raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-21 11:26:56 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: port 64-bit ALU and atomic opcodes

Browse Source
This commit is contained in:
Senaxx
2026-10-01 18:09:22 +04:00
committed by nmzik
parent 11001c97f7
commit 0d0b215fab
20 changed files with 687 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.cpp
+6
View File
@@ -123,6 +123,7 @@ BufferAccess BufferAccessOf(ValueOpcode opcode) {
case ValueOpcode::BufferAtomicSMax32:
case ValueOpcode::BufferAtomicUMax32:
case ValueOpcode::BufferAtomicAnd32:
case ValueOpcode::BufferAtomicAnd64:
case ValueOpcode::BufferAtomicOr32:
case ValueOpcode::BufferAtomicOr64:
case ValueOpcode::BufferAtomicXor32:
@@ -135,6 +136,7 @@ BufferAccess BufferAccessOf(ValueOpcode opcode) {
uint32_t BufferComponentCount(ValueOpcode opcode) {
switch (opcode) {
case ValueOpcode::BufferAtomicSwap64:
case ValueOpcode::BufferAtomicAnd64:
case ValueOpcode::BufferAtomicOr64:
case ValueOpcode::LoadBufferU32x2:
case ValueOpcode::StoreBufferU32x2: return 2u;
@@ -164,6 +166,7 @@ SharedAccess SharedAccessOf(ValueOpcode opcode) {
case ValueOpcode::SharedAtomicFMax32:
case ValueOpcode::SharedAtomicSwap32:
case ValueOpcode::SharedAtomicIAdd32:
case ValueOpcode::SharedAtomicIAdd64:
case ValueOpcode::SharedAtomicISub32:
case ValueOpcode::SharedAtomicInc32:
case ValueOpcode::SharedAtomicDec32:
@@ -183,6 +186,7 @@ SharedAccess SharedAccessOf(ValueOpcode opcode) {
uint32_t SharedComponentCount(ValueOpcode opcode) {
switch (opcode) {
case ValueOpcode::SharedAtomicIAdd64:
case ValueOpcode::SharedAtomicOr64:
case ValueOpcode::LoadSharedU32x2:
case ValueOpcode::WriteSharedU32x2: return 2u;
@@ -218,7 +222,9 @@ ImageOpcodeInfo ImageOpcodeInfoOf(ValueOpcode opcode) {
return {ImageAccess::Write, ImageResourceClass::Storage, false};
case ValueOpcode::ImageAtomicSwap32:
case ValueOpcode::ImageAtomicIAdd32:
case ValueOpcode::ImageAtomicSMin32:
case ValueOpcode::ImageAtomicUMin32:
case ValueOpcode::ImageAtomicSMax32:
case ValueOpcode::ImageAtomicUMax32:
case ValueOpcode::ImageAtomicAnd32:
case ValueOpcode::ImageAtomicOr32:
src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.inc
+10
View File
@@ -137,6 +137,9 @@ VALUE_OPCODE(SMedTri32, U32, U32, U32, U32)
VALUE_OPCODE(UMedTri32, U32, U32, U32, U32)
VALUE_OPCODE(SLessThan32, U1, U32, U32)
VALUE_OPCODE(SLessThan64, U1, U64, U64)
VALUE_OPCODE(SLessThanEqual64, U1, U64, U64)
VALUE_OPCODE(ULessThanEqual64, U1, U64, U64)
VALUE_OPCODE(UGreaterThanEqual64, U1, U64, U64)
VALUE_OPCODE(ULessThan32, U1, U32, U32)
VALUE_OPCODE(ULessThan64, U1, U64, U64)
VALUE_OPCODE(IEqual32, U1, U32, U32)
@@ -156,6 +159,8 @@ VALUE_OPCODE(LogicalXor, U1, U1, U1)
VALUE_OPCODE(LogicalNot, U1, U1)
VALUE_OPCODE(FPOrdEqual32, U1, F32, F32)
VALUE_OPCODE(FPOrdEqual64, U1, F64, F64)
VALUE_OPCODE(FPOrdLessThanEqual64, U1, F64, F64)
VALUE_OPCODE(FPOrdGreaterThanEqual64, U1, F64, F64)
VALUE_OPCODE(FPUnordEqual32, U1, F32, F32)
VALUE_OPCODE(FPOrdNotEqual32, U1, F32, F32)
VALUE_OPCODE(FPUnordNotEqual32, U1, F32, F32)
