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 fallback path for missing shaderFloat16 support. (#3604)

Browse Source
This commit is contained in:
squidbus
2025-09-18 00:43:47 -07:00
committed by GitHub
parent 1e949c5813
commit 0eff74223a
15 changed files with 31 additions and 421 deletions
Download Patch File Download Diff File Expand all files Collapse all files

56
src/shader_recompiler/ir/ir_emitter.cpp
View File

@@ -708,14 +708,10 @@ Value IREmitter::CompositeConstruct(const Value& e1, const Value& e2) {
return Inst(Opcode::CompositeConstructU32x2, e1, e2);
case Type::U32x2:
return Inst(Opcode::CompositeConstructU32x2x2, e1, e2);
case Type::F16:
return Inst(Opcode::CompositeConstructF16x2, e1, e2);
case Type::F32:
return Inst(Opcode::CompositeConstructF32x2, e1, e2);
case Type::F32x2:
return Inst(Opcode::CompositeConstructF32x2x2, e1, e2);
case Type::F64:
return Inst(Opcode::CompositeConstructF64x2, e1, e2);
default:
ThrowInvalidType(e1.Type());
}
@@ -728,12 +724,8 @@ Value IREmitter::CompositeConstruct(const Value& e1, const Value& e2, const Valu
switch (e1.Type()) {
case Type::U32:
return Inst(Opcode::CompositeConstructU32x3, e1, e2, e3);
case Type::F16:
return Inst(Opcode::CompositeConstructF16x3, e1, e2, e3);
case Type::F32:
return Inst(Opcode::CompositeConstructF32x3, e1, e2, e3);
case Type::F64:
return Inst(Opcode::CompositeConstructF64x3, e1, e2, e3);
default:
ThrowInvalidType(e1.Type());
}
@@ -748,12 +740,8 @@ Value IREmitter::CompositeConstruct(const Value& e1, const Value& e2, const Valu
switch (e1.Type()) {
case Type::U32:
return Inst(Opcode::CompositeConstructU32x4, e1, e2, e3, e4);
case Type::F16:
return Inst(Opcode::CompositeConstructF16x4, e1, e2, e3, e4);
case Type::F32:
return Inst(Opcode::CompositeConstructF32x4, e1, e2, e3, e4);
case Type::F64:
return Inst(Opcode::CompositeConstructF64x4, e1, e2, e3, e4);
default:
ThrowInvalidType(e1.Type());
}
@@ -787,24 +775,12 @@ Value IREmitter::CompositeExtract(const Value& vector, size_t element) {
return read(Opcode::CompositeExtractU32x3, 3);
case Type::U32x4:
return read(Opcode::CompositeExtractU32x4, 4);
case Type::F16x2:
return read(Opcode::CompositeExtractF16x2, 2);
case Type::F16x3:
return read(Opcode::CompositeExtractF16x3, 3);
case Type::F16x4:
return read(Opcode::CompositeExtractF16x4, 4);
case Type::F32x2:
return read(Opcode::CompositeExtractF32x2, 2);
case Type::F32x3:
return read(Opcode::CompositeExtractF32x3, 3);
case Type::F32x4:
return read(Opcode::CompositeExtractF32x4, 4);
case Type::F64x2:
return read(Opcode::CompositeExtractF64x2, 2);
case Type::F64x3:
return read(Opcode::CompositeExtractF64x3, 3);
case Type::F64x4:
return read(Opcode::CompositeExtractF64x4, 4);
default:
ThrowInvalidType(vector.Type());
}
@@ -824,24 +800,12 @@ Value IREmitter::CompositeInsert(const Value& vector, const Value& object, size_
return insert(Opcode::CompositeInsertU32x3, 3);
case Type::U32x4:
return insert(Opcode::CompositeInsertU32x4, 4);
case Type::F16x2:
return insert(Opcode::CompositeInsertF16x2, 2);
case Type::F16x3:
return insert(Opcode::CompositeInsertF16x3, 3);
case Type::F16x4:
return insert(Opcode::CompositeInsertF16x4, 4);
case Type::F32x2:
return insert(Opcode::CompositeInsertF32x2, 2);
case Type::F32x3:
return insert(Opcode::CompositeInsertF32x3, 3);
case Type::F32x4:
return insert(Opcode::CompositeInsertF32x4, 4);
case Type::F64x2:
return insert(Opcode::CompositeInsertF64x2, 2);
case Type::F64x3:
return insert(Opcode::CompositeInsertF64x3, 3);
case Type::F64x4:
return insert(Opcode::CompositeInsertF64x4, 4);
default:
ThrowInvalidType(vector.Type());
}
@@ -862,12 +826,8 @@ Value IREmitter::CompositeShuffle(const Value& vector1, const Value& vector2, si
switch (vector1.Type()) {
case Type::U32x4:
return shuffle(Opcode::CompositeShuffleU32x2);
case Type::F16x4:
return shuffle(Opcode::CompositeShuffleF16x2);
case Type::F32x4:
return shuffle(Opcode::CompositeShuffleF32x2);
case Type::F64x4:
return shuffle(Opcode::CompositeShuffleF64x2);
default:
ThrowInvalidType(vector1.Type());
}
@@ -888,12 +848,8 @@ Value IREmitter::CompositeShuffle(const Value& vector1, const Value& vector2, si
switch (vector1.Type()) {
case Type::U32x4:
return shuffle(Opcode::CompositeShuffleU32x3);
case Type::F16x4:
return shuffle(Opcode::CompositeShuffleF16x3);
case Type::F32x4:
return shuffle(Opcode::CompositeShuffleF32x3);
case Type::F64x4:
return shuffle(Opcode::CompositeShuffleF64x3);
default:
ThrowInvalidType(vector1.Type());
}
@@ -916,12 +872,8 @@ Value IREmitter::CompositeShuffle(const Value& vector1, const Value& vector2, si
switch (vector1.Type()) {
case Type::U32x4:
return shuffle(Opcode::CompositeShuffleU32x4);
case Type::F16x4:
return shuffle(Opcode::CompositeShuffleF16x4);
case Type::F32x4:
return shuffle(Opcode::CompositeShuffleF32x4);
case Type::F64x4:
return shuffle(Opcode::CompositeShuffleF64x4);
default:
ThrowInvalidType(vector1.Type());
}
@@ -934,18 +886,10 @@ Value IREmitter::Select(const U1& condition, const Value& true_value, const Valu
switch (true_value.Type()) {
case Type::U1:
return Inst(Opcode::SelectU1, condition, true_value, false_value);
case Type::U8:
return Inst(Opcode::SelectU8, condition, true_value, false_value);
case Type::U16:
return Inst(Opcode::SelectU16, condition, true_value, false_value);
case Type::U32:
return Inst(Opcode::SelectU32, condition, true_value, false_value);
case Type::U64:
return Inst(Opcode::SelectU64, condition, true_value, false_value);
case Type::F32:
return Inst(Opcode::SelectF32, condition, true_value, false_value);
case Type::F64:
return Inst(Opcode::SelectF64, condition, true_value, false_value);
default:
UNREACHABLE_MSG("Invalid type {}", true_value.Type());
}

