Open kiendangtor opened 2 months ago
Hello @kiendangtor ,
Unfortunately that is currently not supported for a single GPU.
Technically it seems that with modern GPUs it's feasible to run such matrices in a single GPU so it may make sense. AMGX actually instantiates num_rows/num_cols
, num_nnz
, and row_offsets
, column_indices
vector off the template parameter, but I haven't tried it with int64 and I believe in many places we assume int. Let me try to do test build with int64 and see if it's a lot of work to allow it - then maybe you can extend API in your local copy and make it work for your app.
Thanks @marsaev i hope it will be feasible. I only need nnz to be int64 if it helps.
In this call _AMGX_RC AMGX_API AMGX_matrix_upload_all(AMGX_matrix_handle mtx, int n, int nnz, int block_dimx, int block_dimy, const int row_ptrs, const int col_indices, const void data, const void diagdata); the number of nnz is int, and I have a fairly large matrix with nnz larger than the max value of int, how do we solve that issues?