[Nouveau] [PATCH] drm/nvc0/pm: read core/shader/memory clocks (more) correctly

[Martin Peres]

Le 18/04/2011 00:56, Ben Skeggs a écrit :
> On Mon, 2011-04-18 at 00:54 +0200, Martin Peres wrote:
>> Le 18/04/2011 00:32, Ben Skeggs a écrit :
>>> On Sun, 2011-04-17 at 17:10 +0200, Martin Peres wrote:
>>>> ---
>>>>    drivers/gpu/drm/nouveau/nouveau_perf.c |    6 +++---
>>>>    1 files changed, 3 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/nouveau/nouveau_perf.c b/drivers/gpu/drm/nouveau/nouveau_perf.c
>>>> index 950caba..d64a98a 100644
>>>> --- a/drivers/gpu/drm/nouveau/nouveau_perf.c
>>>> +++ b/drivers/gpu/drm/nouveau/nouveau_perf.c
>>>> @@ -177,9 +177,9 @@ nouveau_perf_init(struct drm_device *dev)
>>>>    #define subent(n) entry[perf[2] + ((n) * perf[3])]
>>>>    			perflvl->fanspeed = 0; /*XXX*/
>>>>    			perflvl->voltage = entry[2];
>>>> -			perflvl->core = (ROM16(subent(0))&   0xfff) * 1000;
>>>> -			perflvl->shader = (ROM16(subent(1))&   0xfff) * 1000;
>>>> -			perflvl->memory = (ROM16(subent(2))&   0xfff) * 1000;
>>>> +			perflvl->core = (ROM16(subent(3))&   0xfff) * 1000;
>>>> +			perflvl->shader = (ROM16(subent(3))&   0xfff) / 2 * 1000;
>>>> +			perflvl->memory = (ROM16(subent(5))&   0xfff) * 1000;
>>> Um, NACK.  This is *definitely* very very wrong on NVA8, which uses
>>> version 0x40 of this table too.
>>>
>>> Ben.
>> Crap, I would have sworn this table was introduced for Fermi! Sorry for
>> the noise then, I'll try to find a fix for that (other than looking at
>> the codename).
> I'm highly skeptical that it's correct on fermi too, but, I can't check
> that myself right now.  I shall do it during the week however.
>
> Ben.
I was also highly skeptical about this, but I looked at all the vbios 
and it seemed good on all of them.

Then, I looked at tools like nibitor or Fermi bios editor 
(http://www.xtremesystems.org/forums/showthread.php?t=256367) which also 
highlights the subentries 3 and 5.

I'll look further and try to see if ramcfg could help us select the 
right subentries.