cuda

cuda programming exercise

1. Naive Matrix Multiplication
   • For a 32*32 matrix multiplication with float numbers, elapsed time on Host is 0.000131s
   • Elapsed time on Device is 0.000019s if run with 32*32 threads
   • Size of matrix limited by the number of threads allowed in a thread block, which is 1024 with CUDA toolkit 10
2. Advanced Matrix Multiplication
   • Split the matrix into tiles, with each tile assigned to a block
   • Each tile can access the shared memory instead of accessing the global memory directly
   • For a 10241024 matrix with the tile size of 3232, it takes 14.500724s on the host, and 0.000022s on the device
   • Below is a nsight profile screenshot
   • 
3. Flocking Simulation
   • Based on the Reynolds Boids algorithm
   • With two levels of optimization: a uniform grid, and a uniform grid with semi-coherent memory access
   • Below are some results with 1k, 10k and 100k boids(particles)
     • 
     • 
     •