IERR size should be M x IR?
gillette7 opened this issue · comments
DelaunaySparse/python/delsparse.py
Line 86 in 6a80558
When trying to use the python wrapper on a case with output dimension (IR) larger than 1, I got error code 21: The size of the error array IERR does not match the number of interpolation points M. When I changed the size of IERR to M x IR, it worked. I think the readme has to be changed (in two spots) to reflect this.
@gillette7 I'm not following where the readme is wrong. I believe IERR should always be a vector of size M, never relating to the value IR.
The reason your change worked is that the code only checks the first dimension of any array you pass for IERR because it assumes it will be a 1D input (maybe it shouldn't?). Regardless of input array shape, the only values touched by the Fortran source will be IERR(...,1:M), the first M contiguous memory elements of the given IERR array.
The issue was I had declared ierr = np.zeros([1,m]) but it should have been ierr = np.zeros([m,1]); once I changed it, the error went away. Seems like just a different convention between python and Fortran. Thanks for checking on this!
This is a good point. Instead of checking the first shape dimension, the code could instead check the last (since that is really what matters from a Fortran perspective). That would make more sense, mathematically speaking, and it wouldn't change the behavior when an array with shape (M,) gets passed.
@thchang do you think this is a good idea? @gillette7 would it bother you if I make that change and you have to revert your code to be ierr = np.zeros([1,m]) after the update?
No problem for me if you want to change it - just let me know if you do!
@tchlux I think the documentation should reflect the expectation that users will pass a shape of (M,), but if checking the last dimension will make this more intuitive for Python users, then I support the change Thanks all for raising/solving this issue
@gillette7 this newly merged pull request now causes the code to check the last dimension of incoming vectors. If you are going to send in a vector with len(*.shape) > 1 then only the last dimension size needs to match.
For your purposes, this will require you to revert to using ierr = np.zeros([1,m]).