That's looking good. Would you like me to run it on an unburdened opteron to see how it goes? If you like you can send me a tarball and I'll try it out.
I think our best bet for a significant improvement now is the idea of using two Gray tables of half the size simultaneously. I also realised it possibly improves the cache performance for the A matrix too. I was casually wondering whether Magma might use a highly optimised Winograd's algorithm instead of the naive algorithm. But over GF2 I think it probably actually takes longer, since it basically replaces n^2 full length scalar multiplies by n^2 half length ones and 2*n^2 half row additions, plus a pile of other overhead. Bill. On 17 May, 20:32, Martin Albrecht <[EMAIL PROTECTED]> wrote: > On Saturday 17 May 2008, Martin Albrecht wrote: > > > > I think a better idea would be to explicitly force all matrices and > > > all rows to be 128 bit aligned if the matrices are wide enough to > > > benefit from SSE2, Then the combine function can always use SSE2 and > > > there will be no need to check for alignment. > > > That doesn't seem to make a noticeable difference for me (on C2D). However, > > I realised that the multiplications where the target matrix is a real > > matrix rather than a window (which has bad data locality). Copying > > everything over seems not like a good idea but it at least indicates an > > area for improvements. > > Okay, if I only copy when we crossover to M4RM then the memory overhead is > constant (~ cutoff^2) and the performance still improves. > > Old: 64-bit Debian/GNU Linux, 2.33Ghz Core2Duo > Matrix Dimension Magma 2.14-13 (64-bit) M4RI-20080517 (64-bit) > 10,000 x 10,000 2.920 3.610 > 16,384 x 16,384 11.140 12.120 > 20,000 x 20,000 20.370 24.390 > 32,000 x 32,000 74.290 94.910 > > New: 64-bit Debian/GNU Linux, 2.33Ghz Core2Duo > Matrix Dimension Magma 2.14-13 (64-bit) M4RI-20080517 (64-bit) > 10,000 x 10,000 2.920 2.990 > 16,384 x 16,384 11.140 11.750 > 20,000 x 20,000 20.370 21.180 > 32,000 x 32,000 74.290 86.570 > > On Opteron things don't look this way, but I think sage.math is pretty heavily > used right now such that my benchmarks there are not very telling. > > Martin > > -- > name: Martin Albrecht > _pgp:http://pgp.mit.edu:11371/pks/lookup?op=get&search=0x8EF0DC99 > _www:http://www.informatik.uni-bremen.de/~malb > _jab: [EMAIL PROTECTED] --~--~---------~--~----~------------~-------~--~----~ To post to this group, send email to sage-devel@googlegroups.com To unsubscribe from this group, send email to [EMAIL PROTECTED] For more options, visit this group at http://groups.google.com/group/sage-devel URLs: http://www.sagemath.org -~----------~----~----~----~------~----~------~--~---