November 14, 2018, Dallas, TX, USA
This BoF brought together the community focused on extending the Basic Linear Algebra Software (BLAS). The existing BLAS proved to be effective in assisting portable, efficient software for sequential and the current class of high-performance computers. We continued the investigation into the possibilities of extending the currently accepted standards to provide greater parallelism for small size operations, for reproducibility, and for reduced precision support. This was an open forum to discuss and formalize the details of the future standard and reference implementation.
During the BoF, the vendor representatives described their proposals in terms of hardware and mathematical software for their HPC systems. It was followed by various summary of standardization and specification reports from hardware and software vendors as well as the open source library and application developers on what they need in terms of numerical linear algebra software for today's and future systems. The authors of the Batched, Reproducible, and Reduced Precision BLAS presented the current proposal and various implementations (reference ones and more hardware-specific ones). A discussion on various aspects of the current state of the standard followed in addition to the planing of future activities. Interaction between the audience (the community) and the presenters took place at the end of the BoF.
| Presenter | Affiliation | Title | File |
|---|---|---|---|
| Piotr Luszczek | University of Tennessee | Introduction | Download |
| Cris Cecka | NVIDIA | MatMul CublasLt CUTLASS | Download |
| Jason Riedy | Georgia Tech | Updated Proposal for a Next-Generation BLAS | Download |
| Alexander Kalinkin | Intel | Enhancements: Just-in-Time Compilation, Packed GEMM APIs, and Integer GEMMs | Download |
| Siva Rajamanickam | Sandia National Laboratories | Batched Reproducible and Reduced Precision BLAS Forum - SC'18 | Download |