refactor(shader): preserve native wide shared operations

This commit is contained in:
nmzik
2026-08-19 09:11:25 +02:00
parent c9e942a972
commit c236ba713d
25 changed files with 1486 additions and 986 deletions
@@ -392,19 +392,20 @@ constexpr LowerMap LOWER_OPS[] = {
{Decoder::Opcode::DS_READ_U8, Opcode::DsReadUbyte},
{Decoder::Opcode::DS_READ_I16, Opcode::DsReadSshort},
{Decoder::Opcode::DS_READ_U16, Opcode::DsReadUshort},
{Decoder::Opcode::DS_READ2_B32, Opcode::DsReadB32},
{Decoder::Opcode::DS_READ2_B32, Opcode::DsRead2B32},
{Decoder::Opcode::DS_READ2ST64_B32, Opcode::DsRead2B32},
{Decoder::Opcode::DS_READ_B32, Opcode::DsReadB32},
{Decoder::Opcode::DS_READ_B64, Opcode::DsReadB32},
{Decoder::Opcode::DS_READ2_B64, Opcode::DsReadB32},
{Decoder::Opcode::DS_READ2ST64_B64, Opcode::DsReadB32},
{Decoder::Opcode::DS_READ2_B64, Opcode::DsRead2B32},
{Decoder::Opcode::DS_READ2ST64_B64, Opcode::DsRead2B32},
{Decoder::Opcode::DS_READ_B96, Opcode::DsReadB32},
{Decoder::Opcode::DS_READ_B128, Opcode::DsReadB32},
{Decoder::Opcode::DS_WRITE_B8, Opcode::DsWriteByte},
{Decoder::Opcode::DS_WRITE_B16, Opcode::DsWriteShort},
{Decoder::Opcode::DS_WRITE2_B32, Opcode::DsWriteB32},
{Decoder::Opcode::DS_WRITE2ST64_B32, Opcode::DsWriteB32},
{Decoder::Opcode::DS_WRITE2_B64, Opcode::DsWriteB32},
{Decoder::Opcode::DS_WRITE2ST64_B64, Opcode::DsWriteB32},
{Decoder::Opcode::DS_WRITE2_B32, Opcode::DsWrite2B32},
{Decoder::Opcode::DS_WRITE2ST64_B32, Opcode::DsWrite2B32},
{Decoder::Opcode::DS_WRITE2_B64, Opcode::DsWrite2B32},
{Decoder::Opcode::DS_WRITE2ST64_B64, Opcode::DsWrite2B32},
{Decoder::Opcode::DS_WRITE_B32, Opcode::DsWriteB32},
{Decoder::Opcode::DS_WRITE_B64, Opcode::DsWriteB32},
{Decoder::Opcode::DS_WRITE_B96, Opcode::DsWriteB32},
@@ -331,9 +331,11 @@ IR_OPCODE(DsReadSbyte, General, Memory)
IR_OPCODE(DsReadUshort, General, Memory)
IR_OPCODE(DsReadSshort, General, Memory)
IR_OPCODE(DsReadB32, General, Memory)
IR_OPCODE(DsRead2B32, General, Memory)
IR_OPCODE(DsWriteByte, General, Memory)
IR_OPCODE(DsWriteShort, General, Memory)
IR_OPCODE(DsWriteB32, General, Memory)
IR_OPCODE(DsWrite2B32, General, Memory)
IR_OPCODE(DsMinF32, General, Memory)
IR_OPCODE(DsMaxF32, General, Memory)
IR_OPCODE(DsSwizzleB32, General, Memory)
@@ -42,8 +42,6 @@ constexpr Type BufferResource = Type::BufferResource;
constexpr Type AddressResource = Type::AddressResource;
constexpr Type ImageResource = Type::ImageResource;
constexpr Type SamplerResource = Type::SamplerResource;
constexpr Type LdsResource = Type::LdsResource;
constexpr Type GdsResource = Type::GdsResource;
constexpr Type ImageAddress = Type::ImageAddress;
constexpr std::array<OpcodeMeta, static_cast<size_t>(ValueOpcode::Count)> MetaTable = {{
@@ -76,52 +74,23 @@ bool HasSideEffects(ValueOpcode opcode) {
if (buffer_access == BufferAccess::Write || buffer_access == BufferAccess::Atomic) {
return true;
}
const auto shared_access = SharedAccessOf(opcode);
