shader: emit native texel offsets using sample execution context

Co-authored-by: netwarex <mail@pety.me>

src/graphics/host_gpu/renderer/cache/textureCache.cpp

@@ -36,7 +36,9 @@ constexpr uint64_t NumFramesBeforeRemoval = 32;
 [[nodiscard]] bool DecodeColorClear(const TextureCache::ImageDesc& desc, uint8_t code,
                                   vk::ClearColorValue& clear) {
 	const auto& metadata = desc.info.metadata;
-	const auto  format   = desc.view_info.format;
+	const auto  format   = desc.info.IsDepth()
+	                           ? TextureGetSurfaceFormatInfo(desc.info.guest_format).vk_format
+	                           : desc.view_info.format;
 	const bool  cmask    = metadata.kind == ImageMetadataKind::Cmask;
 	if (cmask ? code == 0 : code == 0x20) {
 		// Register clears belong to the color buffer; the texture pipe cannot decode them.
@@ -1059,8 +1061,15 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 		}
 		{
 			std::scoped_lock lock {m_lock};
-			ClearImage(m_scheduler.Current(), id, view.format,
-			           {vk::ImageAspectFlagBits::eColor, view.base_level, view.level_count,
+			const auto& image = m_slot_images[id];
+			const bool depth = image.info.IsDepth();
+			if (depth) {
+				const auto value = clear.color.float32[0];
+				clear.depthStencil = {value, 0u};
+			}
+			ClearImage(m_scheduler.Current(), id, depth ? image.backing.format : view.format,
+			           {depth ? vk::ImageAspectFlagBits::eDepth : vk::ImageAspectFlagBits::eColor,
+			            view.base_level, view.level_count,
 			            image_first + slice, 1}, clear);
 		}
 		// Publish the conversion's expanded keys without treating them as guest writes

src/graphics/host_gpu/renderer/pipeline/descriptors.cpp

@@ -179,7 +179,9 @@ static void ValidateSampledDepthBinding(const ShaderRecompiler::IR::ImageResourc
                                         const ShaderTextureResource& descriptor, const Image& image,
                                         vk::Format view_format, uint64_t size) {
 	const bool resource_ok = IsSupportedSampledDepthResource(resource);
-	const bool encoding_ok = IsSupportedDepthTextureEncoding(descriptor, resource.r128);
+	// Comparison sampling changes the host format, not the guest metadata encoding.
+	const bool encoding_ok = descriptor.TileMode() != Prospero::TileMode::kDepth ||
+	                         IsSupportedDepthTextureEncoding(descriptor, resource.r128);
 	const bool view_ok =
 	    IsSupportedSampledDepthView(image.info.pixel_format, view_format, descriptor.DstSelXYZW());
 	if (resource_ok && encoding_ok && view_ok) {
@@ -622,8 +624,7 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 	if (storage) {
 		ValidateStorageTexture(resource, descriptor, desc.info.data.size);
 	}
-	if (!resource.r128 && descriptor.MetaCompress() && tile != Prospero::TileMode::kDepth &&
-	    !desc.info.IsDepth()) {
+	if (!resource.r128 && descriptor.MetaCompress() && tile != Prospero::TileMode::kDepth) {
 		TileSizeAlign metadata_size {};
 		(void)TileGetDccSize(width, height, volume ? depth : image_layers,
 		                     desc.info.bytes_per_block, physical_levels, tile, metadata_size,

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

@@ -323,25 +323,25 @@ uint32_t QueryDimensions(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
 	return vector;
 }
 
