James Tursa

@ Jonathan Sullivan: Your first example without repmat should have worked. Looks like I introduced a bug in the latest release. I will look into this right away ...

@ Michael Völker: Thanks for the comments. I will look into your request. There is a BLAS routine called DSDOT that does what you request. I will look into how easily I can incorporate this into MTIMESX for BLAS specific results. For the SPEED modes MTIMESX sometimes uses custom code to calculate the dot product (e.g., see function RealKindDotProduct), and in fact the accumulation is done in double precision regardless of input type (calculations themselves are single for single inputs), so it might not be too difficult to adapt this to return the double precision output if requested instead of always converting back to single for single inputs. Thanks for the suggestion.

@ James
The array sizes I'm working with are much larger what I posted. The size of A is 100x33x3. The size of B is 33x1x3x5000x5. That might be why you had trouble recreating it. I'm running on Windows 7, 64 bit, MATLAB 2011b, compiler is cl. If it makes a difference, some of the values in the first matrix are -inf.

@ James
So the problem, at least in part, is caused by using the 'SPEEDOMP' flag. I removed the flag, and now it works. Weird, huh?

James,

Fantastic code. It has come in handy. It is well documented, incredibly fast, and extremely useful, especially for N-D arrays.

For the N-D case, this code would be even more powerful if the singleton dimension capability were expanded. Currently, all the dimensions from 3:end must be either the same size (A to B), or must all be singleton for one of the variables.

For example
A = rand(100,100,3);
B = rand(100,100,3,5);
C = mtimesx(A,B); % Does not work. Crashes MATLAB.

You can get around this by using repmat:
A = rand(100,100,3);
B = rand(100,100,3,5);
C = mtimesx(repmat(A,[1 1 1 5]),B); % Works, but slow

Unfortunately, explicitly expanding out large arrays has overhead which is more than I'd like.

Overall code. A+.