[Mesa-dev] [PATCH] i965: Use NIR by default for vertex shaders on GEN8+

[Matt Turner]

On Sat, May 16, 2015 at 12:45 PM, Jason Ekstrand <jason at jlekstrand.net> wrote:
> On Sat, May 16, 2015 at 12:12 PM, Matt Turner <mattst88 at gmail.com> wrote:
>> On Fri, May 8, 2015 at 3:27 AM, Kenneth Graunke <kenneth at whitecape.org> wrote:
>>> Looking at a couple of the shaders that are still worse off...it looks
>>> like a ton of Source shaders used to do MUL/ADD with an attribute and
>>> two immediates, and now are doing MOV/MOV/MAD.
>>
>> I just looked, and thought that too for a minute, but it actually
>> shouldn't be doing that. Take for instance:
>>
>> shaders/closed/steam/dota-2/498.shader_test VS SIMD8: 47 -> 53 (12.77%)
>>
>> It indeed replaces 6x MUL/ADD pairs with MOV/MAD (introducing 6 extra
>> MOVs), but....
>>
>> Without NIR we have
>>
>> mul(8)          g15<1>F         g6<8,8,1>F      6F
>> ...
>> add(8)          g16<1>F         g15<8,8,1>F     2.1F
>> add(8)          g35<1>F         g15<8,8,1>F     3.1F
>> add(8)          g42<1>F         g15<8,8,1>F     4.1F
>> add(8)          g45<1>F         g15<8,8,1>F     5.1F
>> add(8)          g48<1>F         g15<8,8,1>F     0.1F
>> add(8)          g51<1>F         g15<8,8,1>F     1.1F
>>
>> That is, one multiply is consumed by 6 adds.
>>
>> With NIR we have
>>
>> mov(1)          g22<1>F         2.1F
>> mov(1)          g22.1<1>F       6F
>> mad(8)          g16<1>F         g22<0,1,0>.xF   g22.1<0,1,0>.xF g6<4,4,1>F
>> mov(1)          g22.2<1>F       3.1F
>> mad(8)          g23<1>F         g22.2<0,1,0>.xF g22.1<0,1,0>.xF g6<4,4,1>F
>> mov(1)          g22.3<1>F       4.1F
>> mad(8)          g30<1>F         g22.3<0,1,0>.xF g22.1<0,1,0>.xF g6<4,4,1>F
>> mov(1)          g22.4<1>F       5.1F
>> mad(8)          g33<1>F         g22.4<0,1,0>.xF g22.1<0,1,0>.xF g6<4,4,1>F
>> mov(1)          g22.5<1>F       0.1F
>> mad(8)          g36<1>F         g22.5<0,1,0>.xF g22.1<0,1,0>.xF g6<4,4,1>F
>> mov(1)          g22.6<1>F       1.1F
>> mad(8)          g39<1>F         g22.6<0,1,0>.xF g22.1<0,1,0>.xF g6<4,4,1>F
>>
>> So we're doing the g6 * 6F operation 6 times! We see this in the NIR as well:
>>
>>         vec1 ssa_419 = ffma ssa_384, ssa_132, ssa_133
>>         vec1 ssa_423 = ffma ssa_384, ssa_132, ssa_135
>>         vec1 ssa_427 = ffma ssa_384, ssa_132, ssa_137
>>         vec1 ssa_428 = ffma ssa_384, ssa_132, ssa_139
>>         vec1 ssa_429 = ffma ssa_384, ssa_132, ssa_141
>>         vec1 ssa_430 = ffma ssa_384, ssa_132, ssa_144
>>
>> Whoops. Ideas for fixing that? I'm guessing that this accounts for
>> nearly all of the remaining 1120 hurt programs.
>
> Ugh... We've been tacitly assuming that your constant combine stuff
> will magically make immediates not a problem.  In this case, they are
> a problem.  I guess we could do something different for 1 vs. 2
> immediates.

That's not really the problem as far as I see. I mean, we could split
MADs that do x * imm + imm, but I would think NIR shouldn't be
combining these operations if the multiply is used in a bunch of
places.

The current code in the ffma peephole in does... to quote the comment:

      /* Only absorb a fmul into a ffma if the fmul is is only used in fadd
       * operations.  This prevents us from being too aggressive with our
       * fusing which can actually lead to more instructions.
       */

Can't we pretty trivially modify that to count the number of uses as
well and only combine if it's used in one place?

To be honest, before I looked in the code I thought that's what it was doing.