8:30 AM - 9:15 AM
Room: Elwha Ballroom
Chair:
George Biros, University of Texas at Austin, U.S.
Over the past decade, the BLAS-like Library Instantiation Software (BLIS) project has carefully revisited past progress on how to structure the implementation of the level-3 BLAS-like operations (matrix-matrix computations) in particular and all basic linear algebra operations in general. The paper “The BLIS Framework: Experiments in Portability" demonstrates how a refactoring of prior approaches yields a more flexible, more easily maintained, highly portable, yet high-performing and scalable software library.
Casting level-3 BLAS functionality in terms of multiplication with submatrices was proposed in the works of Kågström, Ling, and Van Loan. This led to efforts to auto-generate and tune such as PHiPAC and ATLAS. A dramatic breakthrough came circa 2000 when Goto proposed “Goto's algorithm" (now at the heart of most high-performance BLAS) for implementing matrix-matrix multiplication.
BLIS casts Goto's algorithm in terms of five portable loops (written in C99) around a “microkernel" that updates a small submatrix of C that fits in registers. It is only this microkernel that needs to be customized for a new architecture when implementing matrix multiplication. The refactoring exposed in BLIS drastically reduced the size, complexity, and number of assembly kernels necessary for supporting high-performance across all datatypes and level-3 operations.
The prize-winning paper was coauthored with Tyler Smith, Bryan Marker, Tze Meng Low, Francisco Igual, Mikhael Smelyanskiy, Xianyi Zhang, Michael Kistler, Vernon Austel, John Gunnels, and Lee Killough.
Field G. Van Zee
Robert A. van de Geijn
The University of Texas at Austin, U.S.
| PP20 Home 2020 | Program | Speaker Index |