mirror of
https://github.com/KytyPS5/KytyPS5.git
synced 2026-10-20 19:07:01 +02:00
shader: add missing 64-bit integer image atomics
Co-authored-by: shadowbeat070 <70843316+shadowbeat070@users.noreply.github.com>
This commit is contained in:
@@ -321,15 +321,21 @@ inline constexpr auto EmitImageSampleRaw = EmitImage;
|
||||
inline constexpr auto EmitImageGatherRaw = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicCompareSwap32 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicSwap32 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicSwap64 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicIAdd32 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicIAdd64 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicSMin32 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicUMin32 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicUMin64 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicSMax32 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicUMax32 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicUMax64 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicAnd32 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicAnd64 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicOr32 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicOr64 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicXor32 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicXor64 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicFMin32 = EmitImage;
|
||||
inline constexpr auto EmitImageAtomicFMax32 = EmitImage;
|
||||
void EmitUnreachable(ValueEmitContext& ctx, const IR::Inst& inst);
|
||||
|
||||
@@ -361,10 +361,12 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
|
||||
case IR::ValueOpcode::ImageAtomicCompareSwap32:
|
||||
case IR::ValueOpcode::BufferAtomicCmpSwap32: return spv::OpAtomicCompareExchange;
|
||||
case IR::ValueOpcode::ImageAtomicSwap32:
|
||||
case IR::ValueOpcode::ImageAtomicSwap64:
|
||||
case IR::ValueOpcode::BufferAtomicSwap32:
|
||||
case IR::ValueOpcode::BufferAtomicSwap64:
|
||||
case IR::ValueOpcode::SharedAtomicSwap32: return spv::OpAtomicExchange;
|
||||
case IR::ValueOpcode::ImageAtomicIAdd32:
|
||||
case IR::ValueOpcode::ImageAtomicIAdd64:
|
||||
case IR::ValueOpcode::BufferAtomicIAdd32:
|
||||
case IR::ValueOpcode::SharedAtomicIAdd64:
|
||||
case IR::ValueOpcode::SharedAtomicIAdd32: return spv::OpAtomicIAdd;
|
||||
@@ -374,6 +376,7 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
|
||||
case IR::ValueOpcode::BufferAtomicSMin32:
|
||||
case IR::ValueOpcode::SharedAtomicSMin32: return spv::OpAtomicSMin;
|
||||
case IR::ValueOpcode::ImageAtomicUMin32:
|
||||
case IR::ValueOpcode::ImageAtomicUMin64:
|
||||
case IR::ValueOpcode::BufferAtomicUMin32:
|
||||
case IR::ValueOpcode::SharedAtomicUMin32: return spv::OpAtomicUMin;
|
||||
case IR::ValueOpcode::ImageAtomicSMax32:
|
||||
@@ -384,15 +387,18 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
|
||||
case IR::ValueOpcode::BufferAtomicUMax32:
|
||||
case IR::ValueOpcode::SharedAtomicUMax32: return spv::OpAtomicUMax;
|
||||
case IR::ValueOpcode::ImageAtomicAnd32:
|
||||
case IR::ValueOpcode::ImageAtomicAnd64:
|
||||
case IR::ValueOpcode::BufferAtomicAnd32:
|
||||
case IR::ValueOpcode::BufferAtomicAnd64:
|
||||
case IR::ValueOpcode::SharedAtomicAnd32: return spv::OpAtomicAnd;
