dimimal / high_performance_computing

Experiments with Image processing algorithms with CUDA

High Performance Computing with CUDA

Tiled and Non tiled convolution algorithms comparison in CUDA and histogram equalization technique for image processing.

Convolutions

Tiled vs Non-tiled Version

None tiled Version

make ./exec_non_tiled

Output:

Enter filter radius : 32
Enter image size. Should be a power of two and greater than 65 : 6400
Image Width x Height = 6400 x 6400

Allocating and initializing host arrays...
CPU computation...
CPU time: 7.05165 seconds

GPU computation...
GPU time: 141.062302 ms.

Tiled Convolution

./tiled_conv

Enter image size. Should be a power of two and greater than 33 : 6400
Image Width x Height = 6400 x 6400

Allocating and initializing host arrays...
CPU computation...
CPU time =    3.54343 seconds
GPU computation...
GPU time: 58.340351 ms.

The results have been taken on a different machine than the results above but they are proportional.

Histogram Equalization

In Histogram Equalization folder run make to create the object file.

Pass an input image to equalize along with its output file you want to save into

./exec input.pgm output.pgm

Check on Histogram_Equalization/images folder for the results of histogram equalization applied on images