sgemm routine

[Kazushige Goto <goto statabo rim or jp>]

Hi,

I made a SGEMM assembler routine.  It runs about 910 to 920 MFlops
constantly(21164A, 600MHz, 2MB 3rd cache, DGEMM routine has not
been changed(about 780MFlops), though...).

   table 1.The ability of FORTRAN SGEMM routine
			(21164A 600MHz + 2MB L3 cache, MFLOPS)
	   ---------------------------------------------
	   Matrix Sizes | N x N | N x T | T x N | T x T
	   -------------+-------+-------+-------+-------
	     200 x  200 |  941  |  941  |  941  |  941
	     400 x  400 |  920  |  920  |  907  |  902
	     600 x  600 |  913  |  909  |  909  |  903
	    1000 x 1000 |  921  |  908  |  909  |  902
	    2000 x 2000 |  917  |  908  |  905  |  898
	    3000 x 3000 |  913  |  910  |  905  |  900
	    4000 x 4000 |  910  |  902  |  905  |  901
	   ---------------------------------------------
                    note: There are maximal value for several times.


This program is available for Fortran and C(original, but similar
interface of Fortran).

ftp://ftp.eni.co.jp/.2/Linux-Alpha-JP/ftp.statabo.rim.or.jp/BLAS/libgemm-980901.tar.gz

Thanks,
  goto@statabo.rim.or.jp