🚀 go-pugleaf

In article <EIJKHOUT.93Jan23164314@cupid.cs.utk.edu> eijkhout@cupid.cs.utk.edu (Victor Eijkhout) writes:
>
>Assembly coded benchmarks, maybe. But the BLAS routines (Basic Linear
>Algebra Subroutines) exist on many machines in assembler, and
>precisely because they are 1/ standardized across all platforms
>2/ optimized for each, you can write real programs in Fortran or C,
>with a reasonable speed if you need the BLAS often enough.

This is a good idea, but as I pointed out, you will not get more than
11 or 12 mflops on daxpy if you stick to Blas-1 (unless you have some
kind of special daxpy [ maybe KAI? ] that assumes that vector 1 stays
in cache and vector 2 comes from main memory, or vice versa).

>
>Just interface your program to assembler *kernels*.
>
>I know of people who write distributed memory linear system
>solvers, that get over 50% performance out of *500 i860's*
>using those assembler BLAS.

These people must be using the Delta, huh?  I don't know anyone else
that has 500 i860s. Are they calling blas-2
routines?  If so, I'd definitely believe you, as I know people who get
90%+ performance on 80 or so shared memory i860-XRs using similar
techniques.
-- 
Larry Meadows		The Portland Group
lfm@pgroup.com