aco: fix half_pi constant for 16-bit fsin/fcos

This worked because the optimizer didn't consider that the 16-bit
instruction would interpret the inline constant differently. This will
change in the next commit.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5245>

diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index 06b8cab7b726570d6f7454a91b3ce03805ad49e6..f256ec6eb3a9aa9999b0966551fef0fb89771b72 100644 (file)
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -2114,12 +2114,13 @@ void visit_alu_instr(isel_context *ctx, nir_alu_instr *instr)
    case nir_op_fcos: {
       Temp src = as_vgpr(ctx, get_alu_src(ctx, instr->src[0]));
       aco_ptr<Instruction> norm;
-      Temp half_pi = bld.copy(bld.def(s1), Operand(0x3e22f983u));
       if (dst.regClass() == v2b) {
+         Temp half_pi = bld.copy(bld.def(s1), Operand(0x3118u));
          Temp tmp = bld.vop2(aco_opcode::v_mul_f16, bld.def(v1), half_pi, src);
          aco_opcode opcode = instr->op == nir_op_fsin ? aco_opcode::v_sin_f16 : aco_opcode::v_cos_f16;
          bld.vop1(opcode, Definition(dst), tmp);
       } else if (dst.regClass() == v1) {
+         Temp half_pi = bld.copy(bld.def(s1), Operand(0x3e22f983u));
          Temp tmp = bld.vop2(aco_opcode::v_mul_f32, bld.def(v1), half_pi, src);
 
          /* before GFX9, v_sin_f32 and v_cos_f32 had a valid input domain of [-256, +256] */