[Intel-gfx] [PATCH 4/6] drm/i915: set ILK_DPFC_FENCE_YOFF to 0 on SNB

[Paulo Zanoni]

2015-07-09 14:39 GMT-03:00 Ville Syrjälä <ville.syrjala at linux.intel.com>:
> On Thu, Jul 09, 2015 at 02:31:15PM -0300, Paulo Zanoni wrote:
>> 2015-07-09 14:22 GMT-03:00 Ville Syrjälä <ville.syrjala at linux.intel.com>:
>> > On Thu, Jul 09, 2015 at 07:10:04PM +0200, Daniel Vetter wrote:
>> >> On Wed, Jul 08, 2015 at 05:58:57PM -0300, Paulo Zanoni wrote:
>> >> > From: Paulo Zanoni <paulo.r.zanoni at intel.com>
>> >> >
>> >> > The doc is pretty clear that this register should be set to 0 on SNB.
>> >> > We already write y_offset to DPFC_CPU_FENCE_OFFSET a few lines below.
>> >> >
>> >> > Signed-off-by: Paulo Zanoni <paulo.r.zanoni at intel.com>
>> >>
>> >> Hm, do we have testcases where we have a sufficiently big y offset? We can
>> >> just allocate 128 lines more and use that as the offset, that should be
>> >> big enough everywhere. Actually make that 129 lines to check the tile-size
>> >> rounding ;-)
>> >>
>> >> Ofc this means we need to have two sets of testcases for all the affected
>> >> tests (i.e. everything that tries to test the gtt hw tracking).
>> >>
>> >> Another funny corner case (which we're getting wrong on skl even without
>> >> fbc) is x offsets > 2048 pixels (since x/y offset registers don't hold
>> >> bigger values and then it wraps).
>> >>
>> >> I.e. I'd like this patch (and the others) to be augmented with a Testcase:
>> >> tag.
>> >
>> > I think the entire Y offset thing is currently being misprogrammed.
>> > IIRC the offset is from the display base address but we program in
>> > the offset from the start of the FB.
>>
>> After patch 3, all the current tests pass on BDW. Can you suggest a
>> different test that won't pass?
>
> Ah patch 3 tries to fix it. It's not entirely accurate though since it
> simply relies on an implementation detail of intel_gen4_compute_page_offset().
> Well, assuming my recollection of the hardware details is correct.
>
> Also IIRC intel_gen4_compute_page_offset() isn't even used on SKL/BXT
> currently, so it should fail on those platforms.

Daniel clarified the problem to me, so I implemented the following test case:
http://cgit.freedesktop.org/xorg/app/intel-gpu-tools/commit/?id=04d1311fc3d2127d609b5c5e670bf9887652cb17

I hope this exercises the problem you're mentioning. So far I only
tested BDW and it passes.

>
> --
> Ville Syrjälä
> Intel OTC



-- 
Paulo Zanoni