Project 2: Vismay Churiwala

[vismaychuriwala]

• https://github.com/vismaychuriwala/CUDA-Stream-Compaction
• Here are some cool features I have implemented:
  • Using Shared Memory (SM) for efficient memory access on the GPU (as opposed to using global memory, which is more than 100× slower.
  • Hardware optimization via bank conflicts prevention.
  • Recursive scanning to scan arrays of arbitrary sizes (tested up to 1B elements (2^30), which took 3470.59 ms).
  • Radix Sort using Parallel Scan
  • Naive and CPU-based implementations to compare and benchmark techniques.
  • Customized testing code to collect, average, and plot GPU and CPU timings.
• Feedback : A tip about how the main crux of the assignment is multi-block compaction would have been good, I thought I had finished this assignment days ago, until the last day when I realized that my code doesn't work on large arrays. I guess I could have known, it just didn't occur to me. Also, making SM compulsory isn't too bad, can do that maybe.