Matlab on NVidia GPUs

From: andrew cooke <andrew@...>

Date: Sat, 17 Oct 2009 07:35:17 -0300

I can't find any notes of what used to exist, but there are now 2
companies - one appears to be offering freeware (but not yet on
Linux).

http://gp-you.org/index.php (free)

http://www.accelereyes.com/ (commercial)

Andrew

Calling OpenCL Directly

From: andrew cooke <andrew@...>

Date: Mon, 28 Dec 2009 21:21:53 -0300

Instead of using those libraries, I ended up writing OpenCL (which is
very similar to CUDA, but cross-platform) directly.  The final speedup
was a factor of 100.

The reason I didn't use the libraries is that I couldn't see how they
could work efficiently for our case.  There were two reasons for this.

First, I tried vectorizing the Matlab code so that it might use
multiple cores on the CPU.  This was a failure, which made me suspect
that the code didn't have the form necessary to be auto-parallelized
(in the very core of the loop is an indirection which makes things
non-trivial).

Second, I was dealing with large data sets and needed to have qute
detailed control over what data was moved onto the GPU at what time
(again related to the indirection).

The entire project (including rewriting code in plain C as well as in
OpenCL) was 1-2 months, and got a 100x speedup.  An awful lot of that
time was spent (1) reverse engineering the algorithm I was porting and
(2) learning the details of OpenCL/CUDA.  Knowing what I know now, I
could repeat the work in a week.

The 100x speedup sounds impressive, but plain C was 6x faster than
Matlab, so the GPU only adds another factor of 17.  This is, I think,
because of the indirection I discussed earlier - the memory accesses
cannot be coallesced, which makes the entire calculation memory bound
(although I have not yet profiled the app to check this).

On the other hand, the original code was running on a rather
expensive, new dual Xeon box.  The GPU was a sub-$200 generic card from newegg. So "speed per$" is over x1000....

Andrew

Matlab/OpenCL Cross Reference

From: andrew cooke <andrew@...>

Date: Mon, 28 Dec 2009 21:22:43 -0300

PS See http://www.acooke.org/cute/Experience0.html