aco: use p_split_vector for nir_op_unpack_half_*

This enables the use of SDWA if possible

Totals from 9933 (7.27% of 136546) affected shaders (RAVEN):
VGPRs: 731764 -> 731772 (+0.00%); split: -0.00%, +0.00%
CodeSize: 90944852 -> 90671472 (-0.30%); split: -0.30%, +0.00%
Instrs: 17881885 -> 17867831 (-0.08%); split: -0.08%, +0.00%
Cycles: 1597904072 -> 1597771260 (-0.01%); split: -0.01%, +0.00%
VMEM: 1702328 -> 1697383 (-0.29%); split: +0.13%, -0.42%
SMEM: 659583 -> 659049 (-0.08%); split: +0.01%, -0.09%
VClause: 318024 -> 318025 (+0.00%); split: -0.00%, +0.00%
SClause: 631670 -> 631707 (+0.01%); split: -0.01%, +0.01%
Copies: 1504107 -> 1504626 (+0.03%); split: -0.01%, +0.04%
PreVGPRs: 683153 -> 683180 (+0.00%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6777>

diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index e751107..87ff6e4 100644 (file)
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -2629,9 +2629,12 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    }
    case nir_op_unpack_half_2x16_split_x_flush_to_zero:
    case nir_op_unpack_half_2x16_split_x: {
+      Temp src = get_alu_src(ctx, instr->src[0]);
+      if (src.regClass() == v1)
+         src = bld.pseudo(aco_opcode::p_split_vector, bld.def(v2b), bld.def(v2b), src);
       if (dst.regClass() == v1) {
          assert(ctx->block->fp_mode.must_flush_denorms16_64 == (instr->op == nir_op_unpack_half_2x16_split_x_flush_to_zero));
-         bld.vop1(aco_opcode::v_cvt_f32_f16, Definition(dst), get_alu_src(ctx, instr->src[0]));
+         bld.vop1(aco_opcode::v_cvt_f32_f16, Definition(dst), src);
       } else {
          isel_err(&instr->instr, "Unimplemented NIR instr bit size");
       }
@@ -2639,11 +2642,14 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    }
    case nir_op_unpack_half_2x16_split_y_flush_to_zero:
    case nir_op_unpack_half_2x16_split_y: {
+      Temp src = get_alu_src(ctx, instr->src[0]);
+      if (src.regClass() == s1)
+         src = bld.sop2(aco_opcode::s_lshr_b32, bld.def(s1), bld.def(s1, scc), src, Operand(16u));
+      else
+         src = bld.pseudo(aco_opcode::p_split_vector, bld.def(v2b), bld.def(v2b), src).def(1).getTemp();
       if (dst.regClass() == v1) {
          assert(ctx->block->fp_mode.must_flush_denorms16_64 == (instr->op == nir_op_unpack_half_2x16_split_y_flush_to_zero));
-         /* TODO: use SDWA here */
-         bld.vop1(aco_opcode::v_cvt_f32_f16, Definition(dst),
-                  bld.vop2(aco_opcode::v_lshrrev_b32, bld.def(v1), Operand(16u), as_vgpr(ctx, get_alu_src(ctx, instr->src[0]))));
+         bld.vop1(aco_opcode::v_cvt_f32_f16, Definition(dst), src);
       } else {
          isel_err(&instr->instr, "Unimplemented NIR instr bit size");
       }