Hi, core NumPy doesn't seem to support a lot of output arguments, or common composite operations. For example, a common operation is something like a = outer(b,c) or a += outer(b,c) There are some workarounds, but they aren't pretty. Consistently providing output arguments throughout NumPy would help; is there any reason this isn't being done? For example, it would be nice if "outer" supported: outer(b,c,output=a) outer(b,c,increment=a) outer(b,c,increment=a,scale=eps) or maybe one could specify an accumulation ufunc, with addition, multiplication, min, and max being fast, and with an optional scale parameter. Another approach might be to provide, in addition to the convenient high-level NumPy operations, direct bindings for BLAS and/or similar libraries, with Fortran-like procedural interfaces, but I can't find any such libraries in NumPy or SciPy. Am I missing something? It seems like writing native code to speed up these kinds of operations isn't really so great because (1) it unnecessarily complicates development and packaging, and (2) is likely to perform worse than BLAS and similar libraries. How are other people handling these cases? Thanks, Tom
Thomas Breuel wrote:
Hi,
core NumPy doesn't seem to support a lot of output arguments, or common composite operations. For example, a common operation is something like
a = outer(b,c)
or
a += outer(b,c)
There are some workarounds, but they aren't pretty. Consistently providing output arguments throughout NumPy would help; is there any reason this isn't being done?
Time.
For example, it would be nice if "outer" supported:
outer(b,c,output=a) outer(b,c,increment=a) outer(b,c,increment=a,scale=eps)
or maybe one could specify an accumulation ufunc, with addition, multiplication, min, and max being fast, and with an optional scale parameter.
What would the increment and scale parameters do? Why should they be part of the outer() interface? Have you looked at the .outer() method on the add, multiply, minimum, and maximum ufuncs?
Another approach might be to provide, in addition to the convenient high-level NumPy operations, direct bindings for BLAS and/or similar libraries, with Fortran-like procedural interfaces, but I can't find any such libraries in NumPy or SciPy. Am I missing something?
scipy.linalg.flapack, etc. -- Robert Kern "I have come to believe that the whole world is an enigma, a harmless enigma that is made terrible by our own mad attempt to interpret it as though it had an underlying truth." -- Umberto Eco
For example, it would be nice if "outer" supported:
outer(b,c,output=a) outer(b,c,increment=a) outer(b,c,increment=a,scale=eps)
or maybe one could specify an accumulation ufunc, with addition, multiplication, min, and max being fast, and with an optional scale parameter.
What would the increment and scale parameters do? Why should they be part of the outer() interface? Have you looked at the .outer() method on the add, multiply, minimum, and maximum ufuncs?
output=a would put the output of the operation into a increment=a would increment the values in a by the result of the operation increment=a,scale=eps would increment the values in a by the result of multiplying the result by eps. I think a good goal for a numerical scripting language should be to allow people to express common numerical algorithms without extraneous array allocations for intermediate results; with a few dozen such primitives, a lot of native code can be avoided in my experience.
Another approach might be to provide, in addition to the convenient
high-level NumPy operations, direct bindings for BLAS and/or similar libraries, with Fortran-like procedural interfaces, but I can't find any such libraries in NumPy or SciPy. Am I missing something?
scipy.linalg.flapack, etc.
Thanks; I'll have to take a look. I don't remember whether LAPACK contains these low-level ops itself. Cheers, Thomas.
participants (2)
-
Robert Kern -
Thomas Breuel