Subject: Re: [PATCH 18/18] 2.6.17.9 perfmon2 patch for review: new x86_64 files - msg#00052

Previous Message by Date: click to view message preview

Re: [PATCH 18/18] 2.6.17.9 perfmon2 patch for review: new x86_64 files

On Wednesday 23 August 2006 14:14, Stephane Eranian wrote: [adding discuss@xxxxxxxxxx so that possibly AMD people can comment] > On Wed, Aug 23, 2006 at 01:22:44PM +0200, Andi Kleen wrote: > > > > > I have a second thought on this. AMD has architected the performance > > > counters. > > > > Quote: > > >> > > Implementations are not required to support the performance > > c o u n t e rs and the event-select registers, or the time-stamp > > counter. The presence of these features can be determined by > > << > > > At the end of this paragraph then mention using CPUID to determine > the presence of the counters. AFAIK, there is no feature bit > covering performance monitoring. Does that mean we are left > with having to check the family and model number just like on > Intel? Yes I puzzled over that too. Maybe they meant the MSR CPUID bits, but most likely it was a mistake by the tech writer. Yes I think you have to. Only checking vendor/family should be fine though -- i am not aware of performance counter variations between models. Perhaps add a force argument again that disables the family check too. > Ok, I think I understand now: > 1/ Bios and Kernel Developer Guide from Ahtlon64 and Opteron 64 is > what you are talking about with K7/K8 Well K8. K7 has a different one. But ok. I think you don't try to support K7 at all currently (it has the same register format as K8, but the list of counters is different) > 2/ AMD64 Architecture Programmer's Manual is the generic AMD64 > description Yep -Andi

Next Message by Date: click to view message preview

Re: [PATCH 18/18] 2.6.17.9 perfmon2 patch for review: new x86_64 files

> If they have variations it could be with > the low power (laptop) models where counters may not be present at all. It would surprise me if they ever released anything again with no counters. It's mainly simulators where the counters are missing (e.g. SimNow is a pretty accurate simulation otherwise, but doesn't have performance counters) > > > Perhaps add a force argument again that disables the family check too. > > > > > Ok, I think I understand now: > > > 1/ Bios and Kernel Developer Guide from Ahtlon64 and Opteron 64 is > > > what you are talking about with K7/K8 > > > > Well K8. > > > > K7 has a different one. But ok. I think you don't try to support K7 at all > > currently (it has the same register format as K8, but the list of counters > > is different) > > > What is the "commercial name" of K7 processors? Athlon/Athlon XP/Athlon MP/Sempron (early models, later are K8)/Duron and some of the Geodes are K7 based too. -Andi

Previous Message by Thread: click to view message preview

Re: [PATCH 18/18] 2.6.17.9 perfmon2 patch for review: new x86_64 files

On Wednesday 23 August 2006 14:14, Stephane Eranian wrote: [adding discuss@xxxxxxxxxx so that possibly AMD people can comment] > On Wed, Aug 23, 2006 at 01:22:44PM +0200, Andi Kleen wrote: > > > > > I have a second thought on this. AMD has architected the performance > > > counters. > > > > Quote: > > >> > > Implementations are not required to support the performance > > c o u n t e rs and the event-select registers, or the time-stamp > > counter. The presence of these features can be determined by > > << > > > At the end of this paragraph then mention using CPUID to determine > the presence of the counters. AFAIK, there is no feature bit > covering performance monitoring. Does that mean we are left > with having to check the family and model number just like on > Intel? Yes I puzzled over that too. Maybe they meant the MSR CPUID bits, but most likely it was a mistake by the tech writer. Yes I think you have to. Only checking vendor/family should be fine though -- i am not aware of performance counter variations between models. Perhaps add a force argument again that disables the family check too. > Ok, I think I understand now: > 1/ Bios and Kernel Developer Guide from Ahtlon64 and Opteron 64 is > what you are talking about with K7/K8 Well K8. K7 has a different one. But ok. I think you don't try to support K7 at all currently (it has the same register format as K8, but the list of counters is different) > 2/ AMD64 Architecture Programmer's Manual is the generic AMD64 > description Yep -Andi

Next Message by Thread: click to view message preview

Re: [PATCH 18/18] 2.6.17.9 perfmon2 patch for review: new x86_64 files

> If they have variations it could be with > the low power (laptop) models where counters may not be present at all. It would surprise me if they ever released anything again with no counters. It's mainly simulators where the counters are missing (e.g. SimNow is a pretty accurate simulation otherwise, but doesn't have performance counters) > > > Perhaps add a force argument again that disables the family check too. > > > > > Ok, I think I understand now: > > > 1/ Bios and Kernel Developer Guide from Ahtlon64 and Opteron 64 is > > > what you are talking about with K7/K8 > > > > Well K8. > > > > K7 has a different one. But ok. I think you don't try to support K7 at all > > currently (it has the same register format as K8, but the list of counters > > is different) > > > What is the "commercial name" of K7 processors? Athlon/Athlon XP/Athlon MP/Sempron (early models, later are K8)/Duron and some of the Geodes are K7 based too. -Andi