1) scipy is as fast as our implementation now (+- 10%)
2) the optimized graph hides shape mismatches (and currently can cause a
segfault)