Prev: Moving Minimum/Max?
Next: Optimization of Location Allocation
From: James Tursa on 4 Dec 2009 13:00
Jan,
I wonder if you could do me a favor and test some C-mex code with your Borland C compiler. I wrote some matrix multiply code and posted it to the FEX. It can be found here:

http://www.mathworks.com/matlabcentral/fileexchange/25977-mtimesx-fast-matrix-multiply-with-multi-dimensional-support

For general matrix multiplication it does the same thing as MATLAB and calls appropriate BLAS routines, so it matches MATLAB results exactly and runs in the same amount of time. However, there are dozens of special cases involving transposes and conjugates where my code beats MATLAB for speed ... sometimes by 300% - 400%. I have only tested this with the included lcc compiler and with MSVC 8 on 32-bit WinXP. I would like to test it with a Borland C compiler as well to see if the results hold, but unfortunately don't have one. In the past I have noted that you do have one and appear to be fairly adept at using it for mex applications, hence my request. The test file in question that I am particularly interested in having you run is the mtimesx_test_ddspeed.m file. This can take quite awhile depending on your computer & MATLAB version (1 hr - several hrs), so best to run it overnight if
possible. Any help here would be greatly appreciated. Thanks.

James Tursa
From: James Tursa on 5 Dec 2009 01:06

P.S. Jan, if you do manage to run this test, issue the following command first to make sure mtimesx is running in speed mode:

mtimesx('speed');

James Tursa
From: James Tursa on 5 Dec 2009 13:33
Jan,
Bug in the code. See my post to Bruno.
From: Jan Simon on 7 Dec 2009 19:10
Dear James!

> Jan,
> Bug in the code. See my post to Bruno.

Back again.
I'll compile it with BCC5.5, LCC2.4 (shipped with Matlab), LCC 3.8 and OpenWatcom 1.8 as soon as possible.

Jan
 | 
Pages: 1
Prev: Moving Minimum/Max?
Next: Optimization of Location Allocation