-uint32_t PackedOffset(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR::Inst& address,
-                      const ImageSampleLayout& layout, ImageDimension dimension) {
+uint32_t PackedOffset(ValueEmitContext& ctx, IR::Value value, ImageDimension dimension,
+                      uint32_t& mask) {
 	const auto components = ImageDimensionInfoFor(dimension).spatial_components;
-	const auto zero       = ConstantI32(ctx.state, 0);
-	if (layout.offset == NoImageComponent || mem.image_address_components <= layout.offset) {
-		if (components == 1u) return zero;
-		const auto result = ctx.state.builder.AllocateId();
-		if (components == 3u) {
-			ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 3),
-			                              result, zero, zero, zero);
-		} else {
-			ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 2),
-			                              result, zero, zero);
+	uint32_t values[3] {};
+	if (value.IsImmediate()) {
+		// RDNA2 packs signed six-bit texel offsets into the low bits of each byte.
+		for (uint32_t index = 0; index < components; index++) {
+			const auto bits = (value.U32() >> (index * 8u)) & 0x3fu;
+			values[index] = ConstantI32(ctx.state, static_cast<int32_t>(bits ^ 0x20u) - 0x20);
 		}
-		return result;
+		mask |= spv::ImageOperandsConstOffsetMask;
+		return components == 1u
+		           ? values[0]
+		           : ctx.state.builder.Constant(spv::OpConstantComposite,
+		                                        TypeI32Vector(ctx.state, components),
+		                                        std::span(values).first(components));
 	}
-	const auto packed    = Unary(ctx.state, spv::OpBitcast, TypeI32(ctx.state),
-	                             AddressU32(ctx, mem, address, layout.offset));
-	uint32_t   values[3] = {zero, zero, zero};
+	mask |= spv::ImageOperandsOffsetMask;
+	const auto packed = Unary(ctx.state, spv::OpBitcast, TypeI32(ctx.state), ctx.Def(value));
 	for (uint32_t index = 0; index < components; index++) {
 		values[index] = ctx.state.builder.AllocateId();
 		ctx.state.builder.AddFunction(spv::OpBitFieldSExtract, TypeI32(ctx.state), values[index],
@@ -350,13 +350,8 @@ uint32_t PackedOffset(ValueEmitContext& ctx, const IR::MemoryInfo& mem, const IR
 	}
 	if (components == 1u) return values[0];
 	const auto result = ctx.state.builder.AllocateId();
-	if (components == 3u) {
-		ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 3),
-		                              result, values[0], values[1], values[2]);
-	} else {
-		ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 2),
-		                              result, values[0], values[1]);
-	}
+	ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, components),
+	                              result, std::span(values).first(components));
 	return result;
 }
 
