Indexing with negative step sizes

[FilipeMaia]

Currently arrays cannot be indexed with negative step sizes.

[FilipeMaia]

The C++ side can almost handle negative step sizes, the only obvious roadblock I see is there being no way to specify seq.begin at the end of the array. calcDims() needs a few changes to take the negative steps into account, in particular in the case when the begin and end are not specified.

[pavanky]

Can this be closed after the PR ?

[FilipeMaia]

I'm afraid not, the issue is deeper. There's no easy way to do array[::-1] because there's no way (that I know) to specify that seq.begin corresponds to the end of the array in the C/C++ API. The fixes I did only handle cases where the begin and end are specified.

[pavanky]

af::seq(end,0, -1) does not work ?

[FilipeMaia]

Nope it gives:

af::exception: ArrayFire Exception(203): Invalid input size
In /Users/filipe/src/arrayfire/src/api/cpp/array.cpp:725

Code:

#include <arrayfire.h>

int main(){
  af::array a = af::randu(3);
  af_print(a);
  af_print(a(af::seq(af::end,0,-1)));
}

[pavanky]

OK, I think I can still handle this in python itself. If you are not working on this and #26 I will give a shot at fixing them.

[FilipeMaia]

If you look at calcDims when end is 0 you end up on:
DIM_ASSERT(1, seq.begin >= -DBL_MIN && seq.begin < parentDim);

which will fail as af::end is smaller than -DBL_MIN

[FilipeMaia]

I'm not working on it, I have workarounds for my code.

[pavanky]

Ok, I think I am going to handle this upstream. The workarounds are ugly.

[pavanky]

This is fixed in 3.1.1 release. I think we can safely close this.