Convert formatted half-float buffer stores from F32

Co-authored-by: TarkusR <90155059+TarkusR@users.noreply.github.com>

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

@@ -710,10 +710,14 @@ void StoreFormattedInBounds(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
                             uint32_t data) {
 	if (component >= plan.info.component_count) return;
 	const auto bits = plan.info.component_bits[component];
-	if (plan.info.type == Format::ComponentType::Snorm && bits == 16u) {
+	if (bits == 16u && (plan.info.type == Format::ComponentType::Snorm ||
+	                    plan.info.type == Format::ComponentType::Float)) {
 		const auto value = EmitBitCastF32U32(ctx.state, data);
-		data = EmitPackSnorm2x16(
-		    ctx.state, EmitCompositeConstructF32x2(ctx.state, value, ConstantF32Value(ctx.state, 0.0f)));
+		const auto pair = EmitCompositeConstructF32x2(ctx.state, value,
+		                                               ConstantF32Value(ctx.state, 0.0f));
+		data = plan.info.type == Format::ComponentType::Float
+		           ? EmitPackHalf2x16(ctx.state, pair)
+		           : EmitPackSnorm2x16(ctx.state, pair);
 	}
 	if (bits == 8u || bits == 16u) {
 		StoreSubwordInBounds(ctx, mem, plan.resource, plan.addresses[component],

tests/ShaderRecompilerComputeTests.cpp

@@ -23706,6 +23706,38 @@ TestCase BufferStoreFormatXResource16UintPreservesAdjacentLanes() {
   return test;
 }
 
+TestCase BufferStoreFormatXyzwFloat16ConvertsComponents() {
+  using O = ShaderOpcode;
+
+  // This finite F32 has low bits 0x7e63, which form an F16 NaN if truncated.
+  constexpr std::array<float, 8> values = {
+      std::bit_cast<float>(0x3f807e63u), -2.0f, 0.5f, 0.25f,
+      -4.0f, 3.0f, 0.0f, 2.0f};
+  std::vector<u32> code;
+  for (u32 record = 0; record < 2; record++) {
+    for (u32 component = 0; component < 4; component++) {
+      AppendVMovLiteral(&code, component,
+                        std::bit_cast<u32>(values[record * 4 + component]));
+    }
+    AppendVMovU32(&code, 20, record);
+    code.push_back(EncodeMubuf0(0x07u, 0, true, false));
+    code.push_back(EncodeMubuf1(0, 0, 20));
+  }
+  AppendEnd(&code);
+
+  TestCase test;
+  test.name = "BufferStoreFormatXyzwFloat16ConvertsComponents";
+  test.code = std::move(code);
+  test.initial = std::vector<u32>(4, 0xdeadbeefu);
+  test.expected = {0xc0003c04u, 0x34003800u, 0x4200c400u, 0x40000000u};
+  test.user_data = MakeStructuredStorageBufferData(
+      8, 2, false, BufferFormat(Prospero::BufferFormat::k16_16_16_16Float));
+  test.has_user_data = true;
+  test.opcodes = {O::V_MOV_B32, O::BUFFER_STORE_FORMAT_XYZW, O::S_ENDPGM};
+  test.required_spirv = {"PackHalf2x16"};
+  return test;
+}
+
 TestCase BufferStoreFormatXyzwSnorm16CapturedSkinningVectors() {
   using O = ShaderOpcode;
 
@@ -28430,6 +28462,7 @@ std::vector<TestCase> MakeCases() {
   AddCase(BufferFormatStoreVariants);
   AddCase(BufferStoreFormatXResource16UintWritesHalfword);
   AddCase(BufferStoreFormatXResource16UintPreservesAdjacentLanes);
+  AddCase(BufferStoreFormatXyzwFloat16ConvertsComponents);
   AddCase(BufferStoreFormatXyzwSnorm16CapturedSkinningVectors);
   AddCase(BufferStoreFormatXSnorm16ClampsRoundsAndPreservesHalfwords);
   AddCase(BufferLoadFormatXResource8UintZeroExtendsByte);
@@ -33158,6 +33191,11 @@ int main(int argc, char **argv) {
     RunCase(&vulkan, BufferFormatStoreVariants());
     return 0;
   }
+  if (argc == 2 && std::strcmp(argv[1], "--buffer-float16-store-only") == 0) {
+    VulkanHarness vulkan;
+    RunCase(&vulkan, BufferStoreFormatXyzwFloat16ConvertsComponents());
+    return 0;
+  }
   if (argc == 2 && std::strcmp(argv[1], "--s-ashr-i64-only") == 0) {
     VulkanHarness vulkan;
     RunCase(&vulkan, ScalarAshrI64Edges(false));

tests/shaderCfgTests.cpp

@@ -6711,9 +6711,9 @@ void TestNewShaderRecompilerFormattedStoreUsesRuntimeArrayLengthOnly() {
   Check(result.resources.buffers.size() == 1 &&
             result.resources.buffers[0].dwords[2] == 5u,
         "formatted store test did not preserve descriptor NumRecords");
-  Check((result.decoded_dump.find("buffer_store_format_x") != std::string::npos),
+  Check((result.decoded_dump.find("BUFFER_STORE_FORMAT_X") != std::string::npos),
         "formatted store regression did not decode buffer_store_format_x");
-  Check((result.ir_dump.find("typed=0 formatted=1") != std::string::npos),
+  Check((result.decoded_dump.find("typed=0 formatted=1") != std::string::npos),
         "formatted store regression did not preserve formatted metadata");
   CheckSpirvBinaryValidates(result.spirv);
 
@@ -6723,6 +6723,15 @@ void TestNewShaderRecompilerFormattedStoreUsesRuntimeArrayLengthOnly() {
   Check(
       !SpirvSourceHasInstructionUsing(source, "OpULessThan", "%uint_5"),
       "formatted store SPIR-V baked descriptor NumRecords into a store guard");
+
+  user_data[1] = 8u << 16u;
+  user_data[3] = static_cast<uint32_t>(Prospero::BufferFormat::k16_16_16_16Float) << 12u;
+  options.user_data = user_data;
+  result = RecompileForTest(shader, options);
+  Check((DisassembleSpirvBinary(result.spirv).find("PackHalf2x16") !=
+         std::string::npos),
+        "formatted half-float store did not convert F32 to F16");
+  CheckSpirvBinaryValidates(result.spirv);
 }
 
 void TestNewShaderRecompilerTypedBufferTranslation() {
@@ -13515,6 +13524,7 @@ int main() {
   TestNewShaderRecompilerNativeWideBufferIr();
   TestNewShaderRecompilerScalarB64LaneTranslation();
   TestNewShaderRecompilerMubufFormatTranslation();
+  TestNewShaderRecompilerFormattedStoreUsesRuntimeArrayLengthOnly();
   TestNewShaderRecompilerTypedBufferTranslation();
   TestNewShaderRecompilerDsReadWrite2Translation();
   TestNewShaderRecompilerDsWideAndAtomicTranslation();