single-gpu-performance

Single GPU performance aspects

In this short exercise we illustrate the single most important performance aspect of GPU usage, i.e. data copies between CPU (host) and GPU (device) memories.

The short program jacobi.F90 solves the the dimensional Laplace equation numerically with the so called Jacobi iteration. The key component of the numerical algorithm is stencil update, which is common operation in numerical methods.

In the algorithm, one performs an iteration where at each iteration the numerical solution is updated. The provided Makefile builds three different versions of the code:

• jacobi_cpu : a pure CPU version
• jacobi_implicit : a GU version where data is copied between the host and device at each iteration.
• jacobi_explicit : a GPU version where data copied only in the beginning and in the end of iterations.

Run the different versions (with single GPU or CPU core) and investigate their performance characteristics. You may also have a look on the source code to get a feel of simple OpenMP offloading code.