shader: read sample IDs from pixel ancillary input

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

@@ -72,7 +72,7 @@ uint32_t EmitBuiltinU32(ValueEmitContext& ctx, IR::StageInputKind kind, uint32_t
 		return bits;
 	}
 	if (kind == IR::StageInputKind::VertexIndex || kind == IR::StageInputKind::InstanceIndex ||
-	    kind == IR::StageInputKind::Layer) {
+	    kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId) {
 		const auto value = state.builder.AllocateId();
 		const auto bits  = state.builder.AllocateId();
 		state.builder.AddFunction({OpLoad, TypeI32(state), value, variable});

src/graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h

@@ -43,6 +43,7 @@ enum : uint32_t {
 	CapabilityImageGatherExtended            = 25,
 	CapabilityClipDistance                   = 32,
 	CapabilityCullDistance                   = 33,
+	CapabilitySampleRateShading              = 35,
 	CapabilitySampled1D                      = 43,
 	CapabilityImage1D                        = 44,
 	CapabilityImageQuery                     = 50,
@@ -92,6 +93,7 @@ enum : uint32_t {
 	BuiltInLayer                     = 9,
 	BuiltInFragCoord                 = 15,
 	BuiltInFrontFacing               = 17,
+	BuiltInSampleId                  = 18,
 	BuiltInSampleMask                = 20,
 	BuiltInFragDepth                 = 22,
 	BuiltInWorkgroupId               = 26,

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

@@ -512,6 +512,7 @@ uint32_t BuiltInForInput(IR::StageInputKind kind) {
 		case IR::StageInputKind::FragCoord: return BuiltInFragCoord;
 		case IR::StageInputKind::FrontFacing: return BuiltInFrontFacing;
 		case IR::StageInputKind::Layer: return BuiltInLayer;
+		case IR::StageInputKind::SampleId: return BuiltInSampleId;
 		case IR::StageInputKind::BaryCoordSmooth: return BuiltInBaryCoordKHR;
 		case IR::StageInputKind::BaryCoordNoPerspective: return BuiltInBaryCoordNoPerspKHR;
 		case IR::StageInputKind::WorkgroupId: return BuiltInWorkgroupId;
@@ -535,7 +536,7 @@ void AddInputAnnotationsAndNames(EmitterState& state) {
 	}
 	for (const auto& input: state.inputs) {
 		state.builder.AddName(input.variable_id, input.debug_name.c_str());
-		if (input.kind == IR::StageInputKind::Layer) {
+		if (input.kind == IR::StageInputKind::Layer || input.kind == IR::StageInputKind::SampleId) {
 			state.builder.AddAnnotation({OpDecorate, input.variable_id, DecorationFlat});
 		}
 		if (input.kind == IR::StageInputKind::Parameter) {
@@ -711,6 +712,9 @@ void DefineModule(EmitterState& state) {
 		state.builder.RequireVersion(0x00010500u);
 		state.builder.RequireCapability(CapabilityShaderLayer);
 	}
+	if (InputVariableForKind(state, IR::StageInputKind::SampleId) != 0) {
+		state.builder.RequireCapability(CapabilitySampleRateShading);
+	}
 	if (state.requirements.image_gather_extended) {
 		state.builder.RequireCapability(CapabilityImageGatherExtended);
 	}
@@ -803,6 +807,7 @@ void DefineModule(EmitterState& state) {
 			case IR::StageInputKind::VertexIndex:
 			case IR::StageInputKind::InstanceIndex:
 			case IR::StageInputKind::Layer:
+			case IR::StageInputKind::SampleId:
 				ptr_type = TypePointer(state, StorageClassInput, TypeI32(state));
 				break;
 			case IR::StageInputKind::WorkgroupId:

src/graphics/shader/recompiler/ir/ShaderIR.h

@@ -162,6 +162,7 @@ enum class StageInputKind {
 	FrontFacing,
 	PackedAncillary,
 	Layer,
+	SampleId,
 	BaryCoordSmooth,
 	BaryCoordNoPerspective,
 	WorkgroupId,

