This series of examples walk you through several implementations of the same Ghost Exchange algorithm at varying stages of optimization. Our starting point is the original CPU-only implementation found at HPCTrainingExamples/MPI-examples/GhostExchange/GhostExchange_ArrayAssign/Orig. In this series, we use the HIP programming model for offloading compute to the GPU. For a good overview of the Ghost Exchange implementation, refer to the documentation here.
Brief descriptions of the various versions of the Ghost Exchange HIP implementation can be found below:
- Ver1: Shows a HIP implementation that uses the Managed memory model to port the code to GPUs using host allocated memory for work buffers and MPI communication
- Ver2: Shows the usage and advantages of using
roctxranges to get more easily readable profiling output from Omnitrace - Ver4: Explores heap-allocating communication buffers once on host
- Ver5: Explores unrolling a 2D array to a 1D array
- Ver6: Explores using explicit memory management directives to reduce data access latency
- Ver8: Under Construction, showcases overlap of compute and communication