GPU PROJECT DESCRIPTION
Group project.  Presentation starts: November 7, Monday, 2014 in class.
Report due: Novembe 24, Monday in class, hard copy.
[Abet Key words: 1-complexity (1, 2), understand inner workings of algorithms (2), Present to peer groups (8)]

Input: Take an image I of size (p x p) pixels.
Create an window W of (n x n) size, where n is an odd integer and n<<p.
Convolve the image I with W (in space domain, not in Fourier domain):
Resulting filtered image R = running window weighted average of I, with weights as in W.

Write a formula/algorithm for the averaging operation, and show me if you have any doubt.
Note that, normally you have to leave a band of (n/2) pixels all around I.
There are many ways to avoid it. However, if you choose to do that, be explicit in your algorithm.

Then re-write the algorithm for parallelization on GPU.

You will measure the timing for the convolution with CPU, and with GPU processors.

You will increase the size n of the window gradually, starting with n=3, and plot the timings against n,
and make observations on how GPU speeds up against CPU, with respect to n, from those studies.

A key aspect of the project is how you distribute portions of I on processors, to parallelize on GPU.
This may depend on your GPU's architecture. Your plots may reflect your distribtions choice, 
i.e. you may vary more parameters for your experiment than what I asked for above.

Clarify, double clarify, with me, so that you do not go off track with the project.

A side issue: apart from the size n, what values you are putting on W will decide what images you get back in R.
However, this will not affect the run times that is the purpose of the project.

Have fun!

PRESENTATION GUIDANCE 
(Prepare for about 10 minutes, it will become 20-30 minutes in class;
Each group member must present some part.)

Source code: describe

Demo: with two types of inputs: 
1. Very small matrices (say, of size 8x8 for image and 3x3 for filter),
where we can manually verify your algorithms product;
2. An input real image, and variable size filter

Slides: A small version of your report including
Algorithm, 
implementation issues, 
your CPU and GPU specs,
timings with varying input sizes, 
any other observations.