Matrix Multiply on the UltraSparc I - large matrix performance

This report shows how to get sustained 250-270 MFLOPs performance for large matrix multiply on an 170 MHz Ultra 1 (the Sun Performance Library 1.2 DGEMM() sustains 220-230 MFLOPS). This slide explains things more succintly.

UltraSparc BLAS (UBLAS)

• double precision routines only; compact code size
• !!! no error checking is performed !!!
• `rectangular routines' only
  • dgemm, dger, dgemv, dcopy, dswap, idamax, ddot, dnrm2, dasum
  A `DGEMM()-based Level 3 BLAS' implementation, tuned for the UltraSPARC, can be used to implement the triangular routines in terms of the UBLAS.
  • best results can be obtained if the effective cache size parameter is set to 256 KB

• for the Level 1 & 2 UBLAS (note: Sun Perf Lib 1.2 often performs relatively better than is indicated here for larger data sizes)
• 1000 x 1000 Matrix Factorization Performance
  • using DBLAS-based codes compiled for a single U170 processor using a blocking factor of 64.
    • uses only the `rectangular BLAS routines'
    • have rather high software overheads, and memory access patterns may not be optimal
  • results in MFLOPS (Sun Perf Lib 1.2 in ()'s)
    • LU: 186 (176), LLT: 198 (131), QR: 207 (175)

1. first obtain a version of the BLAS (eg. from netlib), link that into your application and THOROUGLY test your application.
2. the re-link your application putting the file UBlas.a archive *before* the other BLAS archive.
3. re-run your application!

1. As this is the alpha release of the UBLAS, anyone using these codes does so entirely at their own risk, and should *not* assume that the code is free of bugs, major or minor.
2. The author accepts *no* responsibility for any loss or misfortune of others as a result of using these codes.