shader: implement RDNA2 scalar quad-mask reduction

This commit is contained in:
nmzik
2026-08-20 07:39:56 +02:00
parent bac6bf65af
commit fbdc1e8714
7 changed files with 56 additions and 2 deletions
@@ -28,6 +28,7 @@ constexpr LowerMap LOWER_OPS[] = {
{Decoder::Opcode::S_FF1_I32_B64, Opcode::FindLsbU64},
{Decoder::Opcode::S_FLBIT_I32_B64, Opcode::FindMsbFromHighU64},
{Decoder::Opcode::S_BITREPLICATE_B64_B32, Opcode::BitReplicateB64B32},
{Decoder::Opcode::S_QUADMASK_B64, Opcode::QuadmaskB64},
{Decoder::Opcode::S_NOT_B32, Opcode::BitwiseNotU32},
{Decoder::Opcode::S_NOT_B64, Opcode::BitwiseNotU64},
{Decoder::Opcode::S_ADD_U32, Opcode::IAddU32},
@@ -511,7 +512,8 @@ bool ScalarResultWritesSccNonZero(Decoder::Opcode opcode) {
case Decoder::Opcode::S_LSHL_B64:
case Decoder::Opcode::S_LSHR_B64:
case Decoder::Opcode::S_BFE_U64:
case Decoder::Opcode::S_WQM_B64: return true;
case Decoder::Opcode::S_WQM_B64:
case Decoder::Opcode::S_QUADMASK_B64: return true;
default: return false;
}
}
@@ -530,7 +532,8 @@ bool ScalarResultIs64Bit(Decoder::Opcode opcode) {
case Decoder::Opcode::S_LSHL_B64:
case Decoder::Opcode::S_LSHR_B64:
case Decoder::Opcode::S_BFE_U64:
case Decoder::Opcode::S_WQM_B64: return true;
case Decoder::Opcode::S_WQM_B64:
case Decoder::Opcode::S_QUADMASK_B64: return true;
default: return false;
}
}
@@ -78,6 +78,7 @@ IR_OPCODE(BitReverseU32, General, SimpleInteger)
IR_OPCODE(BitCountU32, General, SimpleInteger)
IR_OPCODE(BitCountU64, General, SimpleInteger)
IR_OPCODE(BitReplicateB64B32, General, ExtendedInteger)
IR_OPCODE(QuadmaskB64, General, ExtendedInteger)
IR_OPCODE(BitCountAddU32, General, ExtendedInteger)
IR_OPCODE(MaskedBitCountLowU32, General, ExtendedInteger)
IR_OPCODE(MaskedBitCountHighU32, General, ExtendedInteger)