Challenge: How to implement a fully-parallelized Jacobi Iteration to support a real-time image fusion?
- Workload/constrains: similar to the 2d-grid example demonstrated in class.
Resources:
- Codebase: https://github.com/Trinkle23897/DIP2018/blob/master/1/image_fusion/image_fusion.cpp, written by Jiayi Weng, one of our group members
- Computation Resources: GHC (CPU, GPU - 2080), PSC (CPU, GPU), laptop (CPU, GPU - 1060)
Goals:
- 75%: implement one parallel version of PIE
- 100%: benchmark the algorithm with OpenMP/MPI/CUDA implementation
- 125%: include a interactive python frontend that can demonstrate the result in a user-friendly style. Real-time computation on a single GTX-1060 Nvidia GPU.
Platform choice:
- OS: Linux, Ubuntu machine
- Language: C++/CUDA for core development, Python for interactive frontend
Schedule:
- 3.28 - 4.9: implement parallel version of PIE (OpenMP, MPI, CUDA)
- 4.11 - 4.22: benchmarking, optimizing, write Python interactive frontend
- 4.25 - 5.5: write report
In Apr. 11th, we have finished all implementations for parallel PIE algorithm. Each implementation contains two different solver: EquSolver and GridSolver. We also create two datasets: one for real image demonstration, another one for benchmark performance.
Most of the results are in the homepage README file. We are going to polish the documentation and running benchmark experiments this week. Then do some ablation study and write report in the next week. If we have extra time, we will implement an interactive frontend for creating mask image.
At the poster session we plan to show a live demo of this project, and on the other hand, briefly describe the benchmark result to demonstrate its high-performance.
See report.