Memory views?

[krober10nd]

The []-operator still uses full Python operations with memoryviews you can access the data buffer directly at C speed.

https://cython.readthedocs.io/en/latest/src/userguide/numpy_tutorial.html#efficient-indexing-with-memoryviews

inpoly-python/inpoly/inpoly_.pyx

Line 75 in 189c63f

bnds[jpos] = 1

Maybe this is useful

[krober10nd]

Turning off boundschecking speeds things up by about 100% on my machine.

@cython.boundscheck(False)  # Deactivate bounds checking

[dengwirda]

Thanks Keith, I tried to test this a little, but didn't find many differences using views, etc. Possibly as the np.arrays are already type tagged there's not more to gain here? I'm not sure...

I did turn off bounds checking + negative indexes, etc though, as well as reorganised the kernel code a little --- for me, this led to marginal gains. It looks like the call to argsort is now taking up around 60% of runtime, so I'm not sure there's a lot more performance to chase here. Let me know if you find anything though --- I've not tested on macOS.

[krober10nd]

Yea, I quickly tried the memory views (which was a new thing for me too) and didn't notice much of a gain either. I think the boundschecking may depend on the C compiler setup and the options passed to the compiler. Best to leave it out and let the user decide.

It looks like the call to argsort is now taking up around 60% of runtime, so I'm not sure there's a lot more performance to chase here. Let me know if you find anything though --- I've not tested on macOS.

I agree. I'll take another look at on my day off tomorrow and then it should be a simple upload. In any case, I'd definitely put a note in the README about exporting the path to the numpy.headers()