Hi Vince, Believe it or not, I actually ran into this problem on the SiCortex machines a year or two back. The issue was an error in one of the patches that crept into one of the .S (or .h) routines and it was directly related to the context switch code in arch/mips. I wish I could tell you where it was but my memory has faded. My recommendation is to do a diff against the arch/mips and look at the .S files. I seem to remember that it was one asm statement guarded by an #ifdef in the context switch code. Let me know if you don't have any luck...
Phil On Jan 18, 2009, at 11:24 PM, Vince Weaver wrote: > Hello > > I've encounterd an odd problem when using perfmon2 on an r12k MIPS > machine > (an SGI Octane). > > The spec2k sixtrack benchmark runs with normal results on the machine. > > However, if I use perfmon2 to collect CYCLES on the same executable, > the > reported floating point results are off, maybe by 5% from what they > should be. > > If I use perfmon2 to collect both CYCLES and INSTRUCTIONS_GRADUATED, > the > program will actually crash due to floating point errors. > > I'm at a bit of a loss as to why this could happen. Perfmon shouldn't > have any affect on floating point results, should it? I'm doing just > plain aggregate counts so the perfmon code should not really be > doing much > except handling overflows of the counters. > > The only thing I can think of is maybe the context-switch code is > somehow > modified in a way that affects the saving of FPU state. > > I'm going to investigate this some more myself, I was just curious if > anyone had any ideas of where to look in a strange case like this. > > Unfortunately I am running an older kernel (2.6.22) on this machine > due to > the SGI Octane hardware support not being in the mainline kernel > yet. So > this might be fixed in later versions of perfmon2. > > Thanks > > Vince > > ------------------------------------------------------------------------------ > This SF.net email is sponsored by: > SourcForge Community > SourceForge wants to tell your story. > http://p.sf.net/sfu/sf-spreadtheword > _______________________________________________ > perfmon2-devel mailing list > perfmon2-devel@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/perfmon2-devel ------------------------------------------------------------------------------ This SF.net email is sponsored by: SourcForge Community SourceForge wants to tell your story. http://p.sf.net/sfu/sf-spreadtheword _______________________________________________ perfmon2-devel mailing list perfmon2-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/perfmon2-devel
