shader: support IMAGE_ATOMIC_SWAP (#456)

src/graphics/shader/recompiler/backend/spirv/spirvEmitterImage.cpp

@@ -576,6 +576,7 @@ uint32_t StoreTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t d
 
 uint32_t ImageAtomicOpcode(IR::ValueOpcode opcode) {
 	switch (opcode) {
+		case IR::ValueOpcode::ImageAtomicSwap32: return OpAtomicExchange;
 		case IR::ValueOpcode::ImageAtomicIAdd32: return OpAtomicIAdd;
 		case IR::ValueOpcode::ImageAtomicUMin32: return OpAtomicUMin;
 		case IR::ValueOpcode::ImageAtomicUMax32: return OpAtomicUMax;

src/graphics/shader/recompiler/frontend/decode/ImageOps.cpp

@@ -202,6 +202,7 @@ constexpr MimgGatherInfo MIMG_GATHER_OPCODE_LIST[] = {
 };
 
 constexpr MimgAtomicInfo MIMG_ATOMIC_OPCODE_LIST[] = {
+    {0x0fu, "image_atomic_swap", Opcode::IMAGE_ATOMIC_SWAP},
     {0x11u, "image_atomic_add", Opcode::IMAGE_ATOMIC_ADD},
     {0x15u, "image_atomic_umin", Opcode::IMAGE_ATOMIC_UMIN},
     {0x17u, "image_atomic_umax", Opcode::IMAGE_ATOMIC_UMAX},

src/graphics/shader/recompiler/frontend/decode/ShaderDecoder.cpp

@@ -535,6 +535,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::IMAGE_SAMPLE:
 		case Opcode::IMAGE_STORE:
 		case Opcode::IMAGE_STORE_MIP:
+		case Opcode::IMAGE_ATOMIC_SWAP:
 		case Opcode::IMAGE_ATOMIC_ADD:
 		case Opcode::IMAGE_ATOMIC_UMIN:
 		case Opcode::IMAGE_ATOMIC_UMAX:

src/graphics/shader/recompiler/frontend/decode/ShaderDecoder.h

@@ -538,6 +538,7 @@ enum class Opcode {
 	IMAGE_LOAD_MIP,
 	IMAGE_STORE,
 	IMAGE_STORE_MIP,
+	IMAGE_ATOMIC_SWAP,
 	IMAGE_ATOMIC_ADD,
 	IMAGE_ATOMIC_UMIN,
 	IMAGE_ATOMIC_UMAX,

src/graphics/shader/recompiler/frontend/translate/Memory.cpp

@@ -223,7 +223,7 @@ Decoder::Operand MemorySourceAt(const Decoder::Instruction& decoded, uint32_t in
 	if (decoded.family == Decoder::Family::MIMG) {
 		const bool store_or_atomic = decoded.opcode == Decoder::Opcode::IMAGE_STORE ||
 		                             decoded.opcode == Decoder::Opcode::IMAGE_STORE_MIP ||
-		                             (decoded.opcode >= Decoder::Opcode::IMAGE_ATOMIC_ADD &&
+		                             (decoded.opcode >= Decoder::Opcode::IMAGE_ATOMIC_SWAP &&
 		                              decoded.opcode <= Decoder::Opcode::IMAGE_ATOMIC_XOR);
 		if (store_or_atomic) {
 			return index == 0u ? decoded.dst : decoded.src0;
@@ -981,6 +981,8 @@ bool Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::DS_WRXCHG_RTN_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicSwap32, true);
 
