[Mesa-dev] [PATCH 3/4] nir: allow nir_intrinsic_load_ubo in opt_peephole_select

[Timothy Arceri]

On 17/07/18 03:41, Eric Anholt wrote:
> Timothy Arceri <tarceri at itsqueeze.com> writes:
> 
>> This makes this opt behave more like the GLSL IR opt
>> lower_if_to_cond_assign(). With this we can disable that GLSL IR
>> opt on drivers with a NIR backend without causing spill
>> regressions.
>>
>> shader-db results for radeonsi (RX580):
>>
>> Totals from affected shaders:
>> SGPRS: 12200 -> 13072 (7.15 %)
>> VGPRS: 13496 -> 11840 (-12.27 %)
>> Spilled SGPRs: 285 -> 290 (1.75 %)
>> Spilled VGPRs: 115 -> 0 (-100.00 %)
>> Private memory VGPRs: 0 -> 0 (0.00 %)
>> Scratch size: 116 -> 0 (-100.00 %) dwords per thread
>> Code Size: 781304 -> 770168 (-1.43 %) bytes
>> LDS: 0 -> 0 (0.00 %) blocks
>> Max Waves: 1558 -> 1586 (1.80 %)
>> Wait states: 0 -> 0 (0.00 %)
> 
> Moving UBO loads out of conditionals seems questionable to me, and
> I don't think shader-db would represent the effect of this kind of code
> transformation well.  I'd love to see some actual performance numbers on
> affected applications.
> 
> That said, lower_if_to_cond_assign() was already doing this, and I'm
> assuming these shader-db stats are just from
> with-glsl-optimization-disabled to with-nir-patch. 

No these results are with lower_if_to_cond_assign() still turned on.

The biggest gain seemed to come from a few shaders with a bunch of 
consecutive if's. It enables a chain of phis to be eliminated and 
reduces a bunch of VGPR spilling. Maybe we should look at that more 
closely on its own.

I'll hold onto this patch for now.

> So, while we should
> probably investigate tuning this, since it lets us dump some GLSL
> optimization:
> 
> Acked-by: Eric Anholt <eric at anholt.net>
>