Project 2: Xuntong Liang

[PacosLelouch]

• Repo Link
• I implemented algorithms include scan, stream compaction, and radix sort. I completed a performance analysis and optimizations.
  • Implemented scan with CPU's implementation, Naive, Work-Efficient, and a shared memory optimization version Work-Efficient-Hierarchical, and compared with thrust::exclusive_scan.
  • Implemented stream compaction with CPU's implementations, Work-Efficient, and a shared memory optimization version Work-Efficient-Hierarchical, and compared with thrust::remove_if.
  • Implemented sort with shared memory optimization Work-Efficient-Hierarchical, and compared with CPU's std::sort and thrust::sort.
  • Completed a performance analysis and guessed what might be happening inside the Thrust's implementations.
• Feedback: It is challenging to optimize the GPU programs with shared memory with the thought of accessing the memory parallel for me, but it is interesting and meaningful. There are few documents and blogs talking about the bank conflict. I think I should take some time to make a clear understanding on it to write better GPU programs. I am also very confused about what happens inside the Thrust's function. I believe the authors use some techniques and optimized for a long time, and maybe there are some magic functions called by them?