src/graphics/shader/recompiler/ir/passes/ConstantPropagation.cpp

@@ -310,17 +310,24 @@ void FoldInstruction(Block& block, Block::iterator instruction) {
 			const auto* source = value.TryInstruction();
 			if (source != nullptr && source->GetOpcode() == ValueOpcode::GetBuiltin &&
 			    source->Arg(0) == Value(static_cast<uint32_t>(StageInputKind::PackedAncillary)) &&
-			    IsImmediate(offset, Type::U32) && IsImmediate(count, Type::U32) &&
+			    IsImmediate(offset, Type::U32) && IsImmediate(count, Type::U32) && count.U32() != 0u) {
-			    offset.U32() >= 16u && offset.U32() < 27u && count.U32() != 0u &&
+				constexpr struct {
-			    count.U32() <= 27u - offset.U32()) {
+					uint32_t       start;
-				// Preserve the extraction (including sign extension) while exposing only its
+					uint32_t       end;
-				// used hardware field. Other live ancillary fields remain unsupported.
+					StageInputKind kind;
-				const auto layer = block.PrependNewInst(
+				} fields[] = {{8u, 12u, StageInputKind::SampleId}, {16u, 27u, StageInputKind::Layer}};
-				    instruction, ValueOpcode::GetBuiltin,
+				for (const auto& field: fields) {
-				    {Value(static_cast<uint32_t>(StageInputKind::Layer)), Value(0u)});
+					if (offset.U32() >= field.start && offset.U32() < field.end &&
-				inst.SetArg(0, Value(&*layer));
+					    count.U32() <= field.end - offset.U32()) {
-				inst.SetArg(1, Value(offset.U32() - 16u));
+						// Preserve extraction and sign extension while exposing only the used field.
-				return;
+						const auto input = block.PrependNewInst(
+						    instruction, ValueOpcode::GetBuiltin,
+						    {Value(static_cast<uint32_t>(field.kind)), Value(0u)});
+						inst.SetArg(0, Value(&*input));
+						inst.SetArg(1, Value(offset.U32() - field.start));
+						return;
+					}
+				}
 			}
 			if (!IsImmediate(value, Type::U32) || !IsImmediate(offset, Type::U32) ||
 			    !IsImmediate(count, Type::U32) || offset.U32() > 32u ||

src/graphics/shader/recompiler/ir/passes/ShaderInfoCollection.cpp

@@ -118,8 +118,9 @@ void ValidateValueReferences(const Program& program, const ShaderInfoOptions& op
 						case StageInputKind::PackedAncillary:
 							return Fail("packed pixel ancillary input has an unsupported live use");
 						case StageInputKind::Layer:
+						case StageInputKind::SampleId:
 							if (program.stage != ShaderType::Pixel || component != 0u) {
-								return Fail("typed layer input is invalid");
+								return Fail("typed pixel scalar input is invalid");
 							}
 							break;
 						case StageInputKind::VertexIndex:
@@ -259,6 +260,7 @@ void CollectBuiltinInputs(const Program& program, ShaderInfo& info) {
 					AddInput(info, kind, 0, 1, "gl_FrontFacing");
 					break;
 				case StageInputKind::Layer: AddInput(info, kind, 0, 1, "gl_Layer"); break;
+				case StageInputKind::SampleId: AddInput(info, kind, 0, 1, "gl_SampleID"); break;
 				case StageInputKind::BaryCoordSmooth:
 					AddInput(info, kind, 0, 3, "gl_BaryCoordKHR");
 					break;

tests/ShaderRecompilerComputeTests.cpp

@@ -1048,6 +1048,7 @@ struct GraphicsCase {
   bool pixel_ancillary = false;
   bool pixel_front_face = false;
   u32 layers = 1;
+  vk::SampleCountFlagBits samples = vk::SampleCountFlagBits::e1;
 };
 