|
||||
case IR::ValueOpcode::ImageAtomicOr32:
|
||||
case IR::ValueOpcode::ImageAtomicOr64:
|
||||
case IR::ValueOpcode::BufferAtomicOr32:
|
||||
case IR::ValueOpcode::BufferAtomicOr64:
|
||||
case IR::ValueOpcode::SharedAtomicOr64:
|
||||
case IR::ValueOpcode::SharedAtomicOr32: return spv::OpAtomicOr;
|
||||
case IR::ValueOpcode::ImageAtomicXor32:
|
||||
case IR::ValueOpcode::ImageAtomicXor64:
|
||||
case IR::ValueOpcode::BufferAtomicXor32:
|
||||
case IR::ValueOpcode::SharedAtomicXor32: return spv::OpAtomicXor;
|
||||
default: return spv::OpNop;
|
||||
|
||||
@@ -20,6 +20,12 @@ struct MimgGatherInfo {
|
||||
uint32_t flags = 0;
|
||||
};
|
||||
|
||||
struct MimgAtomicInfo {
|
||||
uint32_t encoding = 0;
|
||||
Opcode decoded = Opcode::UNSUPPORTED;
|
||||
bool supports_64 = false;
|
||||
};
|
||||
|
||||
constexpr ImageDimension DecodeImageDimension(uint32_t dim) {
|
||||
switch (dim) {
|
||||
case 0u: return ImageDimension::Dim1D;
|
||||
@@ -190,17 +196,17 @@ constexpr MimgGatherInfo MIMG_GATHER_OPCODE_LIST[] = {
|
||||
{0x61u, Opcode::IMAGE_GATHER4H, ImageSampleFlagGatherHorizontal},
|
||||
};
|
||||
|
||||
constexpr Detail::OpcodeMap MIMG_ATOMIC_OPCODE_LIST[] = {
|
||||
{0x0fu, Opcode::IMAGE_ATOMIC_SWAP},
|
||||
constexpr MimgAtomicInfo MIMG_ATOMIC_OPCODE_LIST[] = {
|
||||
{0x0fu, Opcode::IMAGE_ATOMIC_SWAP, true},
|
||||
{0x10u, Opcode::IMAGE_ATOMIC_CMPSWAP},
|
||||
{0x11u, Opcode::IMAGE_ATOMIC_ADD},
|
||||
{0x11u, Opcode::IMAGE_ATOMIC_ADD, true},
|
||||
{0x14u, Opcode::IMAGE_ATOMIC_SMIN},
|
||||
{0x15u, Opcode::IMAGE_ATOMIC_UMIN},
|
||||
{0x15u, Opcode::IMAGE_ATOMIC_UMIN, true},
|
||||
{0x16u, Opcode::IMAGE_ATOMIC_SMAX},
|
||||
{0x17u, Opcode::IMAGE_ATOMIC_UMAX},
|
||||
{0x18u, Opcode::IMAGE_ATOMIC_AND},
|
||||
{0x19u, Opcode::IMAGE_ATOMIC_OR},
|
||||
{0x1au, Opcode::IMAGE_ATOMIC_XOR},
|
||||
{0x17u, Opcode::IMAGE_ATOMIC_UMAX, true},
|
||||
{0x18u, Opcode::IMAGE_ATOMIC_AND, true},
|
||||
{0x19u, Opcode::IMAGE_ATOMIC_OR, true},
|
||||
{0x1au, Opcode::IMAGE_ATOMIC_XOR, true},
|
||||
{0x1eu, Opcode::IMAGE_ATOMIC_FMIN},
|
||||
{0x1fu, Opcode::IMAGE_ATOMIC_FMAX},
|
||||
};
|
||||
@@ -210,7 +216,7 @@ constexpr auto MIMG_GATHER_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_GATHER_OPCO
|
||||
constexpr auto MIMG_ATOMIC_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_ATOMIC_OPCODE_LIST);
|
||||
|
||||
Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const MimgGatherInfo* gather,
|
||||
const Detail::OpcodeMap* atomic) {
|
||||
const MimgAtomicInfo* atomic) {
|
||||
if (sample != nullptr) {
|
||||
return Opcode::IMAGE_SAMPLE;
|
||||
}
|
||||
@@ -246,7 +252,7 @@ uint32_t DecodeMimgSampleFlags(const MimgSampleInfo* sample, const MimgGatherInf
|
||||
|
||||
uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
|
||||
const MimgSampleInfo* sample, const MimgGatherInfo* gather,
|
||||
const Detail::OpcodeMap* atomic) {
|
||||
const MimgAtomicInfo* atomic) {
|
||||
if (sample != nullptr) {
|
||||
return ImageSampleAddressComponents(sample->flags, dimension);
|
||||
}
|
||||
@@ -370,7 +376,7 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
|
||||
SetUnsupported(inst, Family::MIMG, opcode, "MIMG image atomic has invalid DMASK");
|
||||
} else if (inst.dmask == mask64) {
|
||||
inst.data_bits = 64u;
|
||||
if (inst.opcode != Opcode::IMAGE_ATOMIC_UMAX) {
