We decided to parallelize the Image processor written by Christopher Ginac. It is a serial C++ program that supports many functions, such as: negation, rotation, translation, flipping and more. An image processor is perfect to apply a parallel solution to, as it naturally contains a matrix of data. So let's get started!

The first problem we had was with the C++ program itself. Paying respect to the mighty C++, the author of the original code used a class based implementation to store the data members and methods for his image processor. Normally, this would be great! However, with our Cuda implementation, it caused a rather difficult problem. We could not access the private data members within the kernel. Thinking on this problem for a little while, we came up with a rather simple solution! We would copy the data into temporary variable, which we would then pass to the device. This would have worked wonderfully, if we didn't hit road block number 2.

The author stored the pixels in a 2 dimensional array (int ** pixelVal[rows][cols]). In terms of a serial program and in regards to how an image is represented, this makes perfect sense. However, copying that array to the device was a difficult problem that we did not know how to solve. Luckily, we realized that we could serialize this 2D array into a 1D array in Row Major format. Then, our kernel could receive it as it does any other array and all our problems would be solved! Success!

Once we got passed Road Block Number 2, we realized that we wouldn't try to parallelize functions that required row or column manipulation. Therefore, we decided to parallelize 'negation' - that is, inverting each pixels colour. Although serializing the data before and after the kernel call adds two more double loops to the code, we are only interested in the negation portion. The rest, we can optimize in Assignment 3. So, we time-stamped before the kernel call and subtracted that time from the current time after the call finishes to get the 'parallel' time. This would be exactly analogous to the serial program, as it only had one loop. Let's get started!