matmul and matmul3d seem to yield wrong results when used in the form a:matmul(b, a), while a = a:matmul(b) yields the expected result. This is probably because positions are overwritten with their results even if they are still needed for the calculation of other positions.
This could probably be fixed if a:matmul(b, a) calculated results columnwise and buffered results until the column is finished.
2
[Please log in to post a comment]