if (shared_access == SharedAccess::Write || shared_access == SharedAccess::Atomic ||
shared_access == SharedAccess::Append || shared_access == SharedAccess::Consume) {
return true;
}
switch (opcode) {
case ValueOpcode::Reference:
case ValueOpcode::ReferenceU32:
case ValueOpcode::StoreAddressU8:
case ValueOpcode::StoreAddressU16:
case ValueOpcode::StoreAddressU32:
case ValueOpcode::WriteSharedU8:
case ValueOpcode::WriteSharedU16:
case ValueOpcode::WriteSharedU32:
case ValueOpcode::WriteGdsU8:
case ValueOpcode::WriteGdsU16:
case ValueOpcode::WriteGdsU32:
case ValueOpcode::SharedAtomicFMin32:
case ValueOpcode::SharedAtomicFMax32:
case ValueOpcode::GdsAtomicFMin32:
case ValueOpcode::GdsAtomicFMax32:
case ValueOpcode::SharedAtomicSwap32:
case ValueOpcode::SharedAtomicIAdd32:
case ValueOpcode::SharedAtomicISub32:
case ValueOpcode::SharedAtomicSMin32:
case ValueOpcode::SharedAtomicUMin32:
case ValueOpcode::SharedAtomicSMax32:
case ValueOpcode::SharedAtomicUMax32:
case ValueOpcode::SharedAtomicAnd32:
case ValueOpcode::SharedAtomicOr32:
case ValueOpcode::SharedAtomicXor32:
case ValueOpcode::GdsAtomicSwap32:
case ValueOpcode::GdsAtomicIAdd32:
case ValueOpcode::GdsAtomicISub32:
case ValueOpcode::GdsAtomicSMin32:
case ValueOpcode::GdsAtomicUMin32:
case ValueOpcode::GdsAtomicSMax32:
case ValueOpcode::GdsAtomicUMax32:
case ValueOpcode::GdsAtomicAnd32:
case ValueOpcode::GdsAtomicOr32:
case ValueOpcode::GdsAtomicXor32:
case ValueOpcode::ImageAtomicIAdd32:
case ValueOpcode::ImageAtomicUMin32:
case ValueOpcode::ImageAtomicUMax32:
case ValueOpcode::ImageAtomicAnd32:
case ValueOpcode::ImageAtomicOr32:
case ValueOpcode::ImageAtomicXor32:
case ValueOpcode::DataAppend:
case ValueOpcode::DataConsume:
case ValueOpcode::GdsDataAppend:
case ValueOpcode::GdsDataConsume:
case ValueOpcode::ImageWrite:
case ValueOpcode::SetAttribute:
case ValueOpcode::Barrier: return true;
@@ -172,4 +141,48 @@ uint32_t BufferComponentCount(ValueOpcode opcode) {
}
}
SharedAccess SharedAccessOf(ValueOpcode opcode) {
switch (opcode) {
case ValueOpcode::LoadSharedU8:
case ValueOpcode::LoadSharedU16:
case ValueOpcode::LoadSharedU32:
case ValueOpcode::LoadSharedU32x2:
case ValueOpcode::LoadSharedU32x3:
case ValueOpcode::LoadSharedU32x4: return SharedAccess::Read;
case ValueOpcode::WriteSharedU8:
case ValueOpcode::WriteSharedU16:
case ValueOpcode::WriteSharedU32:
case ValueOpcode::WriteSharedU32x2:
case ValueOpcode::WriteSharedU32x3:
case ValueOpcode::WriteSharedU32x4: return SharedAccess::Write;
case ValueOpcode::SharedAtomicFMin32:
case ValueOpcode::SharedAtomicFMax32:
case ValueOpcode::SharedAtomicSwap32:
case ValueOpcode::SharedAtomicIAdd32:
case ValueOpcode::SharedAtomicISub32:
case ValueOpcode::SharedAtomicSMin32:
case ValueOpcode::SharedAtomicUMin32:
case ValueOpcode::SharedAtomicSMax32:
case ValueOpcode::SharedAtomicUMax32:
case ValueOpcode::SharedAtomicAnd32:
case ValueOpcode::SharedAtomicOr32:
case ValueOpcode::SharedAtomicXor32: return SharedAccess::Atomic;
