[Mesa-dev] [PATCH 2/2] nv50/ir: improve POW lowering

[Karol Herbst]

On Thu, Jun 22, 2017 at 9:44 PM, Roland Scheidegger <sroland at vmware.com> wrote:
> Am 22.06.2017 um 21:28 schrieb Karol Herbst:
>> If the exponent is a small integer immediate value, we can lower POW to
>> MULs instead to save a few instructions. Also MUL instructions execute
>> faster than what we lower POW in the default case to.
>>
>> score change for GpuTest /test=pixmark_piano /benchmark /no_scorebox
>> /msaa=0 /benchmark_duration_ms=60000 /width=1024 /height=640:
>> 1045 -> 1060
>>
>> changes in shader-db:
>> total instructions in shared programs : 4350261 -> 4349451 (-0.02%)
>> total gprs used in shared programs    : 525853 -> 525861 (0.00%)
>> total local used in shared programs   : 30081 -> 30081 (0.00%)
>> total bytes used in shared programs   : 39865176 -> 39857712 (-0.02%)
>>
>>                 local        gpr       inst      bytes
>>     helped           0           4         313         313
>>       hurt           0          12           1           1
>>
>> Signed-off-by: Karol Herbst <karolherbst at gmail.com>
>> ---
>>  .../drivers/nouveau/codegen/nv50_ir_build_util.cpp | 49 ++++++++++++++++++++++
>>  1 file changed, 49 insertions(+)
>>
>> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_build_util.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_build_util.cpp
>> index 5756e1b4d4..b31dcec1ab 100644
>> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_build_util.cpp
>> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_build_util.cpp
>> @@ -640,6 +640,55 @@ bool
>>  BuildUtil::lowerPOW(Instruction *i)
>>  {
>>     LValue *val = getScratch();
>> +   ImmediateValue imm;
>> +
>> +   if (i->src(1).getImmediate(imm)) {
>> +      Value *src = i->getSrc(0);
>> +      if (imm.isInteger(0)) {
>> +         i->op = OP_MOV;
>> +         i->setSrc(0, loadImm(NULL, 1));
>> +         i->setSrc(1, NULL);
>> +         return true;
>> +      } else
>> +      if (imm.isInteger(1)) {
>> +         i->op = OP_MOV;
>> +         i->setSrc(1, NULL);
>> +         return true;
>> +      } else
>> +      if (imm.isInteger(2)) {
>> +         i->op = OP_MUL;
>> +         i->setSrc(1, src);
>> +         return true;
>> +      } else
>> +      if (imm.isInteger(3)) {
>> +         mkOp2(OP_MUL, i->dType, val, src, src);
>> +         i->op = OP_MUL;
>> +         i->setSrc(1, val);
>> +         return true;
>> +      } else
>> +      if (imm.isInteger(4)) {
>> +         mkOp2(OP_MUL, i->dType, val, src, src);
>> +         i->op = OP_MUL;
>> +         i->setSrc(0, val);
>> +         i->setSrc(1, val);
>> +         return true;
>> +      } else
>> +      if (imm.isInteger(5)) {
>> +         mkOp2(OP_MUL, i->dType, val, src, src);
>> +         mkOp2(OP_MUL, i->dType, val, val, val);
>> +         i->op = OP_MUL;
>> +         i->setSrc(1, val);
>> +         return true;
>> +      } else
>> +      if (imm.isInteger(8)) {
>> +         mkOp2(OP_MUL, i->dType, val, src, src);
>> +         mkOp2(OP_MUL, i->dType, val, val, val);
>> +         i->op = OP_MUL;
>> +         i->setSrc(0, val);
>> +         i->setSrc(1, val);
>> +         return true;
>> +      }
>> +   }
>>
>>     mkOp1(OP_LG2, TYPE_F32, val, i->getSrc(0));
>>     mkOp2(OP_MUL, TYPE_F32, val, i->getSrc(1), val)->dnz = 1;
>>
>
> You could do integer 6 with just 3 muls too :-).
> But more in general, this looks like it would be useful outside of
> nouveau too.
>
> Roland
>

yeah I know, but the shader-db changes didn't convince me. I think
with proper instruction scheduling it's worth to take another look,
but until now I want to play it safe.