@@ -794,21 +789,28 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const auto  layout         = Layout(mem);
 		const auto  numeric_class  = image.numeric_class;
 		const bool  dref           = HasFlag(mem, Decoder::ImageSampleFlagCompare);
+		IR::Value offset_value;
+		if (layout.offset != NoImageComponent) {
+			const auto value = address->Arg(layout.offset).Resolve();
+			offset_value = IR::ResolveActiveU32(value, inst.Arg(inst.NumArgs() - 1u).Resolve());
+			if (offset_value.IsEmpty()) offset_value = value;
+		}
 		if (dref && state.program.info.images[mem.resource].conversion_format !=
 		                Prospero::BufferFormat::kInvalid) {
 			ctx.Fail(inst, "uses depth comparison with a converted image");
 			return;
 		}
 		if (op == IR::ValueOpcode::ImageGatherRaw) {
-			const auto coord = CoordF32(ctx, mem, *address, layout.coord,
-			                            dimension_info.coordinate_components, image.cube);
 			if (dimension == ImageDimension::Dim1D) {
-				if (dref || !HasFlag(mem, Decoder::ImageSampleFlagLevelZero) ||
+				if (image.indirect_root == mem.resource || dref ||
+				    !HasFlag(mem, Decoder::ImageSampleFlagLevelZero) ||
 				    HasFlag(mem, Decoder::ImageSampleFlagOffset) ||
 				    HasFlag(mem, Decoder::ImageSampleFlagGatherHorizontal)) {
 					ctx.Fail(inst, "has an unsupported 1D gather variant");
 					return;
 				}
+				const auto coord = CoordF32(ctx, mem, *address, layout.coord,
+				                            dimension_info.coordinate_components, image.cube);
 				const auto sample = EmitOneDimensionalGatherLz(ctx, mem, coord, numeric_class);
 				ctx.Define(inst, ResultVector(ctx, UnpackImageGather(ctx, mem, sample),
 				                              numeric_class, false, mem, true));
@@ -829,32 +831,38 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			} else {
 				component_or_dref = ConstantU32(state, ImageGatherSource(state, mem));
 			}
-			uint32_t operand_mask = 0;
-			uint32_t offset = 0;
-			if (HasFlag(mem, Decoder::ImageSampleFlagGatherHorizontal)) {
-				operand_mask = spv::ImageOperandsConstOffsetsMask;
-				offset = HorizontalOffsets(state, dimension);
-			} else if (layout.offset != NoImageComponent) {
-				operand_mask = spv::ImageOperandsOffsetMask;
-				offset = PackedOffset(ctx, mem, *address, layout, dimension);
-			}
 			const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler);
-			const auto EmitGather = [&](uint32_t mip) {
-				const auto sampled = MakeSampledImage(state, mem.resource, sampler_id, mip);
-				const auto sample = state.builder.AllocateId();
-				const std::array operands {operand_mask, offset};
-				state.builder.AddFunction(
-				    dref ? spv::OpImageDrefGather : spv::OpImageGather, result_type,
-				    sample, sampled, coord, component_or_dref,
-				    std::span(operands).first(operand_mask != 0u ? 2u : 0u));
-				return sample;
+			const auto EmitGather = [&](uint32_t resource, uint32_t array_index) {
+				const auto& candidate = state.program.info.images[resource];
+				const auto dimension = candidate.dimension;
+				const auto coord = CoordF32(
+				    ctx, mem, *address, layout.coord,
+				    ImageDimensionInfoFor(dimension).coordinate_components, candidate.cube);
+				uint32_t operand_mask = 0;
+				uint32_t offset = 0;
+				if (HasFlag(mem, Decoder::ImageSampleFlagGatherHorizontal)) {
+					operand_mask = spv::ImageOperandsConstOffsetsMask;
+					offset = HorizontalOffsets(state, dimension);
+				} else if (layout.offset != NoImageComponent) {
+					offset = PackedOffset(ctx, offset_value, dimension, operand_mask);
+				}
+				const auto EmitMip = [&](uint32_t mip) {
+					const auto sampled = MakeSampledImage(state, resource, sampler_id, mip, array_index);
+					const auto sample = state.builder.AllocateId();
+					const std::array operands {operand_mask, offset};
+					state.builder.AddFunction(
+					    dref ? spv::OpImageDrefGather : spv::OpImageGather, result_type,
+					    sample, sampled, coord, component_or_dref,
+					    std::span(operands).first(operand_mask != 0u ? 2u : 0u));
+					return sample;
+				};
+				return candidate.mip_mode == IR::ImageMipMode::Dynamic && candidate.mip_count > 1u
+				           ? EmitIndexSwitch(
+				                 state, GatherMip(ctx, inst, mem, *address, layout, candidate.mip_count),
+				                 candidate.mip_count, result_type, EmitMip)
+				           : EmitMip(0u);
 			};
-			const auto sample = image.mip_mode == IR::ImageMipMode::Dynamic && image.mip_count > 1u
-			                        ? EmitIndexSwitch(
-			                              state, GatherMip(ctx, inst, mem, *address, layout,
-			                                               image.mip_count),
-			                              image.mip_count, result_type, EmitGather)
-			                        : EmitGather(0);
+			const auto sample = EmitImageAccess(ctx, inst, result_type, EmitGather);
 			ctx.Define(inst, ResultVector(ctx, UnpackImageGather(ctx, mem, sample),
 			                              result_numeric_class, false, mem, true));
 			return;
@@ -878,7 +886,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				dref_value = AddressF32(ctx, mem, *address, layout.dref);
 			}
 		}
-		std::array<uint32_t, 3> operands {};
+		std::array<uint32_t, 4> operands {};
 		size_t operand_count = 0;
 		if (HasFlag(mem, Decoder::ImageSampleFlagDerivative)) {
 			operands[0] = spv::ImageOperandsGradMask;
@@ -908,9 +916,20 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			    state, ctx.Def(handle->Arg(0)), static_cast<uint32_t>(samplers.size()),
 			    TypeU32(state), [&](uint32_t ordinal) { return ConstantU32(state, samplers[ordinal]); });
 		}
+		if (layout.offset != NoImageComponent &&
+		    !offset_value.IsImmediate()) {
+			ctx.Fail(inst, "has a nonconstant sample texel offset");
+		}
 		const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler, sampler_index);
 		const auto EmitSample = [&](uint32_t resource, uint32_t array_index) {
 			const auto& candidate = state.program.info.images[resource];
+			auto sample_operands = operands;
+			auto sample_operand_count = operand_count;
+			if (layout.offset != NoImageComponent) {
+				sample_operand_count = std::max(size_t{1}, sample_operand_count);
+				sample_operands[sample_operand_count++] = PackedOffset(
+				    ctx, offset_value, candidate.dimension, sample_operands[0]);
+			}
 			const auto coord =
 			    CoordF32(ctx, mem, *address, layout.coord,
 			             ImageDimensionInfoFor(candidate.dimension).coordinate_components,
@@ -920,7 +939,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			const auto sample = state.builder.AllocateId();
 			state.builder.AddFunction(opcode, result_type, sample, sampled, coord,
 			                          std::span(&dref_value, dref ? 1u : 0u),
-			                          std::span<const uint32_t>(operands).first(operand_count));
+			                          std::span<const uint32_t>(sample_operands).first(sample_operand_count));
 			return sample;
 		};
 		auto result = EmitImageAccess(ctx, inst, result_type, EmitSample);

