For the use case where the number of bins is specified by an integer in np.histogram, the use of searchsorted etc. is very inefficient. It's possible to get at least a factor of 5x speedup by being smarter about this case just with Python + Numpy code, and I can get a factor of 30 speedup with Cython.

Would it be worth implementing a more efficient version of that special (yet common) case? If so, I'd be happy to turn this issue into a pull request.