[Mesa-dev] [PATCH v2 8/9] nvc0/ir: Implement TGSI_OPCODE_BALLOT and TGSI_OPCODE_READ_*

[Ilia Mirkin]

On Sun, Apr 9, 2017 at 8:58 PM, Boyan Ding <boyan.j.ding at gmail.com> wrote:
> ---
>  .../drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp  | 28 ++++++++++++++++++++++
>  1 file changed, 28 insertions(+)
>
> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
> index 1bd01a9a32..2ce6f29905 100644
> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
> @@ -978,6 +978,10 @@ static nv50_ir::operation translateOpcode(uint opcode)
>     NV50_IR_OPCODE_CASE(VOTE_ANY, VOTE);
>     NV50_IR_OPCODE_CASE(VOTE_EQ, VOTE);
>
> +   NV50_IR_OPCODE_CASE(BALLOT, VOTE);
> +   NV50_IR_OPCODE_CASE(READ_INVOC, SHFL);
> +   NV50_IR_OPCODE_CASE(READ_FIRST, SHFL);
> +
>     NV50_IR_OPCODE_CASE(END, EXIT);
>
>     default:
> @@ -3431,6 +3435,30 @@ Converter::handleInstruction(const struct tgsi_full_instruction *insn)
>           mkCvt(OP_CVT, TYPE_U32, dst0[c], TYPE_U8, val0);
>        }
>        break;
> +   case TGSI_OPCODE_BALLOT:
> +      val0 = new_LValue(func, FILE_PREDICATE);
> +      mkCmp(OP_SET, CC_NE, TYPE_U32, val0, TYPE_U32, fetchSrc(0, 0), zero);
> +      mkOp1(op, TYPE_U32, dst0[0], val0)->subOp = NV50_IR_SUBOP_VOTE_ANY;
> +      mkMov(dst0[1], zero, TYPE_U32);

Check that dst[n] isn't masked though before writing to it.

> +      break;
> +   case TGSI_OPCODE_READ_FIRST:
> +      // ReadFirstInvocationARB(src) is implemented as
> +      // ReadInvocationARB(src, findLSB(ballot(true)))
> +      val0 = getScratch();
> +      mkOp1(OP_VOTE, TYPE_U32, val0, mkImm(1))->subOp = NV50_IR_SUBOP_VOTE_ANY;
> +      mkOp2(OP_EXTBF, TYPE_U32, val0, val0, mkImm(0x2000))
> +         ->subOp = NV50_IR_SUBOP_EXTBF_REV;
> +      mkOp1(OP_BFIND, TYPE_U32, val0, val0)->subOp = NV50_IR_SUBOP_BFIND_SAMT;
> +      src1 = val0;
> +      /* fallthrough */

You could, of course, do this as:

if (false)

> +   case TGSI_OPCODE_READ_INVOC:
> +      if (tgsi.getOpcode() == TGSI_OPCODE_READ_INVOC)

And then remove this if statement. (Ain't C fun.)

But don't actually do that :) I'm more pointing it out due to the crazy factor.

I really do hate that if for somewhat irrational reasons though...
can't think of a clean way of getting rid of it. Oh well.

> +         src1 = fetchSrc(1, 0);
> +      FOR_EACH_DST_ENABLED_CHANNEL(0, c, tgsi) {
> +         geni = mkOp3(op, dstTy, dst0[c], fetchSrc(0, c), src1, mkImm(0x1f));
> +         geni->subOp = NV50_IR_SUBOP_SHFL_IDX;
> +      }
> +      break;
>     case TGSI_OPCODE_CLOCK:
>        // Stick the 32-bit clock into the high dword of the logical result.
>        if (!tgsi.getDst(0).isMasked(0))
> --
> 2.12.1
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev