Exercies on how to use Nvidia GPUs using CUDA and learn optimization methods specific for GPU-parallelized computing.
In each chapter folder, there is a readme.md file explaining on how to build and run the template files for each lab.
- Use CUDA APIs to implement vector addition
- Learn basic transfer of data between CPU and GPU and memory allocation
- Implement tiled dense matrix multiplication using CUDA
- Learn how to allocate memory on GPU and transfer data between GPU and CPU
- Use shared memory to optimize computation and memory latency
- Find the difference of performance on usage of shared memory
- Use pinned memory with CUDA streams by implementing vector addition
- Benchmark the performance on the usage of CUDA streams by hiding memory latency
- Apply convolution to a ppm image using CUDA APIs
- Find overhead in using output tiling algorithm
- Evaluate performance of output tiling algorithm with different tiling sizes including extreme ones
- Perform Histogram Reduction using CUDA APIs
- Learn how to use atomics for memory address in GPU through CUDA memory APIs
- Analyze performance impact of using atomics
- Implement 1D inclusive parallel scan using CUDA and work-efficient algorithm (Brent-Kung)
- Learn the restriction of the algorithm based on the commutative feature of binary operator for scan
- Implement sparse matrix-vector (SPMV) mutiplication using CUDA and a transposed JDS (Jagged Diagonal Sparse) formatted matrix
- Implement the conversion from 2D array to JDS-formatted matrix using C++ and STL vector
- Compare the performance difference of the kernel between the version using shared memory and not using shared memory
- Compare the performance between the serial (CPU) and parallel verions of Quick Hull algorithm for Convex Hull problem.
- Introduced parallelized sorting, reduction, and computation to reduce the runtime of the algorithm
- Check submodule for more information