 struct CompiledShader {
@@ -10252,7 +10253,8 @@ public:
                         u32 dwords_per_pixel, vk::ImageLayout final_layout,
                         vk::ImageType image_type, vk::ImageViewType view_type,
                         u32 layers, u32 view_base_layer = 0,
-                        u32 view_layers = 0) {
+                        u32 view_layers = 0,
+                        vk::SampleCountFlagBits samples = vk::SampleCountFlagBits::e1) {
     Image ret;
     ret.format = format;
     ret.width = width;
@@ -10270,7 +10272,7 @@ public:
     image_info.extent.depth = 1;
     image_info.mipLevels = ret.mip_levels;
     image_info.arrayLayers = layers;
-    image_info.samples = vk::SampleCountFlagBits::e1;
+    image_info.samples = samples;
     image_info.tiling = vk::ImageTiling::eOptimal;
     image_info.usage = usage | vk::ImageUsageFlagBits::eTransferDst |
                        vk::ImageUsageFlagBits::eTransferSrc;
@@ -10914,7 +10916,15 @@ public:
                       vk::ImageUsageFlagBits::eColorAttachment, {}, 4,
                       vk::ImageLayout::eGeneral, vk::ImageType::e2D,
                       test.layers > 1 ? vk::ImageViewType::e2DArray : vk::ImageViewType::e2D,
-                      test.layers);
+                      test.layers, 0, 0, test.samples);
+    Image resolved;
+    if (test.samples != vk::SampleCountFlagBits::e1) {
+      resolved = CreateImageMips(
+          test.name, 1, 1, target.format, vk::ImageUsageFlagBits::eColorAttachment,
+          {}, 4, vk::ImageLayout::eGeneral, vk::ImageType::e2D,
+          test.layers > 1 ? vk::ImageViewType::e2DArray : vk::ImageViewType::e2D,
+          test.layers);
+    }
     const std::vector<u32> default_vertices = {
         0xbf800000u, 0xbf800000u, 0x3e800000u, 0x3f000000u, 0x3f400000u,
         0x3f800000u, 0x40400000u, 0xbf800000u, 0x3e800000u, 0x3f000000u,
@@ -11028,7 +11038,7 @@ public:
 
     vk::PipelineMultisampleStateCreateInfo multisample{};
     multisample.sType = vk::StructureType::ePipelineMultisampleStateCreateInfo;
-    multisample.rasterizationSamples = vk::SampleCountFlagBits::e1;
+    multisample.rasterizationSamples = test.samples;
 
     vk::PipelineColorBlendAttachmentState color_attachment{};
     color_attachment.colorWriteMask =
@@ -11069,6 +11079,11 @@ public:
     color.imageLayout = vk::ImageLayout::eGeneral;
     color.loadOp = vk::AttachmentLoadOp::eClear;
     color.storeOp = vk::AttachmentStoreOp::eStore;
+    if (resolved.image != nullptr) {
+      color.resolveMode = vk::ResolveModeFlagBits::eAverage;
+      color.resolveImageView = resolved.view;
+      color.resolveImageLayout = vk::ImageLayout::eGeneral;
+    }
     vk::RenderingInfo rendering{};
     rendering.sType = vk::StructureType::eRenderingInfo;
     rendering.renderArea.extent = {1, 1};
@@ -11091,7 +11106,7 @@ public:
     EndSubmitAndFree(test.name, "graphics", cmd);
     target.layout = vk::ImageLayout::eGeneral;
 
-    auto pixel = ReadImage(test.name, &target);
+    auto pixel = ReadImage(test.name, resolved.image != nullptr ? &resolved : &target);
     pixel.resize(4 * test.layers);
 
     m_device.destroyPipeline(pipeline, nullptr);
@@ -11099,6 +11114,7 @@ public:
     m_device.destroyShaderModule(fragment_module, nullptr);
     m_device.destroyShaderModule(vertex_module, nullptr);
     DestroyBuffer(&vertex_buffer);
+    DestroyImage(&resolved);
     DestroyImage(&target);
     return pixel;
   }
@@ -22449,6 +22465,29 @@ GraphicsCase GraphicsAncillaryLayer(bool front_face) {
   return test;
 }
 
