[Mesa-dev] [PATCH v2] nir: add a comparison simplification

[Timothy Arceri]

On 16/03/18 19:34, Samuel Pitoiset wrote:
> 
> 
> On 03/16/2018 09:15 AM, Timothy Arceri wrote:
>> On 16/03/18 18:45, Samuel Pitoiset wrote:
>>> shader stats?
>>
>> I didn't bother posting them because they are misleading. There are a 
>> few instructions remove here and there but the main changes are due to 
>> loop unrolling.
>>
>> Totals from affected shaders:
>> SGPRS: 640 -> 664 (3.75 %)
>> VGPRS: 352 -> 368 (4.55 %)
>> Spilled SGPRs: 6 -> 8 (33.33 %)
>> Spilled VGPRs: 0 -> 0 (0.00 %)
>> Private memory VGPRs: 0 -> 0 (0.00 %)
>> Scratch size: 0 -> 0 (0.00 %) dwords per thread
>> Code Size: 24200 -> 36216 (49.65 %) bytes
>> LDS: 0 -> 0 (0.00 %) blocks
>> Max Waves: 52 -> 51 (-1.92 %)
>> Wait states: 0 -> 0 (0.00 %)
> 
> Mmmh, the stats don't seem really good to me. Which games are affected 
> by this change?

Just Tomb Raider.

> 
> As you know, we shouldn't *always* trust shader stats. Though, it would 
> quite appreciated to run some benchmarks in this situation.

I've benchmarked it and there was no change. The stats are misleading 
because register and code size increase but that is an expected outcome 
of loop unrolling. At the end of the day this is a clear simplification 
of logic, any bad outcome would indicate we should tune loop unrolling 
rather than reject this patch IMO.

> 
>>
>>
>>>
>>> On 03/16/2018 06:48 AM, Timothy Arceri wrote:
>>>> These can be found in the Tomb Raider shaders, eliminating them
>>>> helps unroll more loops.
>>>> ---
>>>>   src/compiler/nir/nir_opt_algebraic.py | 1 +
>>>>   1 file changed, 1 insertion(+)
>>>>
>>>> diff --git a/src/compiler/nir/nir_opt_algebraic.py 
>>>> b/src/compiler/nir/nir_opt_algebraic.py
>>>> index c9575e6be4..2148bc121b 100644
>>>> --- a/src/compiler/nir/nir_opt_algebraic.py
>>>> +++ b/src/compiler/nir/nir_opt_algebraic.py
>>>> @@ -152,6 +152,7 @@ optimizations = [
>>>>      (('inot', ('ige', a, b)), ('ilt', a, b)),
>>>>      (('inot', ('ieq', a, b)), ('ine', a, b)),
>>>>      (('inot', ('ine', a, b)), ('ieq', a, b)),
>>>> +   (('ine', ('b2i', 'a at bool'), 0), 'a'),
>>>>      # 0.0 >= b2f(a)
>>>>      # b2f(a) <= 0.0
>>>>