W dniu 18 lutego 2010 21:47 użytkownik Alex Deucher
<[email protected]> napisał:
> 2010/2/18 Rafał Miłecki <[email protected]>:
>> W dniu 18 lutego 2010 20:29 użytkownik Alex Deucher
>> <[email protected]> napisał:
>>> 2010/2/17 Rafał Miłecki <[email protected]>:
>>>> We kept requested and current modes in many places, depending on current 
>>>> state.
>>>> That was useless, one place for holding that is enough.
>>>>
>>>> Signed-off-by: Rafał Miłecki <[email protected]>
>>>> ---
>>>> Tested on my RV620, no problems. Alex: can you review this patch? It's your
>>>> code I modify/remove in it.
>>>
>>> NACK.  Why are you replacing pointers with copies of of the power
>>> state structs?  The idea is to keep one array of power states and
>>> pointers to the current one, default one, and requested one.  Then
>>> comparing power states is just comparing pointers and when you change
>>> the power state, you just update the pointer rather then memcpying the
>>> entire struct.
>>
>> Then first of all we would need to reduce modes pointers. Keeping mode
>> pointer in every state struct is/was useless.
>
> What do you mean?   There is not a pointer to every state.  There are
> only 3 (1. current, 2. requested 3. default) and they are not useless.
>
> You have power states which defines the global state of a particular
> power state and then each power state has a set of 1 or more clock
> modes.  So within a power state you can switch between clock modes.
> You need a set of clock pointers to track the current clock mode and
> the requested clock mode.  You also need to know the current and
> requested power state so you need pointers there as well.  We could
> get rid of the default pointers, but that's mostly there for
> convenience to avoid having to look up the default state when we need
> it.
>
> For example, say you had the following power tables:
>
> 1. 3d power state:
> a. low clock mode
> b. medium clock mode
> c. high clock mode
>
> 2. battery power state
> a. low clock mode
> b. medium clock mode
> c. high clock mode
>
> 3. default power state
> a. low clock mode
> b. medium clock mode
> c. high clock mode
>
>
> By default state 3 would be selected so the current state would be '3'
> and the current clock mode would be a say 'a'.  The current state
> pointer would point to '3' and the current clock mode pointer would
> point to 'a'.  If you then went on battery, you'd want power state
> '2'.  At that point, the requested state pointer points to '2' and one
> of the clock modes gets selected, lets say 'c', so the requested clock
> mode pointer points to 'c'.  When we change the power state, we
> compare the current and requested power state pointers, if they are
> the same, no need to change anything.  if they are different, we make
> the changes, then set the current power state pointer to the requested
> power state pointer and the current clock mode pointer to the
> requested clock mode pointer.  Now say we are in power state '2' and
> clock mode 'c' (the high clock for power state '2') and the system
> goes idle; at the point we will want to stay in power state '2', but
> select clock mode 'a' (the low clock for power state '2').  So we
> compare the current and requested clock mode pointers and if they are
> not equal, we change the clock mode, at which time we update the
> current clock mode pointers to point to the requested clock.

Really, you didn't need to explain all of that. I understand current
idea and implementation :)

The thing is that in your implementation we have 8 (eight!) pointers
to current clock mode and 8 pointers to requested clock mode. This
part is useless IMO.

Let me prepare another patch to show reduction without introducing
stuct & memcpy.


>> About introduced solution (keeping struct and memcpy to it) I
>> introduced that to make hacking requested mode possible. Info from
>> AtomBIOS about PCIE lanes seem to be useless (I've never seen anything
>> else than 16) so I want to hack found mode for DPMS OFF to use 1 PCIE
>> lane. Without keeping whole struct it would be impossible without
>> overwriting original entry in array and loosing original info.
>
> That's fine for a local hack, but we don't want that upstream.  If you
> are planning to hack all the power tables, then why even use them?
> I'm not sure we really change the lanes that often if at all on r6xx+.
>  In my experience, a lot of cards lock up when change the lanes or
> fail to change properly which results in getting stuck at lower
> numbers of lanes.

I don't want to hack all power tables, just consider hacking
clocks/etc for minimal (dpms off). Don't want to touch other.


-- 
Rafał

------------------------------------------------------------------------------
Download Intel&#174; Parallel Studio Eval
Try the new software tools for yourself. Speed compiling, find bugs
proactively, and fine-tune applications for parallel performance.
See why Intel Parallel Studio got high marks during beta.
http://p.sf.net/sfu/intel-sw-dev
--
_______________________________________________
Dri-devel mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/dri-devel

Reply via email to