diff options
Diffstat (limited to 'nerv/lib/matrix/generic/cumatrix.c')
| -rw-r--r-- | nerv/lib/matrix/generic/cumatrix.c | 21 |
1 files changed, 19 insertions, 2 deletions
diff --git a/nerv/lib/matrix/generic/cumatrix.c b/nerv/lib/matrix/generic/cumatrix.c index 40a0030..770e503 100644 --- a/nerv/lib/matrix/generic/cumatrix.c +++ b/nerv/lib/matrix/generic/cumatrix.c @@ -315,16 +315,17 @@ void nerv_matrix_(copy_rows_fromh_by_idx)(Matrix *a, const Matrix *b, long nrow = a->nrow; if (!(0 <= b_begin && b_begin + nrow <= idx->ncol)) NERV_EXIT_STATUS(status, MAT_INVALID_COPY_INTERVAL, 0); - long *idx_ptr = idx->data.i; + float *idx_ptr = idx->data.f; int i; if (idx->nrow != 1) NERV_EXIT_STATUS(status, MAT_IDX_VECTOR_EXP, 0); if (a->ncol != b->ncol) NERV_EXIT_STATUS(status, MAT_MISMATCH_DIM, 0); + PROFILE_START cudaStream_t *streams = (cudaStream_t*)malloc(sizeof(cudaStream_t) * nrow); for (i = 0; i < nrow; i++) { - int src_row = idx_ptr[b_begin + i]; + int src_row = (int)idx_ptr[b_begin + i]; if (!(0 <= src_row && src_row < b->nrow)) NERV_EXIT_STATUS(status, MAT_INVALID_IDX, 0); CUDA_SAFE_CALL(cudaStreamCreate(streams + i), status); @@ -339,6 +340,22 @@ void nerv_matrix_(copy_rows_fromh_by_idx)(Matrix *a, const Matrix *b, CUDA_SAFE_CALL(cudaStreamDestroy(streams[i]), status); } free(streams); + PROFILE_STOP + NERV_SET_STATUS(status, NERV_NORMAL, 0); +} + +void nerv_matrix_(copy_rows_fromd_by_idx)(Matrix *a, const Matrix *b, + const Matrix *idx, int b_begin, Status *status) { + long nrow = a->nrow; + if (!(0 <= b_begin && b_begin + nrow <= idx->ncol)) + NERV_EXIT_STATUS(status, MAT_INVALID_COPY_INTERVAL, 0); + if (idx->nrow != 1) + NERV_EXIT_STATUS(status, MAT_IDX_VECTOR_EXP, 0); + if (a->ncol != b->ncol) + NERV_EXIT_STATUS(status, MAT_MISMATCH_DIM, 0); + PROFILE_START + cudak_(cuda_copy_rows_by_idx)(b, a, idx, b_begin); + PROFILE_STOP NERV_SET_STATUS(status, NERV_NORMAL, 0); } |