|
||||
if (!atomic->supports_64) {
|
||||
SetUnsupported(inst, Family::MIMG, opcode,
|
||||
"MIMG 64-bit image atomic opcode is not implemented");
|
||||
}
|
||||
|
||||
@@ -546,7 +546,10 @@ void Translator::BUFFER_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode
|
||||
}
|
||||
}
|
||||
|
||||
void Translator::IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode) {
|
||||
void Translator::IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode32,
|
||||
IR::ValueOpcode opcode64) {
|
||||
const auto opcode = inst.data_bits == 64u ? opcode64 : opcode32;
|
||||
EXIT_IF(opcode == IR::ValueOpcode::Count);
|
||||
const auto memory = MemoryInfoFromDecoded(inst);
|
||||
const auto resource = GetImageResource(memory);
|
||||
const auto address = MakeImageAddress(inst, MemorySourceAt(inst, 1));
|
||||
@@ -1041,24 +1044,30 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_CMPSWAP:
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicCompareSwap32);
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_SWAP:
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSwap32);
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSwap32,
|
||||
IR::ValueOpcode::ImageAtomicSwap64);
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_ADD:
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicIAdd32);
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicIAdd32,
|
||||
IR::ValueOpcode::ImageAtomicIAdd64);
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_SMIN:
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSMin32);
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_UMIN:
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMin32);
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMin32,
|
||||
IR::ValueOpcode::ImageAtomicUMin64);
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_SMAX:
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSMax32);
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_UMAX:
|
||||
return IMAGE_ATOMIC(inst, inst.data_bits == 64u ? IR::ValueOpcode::ImageAtomicUMax64
|
||||
: IR::ValueOpcode::ImageAtomicUMax32);
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMax32,
|
||||
IR::ValueOpcode::ImageAtomicUMax64);
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_AND:
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicAnd32);
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicAnd32,
|
||||
IR::ValueOpcode::ImageAtomicAnd64);
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_OR:
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicOr32);
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicOr32,
|
||||
IR::ValueOpcode::ImageAtomicOr64);
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_XOR:
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicXor32);
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicXor32,
|
||||
IR::ValueOpcode::ImageAtomicXor64);
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_FMIN:
|
||||
return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicFMin32);
|
||||
case Decoder::Opcode::IMAGE_ATOMIC_FMAX:
|
||||
|
||||
@@ -86,7 +86,8 @@ private:
|
||||
void BUFFER_LOAD(const Decoder::Instruction& inst);
|
||||
void BUFFER_STORE(const Decoder::Instruction& inst);
|
||||
void BUFFER_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
|
||||
void IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
|
||||
void IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode32,
|
||||
IR::ValueOpcode opcode64 = IR::ValueOpcode::Count);
|
||||
void DS_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool returns_value);
|
||||
void FLAT_LOAD(const Decoder::Instruction& inst);
|
||||
void FLAT_STORE(const Decoder::Instruction& inst);
|
||||
|
||||
@@ -224,15 +224,21 @@ ImageOpcodeInfo ImageOpcodeInfoOf(ValueOpcode opcode) {
|
||||
return {ImageAccess::Write, ImageResourceClass::Storage, false};
|
||||
case ValueOpcode::ImageAtomicCompareSwap32:
|
||||
case ValueOpcode::ImageAtomicSwap32:
|
||||
case ValueOpcode::ImageAtomicSwap64:
|
||||
case ValueOpcode::ImageAtomicIAdd32:
|
||||
case ValueOpcode::ImageAtomicIAdd64:
|
||||
case ValueOpcode::ImageAtomicSMin32:
|
||||
case ValueOpcode::ImageAtomicUMin32:
|
||||
case ValueOpcode::ImageAtomicUMin64:
|
||||
case ValueOpcode::ImageAtomicSMax32:
|
||||
case ValueOpcode::ImageAtomicUMax32:
|
||||
case ValueOpcode::ImageAtomicUMax64:
|
||||
case ValueOpcode::ImageAtomicAnd32:
|
||||
case ValueOpcode::ImageAtomicAnd64:
|
||||
case ValueOpcode::ImageAtomicOr32:
|
||||
case ValueOpcode::ImageAtomicOr64:
|
||||
case ValueOpcode::ImageAtomicXor32:
|
||||
case ValueOpcode::ImageAtomicXor64:
|
||||
case ValueOpcode::ImageAtomicFMin32:
|
||||
case ValueOpcode::ImageAtomicFMax32:
|
||||
return {ImageAccess::Atomic, ImageResourceClass::Storage, false};
|
||||
|
||||
@@ -310,15 +310,21 @@ VALUE_OPCODE(ImageSampleRaw, U32x4, ImageResource, SamplerResource, ImageAddress
|
||||
VALUE_OPCODE(ImageGatherRaw, U32x4, ImageResource, SamplerResource, ImageAddress, U32, U32, U32, U32)
|
||||
VALUE_OPCODE(ImageAtomicCompareSwap32, U32, ImageResource, ImageAddress, U32, U32, U1)
|
||||
VALUE_OPCODE(ImageAtomicSwap32, U32, ImageResource, ImageAddress, U32, U1)
|
||||
VALUE_OPCODE(ImageAtomicSwap64, U64, ImageResource, ImageAddress, U64, U1)
|
||||
VALUE_OPCODE(ImageAtomicIAdd32, U32, ImageResource, ImageAddress, U32, U1)
|
||||
VALUE_OPCODE(ImageAtomicIAdd64, U64, ImageResource, ImageAddress, U64, U1)
|
||||
VALUE_OPCODE(ImageAtomicSMin32, U32, ImageResource, ImageAddress, U32, U1)
|
||||
VALUE_OPCODE(ImageAtomicUMin32, U32, ImageResource, ImageAddress, U32, U1)
|
||||
VALUE_OPCODE(ImageAtomicUMin64, U64, ImageResource, ImageAddress, U64, U1)
|
||||
VALUE_OPCODE(ImageAtomicSMax32, U32, ImageResource, ImageAddress, U32, U1)
|
||||
VALUE_OPCODE(ImageAtomicUMax32, U32, ImageResource, ImageAddress, U32, U1)
|
||||
VALUE_OPCODE(ImageAtomicUMax64, U64, ImageResource, ImageAddress, U64, U1)
|
||||
VALUE_OPCODE(ImageAtomicAnd32, U32, ImageResource, ImageAddress, U32, U1)
|
||||
VALUE_OPCODE(ImageAtomicAnd64, U64, ImageResource, ImageAddress, U64, U1)
|
||||
VALUE_OPCODE(ImageAtomicOr32, U32, ImageResource, ImageAddress, U32, U1)
|
||||
VALUE_OPCODE(ImageAtomicOr64, U64, ImageResource, ImageAddress, U64, U1)
|
||||
VALUE_OPCODE(ImageAtomicXor32, U32, ImageResource, ImageAddress, U32, U1)
|
||||
VALUE_OPCODE(ImageAtomicXor64, U64, ImageResource, ImageAddress, U64, U1)
|
||||
VALUE_OPCODE(ImageAtomicFMin32, U32, ImageResource, ImageAddress, U32, U1)
|
||||
VALUE_OPCODE(ImageAtomicFMax32, U32, ImageResource, ImageAddress, U32, U1)
|
||||
VALUE_OPCODE(GetAttribute, U32, U32, U32)
|
||||
|
||||
Reference in New Issue
Block a user