28
src/shader_recompiler/ir/opcodes.inc
View File

@@ -165,18 +165,6 @@ OPCODE(CompositeInsertU32x4, U32x4, U32x
OPCODE(CompositeShuffleU32x2, U32x2, U32x2, U32x2, U32, U32, )
OPCODE(CompositeShuffleU32x3, U32x3, U32x3, U32x3, U32, U32, U32, )
OPCODE(CompositeShuffleU32x4, U32x4, U32x4, U32x4, U32, U32, U32, U32, )
OPCODE(CompositeConstructF16x2, F16x2, F16, F16, )
OPCODE(CompositeConstructF16x3, F16x3, F16, F16, F16, )
OPCODE(CompositeConstructF16x4, F16x4, F16, F16, F16, F16, )
OPCODE(CompositeExtractF16x2, F16, F16x2, U32, )
OPCODE(CompositeExtractF16x3, F16, F16x3, U32, )
OPCODE(CompositeExtractF16x4, F16, F16x4, U32, )
OPCODE(CompositeInsertF16x2, F16x2, F16x2, F16, U32, )
OPCODE(CompositeInsertF16x3, F16x3, F16x3, F16, U32, )
OPCODE(CompositeInsertF16x4, F16x4, F16x4, F16, U32, )
OPCODE(CompositeShuffleF16x2, F16x2, F16x2, F16x2, U32, U32, )
OPCODE(CompositeShuffleF16x3, F16x3, F16x3, F16x3, U32, U32, U32, )
OPCODE(CompositeShuffleF16x4, F16x4, F16x4, F16x4, U32, U32, U32, U32, )
OPCODE(CompositeConstructF32x2, F32x2, F32, F32, )
OPCODE(CompositeConstructF32x3, F32x3, F32, F32, F32, )
OPCODE(CompositeConstructF32x4, F32x4, F32, F32, F32, F32, )
@@ -190,27 +178,11 @@ OPCODE(CompositeInsertF32x4, F32x4, F32x
OPCODE(CompositeShuffleF32x2, F32x2, F32x2, F32x2, U32, U32, )
OPCODE(CompositeShuffleF32x3, F32x3, F32x3, F32x3, U32, U32, U32, )
OPCODE(CompositeShuffleF32x4, F32x4, F32x4, F32x4, U32, U32, U32, U32, )
OPCODE(CompositeConstructF64x2, F64x2, F64, F64, )
OPCODE(CompositeConstructF64x3, F64x3, F64, F64, F64, )
OPCODE(CompositeConstructF64x4, F64x4, F64, F64, F64, F64, )
OPCODE(CompositeExtractF64x2, F64, F64x2, U32, )
OPCODE(CompositeExtractF64x3, F64, F64x3, U32, )
OPCODE(CompositeExtractF64x4, F64, F64x4, U32, )
OPCODE(CompositeInsertF64x2, F64x2, F64x2, F64, U32, )
OPCODE(CompositeInsertF64x3, F64x3, F64x3, F64, U32, )
OPCODE(CompositeInsertF64x4, F64x4, F64x4, F64, U32, )
OPCODE(CompositeShuffleF64x2, F64x2, F64x2, F64x2, U32, U32, )
OPCODE(CompositeShuffleF64x3, F64x3, F64x3, F64x3, U32, U32, U32, )
OPCODE(CompositeShuffleF64x4, F64x4, F64x4, F64x4, U32, U32, U32, U32, )
// Select operations
OPCODE(SelectU1, U1, U1, U1, U1, )
OPCODE(SelectU8, U8, U1, U8, U8, )
OPCODE(SelectU16, U16, U1, U16, U16, )
OPCODE(SelectU32, U32, U1, U32, U32, )
OPCODE(SelectU64, U64, U1, U64, U64, )
OPCODE(SelectF32, F32, U1, F32, F32, )
OPCODE(SelectF64, F64, U1, F64, F64, )
// Bitwise conversions
OPCODE(BitCastU16F16, U16, F16, )

