Hi,
I think I've narrowed the problem with your hints, Gabe.
I executed M5 with traces of instructions and it triggered a warning of
unimplemented instruction emms after calling the printf function.
However, it seems that if the printf is not called, it does not carry
the error. So, an approach to avoid using emms is not to print data to
check if the result is correct, but by other methods.
Anyway, I'll test this with a more serious application and check if we
can survive at the moment without the emms instruction.
Best regards,
Cecilia
Quoting Gabriel Michael Black <[email protected]>:
There's a good chance this is a bug in M5's floating point support.
x87 is almost unimplemented, but 64 bit compilers should, generally
speaking, use sse since that's pretty much guaranteed to be available
on anything that has 64 bit mode. SSE is pretty well supported, but
there could be bugs. Does M5 print any warnings about unimplemented
instructions? What compiler flags are you using? If you use fancy,
really new SSE instructions they're less likely to be implemented.
I'll try this myself at some point soon, although I'm going to be
really busy for the next week or two.
If you want to try fixing this yourself, you should check out
http://www.m5sim.org/wiki/index.php/Debugging_M5 for general M5
debugging tools and techniques. One tool not mentioned there, as far
as I know, is called statetrace. There are many nuances to how to use
it which I won't try to explain here, but if you want to try to use it
let me know and I'll try to find some time to put some information on
the wiki.
Gabe
Quoting [email protected]:
Hello everybody,
I'm executing some bioinformatics applications within the x86_SE
mode of M5. I've found some problems on the execution of
applications that use floating point arithmetic. I illustrate this
with an example:
The program:
#include<stdio.h>
int main (void)
{
double self_energy, self_energy2;
double intra, tmp;
intra = 9336.438188;
self_energy = 0.0;
self_energy2 = 6;
tmp = 64.0;
printf ("[ewald] se = %lf\n", self_energy);
printf ("[ewald] intra = %lf\n", intra);
self_energy = tmp * intra;
printf ("[ewald] se = %lf\n", self_energy);
printf ("[ewald] se2 = %lf\n", self_energy2);
return 0;
}
Compiled statically, as a x86-64 binary, and executed on the native
machine, the output of the program:
[ewald] se = 0.000000
[ewald] intra = 9336.438188
[ewald] se = 597532.044032
[ewald] se2 = 6.000000
However, in M5, its output is:
[ewald] se = 0.000000
[ewald] intra = 0000.000008
[ewald] se = 020802.000504
[ewald] se2 = 6.000000
I had tried that with float type instead of double and I got the same.
So, there is some problem with floating point numbers bigger than a
certain figure (which I couldn't find out because sometimes it works
well with, for instance, 300.0 and sometimes not, it depends on the
program).
I really appreciate any suggestion on the problem.
Kind regards,
Cecilia
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
