This one is a reminder that the physical transposed-ness of weight matrices (fortran vs c order) can affect the speed of evaluation. I found that the fortran order made the logistic regression example go faster. But for another thing it didn't help.

If we had a list of standard tricks for optimizing computations, then this would be an addition to it.

We would need to confirm that it is still the case.

We would need to confirm that it is still the case.