Hi Steven,
Yes, you're correct. For completeness, here are the results for the
coarser resolution. Indeed, it seems that a higher resolution reduces
the epsilon discrepancies.
20ppw [C++ finds one duplicate for a total of 12 points, libctl got 11
points. Duplicate was manually removed]
libctl C++ Pct Diff
a b abs(a-b)/a*100%
0.46600 0.46937 0.722231
0.63847 0.63490 0.558944
0.70052 0.69812 0.343073
0.70365 0.70135 0.327254
0.81999 0.81604 0.481117
0.86852 0.86694 0.181721
0.89094 0.89110 0.017461
0.94510 0.94228 0.298920
0.98007 0.97293 0.728422
0.98870 0.98235 0.641928
0.99296 0.99775 0.482598
80ppw (from previous email)
libctl C++ Pct Diff
a b abs(a-b)/a*100%
0.46864 0.46867 0.0080870
0.63829 0.64327 0.7806107
0.64429 0.64847 0.6479023
0.70878 0.70899 0.0293657
0.83324 0.71024 14.7610847
0.83853 0.83288 0.6739258
0.88855 0.88819 0.0403651
0.91237 0.91270 0.0363838
0.95640 0.95648 0.0079315
0.96689 0.96746 0.0589339
0.99507 0.99884 0.3790904
Perhaps a better way to check this is to examine the number of epsilon
discrepancies using h5diff. (eg., 'h5diff ctl.20.h5 cpp.20.h5 -d .1')
for 20ppw, the data set is 20x240 = 4800 points.
for 80ppw, the data set is 80x960 = 76800 points.
20ppw %age of 4800 80ppw %age of 76800
no -d 192 4 % 784 1.02%
-d .1 80 1.67% 348 0.45%
-d 1 24 0.5% 68 0.09%
For this particular example, increasing the resolution by a factor of
four reduced the normalized discrepancy count between the cpp and ctl
versions approximately by a factor of four.
Best Regards,
Matt
On Tue, 17 Oct 2006, Steven G. Johnson wrote:
On Tue, 17 Oct 2006, matt wrote:
Increasing the resolution should minimize the errors you describe, so
let's try changing the resolution of the example from 20 to 80.
Note that the comparison you posted was incomplete, because you didn't show
the differences for a resolution of 20 to see whether 80 was indeed smaller.
_______________________________________________
meep-discuss mailing list
[email protected]
http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/meep-discuss
_______________________________________________
meep-discuss mailing list
[email protected]
http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/meep-discuss
