To use scipy as a complete replacement for Numeric or
> numarray will require work in numerix, in the setup.py build
> system, and possibly in cntr.c, which needs to build
> separate versions for numarray and Numeric. (I expect little
> or no change will actually be needed in cntr.c.)
I talked to Travis about this at scipy and he said that with the new
array interface, we wouldn't need to compile separate extensions for
each of Numeric, numarray and scipy (as we do now, eg with
_nc_transforms, _na_transforms, etc). This is good news because it
implies shorter compile times and smaller binary distributions, but I
can't quite understand how this would work yet, eg, to support the
there packages with one binary *.so). But this may be some of the
black magic of the new array interface. If correct, this approach
would require *either* Numeric 24, numarray (something recent) or
scipy_base.
Perhaps Todd has further insight, as the author of numerix and
presumably the implementor of the array interface for numarray...
JDH
John,
There may be a couple of related questions:
1) For how long do we maintain compatibility with older versions of Numeric and numarray? Even if the array interface has been added, or is being added, to numarray, it won't be present in older versions or on classic Numeric.
2) Even if the array interface is the way to go for the future, it may be easier for now to let cntr.c continue to use the Numeric C API, which I think is maintained intact in scipy. In fact, doing it this way looks very easy indeed.
It may be that I am misunderstanding something fundamental about how the array interface works and is implemented; I haven't looked at it closely.
Eric
John Hunter wrote:
···
"Eric" == Eric Firing <efiring@...202...> writes:
> To use scipy as a complete replacement for Numeric or
> numarray will require work in numerix, in the setup.py build
> system, and possibly in cntr.c, which needs to build
> separate versions for numarray and Numeric. (I expect little
> or no change will actually be needed in cntr.c.)
I talked to Travis about this at scipy and he said that with the new
array interface, we wouldn't need to compile separate extensions for
each of Numeric, numarray and scipy (as we do now, eg with
_nc_transforms, _na_transforms, etc). This is good news because it
implies shorter compile times and smaller binary distributions, but I
can't quite understand how this would work yet, eg, to support the
there packages with one binary *.so). But this may be some of the
black magic of the new array interface. If correct, this approach
would require *either* Numeric 24, numarray (something recent) or
scipy_base.
Perhaps Todd has further insight, as the author of numerix and
presumably the implementor of the array interface for numarray...
JDH