PRF: Speed up point_in_path and friends #6450
Conversation
Since we now know the performance is so bad, this should discourage its use and flag it as something special. Also, use () in more places
|
Thanks to @tacaswell and @QuLogic for the legwork on this one! |
| subtrans(1, 1), | ||
| subtrans(0, 2), | ||
| subtrans(1, 2)); | ||
| int it = i % Ntransforms; |
There was a problem hiding this comment.
here you are using int's, but elsewhere, you use size_t. Any particular reason for that?
|
I am also perplexed about the performance issues with bool vs. uint8_t. Perhaps we should mention it to the numpy folks to see if they have any insights? That is the only change that has me a bit uneasy. Other than that, this looks good to me. |
|
If I recall correctly stl special cases bool and implements it as a As this is stl I do not think numpy would have any input (as I thought they Tom On Thu, May 19, 2016 at 9:30 AM Benjamin Root notifications@github.com
|
|
@tacaswell I remembered something similar to you about Notes from there:
So, going back to |
This is an alternative to #6446. Fix #6444.
point_in_path_implwhich avoids Python ref/deref and is therefore much fasteroperator[]to subarray (which is still useful in some places where speed is less critical) and moves to using (, ,) elsewhere in the codebase where it makes sense.boolfor the results inpoint_in_path_impl, which strangely was a lot slower