[Mesa-dev] [PATCH] gm107/ir: use lane 0 for manual textureGrad handling

[Ilia Mirkin]

On Tue, Dec 19, 2017 at 11:41 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
> This is parallel to the pre-SM50 change which does this. Adjusts the
> shuffles / quadops to make the values correct relative to lane 0, and
> then splat the results to all lanes for the final move into the target
> register.
>
> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
> ---
>
> Entirely untested beyond compilation. Should check
>
> bin/tex-miplevel-selection textureGrad Cube
> bin/tex-miplevel-selection textureGrad CubeShadow
> bin/tex-miplevel-selection textureGrad CubeArray
> KHR-GL45.texture_cube_map_array.sampling
>
> to see if they start passing with this change.
>
>  .../nouveau/codegen/nv50_ir_lowering_gm107.cpp     | 56 ++++++++++++++--------
>  1 file changed, 35 insertions(+), 21 deletions(-)
>
> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp
> index 6b9edd48645..a2427526a81 100644
> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp
> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_gm107.cpp
> @@ -95,18 +95,15 @@ GM107LegalizeSSA::visit(Instruction *i)
>  bool
>  GM107LoweringPass::handleManualTXD(TexInstruction *i)
>  {
> -   static const uint8_t qOps[4][2] =
> -   {
> -      { QUADOP(MOV2, ADD,  MOV2, ADD),  QUADOP(MOV2, MOV2, ADD,  ADD) }, // l0
> -      { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(MOV2, MOV2, ADD,  ADD) }, // l1
> -      { QUADOP(MOV2, ADD,  MOV2, ADD),  QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l2
> -      { QUADOP(SUBR, MOV2, SUBR, MOV2), QUADOP(SUBR, SUBR, MOV2, MOV2) }, // l3
> -   };
> +   // See NVC0LoweringPass::handleManualTXD for rationale. This function
> +   // implements the same logic, but using SM50-friendly primitives.
> +   static const uint8_t qOps[2] =
> +      { QUADOP(MOV2, ADD,  MOV2, ADD),  QUADOP(MOV2, MOV2, ADD,  ADD) };
>     Value *def[4][4];
> -   Value *crd[3];
> +   Value *crd[3], *arr, *shadow;
>     Value *tmp;
>     Instruction *tex, *add;
> -   Value *zero = bld.loadImm(bld.getSSA(), 0);
> +   Value *quad = bld.mkImm(SHFL_BOUND_QUAD);
>     int l, c;
>     const int dim = i->tex.target.getDim() + i->tex.target.isCube();
>     const int array = i->tex.target.isArray();
> @@ -115,35 +112,40 @@ GM107LoweringPass::handleManualTXD(TexInstruction *i)
>
>     for (c = 0; c < dim; ++c)
>        crd[c] = bld.getScratch();
> +   arr = bld.getScratch();
> +   shadow = bld.getScratch();
>     tmp = bld.getScratch();
>
>     for (l = 0; l < 4; ++l) {
>        Value *src[3], *val;
> -      // mov coordinates from lane l to all lanes
> +      Value *lane = bld.mkImm(l);
>        bld.mkOp(OP_QUADON, TYPE_NONE, NULL);
> +      // Make sure lane 0 has the appropriate array/depth compare values
> +      if (l != 0) {
> +         if (array)
> +            bld.mkOp3(OP_SHFL, TYPE_F32, arr, i->getSrc(0), lane, quad);
> +         if (i->tex.target.isShadow())
> +            bld.mkOp3(OP_SHFL, TYPE_F32, shadow, i->getSrc(array + dim), lane, quad);

In the great argument switcheroo between each SM version, the shadow
compare is actually after the indirect handle (which in turn is after
array + dim). So this should become array + dim + indirect (and
similarly below).

> +      }
> +
> +      // mov coordinates from lane l to all lanes
>        for (c = 0; c < dim; ++c) {
> -         bld.mkOp3(OP_SHFL, TYPE_F32, crd[c], i->getSrc(c + array),
> -                   bld.mkImm(l), bld.mkImm(SHFL_BOUND_QUAD));
> -         add = bld.mkOp2(OP_QUADOP, TYPE_F32, crd[c], crd[c], zero);
> -         add->subOp = 0x00;
> -         add->lanes = 1; /* abused for .ndv */
> +         bld.mkOp3(OP_SHFL, TYPE_F32, crd[c], i->getSrc(c + array), lane, quad);
>        }
>
>        // add dPdx from lane l to lanes dx
>        for (c = 0; c < dim; ++c) {
> -         bld.mkOp3(OP_SHFL, TYPE_F32, tmp, i->dPdx[c].get(), bld.mkImm(l),
> -                   bld.mkImm(SHFL_BOUND_QUAD));
> +         bld.mkOp3(OP_SHFL, TYPE_F32, tmp, i->dPdx[c].get(), lane, quad);
>           add = bld.mkOp2(OP_QUADOP, TYPE_F32, crd[c], tmp, crd[c]);
> -         add->subOp = qOps[l][0];
> +         add->subOp = qOps[0];
>           add->lanes = 1; /* abused for .ndv */
>        }
>
>        // add dPdy from lane l to lanes dy
>        for (c = 0; c < dim; ++c) {
> -         bld.mkOp3(OP_SHFL, TYPE_F32, tmp, i->dPdy[c].get(), bld.mkImm(l),
> -                   bld.mkImm(SHFL_BOUND_QUAD));
> +         bld.mkOp3(OP_SHFL, TYPE_F32, tmp, i->dPdy[c].get(), lane, quad);
>           add = bld.mkOp2(OP_QUADOP, TYPE_F32, crd[c], tmp, crd[c]);
> -         add->subOp = qOps[l][1];
> +         add->subOp = qOps[1];
>           add->lanes = 1; /* abused for .ndv */
>        }
>
> @@ -164,8 +166,20 @@ GM107LoweringPass::handleManualTXD(TexInstruction *i)
>
>        // texture
>        bld.insert(tex = cloneForward(func, i));
> +      if (l != 0) {
> +         if (array)
> +            tex->setSrc(0, arr);
> +         if (i->tex.target.isShadow())
> +            tex->setSrc(array + dim, shadow);
> +      }
>        for (c = 0; c < dim; ++c)
>           tex->setSrc(c + array, src[c]);
> +      // broadcast results from lane 0 to all lanes
> +      if (l != 0) {
> +         Value *lane = bld.mkImm(l);

This should of course be bld.mkImm(0), not l, since we're broadcasting
from lane *0* to all lanes.

These are all fixed up in
https://github.com/imirkin/mesa/commit/618b99d86396417e31551dc464ab2ca5d038151f

> +         for (c = 0; i->defExists(c); ++c)
> +            bld.mkOp3(OP_SHFL, TYPE_F32, tex->getDef(c), tex->getDef(c), lane, quad);
> +      }
>        bld.mkOp(OP_QUADPOP, TYPE_NONE, NULL);
>
>        // save results
> --
> 2.13.6
>