Repository for Performance Benchmarks for OpenMP Features

This is a repository for benchmarks for performance experimentation of features of OpenMP 5.0 for United States Department of Energy's Exascale Computing Project's (ECP's) OpenMP project called SOLLVE. See sollve.github.io for more information on the SOLLVE project.

The benchmarks are based on SOLLVE's efforts to undertand performance of OpenMP features, particularly OpenMP offload features. These benchmarks have been experimented on various systems as shown in the table below.

System	LLVM 11 - Summit	LLVM 11 - Cori-GPU	LLVM 11 - Theta	Spock
Stencil - 1D	tested, works	tested, works	tested, works
Stencil - 2D	tested, works	tested, works	tested, works
MatMul	tested, works	tested, works	tested, works
Load Imbalanced MatMul	tested, works	tested	tested, works
Dot Product	tested, works	tested	tested, works
Stream	tested
Square Rooted Vec Prod	tested, works	tested, work	tested, works

System	LLVM 12 - Summit	LLVM 12 - Cori-GPU	LLVM 12 - Theta	Spock
Stencil - 1D	tested, works	tested, works	tested, works
Stencil - 2D	tested, works	tested, works	tested, works
MatMul	tested, works	tested, works	tested, works
Dot Product	tested, works	tested	tested, works
Stream	tested
Square Rooted Vec Prod	tested, works	tested, work	tested, works
gauss-seidel	tested, works	tested, works	tested, works
lu factorization	tested, works	tested, works	tested, works	tested, works
spmv	tested, works	tested, works	tested, works	tested, works
quicksort	tested, works	tested, works	tested, works	tested, works
barnes-hut	tested, works	tested, works	tested, works	tested, works
Load Imbalanced MatMul	tested, works	tested	tested, works

System	LLVM 12 - Summit	LLVM 12 - Cori-GPU	LLVM 12 - Theta	Spock
Stencil - 1D	tested, works	tested, works	tested, works	tested, works
Stencil - 2D	tested, works	tested, works	tested, works	tested, works
MatMul	tested, works	tested, works	tested, works	tested, works
Dot Product	tested, works	tested	tested, works	tested, works
Stream	tested
Square Rooted Vec Prod	tested, works	tested, works	tested, works
gauss-seidel	tested, works	tested, works	tested, works
lu factorization	tested, works	tested, works	tested, works	tested, works
spmv	tested, works	tested, works	tested, works	tested, works
quicksort	tested, works	tested, works	tested, works	tested, works
barnes-hut	tested, works	tested, works	tested, works	tested, works
Load Imbalanced MatMul	tested, works	tested	tested, works	tested, works

Each benchmark uses a subset of OpenMP offload features. The stencil, matmul, dot product, square rooted vec prod use a target offload construct. The load imbalanced matrix multiplication uses a device clause with the target construct. The stream benchmark uses an experimental features of target spread for multi-GPUs along with the target construct and devices clause.