[Mesa-dev] [PATCH 2/2] i965: Use IVB specific formula for depthbuffer

[Ben Widawsky]

On Thu, Sep 26, 2013 at 01:28:47PM -0700, Chad Versace wrote:
> On 09/22/2013 10:37 AM, Ben Widawsky wrote:
> >After the last patch, we can replace the region allocated in the miptree
> >creation with a more straightforward (and hopefully smaller resulting)
> >buffer based on the bspec's allocation formula.
> >
> >Since I am relatively new to this part of the bspec, I would very much
> >appreciate scrutiny during review of this. There were some ambiguities
> >to me which are likely obvious to others.
> >
> >To prove the reduced [GPU] memory usage I created a simple script which
> >polls the memory usage of the process through debugfs ever .1 seconds.
> >The following results show the memory usage difference over 5 runs of
> >xonotic-glx with ultra settings.
> >
> >The data suggests a 10MB savings on average. I've not measured the
> >savings on the CPU side, but I imagine some amount of savings would be
> >present there as well.
> >
> >x master/mem_usage.txt
> >+ mine/mem_usage.txt
> >     N           Min           Max        Median           Avg Stddev
> >x 17121      98959360 7.3394995e+08 7.2782234e+08 7.2209615e+08 43633222
> >+ 17166 1.2538266e+08 7.2241562e+08   7.16288e+08 7.1071472e+08 42964578
> >
> >Below is the FPS data over those same 5 tests. I'm not sure if the
> >decrease is statistically significant to y'all. I don't have any
> >theories about it.
> >
> >x master/xonotic.fps
> >+ mine/xonotic.fps
> >     N           Min           Max        Median           Avg Stddev
> >x   5     27.430746     27.524985      27.50568     27.487017 0.039439874
> >+   5     27.409173     27.461715     27.441207     27.440883 0.021086805
> >
> >NOTE: There were a couple of places in the arithmetic where I could have
> >taken some shortcuts. In order to make the code match with the spec as
> >much as possible, I've decided not to do this. One shortcut I did make
> >was the tiling type. Digging through the code it looks like you always
> >want Y-tiled, except when it won't fit, in which case you want X-tiled.
> >I wasn't a fan of the existing helper function that's there since it has
> >a few irrelevant parameters for this operation. I suspect people
> >reviewing this might ask me to change this, which is fine; I just wanted
> >to explain the motivation.
> >
> >CC: Chad Versace <chad.versace at linux.intel.com>
> >Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=67564
> >Signed-off-by: Ben Widawsky <ben at bwidawsk.net>
> >---
> >  src/mesa/drivers/dri/i965/intel_mipmap_tree.c | 149 +++++++++++++++++++++++---
> >  1 file changed, 133 insertions(+), 16 deletions(-)
> 
> 
> I began reviewing the equations in this patch, and hit a nasty roadblock.
> The all-important variables z_width, z_height, and z_depth are the problem.
> My interpretation of what those variables should be *differs* from how
> the driver currently programs them in 3DSTATE_DEPTH_BUFFER. I believe the
> driver has a bug regarding this.
> 
> Until I determine who is correct, the driver or me, I'm unable to
> review your implementation of the equations.
> 
> Let's put this patch on hold for several days while I write a Piglit
> test to determine the truth.


I've fixed the issues mentioned on the mails thus far. I'll look at
integrating your patch. Here is the data with just the fixes (on a
different machine with a larger display). I thought I had used the same
quality settings, but I must be mistaken given that the FPS went from
27->380. In any case, the FPS was just background info.

x HEAD/memusage
+ mine/memusage
    N           Min           Max        Median           Avg Stddev
    x 17817      89432064 7.9064678e+08 7.8487552e+08 7.7993921e+08 39479219
    + 17792 1.8873139e+08  7.812096e+08 7.7573734e+08 7.7078243e+08 38542321


x HEAD/xonotic
+ mine/xonotic
    N           Min           Max        Median           Avg Stddev
    x   5     382.93275     386.07911     383.08233     383.72959 1.3291431
    +   5     382.85601     383.86606     383.10943     383.23247 0.38605285

I was expecting more memory savings on average with the larger display -
but it's gone down from 11M->9M. Also surprising to me is the min is 2x
greater, which leads me to believe I missed a shortcut somewhere, or I
am not collecting samples well enough.  Looking back at the previous
data, the min was also quite significantly increased. I believe
regardless of your test patch, I should be able to get min under
control, and that should greatly improve the numbers.

I will also look at the test patch you gave me and see if I can do
something with it before you do.


-- 
Ben Widawsky, Intel Open Source Technology Center