[Mesa-dev] [PATCH] nir: remove some redundant bcsel instructions

[Timothy Arceri]

On 17/10/18 8:49 pm, Bas Nieuwenhuizen wrote:
> On Wed, Oct 17, 2018 at 5:49 AM Timothy Arceri <tarceri at itsqueeze.com> wrote:
>>
>> For example:
>>
>>     vec1 32 ssa_386 = feq ssa_333.x, ssa_6
>>     vec1 32 ssa_387 = feq ssa_333.x, ssa_2
>>     vec1 32 ssa_391 = bcsel ssa_387, ssa_388, ssa_324
>>     vec1 32 ssa_396 = bcsel ssa_386, ssa_324, ssa_391
>>
>> Can be simplified to:
>>
>>     vec1 32 ssa_386 = feq ssa_333.x, ssa_6
>>     vec1 32 ssa_391 = bcsel ssa_387, ssa_388, ssa_324
>>
>> There are a bunch of these in Rise of The Tomb Raiders Vulkan
>> shaders. There are also a hadful of shaders helped in shader-db
>> but the changes there are smaller.
>>
>> For RADV:
>>
>> Totals from affected shaders:
>> SGPRS: 11184 -> 11168 (-0.14 %)
>> VGPRS: 11484 -> 11484 (0.00 %)
>> Spilled SGPRs: 1119 -> 1116 (-0.27 %)
>> Spilled VGPRs: 0 -> 0 (0.00 %)
>> Private memory VGPRs: 0 -> 0 (0.00 %)
>> Scratch size: 0 -> 0 (0.00 %) dwords per thread
>> Code Size: 1210856 -> 1210372 (-0.04 %) bytes
>> LDS: 0 -> 0 (0.00 %) blocks
>> Max Waves: 360 -> 360 (0.00 %)
>> Wait states: 0 -> 0 (0.00 %)
>> ---
>>   src/compiler/nir/nir_opt_algebraic.py | 4 ++++
>>   1 file changed, 4 insertions(+)
>>
>> diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
>> index cc747250ba5..7530710cbe0 100644
>> --- a/src/compiler/nir/nir_opt_algebraic.py
>> +++ b/src/compiler/nir/nir_opt_algebraic.py
>> @@ -34,6 +34,7 @@ a = 'a'
>>   b = 'b'
>>   c = 'c'
>>   d = 'd'
>> +e = 'e'
>>
>>   # Written in the form (<search>, <replace>) where <search> is an expression
>>   # and <replace> is either an expression or a value.  An expression is
>> @@ -525,6 +526,9 @@ optimizations = [
>>      # The result of this should be hit by constant propagation and, in the
>>      # next round of opt_algebraic, get picked up by one of the above two.
>>      (('bcsel', '#a', b, c), ('bcsel', ('ine', 'a', 0), b, c)),
>> +   # Remove redundant bcsel
>> +   (('bcsel', ('ieq', '#a', b), c, ('bcsel', ('ieq', '#d', b), e, c)), ('bcsel', ('ieq', d, b), e, c)),
> 
> I think this only works if the value of a is not equal to the value of
> d? if a is equal to d, then the expression on the left is always c,
> while the expression on the right is e sometimes?

Hmm. I though the search/matching code was smart enough to handle this, 
but looking at it it seems I was wrong.

I'll take a look tomorrow to see how hard it would be to handle this safely.

> 
> 
>> +   (('bcsel', ('feq', '#a', b), c, ('bcsel', ('feq', '#d', b), e, c)), ('bcsel', ('feq', d, b), e, c)),
> 
>>
>>      (('bcsel', a, b, b), b),
>>      (('fcsel', a, b, b), b),
>> --
>> 2.17.1
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev