Mathematical morphological operators implementation in CUDA (parallel version), three versions are avaiable: check out branches. A report is included.

An implementation of morfological operators, in order to compare it with the sequential C++ version (here) and measure speedup.

Code works with .ppm B/W images, extracting only the first channel if multiple channels are detected. Results are also .ppm images, timings are always saved in a .csv file.

Regarding operators, multiple shapes are avaiable: you can even set a custom sizes. Check out StructuringElement.h.

MorphableOperators.h contains implementations of some well know tranformations: erosion, dilatation, opening, closing, top-hat and bottom-hat transform. Derived transformations were implemented calling base transformations methods.

Every version stores the structuring element in constant memory.

master: branch use shared memory for store locally each tile, compute the transformation on that tile and writes back in global memory. Tile's paddin is managed by two threads.

optimized-padding: works as above, except for padding's loading: this time all threads on the border work to load the padding.

no-shared-memory: pretty explicative, works without shared memory. Every read and write operation is done on global memory.

GPU timings were recorded on a Nvidia K80, CPU timings on a AMD Ryzen 1600; every GPU version was tested varing input tile's size from 8 to 32. 5 images from 400x400 5000x5000 were used for the test, below a table with three images and tile's size set to 16, full data can be found on the report.

Timings are in seconds.