+GraphicsCase GraphicsAncillarySampleId() {
+  GraphicsCase test;
+  test.name = "GraphicsAncillarySampleId";
+  test.pixel_ancillary = true;
+  test.samples = vk::SampleCountFlagBits::e4;
+  AppendVop3(&test.fragment_code, 0x148, 4, Vgpr(2), InlineU32(8), InlineU32(4));
+  AppendVop3(&test.fragment_code, 0x149, 5, Vgpr(2), InlineU32(8), InlineU32(2));
+  AppendVop3(&test.fragment_code, 0x148, 6, Vgpr(2), InlineU32(9), InlineU32(1));
+  test.fragment_code.push_back(EncodeVop1(0x06, 0, Vgpr(4)));
+  test.fragment_code.push_back(EncodeVop1(0x05, 1, Vgpr(5)));
+  test.fragment_code.push_back(EncodeVop1(0x06, 2, Vgpr(6)));
+  AppendVMovLiteral(&test.fragment_code, 3, 0x3f800000u);
+  test.fragment_code.push_back(EncodeExp0(0x00, 0xf));
+  test.fragment_code.push_back(EncodeExp1(0, 1, 2, 3));
+  AppendEnd(&test.fragment_code);
+  test.opcodes = {ShaderOpcode::V_BFE_U32, ShaderOpcode::V_BFE_I32,
+                  ShaderOpcode::V_CVT_F32_U32, ShaderOpcode::V_CVT_F32_I32,
+                  ShaderOpcode::V_MOV_B32, ShaderOpcode::EXP, ShaderOpcode::S_ENDPGM};
+  // Average samples 0..3, their signed low two bits, and their second bit.
+  test.expected_pixel = {0x3fc00000u, 0xbf000000u, 0x3f000000u, 0x3f800000u};
+  return test;
+}
+
 GraphicsCase GraphicsFlatInterpolatorExport() {
   using O = ShaderOpcode;
 
@@ -23012,6 +23051,7 @@ std::vector<GraphicsCase> MakeGraphicsCases() {
       GraphicsInterpolationExport(),
       GraphicsAncillaryLayer(false),
       GraphicsAncillaryLayer(true),
+      GraphicsAncillarySampleId(),
       GraphicsFlatInterpolatorExport(),
       GraphicsDsAddtidScratchExport(),
       GraphicsDirectSgprPushConstantExport(),

tests/shaderCfgTests.cpp

@@ -6011,8 +6011,22 @@ void TestPixelAncillaryLayerInput() {
             unused_result.spirv[1] == 0x00010300u &&
             !SpirvContainsCapability(unused_result.spirv, 69u),
         "unused ancillary input changed the module requirements");
-#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
+
   shader[1] = EncodeVop3Word1(5 + 256, 128 + 8, 128 + 4);
+  auto sample_result = RecompileForTest(shader, options);
+  Check(ProgramHasInput(sample_result.program, StageInputKind::SampleId) &&
+            ProgramHasInput(sample_result.program, StageInputKind::Layer) &&
+            !ProgramHasInput(sample_result.program, StageInputKind::PackedAncillary),
+        "sample index extraction discarded another live ancillary field");
+  Check(SpirvHasDecorationValueWithDecoration(sample_result.spirv, 11u, 18u, 14u) &&
+            SpirvContainsCapability(sample_result.spirv, 35u),
+        "sample index lacks its flat builtin or sample-rate capability");
+  const auto sample_source = DisassembleSpirvBinary(sample_result.spirv);
+  Check(SpirvSourceHasInstructionUsing(sample_source, "OpLoad", "%int %gl_SampleID"),
+        "sample index was not loaded from its scalar integer input");
+  CheckSpirvBinaryValidates(sample_result.spirv);
+#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
+  shader[1] = EncodeVop3Word1(5 + 256, 128 + 12, 128 + 4);
   ExpectFatal([&] { (void)RecompileForTest(shader, options); },
               "unsupported live ancillary field was silently replaced");
 #endif