On Monday 11 July 2011, Simon King wrote:
> Hi Martin,
Hi Simon,
> On 10 Jul., 20:38, Martin Albrecht <martinralbre...@googlemail.com>
>
> wrote:
> > Which makes sense because we always return a zero matrix when we allocate
> > in M4RI, so copy() is alloc + memcpy while just creating it is just an
> > alloc.
>
> How does that work? I learnt the hard way: Allocating memory does not
> necessarily mean that the memory is initialised by zeroes. Perhaps the
> other matrix classes can benefit from M4RI's way to create initialised
> memory?
We are not doing anything special except for calling calloc which zeros the
memory or we memset() the memory to zero which should still be faster than a
memcpy().
> > I guess one approach could be a class attribute which determines whether
> > a cache should be used or not (i.e. by type). Another one would decide
> > based on some dimension (I don't know which since that depends on the
> > implementation heavily).
>
> According to Robert, the fastest way to create an empty matrix depends
> not only on the class but on the dimensions. My main concern is that
> doing an elaborate test - repeated each time a zero matrix is created
> - would considerably slow things down. Therefore, my first suggestion
> was to introduce a static method of the matrix classes (see post
> above).
>
> But I forgot that the test actually needs to be done only once for
> each matrix class: It is an ideal candidate for a lazy attribute of
> MatrixSpace_generic! Hence, we may do something like
> @lazy_attribute
> def _copy_zero(self):
> if self.__is_sparse:
> return False
> if self.__matrix_class ==
> sage.matrix.matrix_mod2_dense.Matrix_mod2_dense:
> return False
> if self.__matrix_class ==
> sage.matrix.matrix_modn_dense.Matrix_modn_dense:
> if self._nrows>... and self._ncols>...:
> return False
> else:
> return True
> ...<do something for matrices over ZZ and so on>...
> return True
>
> def __call__(self,...):
> ...
> if entries is None or entries == 0:
> if self._copy_zero: # faster to copy than to create a new
> one.
> return self.zero_matrix().__copy__()
> else:
> return self.__matrix_class(self, {}, coerce=coerce,
> copy=copy)
Well, you are still adding one conditional jump (if self._copy_zero), but I'll
give this approach a try.
Cheers,
Martin
--
name: Martin Albrecht
_pgp: http://pgp.mit.edu:11371/pks/lookup?op=get&search=0x8EF0DC99
_otr: 47F43D1A 5D68C36F 468BAEBA 640E8856 D7951CCF
_www: http://martinralbrecht.wordpress.com/
_jab: martinralbre...@jabber.ccc.de
--
To post to this group, send an email to sage-devel@googlegroups.com
To unsubscribe from this group, send an email to
sage-devel+unsubscr...@googlegroups.com
For more options, visit this group at http://groups.google.com/group/sage-devel
URL: http://www.sagemath.org