case ValueOpcode::DataAppend: return SharedAccess::Append;
case ValueOpcode::DataConsume: return SharedAccess::Consume;
default: return SharedAccess::None;
}
}
uint32_t SharedComponentCount(ValueOpcode opcode) {
switch (opcode) {
case ValueOpcode::LoadSharedU32x2:
case ValueOpcode::WriteSharedU32x2: return 2u;
case ValueOpcode::LoadSharedU32x3:
case ValueOpcode::WriteSharedU32x3: return 3u;
case ValueOpcode::LoadSharedU32x4:
case ValueOpcode::WriteSharedU32x4: return 4u;
default: return SharedAccessOf(opcode) == SharedAccess::None ? 0u : 1u;
}
}
} // namespace Libs::Graphics::ShaderRecompiler::IR
@@ -17,6 +17,7 @@ enum class ValueOpcode {
};
enum class BufferAccess { None, Read, Write, Atomic };
enum class SharedAccess { None, Read, Write, Atomic, Append, Consume };
struct DppMoveFlags {
uint16_t control = 0;
@@ -56,6 +57,8 @@ static_assert(std::is_trivially_copyable_v<ExportFlags>);
[[nodiscard]] bool HasSideEffects(ValueOpcode opcode);
[[nodiscard]] BufferAccess BufferAccessOf(ValueOpcode opcode);
[[nodiscard]] uint32_t BufferComponentCount(ValueOpcode opcode);
[[nodiscard]] SharedAccess SharedAccessOf(ValueOpcode opcode);
[[nodiscard]] uint32_t SharedComponentCount(ValueOpcode opcode);
[[nodiscard]] std::string_view ValueOpcodeName(ValueOpcode opcode);
} // namespace Libs::Graphics::ShaderRecompiler::IR
@@ -202,8 +202,6 @@ VALUE_OPCODE(GetBufferResource, BufferResource, U32, U32, U32, U32)
VALUE_OPCODE(GetAddressResource, AddressResource, U32, U32)
VALUE_OPCODE(GetImageResource, ImageResource, U32, U32, U32, U32, U32, U32, U32, U32)
VALUE_OPCODE(GetSamplerResource, SamplerResource, U32, U32, U32, U32)
VALUE_OPCODE(GetLdsResource, LdsResource)
VALUE_OPCODE(GetGdsResource, GdsResource)
VALUE_OPCODE(MakeImageAddress, ImageAddress, U32, U32, U32, U32, U32, U32, U32, U32, U32, U32, U32,
U32, U32)
@@ -240,46 +238,32 @@ VALUE_OPCODE(BufferAtomicXor32, U32, BufferResource, U32, U32, U32, U32, U1)
VALUE_OPCODE(BufferAtomicFMin32, U32, BufferResource, U32, U32, U32, U32, U1)
VALUE_OPCODE(BufferAtomicFMax32, U32, BufferResource, U32, U32, U32, U32, U1)
VALUE_OPCODE(LoadSharedU8, U8, LdsResource, U32, U1)
VALUE_OPCODE(LoadSharedU16, U16, LdsResource, U32, U1)
VALUE_OPCODE(LoadSharedU32, U32, LdsResource, U32, U1)
VALUE_OPCODE(WriteSharedU8, Void, LdsResource, U32, U8, U1)
VALUE_OPCODE(WriteSharedU16, Void, LdsResource, U32, U16, U1)
VALUE_OPCODE(WriteSharedU32, Void, LdsResource, U32, U32, U1)
VALUE_OPCODE(LoadGdsU8, U8, GdsResource, U32, U1)
VALUE_OPCODE(LoadGdsU16, U16, GdsResource, U32, U1)
VALUE_OPCODE(LoadGdsU32, U32, GdsResource, U32, U1)
VALUE_OPCODE(WriteGdsU8, Void, GdsResource, U32, U8, U1)
VALUE_OPCODE(WriteGdsU16, Void, GdsResource, U32, U16, U1)
VALUE_OPCODE(WriteGdsU32, Void, GdsResource, U32, U32, U1)
VALUE_OPCODE(SharedAtomicFMin32, Void, LdsResource, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicFMax32, Void, LdsResource, U32, U32, U32, U1)
VALUE_OPCODE(GdsAtomicFMin32, Void, GdsResource, U32, U32, U32, U1)
