Video Geeks Club

I presented a presentation containing introduction and motivation of GPU computing and a simple CUDA  code walk-though. The presentation file is available here.

We planned the weekend project for this coming weekend: a multi touch interface for our sand box. Please stay tuned!

For grads like me, who are running simulations everyday,  one CPU is far from enough. How can we make the best of  multiple cores in our machine?

For those who are more familiar with Matlab, parallel computing toolbox provides an extremely simple solution. Here is a simple way to use multiple cores in Matlab:

matlabpool(4) % creating 4 threads, which is the maximum cores on my machine.

% your code

matlabpool close

In this case, Matlab will launch 4 threads to run your code; if your code has parfor, it will allocate each thread a certain (toughly equal) amount of effort in that loop.

For those who really want to get something done in real-time,  here I provide the simplest way to use the cores that are idle for most of the time:

//make sure that your programming environment support openMP

#include “omp.h”

...

omp_set_num_threads(4); // again, creating 4 threads

#pragma omp parallel

{

unsigned int cpu_thread_id = omp_get_thread_num();

unsigned int num_cpu_threads = omp_get_num_threads();

// your code

}

In this case, you have to use the thread id to distinguish which specific code or data that thread will run. Well, if you meet any problem, feel free to contact me.  See you on Friday.

In preparation for the launch of the Video Geeks Club, I prepared a demo to demonstrate the power and flexibility of GPU programming using CUDA. I hacked a fluid simulation demo in CUDA SDK. It solves PDEs that model fluid dynamics in real time on a GPU (image on right). Then, I computed sparse optical flow using SIFT (scale-invariant feature transform), a robust feature extractor, and inter-frame SIFT-feature matching. The SIFT operator is a perfect tool for single-object matching, and other applications.

I used the flow vectors to animate the fluid. Since the complexity is high, all this was programmed in CUDA and executed on a CUDA-enabled GPU. On left is a screen shot from CUDA/GPU  SIFT calculation. The impact of optical-flow 'force' on the fluid is clearly visible in the fluid shot above. Once connected to a video camera, this is an impressive live demonstration. A hand motion in front of the camera permits fascinating interaction with fluid particles on the screen.

I would like to thank Changchang Wu from UNC for  SIFT library used in this project.