A starter CUDA optimization project. It exercises many critical paths like reducing global memory loads, sizing grids appropriately, evaluating performance of scattered writes and so on. It is a simple algorithm that presents a rich surface for optimization.

Run run.sh. It won't work unless you have nvcc installed in a specific place.

• learn about grid.sync() in cooperative kernel launch
• reading and writing the array best[] in a kernel with different indices requires some synchronization coordination. Seeing wrong results because of this.
• kernel launch is fast; but it is noticeable if you launch 100K times.

See results folder.