raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-28 05:46:56 +01:00
Code Issues Packages Projects Releases Wiki Activity

refactor(shader): compact whole-quad mask lowering

Browse Source
This commit is contained in:
nmzik
2026-08-19 08:21:18 +02:00
parent 5383972e96
commit c9e942a972
9 changed files with 186 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/ValueProgram.cpp
+13
View File
@@ -304,6 +304,13 @@ bool ValidateValueProgram(const ValueProgram& program, bool require_ssa, std::st
return Fail(error, fmt::format("untyped opcode {} survived translation",
ValueOpcodeName(inst.GetOpcode())));
}
const bool fixed_signature =
inst.GetOpcode() != ValueOpcode::Phi && inst.GetOpcode() != ValueOpcode::Identity;
if (fixed_signature && inst.NumArgs() != NumArgsOf(inst.GetOpcode())) {
return Fail(error, fmt::format("{} has {} arguments, expected {}",
ValueOpcodeName(inst.GetOpcode()), inst.NumArgs(),
NumArgsOf(inst.GetOpcode())));
}
const auto buffer_components = BufferComponentCount(inst.GetOpcode());
if (buffer_components != 0u) {
const auto memory_index = inst.Flags<MemoryFlags>().index;
@@ -350,6 +357,12 @@ bool ValidateValueProgram(const ValueProgram& program, bool require_ssa, std::st
return Fail(error, fmt::format("{} has an empty argument",
ValueOpcodeName(inst.GetOpcode())));
}
if (fixed_signature && arg.GetType() != ArgTypeOf(inst.GetOpcode(), arg_index)) {
return Fail(error, fmt::format("{} argument {} has type {}, expected {}",
ValueOpcodeName(inst.GetOpcode()), arg_index,
TypeName(arg.GetType()),
TypeName(ArgTypeOf(inst.GetOpcode(), arg_index))));
}
if (const auto* definition = arg.TryInstruction();
definition != nullptr && !instructions.contains(definition)) {
return Fail(error, "value IR argument has a foreign definition");
src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.inc
+1
View File
@@ -79,6 +79,7 @@ VALUE_OPCODE(BitFieldUExtract, U32, U32, U32, U32)
VALUE_OPCODE(BitFieldSExtract, U32, U32, U32, U32)
VALUE_OPCODE(DppMoveU32, U32, U32, U1)
VALUE_OPCODE(DppUpdateU32, U32, U32, U32, U1)
VALUE_OPCODE(WqmU64, U64, U64)
VALUE_OPCODE(WqmMask, U1, U1)
// Core value operations used to expand architectural side effects.
src/graphics/shader/recompiler/ir/passes/ConstantPropagation.cpp
+15
View File
@@ -357,6 +357,21 @@ void FoldInstruction(Inst& inst) {
ReplaceBinaryIdentity(inst, Type::U64, 1u);
}
return;
case ValueOpcode::WqmU64: {
const auto value = Arg(inst, 0);
if (IsImmediate(value, Type::U64)) {
const auto expand = [](uint32_t word) {
auto quads = word | (word >> 1u);
quads |= quads >> 2u;
return (quads & 0x11111111u) * 0x0fu;
};
const auto low = expand(static_cast<uint32_t>(value.U64()));
const auto high = expand(static_cast<uint32_t>(value.U64() >> 32u));
Replace(inst, Value(static_cast<uint64_t>(low) |
(static_cast<uint64_t>(high) << 32u)));
}
return;
}
case ValueOpcode::UDiv32: {
const auto rhs = Arg(inst, 1);
if (IsImmediate(rhs, Type::U32) && rhs.U32() == 1u) {