mirrors/shadPS4
1
0
Fork 0
mirror of https://github.com/shadps4-emu/shadPS4.git synced 2025-12-12 14:48:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

shader_recompiler: Implement ff1 with subgroup ops (#3225)

Browse Source
This commit is contained in:
TheTurtle
2025-07-10 21:52:56 +03:00
committed by GitHub
parent 88abb93669
commit 8bc30270c8
8 changed files with 27 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/shader_recompiler/ir/ir_emitter.cpp
View File

@@ -660,6 +660,14 @@ U32 IREmitter::WriteLane(const U32& value, const U32& write_value, const U32& la
return Inst<U32>(Opcode::WriteLane, value, write_value, lane);
}
Value IREmitter::Ballot(const U1& bit) {
return Inst(Opcode::Ballot, bit);
}
U32 IREmitter::BallotFindLsb(const Value& mask) {
return Inst<U32>(Opcode::BallotFindLsb, mask);
}
F32F64 IREmitter::FPAdd(const F32F64& a, const F32F64& b) {
if (a.Type() != b.Type()) {
UNREACHABLE_MSG("Mismatching types {} and {}", a.Type(), b.Type());

2
src/shader_recompiler/ir/ir_emitter.h
View File

@@ -176,6 +176,8 @@ public:
[[nodiscard]] U32 ReadFirstLane(const U32& value);
[[nodiscard]] U32 ReadLane(const U32& value, const U32& lane);
[[nodiscard]] U32 WriteLane(const U32& value, const U32& write_value, const U32& lane);
[[nodiscard]] Value Ballot(const U1& bit);
[[nodiscard]] U32 BallotFindLsb(const Value& mask);
[[nodiscard]] Value CompositeConstruct(const Value& e1, const Value& e2);
[[nodiscard]] Value CompositeConstruct(const Value& e1, const Value& e2, const Value& e3);

2
src/shader_recompiler/ir/opcodes.inc
View File

@@ -472,5 +472,7 @@ OPCODE(QuadShuffle, U32, U32,
OPCODE(ReadFirstLane, U32, U32, )
OPCODE(ReadLane, U32, U32, U32 )
OPCODE(WriteLane, U32, U32, U32, U32 )
OPCODE(Ballot, U32x4, U1, )
OPCODE(BallotFindLsb, U32, U32x4, )
OPCODE(DataAppend, U32, U32, U32 )
OPCODE(DataConsume, U32, U32, U32 )

12
src/shader_recompiler/ir/passes/readlane_elimination_pass.cpp
View File

@@ -95,17 +95,7 @@ void ReadLaneEliminationPass(IR::Program& program) {
if (inst.GetOpcode() != IR::Opcode::ReadLane) {
continue;
}
// Check for the following pattern and replace it with ReadFirstLane
// s_ff1_i32_b64 sgpr, exec
// v_readlane_b32 sdst, vgpr, sgpr
if (const auto lane = inst.Arg(1); !lane.IsImmediate()) {
if (lane.InstRecursive()->GetOpcode() == IR::Opcode::FindILsb64) {
const auto value = inst.Arg(0);
inst.ReplaceOpcode(IR::Opcode::ReadFirstLane);
inst.ClearArgs();
inst.SetArg(0, value);
}
if (!inst.Arg(1).IsImmediate()) {
continue;
}