukernel.cu @@ -328,6 +328,15 @@ __global__ void cudak_(rearrange_frm)(const MATRIX_ELEM *a, MATRIX_ELEM *b, b[j + i * stride] = a[j / step + (j % step) * orig_dim + i * stride]; } +__global__ void cudak_(set_values_by_mask)(const MATRIX_ELEM *a, MATRIX_ELEM *b, + int nrow, int ncol, + int astride, int bstride, double val) { + int j = blockIdx.x * blockDim.x + threadIdx.x; + int i = blockIdx.y * blockDim.y + threadIdx.y; + if (i >= nrow || j >= ncol || a[i * astride] != 0.0) return; + b[j + i * bstride] = val; +} + __global__ void cudak_(scale_rows_by_col)(const MATRIX_ELEM *a, MATRIX_ELEM *b, int nrow, int ncol, int astride, int bstride) { @@ -772,6 +781,19 @@ extern "C" { cudaStreamSynchronize(0); } + void cudak_(cuda_set_values_by_mask)(const Matrix *a, Matrix *b, double val) { + dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N); + dim3 numBlocks(CEIL_DIV(b->ncol, threadsPerBlock.x), + CEIL_DIV(b->nrow, threadsPerBlock.y)); + cudak_(set_values_by_mask)<<<numBlocks, threadsPerBlock>>> \ + (MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(b), + b->nrow, b->ncol, + a->stride / sizeof(MATRIX_ELEM), + b->stride / sizeof(MATRIX_ELEM), + val); + cudaStreamSynchronize(0); + } + void cudak_(cuda_scale_rows_by_row)(const Matrix *a, Matrix *b) { dim3 threadsPerBlock(CUDA_THREADS_N, CUDA_THREADS_N); dim3 numBlocks(CEIL_DIV(b->ncol, threadsPerBlock.x), |