13
src/shader_recompiler/ir/passes/constant_propagation_pass.cpp
View File

@@ -403,12 +403,8 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
case IR::Opcode::UnpackSint2_10_10_10:
return FoldInverseFunc(inst, IR::Opcode::PackSint2_10_10_10);
case IR::Opcode::SelectU1:
case IR::Opcode::SelectU8:
case IR::Opcode::SelectU16:
case IR::Opcode::SelectU32:
case IR::Opcode::SelectU64:
case IR::Opcode::SelectF32:
case IR::Opcode::SelectF64:
return FoldSelect(inst);
case IR::Opcode::FPNeg32:
FoldWhenAllImmediates(inst, [](f32 a) { return -a; });
@@ -563,15 +559,6 @@ void ConstantPropagation(IR::Block& block, IR::Inst& inst) {
case IR::Opcode::CompositeExtractF32x4:
return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF32x4,
IR::Opcode::CompositeInsertF32x4);
case IR::Opcode::CompositeExtractF16x2:
return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF16x2,
IR::Opcode::CompositeInsertF16x2);
case IR::Opcode::CompositeExtractF16x3:
return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF16x3,
IR::Opcode::CompositeInsertF16x3);
case IR::Opcode::CompositeExtractF16x4:
return FoldCompositeExtract(inst, IR::Opcode::CompositeConstructF16x4,
IR::Opcode::CompositeInsertF16x4);
case IR::Opcode::ConvertF32F16:
return FoldConvert(inst, IR::Opcode::ConvertF16F32);
case IR::Opcode::ConvertF16F32:

26
src/shader_recompiler/ir/passes/lower_fp64_to_fp32.cpp
View File

@@ -50,32 +50,6 @@ IR::Value F32ToPackedF64(IR::IREmitter& ir, const IR::Value& raw) {
static IR::Opcode Replace(IR::Opcode op) {
switch (op) {
case IR::Opcode::CompositeConstructF64x2:
return IR::Opcode::CompositeConstructF32x2;
case IR::Opcode::CompositeConstructF64x3:
return IR::Opcode::CompositeConstructF32x3;
case IR::Opcode::CompositeConstructF64x4:
return IR::Opcode::CompositeConstructF32x4;
case IR::Opcode::CompositeExtractF64x2:
return IR::Opcode::CompositeExtractF32x2;
case IR::Opcode::CompositeExtractF64x3:
return IR::Opcode::CompositeExtractF32x3;
case IR::Opcode::CompositeExtractF64x4:
return IR::Opcode::CompositeExtractF32x4;
case IR::Opcode::CompositeInsertF64x2:
return IR::Opcode::CompositeInsertF32x2;
case IR::Opcode::CompositeInsertF64x3:
return IR::Opcode::CompositeInsertF32x3;
case IR::Opcode::CompositeInsertF64x4:
return IR::Opcode::CompositeInsertF32x4;
case IR::Opcode::CompositeShuffleF64x2:
return IR::Opcode::CompositeShuffleF32x2;
case IR::Opcode::CompositeShuffleF64x3:
return IR::Opcode::CompositeShuffleF32x3;
case IR::Opcode::CompositeShuffleF64x4:
return IR::Opcode::CompositeShuffleF32x4;
case IR::Opcode::SelectF64:
return IR::Opcode::SelectF64;
case IR::Opcode::FPAbs64:
return IR::Opcode::FPAbs32;
case IR::Opcode::FPAdd64:

1
src/shader_recompiler/ir/passes/shader_info_collection_pass.cpp
View File

@@ -73,6 +73,7 @@ void Visit(Info& info, const IR::Inst& inst) {
break;
case IR::Opcode::ConvertF16F32:
case IR::Opcode::ConvertF32F16:
case IR::Opcode::BitCastU16F16:
case IR::Opcode::BitCastF16U16:
info.uses_fp16 = true;
break;