[Mesa-dev] [PATCH v2] nvc0/ir: use levelZero flag when the lod is set to 0

[Samuel Pitoiset]

On 11/14/2016 06:53 PM, Ilia Mirkin wrote:
> On Mon, Nov 14, 2016 at 12:39 PM, Samuel Pitoiset
> <samuel.pitoiset at gmail.com> wrote:
>>
>>
>> On 11/10/2016 03:42 AM, Ilia Mirkin wrote:
>>>
>>> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
>>> ---
>>>
>>> v1 -> v2:
>>>   Move to handling this at SSA time. This is a lot more fragile since the
>>>   texture arguments have been reordered already, but it's still easy
>>> enough
>>>   to find the LOD argument.
>>>
>>>  .../nouveau/codegen/nv50_ir_lowering_nvc0.cpp      | 49
>>> +++++++++++++++++++---
>>>  .../nouveau/codegen/nv50_ir_lowering_nvc0.h        |  1 +
>>>  2 files changed, 44 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
>>> b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
>>> index 68f2b15..68bae08 100644
>>> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
>>> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
>>> @@ -115,6 +115,39 @@ NVC0LegalizeSSA::handleFTZ(Instruction *i)
>>>     i->ftz = true;
>>>  }
>>>
>>> +void
>>> +NVC0LegalizeSSA::handleTEXLOD(TexInstruction *i)
>>> +{
>>> +   if (i->tex.target.isMS())
>>> +      return;
>>> +
>>> +   ImmediateValue lod;
>>> +
>>> +   // The LOD argument comes right after the coordinates (before depth
>>> bias,
>>> +   // offsets, etc).
>>> +   int arg = i->tex.target.getArgCount();
>>> +
>>> +   // SM30+ stores the indirect handle as a separate arg, which comes
>>> before
>>> +   // the LOD.
>>> +   if (prog->getTarget()->getChipset() >= NVISA_GK104_CHIPSET &&
>>> +       i->tex.rIndirectSrc >= 0)
>>> +      arg++;
>>> +   // SM20 stores indirect handle combined with array coordinate
>>> +   if (prog->getTarget()->getChipset() < NVISA_GK104_CHIPSET &&
>>> +       !i->tex.target.isArray() &&
>>> +       i->tex.rIndirectSrc >= 0)
>>
>>
>> This could be on the same line.
>
> I prefer either all on the same line or 1 per line.

Your call.

>
>>
>>> +      arg++;
>>> +
>>> +   if (!i->src(arg).getImmediate(lod) || !lod.isInteger(0))
>>> +      return;
>>> +
>>> +   if (i->op == OP_TXL)
>>> +      i->op = OP_TEX;
>>> +   i->tex.levelZero = true;
>>> +   i->moveSources(arg + 1, -1);
>>> +}
>>> +
>>> +
>>
>>
>> Extra empty line.
>
> Will fix.
>
>>
>>>  bool
>>>  NVC0LegalizeSSA::visit(Function *fn)
>>>  {
>>> @@ -128,21 +161,25 @@ NVC0LegalizeSSA::visit(BasicBlock *bb)
>>>     Instruction *next;
>>>     for (Instruction *i = bb->getEntry(); i; i = next) {
>>>        next = i->next;
>>> -      if (i->sType == TYPE_F32) {
>>> -         if (prog->getType() != Program::TYPE_COMPUTE)
>>> -            handleFTZ(i);
>>> -         continue;
>>> -      }
>>> +
>>> +      if (i->sType == TYPE_F32 && prog->getType() !=
>>> Program::TYPE_COMPUTE)
>>> +         handleFTZ(i);
>>> +
>>
>>
>> Why do you need to change this?
>
> I had to remove the continue, and having
>
> if (a) {
>  if (b)
>    c;
> }
>
> seemed silly.

Okay.

No need to send a v3 for that empty line. :-)

Reviewed-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>

>
>>
>>
>>>        switch (i->op) {
>>>        case OP_DIV:
>>>        case OP_MOD:
>>> -         handleDIV(i);
>>> +         if (i->sType != TYPE_F32)
>>> +            handleDIV(i);
>>>           break;
>>>        case OP_RCP:
>>>        case OP_RSQ:
>>>           if (i->dType == TYPE_F64)
>>>              handleRCPRSQ(i);
>>>           break;
>>> +      case OP_TXL:
>>> +      case OP_TXF:
>>> +         handleTEXLOD(i->asTex());
>>> +         break;
>>>        default:
>>>           break;
>>>        }
>>> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h
>>> b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h
>>> index 6f4da8c..d91b6aa 100644
>>> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h
>>> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.h
>>> @@ -35,6 +35,7 @@ private:
>>>     void handleDIV(Instruction *); // integer division, modulus
>>>     void handleRCPRSQ(Instruction *); // double precision float
>>> recip/rsqrt
>>>     void handleFTZ(Instruction *);
>>> +   void handleTEXLOD(TexInstruction *);
>>>
>>>  protected:
>>>     BuildUtil bld;
>>>
>>