Re: [Matlisp-users] Matlisp subset in pure CL

[Nicolas N. <Nic...@iw...>]

Nicolas Neuss <Nic...@iw...> writes:

> Example (P4, 2.4GHz), M+!-N adds two NxN-matrices:
> 
> MATLISP (2_0beta-2003-10-14, with ATLAS)
> 
> M+!-1: 0.27 MFLOPS
> M+!-2: 1.25 MFLOPS
> M+!-4: 4.99 MFLOPS
> M+!-8: 15.25 MFLOPS
> M+!-16: 71.39 MFLOPS
> M+!-32: 186.41 MFLOPS
> M+!-64: 353.20 MFLOPS
> M+!-128: 432.96 MFLOPS
> M+!-256: 78.03 MFLOPS
> M+!-512: 72.94 MFLOPS

And these numbers deserve more explanation.  They are measured by doing an
M+!-call repeatedly on matrices of size NxN.  The large values are achieved
if the method call is negligible and the matrices are already in the cache
from the previous call.  If M+! is applied to fresh data which is not in
the cache, the performance should remain below the 70-80 MFLOPS measured for
large matrices above.

Nicolas.