Open folshost opened 4 years ago
Either this needs to be changed to be unsupported, or we should document it as a requirement that all tile rows for the LHS must have homogeneous boundaries, and that the tile columns for the RHS must also have homogeneous boundaries.
@folshost We should at least diagnose this case and throw and exception if detected.
The dot_d::2d2d result matrix is the wrong size. I did some premature optimization, and there's an edge case which is currently not accounted for. That is when the tiles of the LHS do not form tile rows with consistent row boundaries, i.e., one tile may have from row [25, 50), while another tile may have from row [24, 49). In fact, as it stands, the case when the local tile of the LHS does not possess all columns may not work at all, as there may very well be a mismatch between the sizes of the result matrices between localities.