In Matlab the buffer of matrix is continuous in column . So what about the numpy array of Python. which one is beter between numpy.empty((n,1)) and numpy.empty((1,n))
Add a comment
|
2 Answers
In numpy you can choose between Fortran-contiguous (along the column, like in Matlab) and C-contiguous (along the row, which is the default in numpy) order, passing the order argument when you create an array, so you have more flexibility.
As @user2357112 already said, for a 1xN or Nx1 array it does not matter, but for a MXN array it does matter and you should be aware of that.
answered Aug 30, 2013 at 7:43
Saullo G. P. Castro
59.4k28 gold badges191 silver badges244 bronze badges
1 Comment
Samuel
Is the C-contiguous the default
order argument.They do different things. One makes an Nx1 array; the other makes a 1xN array. Neither is "better". (In fact, the memory layout will be identical for both arrays, even if you specify column-major storage.)
To answer the question about storage layout, though, numpy defaults to row-major layout, a.k.a. C-contiguous. You can see this clearly reflected in the docs.
4 Comments
usethedeathstar
And if you specify fortran-layout, whatever numpy routine you use will still return you something in default order (so fortran-layout seems pretty pointless to me)
Saullo G. P. Castro
@usethedeathstar you need Fortran layout when you are using for example
BLAS routines programmed in Fortran, so this layout can be very usefulusethedeathstar
@SaulloCastro how do you access those routines from numpy? Is there a fortran version of things like numpy.sqrt(x) or whatever you wish to do? Or do you mean if you would work with f2py or so?
Saullo G. P. Castro
@usethedeathstar you can have a look at this other answer for futher information... this can considerably speed up your code, but you should worry about this only for heavy tasks...I adapted this from pv for a faster
dot() when calling from Cython