VALUE_OPCODE(GdsAtomicFMax32, Void, GdsResource, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicSwap32, U32, LdsResource, U32, U32, U1)
VALUE_OPCODE(SharedAtomicIAdd32, U32, LdsResource, U32, U32, U1)
VALUE_OPCODE(SharedAtomicISub32, U32, LdsResource, U32, U32, U1)
VALUE_OPCODE(SharedAtomicSMin32, U32, LdsResource, U32, U32, U1)
VALUE_OPCODE(SharedAtomicUMin32, U32, LdsResource, U32, U32, U1)
VALUE_OPCODE(SharedAtomicSMax32, U32, LdsResource, U32, U32, U1)
VALUE_OPCODE(SharedAtomicUMax32, U32, LdsResource, U32, U32, U1)
VALUE_OPCODE(SharedAtomicAnd32, U32, LdsResource, U32, U32, U1)
VALUE_OPCODE(SharedAtomicOr32, U32, LdsResource, U32, U32, U1)
VALUE_OPCODE(SharedAtomicXor32, U32, LdsResource, U32, U32, U1)
VALUE_OPCODE(GdsAtomicSwap32, U32, GdsResource, U32, U32, U1)
VALUE_OPCODE(GdsAtomicIAdd32, U32, GdsResource, U32, U32, U1)
VALUE_OPCODE(GdsAtomicISub32, U32, GdsResource, U32, U32, U1)
VALUE_OPCODE(GdsAtomicSMin32, U32, GdsResource, U32, U32, U1)
VALUE_OPCODE(GdsAtomicUMin32, U32, GdsResource, U32, U32, U1)
VALUE_OPCODE(GdsAtomicSMax32, U32, GdsResource, U32, U32, U1)
VALUE_OPCODE(GdsAtomicUMax32, U32, GdsResource, U32, U32, U1)
VALUE_OPCODE(GdsAtomicAnd32, U32, GdsResource, U32, U32, U1)
VALUE_OPCODE(GdsAtomicOr32, U32, GdsResource, U32, U32, U1)
VALUE_OPCODE(GdsAtomicXor32, U32, GdsResource, U32, U32, U1)
VALUE_OPCODE(DataAppend, U32, LdsResource, U32, U1, U32, U32)
VALUE_OPCODE(DataConsume, U32, LdsResource, U32, U1, U32, U32)
VALUE_OPCODE(GdsDataAppend, U32, GdsResource, U32, U1, U32, U32)
VALUE_OPCODE(GdsDataConsume, U32, GdsResource, U32, U1, U32, U32)
VALUE_OPCODE(LoadSharedU8, U8, U32, U1)
VALUE_OPCODE(LoadSharedU16, U16, U32, U1)
VALUE_OPCODE(LoadSharedU32, U32, U32, U1)
VALUE_OPCODE(LoadSharedU32x2, U32x2, U32, U1)
VALUE_OPCODE(LoadSharedU32x3, U32x3, U32, U1)
VALUE_OPCODE(LoadSharedU32x4, U32x4, U32, U1)
VALUE_OPCODE(WriteSharedU8, Void, U32, U8, U1)
VALUE_OPCODE(WriteSharedU16, Void, U32, U16, U1)
VALUE_OPCODE(WriteSharedU32, Void, U32, U32, U1)
VALUE_OPCODE(WriteSharedU32x2, Void, U32, U32, U32, U1)
VALUE_OPCODE(WriteSharedU32x3, Void, U32, U32, U32, U32, U1)
VALUE_OPCODE(WriteSharedU32x4, Void, U32, U32, U32, U32, U32, U1)
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(SharedAtomicISub32, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicSMin32, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicUMin32, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicSMax32, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicUMax32, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicAnd32, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicOr32, U32, U32, U32, U1)
VALUE_OPCODE(SharedAtomicXor32, U32, U32, U32, U1)
VALUE_OPCODE(DataAppend, U32, U32, U1, U32, U32)
VALUE_OPCODE(DataConsume, U32, U32, U1, U32, U32)
VALUE_OPCODE(SwizzleU32, U32, U32, U32, U1)
VALUE_OPCODE(ImageQueryDimensions, U32x4, ImageResource, ImageAddress)