Abstract | ||
---|---|---|
The HPGMG benchmark is a non-trivial Multigrid benchmark used to evaluate system performance. We ported this benchmark from CUDA to OpenMP target offload and added the capability to use explicit data management rather than managed memory. Our optimized OpenMP target offload implementation obtains a performance of 0.73x and 2.04x versus the baseline CUDA version on two different node architectures with NVIDIA Volta GPUs. We explain how we successfully used OpenMP target offload, including the code refactoring required, and how we improved upon our initial performance with LLVM/Clang by 97x. |
Year | DOI | Venue |
---|---|---|
2020 | 10.1007/978-3-030-58144-2_3 | IWOMP |
DocType | Citations | PageRank |
Conference | 1 | 0.37 |
References | Authors | |
0 | 4 |
Name | Order | Citations | PageRank |
---|---|---|---|
Christopher S. Daley | 1 | 1 | 0.37 |
Hadia Ahmed | 2 | 1 | 0.37 |
Samuel Williams | 3 | 1282 | 98.56 |
Nicholas J. Wright | 4 | 408 | 27.79 |