Dilation and Erosion filters in CUDA

Several implementations of the dilation and erosion filters are shown:

CPU:
1. Using separable filters. (erosionCPU.cpp)
GPU:
2. Naïve implementation, one thread loading n^2 elements per each pixel. (erosion.cu) 3. Separable filter implementation, the processing is divided in two steps and only 2*n elements are loaded per each pixel. (erosion.cu) 4. Shared memory implementation, a tiling approach is used. (erosion.cu) 5. The radio of the filter is templatized to enable unrolling of the main loop. (erosion.cu) 6. The filtering operation is templatized to reuse the same code for the erosion and the dilation. (erosionFuncTemplate.cu)

Performance

I have performed some tests on a Nvidia GTX 760.

With an image of 1280x1024 and a radio ranging from 2 to 15:

Radio / Implementation	Speed-up	CPU	Naïve	Separable	Shared mem.	Radio templatized	Filter op. templatized
2	34x	0.07057s	0.00263s	0.00213s	0.00209s	0.00207s	0.00207s
3	42x	0.08821s	0.00357s	0.00229s	0.00213s	0.00211s	0.00210s
4	48x	0.10283s	0.00465s	0.00240s	0.00213s	0.00221s	0.00213s
5	56x	0.12405s	0.00604s	0.00258s	0.00219s	0.00219s	0.00221s
10	85x	0.20183s	0.01663s	0.00335s	0.00234s	0.00237s	0.00237s
15	95x	0.26114s	0.03373s	0.00433s	0.00287s	0.00273s	0.00274s