BLAS Based on Block Data Structures
Greg Henry · eCommons (Cornell University) · 1992
The optimization of the BLAS is discussed, with examples given for the IBM superscalar RISC S/6000. The approach suggested is to use block data structures based on store-by-block schemes. We give results and analysis of the optimization of DGEMM. We also suggest how these results can be applied to the higher level factorizations and the other BLAS. Results are given to show the advantages of using block data structures.