uBLA performance

From: Matt Hurd (matt_at_[hidden])
Date: 2003-06-06 19:12:37

Hi,

Ran off a simple test of the uBLA performance for sqare matrix
multiplication.  Looks quite good.  Put in Intel’s Math Kernel Library
for good measure using uBLA as the data holder but with a call to dgemm.
 
Do you think these are the kind of numbers I should be expecting?

Numbers below.  Added a chart to the files section
http://groups.yahoo.com/group/Boost-Users/files/  called “uBLA
Performance.pdf”
 
Used head of CVS from today.
 
Regards,
 
Matthew Hurd

___________________________________________
Figures in the raw:

Matrix Size	 
    intel7.1 	 
            vc7.0
                 mingw
                       intel_mkl6.0_vc7.0 
___________________________________________

1	 2 	 2 	 2 	 1 
2	 13 	 12 	 13 	 11 
3	 32 	 30 	 34 	 31 
4	 51 	 46 	 59 	 70 
5	 67 	 59 	 85 	 109 
7	 70 	 78 	 92 	 180 
9	 85 	 94 	 113 	 262 
12	 99 	 115 	 139 	 304 
16	 110 	 131 	 159 	 362 
20	 118 	 127 	 178 	 355 
26	 125 	 135 	 196 	 334 
33	 132 	 148 	 209 	 390 
42	 134 	 158 	 217 	 389 
53	 130 	 163 	 228 	 451 
67	 133 	 171 	 233 	 459 
84	 138 	 177 	 241 	 467 
105	 133 	 176 	 239 	 473 
132	 133 	 178 	 241 	 468 
166	 106 	 174 	 233 	 448 
208	 90 	 148 	 196 	 445 
261	 40 	 63 	 63 	 427 
327	 24 	 43 	 45 	 434 
409	 24 	 44 	 45 	 421 
512	 11 	 12 	 12 	 404 
640	 12 	 12 	 12 	 459 
800	 36 	 38 	 38 	 455 
1000	 36 	 37 	 38 	 448
___________________________________________