United Kingdom, January 22, 2018

This five-day continuing professional development (CPD) programme provides an introduction to programming GPUs for general computing tasks. On completion you will be able to: judge when and where to apply GPU computing; use existing development tools and frameworks to best advantage; adapt existing algorithms; understand performance issues.

Core content:

Introduction to general purpose GPU; the GPU architecture; CUDA threading model (threads, blocks, warps); CUDA synchronisation (intra-block and host-device); CUDA memory model (registers, global, and shared memory) and memory management API; performance considerations (memory coalescing, shared memory bank conflicts, warp divergence); running the Hello World example (kernel dispatch, CPU-GPU data transfers); examples from linear algebra, outlining: 2D thread blocks, reduction operations, shared memory, synchronisation, coalescing; GPU performance issues (compute capability, device occupancy, peak performance, performance measurement); development tools (Nsight, Visual profiler, Occupancy calculator, cuda-memcheck); GPU libraries (Thrust, CUBLAS, CUSPARSE, CUFFT); advanced optimisation techniques.