[Mesa-dev] i965: Turn UBOs as push constants

Abdiel Janulgue abdiel.janulgue at linux.intel.com
Mon Sep 14 06:03:10 PDT 2015



On 09/12/2015 01:42 AM, Ben Widawsky wrote:
> On Fri, Sep 11, 2015 at 11:33:14AM +0300, Abdiel Janulgue wrote:
>> Some updated figures first:
>>
>> Ue4 Reflections Subway demo
>> x  fs gather constants disabled
>> +  fs gather constants enabled
>>
>> HSW:
>>     N           Min           Max        Median           Avg        Stddev
>> x  10       5.09739       6.48963       6.33357      6.197985    0.40742969
>> +  10       6.56139       6.86579       6.80952      6.758516    0.10267153
>> Difference at 95.0% confidence
>>         0.560531 ± 0.279157
>>         9.04376% ± 4.50399%
>>         (Student's t, pooled s = 0.297103)
>>
>> BDW:
>>     N           Min           Max        Median           Avg        Stddev
>> x  10       3.64474       3.69746       3.68596       3.67842   0.015452422
>> +  10       3.77998       3.80967       3.79824      3.795633  0.0079241642
>> Difference at 95.0% confidence
>>         0.117213 ± 0.0115377
>>         3.1865% ± 0.31366%
>>         (Student's t, pooled s = 0.0122794)
> 
> Does this imply that's the only benchmark which improved, or that it was the
> only one which was run?

My bad, I should have included more:

This is from the another demo featuring heavy usage of UBOs called
"EffectsCave" on Broadwell:

x Gather constants disabled
+ Gather constants Enabled

    N      Min        Max        Median           Avg
x  10      2.7031     2.72984    2.71062      2.713888  0.0089086734
+  10      2.80884    2.8562     2.84192      2.839292   0.014203021
Difference at 95.0% confidence
        0.125404 +/- 0.0111391
        4.62082% +/- 0.410447%
        (Student's t, pooled s = 0.0118552)

and when run on Haswell...

    N     Min         Max        Median           Avg       Stddev
x  10     4.09309     4.85978    4.69503      4.611491    0.23659143
+  10       4.79169   5.25551    5.19242      5.127448    0.16535407
Difference at 95.0% confidence
        0.515957 +/- 0.191776
        11.1885% +/- 4.15865%
        (Student's t, pooled s = 0.204105)


I'll throw in public shader-db results to the mix as well:

total instructions in shared programs: 130441 -> 130357 (-0.06%)
instructions in affected programs:     688 -> 604 (-12.21%)
helped:                                12
HURT:                                  0
GAINED:                                0
LOST:


- Abdiel


More information about the mesa-dev mailing list