diff options
Diffstat (limited to 'nerv/lib/matrix/generic/cukernel.cu')
| -rw-r--r-- | nerv/lib/matrix/generic/cukernel.cu | 25 |
1 files changed, 25 insertions, 0 deletions
diff --git a/nerv/lib/matrix/generic/cukernel.cu b/nerv/lib/matrix/generic/cukernel.cu index 9244783..2b696d5 100644 --- a/nerv/lib/matrix/generic/cukernel.cu +++ b/nerv/lib/matrix/generic/cukernel.cu @@ -363,6 +363,20 @@ __global__ void cudak_(copy_rows_by_idx)(const MATRIX_ELEM *a, MATRIX_ELEM *b, b[j + i * stride] = a[j + k * stride]; } +__global__ void cudak_(copy_rows_by_colidx)(const MATRIX_ELEM *a, MATRIX_ELEM *b, + const MATRIX_ELEM *idx, + int nrow, int ncol, int a_nrow, int stride, int idx_stride) { + int j = blockIdx.x * blockDim.x + threadIdx.x; + int i = blockIdx.y * blockDim.y + threadIdx.y; + if (i >= nrow || j >= ncol) return; + int k = lrintf(idx[i * idx_stride]); + if (k < 0 || k >= a_nrow) { + printf("error in kernel copy_rows_by_colidx k(%d) out of range\n", k); + } + b[j + i * stride] = a[j + k * stride]; +} + + extern "C" { #include "../cukernel.h" void cudak_(cuda_log_elem)(const Matrix *a, Matrix *b) { @@ -744,5 +758,16 @@ extern "C" { b->nrow, b->ncol, a->nrow, b->stride / sizeof(MATRIX_ELEM)); cudaStreamSynchronize(0); } + + void cudak_(cuda_copy_rows_by_colidx)(const Matrix *a, Matrix *b, + const Matrix *idx, int idx_begin) { + dim3 threadsPerBlock(CUDA_THREADS_NN, 1); + dim3 numBlocks(CEIL_DIV(b->ncol, threadsPerBlock.x), b->nrow); + cudak_(copy_rows_by_colidx)<<<numBlocks, threadsPerBlock>>> \ + (MATRIX_ELEM_PTR(a), MATRIX_ELEM_PTR(b), + MATRIX_ELEM_PTR(idx) + idx_begin, + b->nrow, b->ncol, a->nrow, b->stride / sizeof(MATRIX_ELEM), idx->stride / sizeof(MATRIX_ELEM)); + cudaStreamSynchronize(0); + } } #endif |