Install Ginkgo library to run the benchmark suite

fnrizzi / kokkos-kernels

Kokkos C++ Performance Portability Programming EcoSystem: Math Kernels - Provides BLAS, Sparse BLAS and Graph Kernels

Other

0 stars 0 forks source link

Install Ginkgo library to run the benchmark suite #3

Open mzuzek opened 3 years ago

mzuzek commented 3 years ago

https://ginkgo-project.github.io

https://github.com/ginkgo-project/ginkgo/blob/develop/BENCHMARKING.md

Reference paper

Asignee: @MikolajZuzek

mzuzek commented 3 years ago

@uhetmaniuk @fnrizzi

Got first results from Ginkgo benchmark on my system: Ginkgo-benchmark_NGA-MZ-P2200.zip

fnrizzi commented 3 years ago

@MikolajZuzek i would suggest also running on the host too

uhetmaniuk commented 3 years ago

@MikolajZuzek

It looks like fbcrs (the block CRS format) is not used.
Can you confirm that this is not due to a compilation flag? If fbcrs is never used, can you select for the moment a few ( < 5 ) structural problems (look for "kind" = "structural problem") ?

For 3D structural problems, we should be able to store the matrix in BlockCrsMatrix with blocks of size 3 x 3.

mzuzek commented 3 years ago

I can't find fbcrs format mentioned anywhere in Ginkgo repository...

uhetmaniuk commented 3 years ago

Definition is here https://github.com/ginkgo-project/ginkgo/blob/develop/include/ginkgo/core/matrix/fbcsr.hpp

Implementation details are partly here https://github.com/ginkgo-project/ginkgo/blob/develop/core/matrix/fbcsr.cpp

uhetmaniuk commented 3 years ago

Let us focus on these matrices for the moment

They were discussed in this paper.

1366653.pdf

mzuzek commented 3 years ago

@uhetmaniuk

Thanks! Unfortunately, fbcsr is not connected to the benchmark in Ginkgo (yet?): check benchmark/utils/formats.hpp61:81 for all available formats). I found no related issue.

Here are my Ginkgo benchmark results (CUDA and OpenMP) for our six matrices: ginkgo_5mtx_cuda+omp.zip