+		case Decoder::Opcode::IMAGE_ATOMIC_SWAP:
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSwap32);
 		case Decoder::Opcode::IMAGE_ATOMIC_ADD:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicIAdd32);
 		case Decoder::Opcode::IMAGE_ATOMIC_UMIN:

src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.cpp

@@ -204,6 +204,7 @@ ImageOpcodeInfo ImageOpcodeInfoOf(ValueOpcode opcode) {
 			return {ImageAccess::Read, ImageResourceClass::Sampled, true};
 		case ValueOpcode::ImageWrite:
 			return {ImageAccess::Write, ImageResourceClass::Storage, false};
+		case ValueOpcode::ImageAtomicSwap32:
 		case ValueOpcode::ImageAtomicIAdd32:
 		case ValueOpcode::ImageAtomicUMin32:
 		case ValueOpcode::ImageAtomicUMax32:

src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.inc

@@ -268,6 +268,7 @@ VALUE_OPCODE(ImageRead, U32x4, ImageResource, ImageAddress, U1)
 VALUE_OPCODE(ImageWrite, Void, ImageResource, ImageAddress, U32x4, U1)
 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(ImageAtomicUMin32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicUMax32, U32, ImageResource, ImageAddress, U32, U1)

tests/ShaderRecompilerComputeTests.cpp

@@ -12206,6 +12206,7 @@ CoverageClass ClassifyOpcode(ShaderOpcode opcode,
   case Opcode::IMAGE_LOAD_MIP:
   case Opcode::IMAGE_STORE:
   case Opcode::IMAGE_STORE_MIP:
+  case Opcode::IMAGE_ATOMIC_SWAP:
   case Opcode::IMAGE_ATOMIC_ADD:
   case Opcode::IMAGE_ATOMIC_UMIN:
   case Opcode::IMAGE_ATOMIC_AND:
@@ -20897,6 +20898,39 @@ TestCase ImageStoreAndAtomicShareTypedBinding() {
   return test;
 }
 
+TestCase ImageAtomicSwapReturnsPreviousTexel() {
+  using O = ShaderOpcode;
+
+  std::vector<u32> code;
+  AppendVMovU32(&code, 20, 2);
+  AppendVMovU32(&code, 21, 1);
+  AppendVMovU32(&code, 22, 0);
+  AppendVMovU32(&code, 0, 5);
+  code.push_back(EncodeMimg0(0x08, 0x1));
+  code.push_back(EncodeMimg1(0, 20));
+  AppendVMovU32(&code, 0, 3);
+  code.push_back(EncodeMimg0(0x0f, 0x1, 0, true));
+  code.push_back(EncodeMimg1(0, 20));
+  AppendStoreVgpr(&code, 0, 0);
+  AppendEnd(&code);
+
+  std::vector<u32> expected_image(16, 0);
+  expected_image[1 * 4 + 2] = 3;
+
+  TestCase test;
+  test.name = "ImageAtomicSwapReturnsPreviousTexel";
+  test.code = std::move(code);
+  test.expected = {5};
+  test.opcodes = {O::V_MOV_B32, O::IMAGE_STORE, O::IMAGE_ATOMIC_SWAP,
+                  O::BUFFER_STORE_DWORD, O::S_ENDPGM};
+  test.user_data = MakeStorageTextureData(Prospero::BufferFormat::k32UInt);
+  test.has_user_data = true;
+  test.storage_image_r32ui = std::vector<u32>(16, 0);
+  test.expected_storage_image_r32ui = std::move(expected_image);
+  test.required_spirv = {"OpAtomicExchange", "OpImageTexelPointer"};
+  return test;
+}
+
 TestCase ImageStoreAndAtomicUseSeparateBindings() {
   using O = ShaderOpcode;
 
@@ -21566,6 +21600,7 @@ std::vector<TestCase> MakeCases() {
   AddCase(ImageStorePackedUintHonorsSparseDmask);
   AddCase(ComputeTgSizeSgprUsesWaveMetadata);
   AddCase(ImageStoreAndAtomicShareTypedBinding);
+  AddCase(ImageAtomicSwapReturnsPreviousTexel);
   AddCase(ImageStoreAndAtomicUseSeparateBindings);
   AddCase(ImageAtomicVariants);
   AddCase(ImageAtomicGlc0DoesNotReturnOldValue);