[Intel-gfx] [PATCH] drm/i915: set IDICOS to medium uncore resources

[Eric Anholt]

On Fri,  4 May 2012 18:58:59 -0700, Ben Widawsky <ben at bwidawsk.net> wrote:
> I'm seeing about a 5% FPS improvement across various benchmarks on my
> IVB i3. Rumor has it that the higher end parts show even more benefit.
> 
> This derives from a patch originally given to me by Bernard. The docs
> are  confusing about the definition names (ie. medium really seems like
> max), but it would seem it gives more cache to the GT at the expense of
> uncore. This configuration makes the split most in favor of the GT. I've
> not tried the other IDICOS values.

My IVB system started with an SNBPCR value of 0x1b40.  To test the effect of
this patch, I just smashed the value of the reg at runtime:

Results:
x intel_reg_write 0x900c 0x001b40
+ intel_reg_write 0x900c 0x201b40

citybench:
+------------------------------------------------------------------------------+
|                   xxx                          +                             |
|   x               xxx          +               + +                           |
|   x               xxx          +               +++                           |
|   x               xxx          +            + ++++                           |
|   xx             xxxx          +            + ++++                           |
|x xxx           x xxxx        + +            ++++++ ++                        |
|xxxxx          xxxxxxxx       + + + +        +*++++ ++                        |
|xxxxxxxxx      xxxxxxxx x     +++++ ++++     +*+++++++                        |
|xxxxxxxxx    xxxxxxxxxx x     ++++++++++   ++**+++++++ +          +           |
|xxxxxxxxx  xxxxxxxxxxxxxx    *++++++++++++ ****+*++++++++      +  ++++++     +|
|    |___________AM_________|       |_________A_M_______|                      |
+------------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x 118       335.779       387.456       354.306     352.63131     12.462796
+ 118       367.324        419.02        386.14     384.47139     10.970918
Difference at 95.0% confidence
	31.8401 +/- 2.99585
	9.02928% +/- 0.849569%
	(Student's t, pooled s = 11.7406)

openarena:
+------------------------------------------------------------------------------+
|         +                ++                                xx xx x  x xx     |
|         +      +         ++ + +                            xxxxxxx xxxxx     |
|        +++  ++++++++++ ++++++ ++ +                    x  xxxxxxxxxxxxxxx   xx|
|++ + ++++++++++++++++++++++++++++ + + +             x  x xxxxxxxxxxxxxxxxxx xx|
|          |_________A________|                              |_____A____|      |
+------------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x  65         273.5         292.2           284     283.73385     4.2561822
+  65         234.8         263.2         249.9         249.6     6.9873323
Difference at 95.0% confidence
	-34.1338 +/- 1.989
	-12.0302% +/- 0.70101%

nexuiz:
+------------------------------------------------------------------------------+
|                                                                 +            |
|                                                   x        +    +            |
|                                             +     x   x xx *x+  +x           |
|+     +         +        x   *x+x xxx    + x++ + x+*xxxx+x***x+++*x**+x ++  *+|
|                                   |____|____________AAM____M_____|______|    |
+------------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x  33     74.314888     76.177197     75.424764     75.332161    0.48205891
+  33     73.397254     76.219868     75.584749     75.363209    0.69408523

Using these 3 tests, this patch does not look like a win.  Note that OA
is the most CPU-bound of these, while nexuiz is usually on the edge
between being CPU and GPU-bound, and citybench is definitely GPU
limited.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 197 bytes
Desc: not available
URL: <http://lists.freedesktop.org/archives/intel-gfx/attachments/20120506/9c589ffb/attachment.sig>