[Intel-gfx] [PATCH] drm/i915: Fix correct FIFO size for Baytrail

[Ville Syrjälä]

On Fri, Feb 07, 2014 at 09:56:29PM +0530, Vijay Purushothaman wrote:
> On 2/7/2014 9:28 PM, Ville Syrjälä wrote:
> > On Fri, Feb 07, 2014 at 08:43:12PM +0530, Vijay Purushothaman wrote:
> >> B-spec says the FIFO total size is 512. So fix this to 512.
> >>
> >> Signed-off-by: Vijay Purushothaman <vijay.a.purushothaman at intel.com>
> >> ---
> >>   drivers/gpu/drm/i915/i915_reg.h |    2 +-
> >>   1 file changed, 1 insertion(+), 1 deletion(-)
> >>
> >> diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h
> >> index cc3ea04..fb73031 100644
> >> --- a/drivers/gpu/drm/i915/i915_reg.h
> >> +++ b/drivers/gpu/drm/i915/i915_reg.h
> >> @@ -3395,7 +3395,7 @@
> >>   #define I915_FIFO_LINE_SIZE	64
> >>   #define I830_FIFO_LINE_SIZE	32
> >>
> >> -#define VALLEYVIEW_FIFO_SIZE	255
> >> +#define VALLEYVIEW_FIFO_SIZE	511
> >>   #define G4X_FIFO_SIZE		127
> >>   #define I965_FIFO_SIZE		512
> >>   #define I945_FIFO_SIZE		127
> >
> > Reviewed-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> 
> Thanks for the review.
> 
> >
> > Not that we actually use the value anywhere at the moment.
> 
> This value is used when the display controller is configured in Max FIFO 
> mode. This is working fine in the android tree. At this moment i am 
> rewriting some logic related to this Max FIFO, memory arbiter credits 
> and drain latency handling for the sprite planes. I should have the 
> patches ready over the week end, will test it on monday once i get to 
> office.

Oh right, we do use it in g4x_compute_wm0() which would mean this patch
is actually wrong when not in maxfifo mode. We'd need to dynamically
select the correct value. When not in maxfifo read the FIFO size from
the DSPARB, and when in maxfifo use the 511 value.

> >
> > As a side note the FIFO sizing for gmch platforms seems to be a place
> > where the documentation is rather poor. It kind of looks like there
> > are off by one errors in the text, and yet when I was playing around
> > with this stuff on gen2/gen4 machines it kind of looks like the
> > hardware has the same off by one issues too. IIRC my conlusion was
> > that the last cacheline in the FIFO can't actually be used. So
> > specifying 511 matches with my conclusion.
> 
> I agree.. I was thrown off by this oddity as well and it took some time 
> for me to understand this completely. The display block in Baytrail is a 
> mix and match of features from gen4 & gen5 (Cantiga, Crestline and 
> Ironlake). No wonder this chip has the same h/w issues..
> 
> >
> > One other thing I did notice now that I look at our g4x/vlv watermark
> > code. We seem to assume the watermarks for g4x/vlv work the same way
> > as pch platforms. Ie. you specify the minimum level of data left in the
> > FIFO before it needs to start fetching more. But the documentation
> > suggests that it's the other way around, where you specify the max
> > amount of free space allowed in the FIFO before more data needs to be
> > fetched. We use the latter logic for gen2-gen4. I wonder if the spec
> > is wrong, or if your code is wrong. I guess I just need to verify it
> > on real hardware at some point...
> >
> I think this is implemented correctly in the android kernel.. On a high 
> level the split is something like 32 KB FIFO per pipe - 16 KB for 
> primary plane and 8 KB for each sprite. When we are in single display 
> mode we can configure the entire 64KB FIFO for the same pipe.

I'm not talking about the split, but rather the actual watermark for any
single plane.

This is what the spec says:
"Number in 64Bs of space in FIFO above which the Display A Stream
 will generate requests to Memory"

That to me sounds like it wants the opposite value than what we program
in. But I think I need to verify this at some point on some g4x machine.

> There is 
> another trick to enable trickle feed.. With all these tricks i am seeing 
> good memory self refresh numbers - almost on par with the theoretical 
> target. I should be able to post the patch series on monday once i do 
> some sanity testing..

Enable trickle feed to get better power savings? Sounds strange.
Trickle feed isn't really explained all that well anywhere but based
on what I've gathered, my idea has been that it would keep the memory
awake almost constantly by trickling in small amounts of data whenever
the FIFO has a little bit of room. IIRC I did some experiments which
supported that conclusion. I can't recall the specifics anymore but I
think with trickle feed enabled, the watermarks were more or less
useless. Only when I disabled trickle feed, the watermark level
started to have a real effect (ie. set it too low and you get underruns).

-- 
Ville Syrjälä
Intel OTC