Performance and debugging

[lucaparisi91]

Base material at https://epcced.github.io/APT-CUDA/lectures/optimisation.html .

Possible list of topics below. Some of these might be in the practical information sheet instead of the practicals.

• parallelism : enough tasks to fill the GPU
• occupancy limiting factors ( stretch ? ): registers, shared memory, max number of blocks/threads per sm
• memory coalescing: have contiguous threads access contiguous elements
• branching: avoid different warps to take different paths in the code
• profiling tools ( mention ) : nvidia nsights compute & systems, rocprof , scalasca
• openmp specific tips: see Best practices for OpenMP #14 .
• Environment variables for debugging ( CRAY_ACC_DEBUG etc.. )