Stencil Construction Profiling - Speed up with GPU?

[JanGaertner]

Profiling Stencil Collection

The stencil collection process during the construction of the WENOBase class is quite expensive and takes a lot of time. Therefore, to improve start-up times the construction of the WENOBase class is profiled with the AMDuProf version 3.5.671.0 in single core on a AMD Ryzen 5 PRO 3500U.

The hottest function is the calculation of the Gauss quadrature in

WENOEXT/libWENOEXT/WENOBase/geometryWENO/geometryWENO.C

Line 346 in f45593a

Foam::scalar Foam::geometryWENO::gaussQuad

Even though it is already attempted to improve performance using manual AVX instructions, the calculation of the power still takes a long time. That the vector operations are the ones taking up most of the time can also be seen in the assembly code analysis of the profiler.

Possibly the calculation of the quadrature can be reformulated as a single matrix multiplication or as multiple matrix multiplication which could be executed on a GPU to improve performance.

Speed up with GPU

• Use openACC to quickly test if GPU operations would improve performance or if copying the data of the small matrices back and forth lead to only a small performance increase or even decrease
• Use CUDA to manually control the operation, possibly even rewrite the transformIntegral function which uses the Guass quadrature functions