@@ -174,6 +179,7 @@ VALUE_OPCODE(FPAdd32, F32, F32, F32)
VALUE_OPCODE(FPSub32, F32, F32, F32)
VALUE_OPCODE(FPFma32, F32, F32, F32, F32)
VALUE_OPCODE(FPMul32, F32, F32, F32)
VALUE_OPCODE(FPAdd64, F64, F64, F64)
VALUE_OPCODE(FPMul64, F64, F64, F64)
VALUE_OPCODE(FPFma64, F64, F64, F64, F64)
VALUE_OPCODE(FPRecip64, F64, F64)
@@ -246,6 +252,7 @@ VALUE_OPCODE(BufferAtomicUMin32, U32, BufferResource, U32, U32, U32, U32, U1)
VALUE_OPCODE(BufferAtomicSMax32, U32, BufferResource, U32, U32, U32, U32, U1)
VALUE_OPCODE(BufferAtomicUMax32, U32, BufferResource, U32, U32, U32, U32, U1)
VALUE_OPCODE(BufferAtomicAnd32, U32, BufferResource, U32, U32, U32, U32, U1)
VALUE_OPCODE(BufferAtomicAnd64, U64, BufferResource, U32, U32, U32, U64, U1)
VALUE_OPCODE(BufferAtomicOr32, U32, BufferResource, U32, U32, U32, U32, U1)
VALUE_OPCODE(BufferAtomicOr64, U64, BufferResource, U32, U32, U32, U64, U1)
VALUE_OPCODE(BufferAtomicXor32, U32, BufferResource, U32, U32, U32, U32, U1)
@@ -268,6 +275,7 @@ VALUE_OPCODE(SharedAtomicFMin32, Void, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicFMax32, Void, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicSwap32, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicIAdd32, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicIAdd64, Void, U32, U64, U1)
VALUE_OPCODE(SharedAtomicISub32, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicInc32, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicDec32, U32, U32, U32, U1)
@@ -292,7 +300,9 @@ VALUE_OPCODE(ImageSampleRaw, U32x4, ImageResource, SamplerResource, ImageAddress
VALUE_OPCODE(ImageGatherRaw, U32x4, ImageResource, SamplerResource, ImageAddress)
VALUE_OPCODE(ImageAtomicSwap32, U32, ImageResource, ImageAddress, U32, U1)
VALUE_OPCODE(ImageAtomicIAdd32, U32, ImageResource, ImageAddress, U32, U1)
VALUE_OPCODE(ImageAtomicSMin32, U32, ImageResource, ImageAddress, U32, U1)
VALUE_OPCODE(ImageAtomicUMin32, U32, ImageResource, ImageAddress, U32, U1)
VALUE_OPCODE(ImageAtomicSMax32, U32, ImageResource, ImageAddress, U32, U1)
VALUE_OPCODE(ImageAtomicUMax32, U32, ImageResource, ImageAddress, U32, U1)
VALUE_OPCODE(ImageAtomicAnd32, U32, ImageResource, ImageAddress, U32, U1)
VALUE_OPCODE(ImageAtomicOr32, U32, ImageResource, ImageAddress, U32, U1)
src/graphics/shader/recompiler/ir/passes/ConstantPropagation.cpp
+11
View File
@@ -575,6 +575,17 @@ void FoldInstruction(Block& block, Block::iterator instruction,
case ValueOpcode::UGreaterThan64:
FoldU64Compare(inst, [](uint64_t a, uint64_t b) { return a > b; });
return;
case ValueOpcode::ULessThanEqual64:
FoldU64Compare(inst, [](uint64_t a, uint64_t b) { return a <= b; });
return;
case ValueOpcode::UGreaterThanEqual64:
FoldU64Compare(inst, [](uint64_t a, uint64_t b) { return a >= b; });
return;
case ValueOpcode::SLessThanEqual64:
FoldU64Compare(inst, [](uint64_t a, uint64_t b) {
return std::bit_cast<int64_t>(a) <= std::bit_cast<int64_t>(b);
});
return;
case ValueOpcode::SLessThan64:
FoldU64Compare(inst, [](uint64_t a, uint64_t b) {
return std::bit_cast<int64_t>(a) < std::bit_cast<int64_t>(b);