Gaussian-Blur-CUDA

• We implement Gaussian Blur with CUDA
• We used different size image to compare their performance

Result

Before	After

Experiment

• CPU: Intel(R) Xeon(R) CPU X5670 @ 2.93GHz
• GPU: GEFORCE GTX 1080 Ti

Computation Performance Comparison between Sequential, OpenMP, CUDA

• Filter Matrix size: 53*53
• Multiple threads: 8
• Multiple GPU number: 2
• V3 version sequential implementation

Filename	Size	Single thread	Multiple threads	Single GPU	Multiple GPU
iceberg.png	0.8 MB	12.6372(s)	1.44974(s)	0.528766(s)	0.204902(s)
mountain.png	1.9 MB	23.9481(s)	2.5741(s)	0.594938(s)	0.238576(s)
bridge.png	2.1 MB	35.1401(s)	3.69399(s)	0.647862(s)	0.369114(s)
view.png	2.4 MB	46.6573(s)	4.82201(s)	0.701165(s)	0.300363(s)
candy.png	13.9 MB	362.181(s)	36.6829(s)	2.43379(s)	1.1862(s)

Time Profile	Speedup Factor

GPU performance on Big Size Data

Filename	Size	One GPU	Two GPU
candy.png	13.9 MB	2.43379(s)	1.1862(s)
jerry.png	27.6 MB	8.07605(s)	4.20553(s)
large-candy.png	79 MB	21.5988 (s)	11.0503 (s)

Time Profile	Speedup Factor

Acknowledgments

Our Gaussian Blur implementation is mainly based on the following two website.

• Gaussian Blur introduction
• Gaussian Blur implementation