Differences between revisions 10 and 11
Revision 10 as of 2007-09-30 21:07:48
Size: 3874
Editor: DavidHarvey
Revision 11 as of 2008-11-14 13:42:07
Size: 3874
Editor: localhost
Comment: converted to 1.6 markup
No differences found!

NTL wrapper

  • NTL wrapper -- Joel Mohler, Robert Bradshaw, Craig Citro, David Roe, William Stein, Martin Albrecht, Gonzalo Tornaria

Properties the wrapper should have

  • There are really three wrappers going on:
    • The lowest level is all the code in ntl_wrap.cpp, and the string transformations in decl.pxi. These basically make NTL available at the cython level, with zero or minimal overhead.
    • Building on top of that is the other cython files in sage.libs.ntl, e.g. ntl_ZZ.pyx. These define python classes which mimic the underlying NTL types.
    • Also building on top of the low level are classes like Polynomial_integer_dense, which are objects with SAGE semantics, that happen to use NTL for the underlying implementation.
  • Wrapper functions should never returns pointers to `new' objects. We should always pass in a reference to the object where the result will be stored. (This prevents unnecessary copies in the calling code.)
  • NTL classes are ctypedef'd with the same name with a _c suffix, e.g. the NTL type ZZ is imported to cython as ZZ_c. This happens in decl.pxi.

  • Function naming conventions. The rules are hard to specify precisely, but here are some examples.
    • A function that adds two ZZ's is called ZZ_add. A function that adds a ZZ and a long is called ZZ_add_long. The prefix is the class that this would be a method of, had NTL been written that way :-)

    • Capitalisation of function names is the same as in NTL, even when these are inconsistent within NTL itself.
  • Regarding filenames for ZZ_p vs zz_p. We use the same rules as NTL, i.e. the class name comes from zz_p, but the filename is lzz_p, to avoid capitalisation problems on the filesystem.
  • Contexts (ZZ_pContext, zz_pContext, ZZ_pEContext, zz_pEContext, GF2EContext):

an email

Call to arms email from David Harvey:

Last night Craig suggested that we should maybe devote some serious  
SD5 time to working on the NTL wrapper(s).

I don't think there's going to be enough time to actually do all the  
coding, and I'm not sure that it should be displacing other things,  
but definitely I think it would be an excellent idea to have a  
discussion meeting. Perhaps Monday afternoon or evening so that  
Robert can join in.

The thing is, we currently have *five* developers pretty actively  
working on NTL-related code: Joel, Robert, Craig, David Roe, and  
myself. This is a historic opportunity to finally Get It Right, while  
we have everyone's attention focussed.

So I propose the following: we should have a meeting, with the  
following goals:

(a) get some idea of what is the current status of NTL-related code
(b) figure out where we want it all to be
(c) come up with a rough plan of how we're going to get there, over  
the next few weeks perhaps
(d) get some of (c) and (d) *written down*

Here are some of the issues I identify:

(1) a consistent naming schemes for type, functions in the wrapper  
(Joel has already been doing a splendid job with this, but I think  
there's still a lot to do)
(2) a uniform interface for conversions to/from NTL objects, and  
efficient underlying conversion code
(3) I claim that ZZ_p::init, ZZ_pE::init should *never* be called, it  
should all go through Context objects
(4) what is the point of a zz_p wrapper?
(5) pointers or references? What's the best way to wrap? Can we do  
this consistently?
(6) how to minimise code duplication between the libs.ntl wrapper and  
SAGE classes that use NTL directly (or that soon will use NTL directly)
(7) memory management? When to use new? Placement new?

I think the most import thing that can come out of this meeting is an  
RFC-like document... "the NTL wrapper MUST do XYZ..... the NTL  
wrapper SHOULD do ABC....."

days5/proj/ntl (last edited 2008-11-14 13:42:07 by localhost)