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

@@ -687,7 +687,8 @@ void Translator::IMAGE_SAMPLE(const Decoder::Instruction& inst) {
 	const auto resource = GetImageResource(memory);
 	const auto sampler  = GetSamplerResource(memory);
 	const auto address  = MakeImageAddress(inst, MemorySourceAt(inst, 0));
-	const auto result   = ir.Emit(IR::ValueOpcode::ImageSampleRaw, {resource, sampler, address},
+	const auto result   = ir.Emit(IR::ValueOpcode::ImageSampleRaw,
+	                              {resource, sampler, address, ir.GetExec()},
 	                              AddMemoryInfo(memory, inst.pc));
 	const bool dref     = (memory.image_sample_flags & Decoder::ImageSampleFlagCompare) != 0u;
 	if (dref && memory.data_bits != 16u) {
@@ -714,7 +715,7 @@ void Translator::IMAGE_GATHER(const Decoder::Instruction& inst) {
 	     has_lod ? resource.Instruction()->Arg(1) : IR::Value(0u),
 	     has_lod ? resource.Instruction()->Arg(3) : IR::Value(0u),
 	     has_lod ? sampler.Instruction()->Arg(1) : IR::Value(0u),
-	     has_lod ? sampler.Instruction()->Arg(2) : IR::Value(0u)},
+	     has_lod ? sampler.Instruction()->Arg(2) : IR::Value(0u), ir.GetExec()},
 	    AddMemoryInfo(memory, inst.pc));
 	for (uint32_t index = 0; index < memory.data_dwords; index++) {
 		WriteOperand(OffsetOperand(inst.dst, index),

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

@@ -317,8 +317,8 @@ VALUE_OPCODE(BvhIntersect, U32x4, U32x4, ImageAddress, U1)
 VALUE_OPCODE(ImageQueryLod, U32x4, ImageResource, SamplerResource, ImageAddress)
 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, U32, U32, U32, U32)
+VALUE_OPCODE(ImageSampleRaw, U32x4, ImageResource, SamplerResource, ImageAddress, U1)
+VALUE_OPCODE(ImageGatherRaw, U32x4, ImageResource, SamplerResource, ImageAddress, U32, U32, U32, U32, U1)
 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)

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

@@ -989,23 +989,29 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	// Keep only that Boolean structure, never the varying shader dependency graph.
 	Value unknown;
 	bool capture_control_reads = false;
+	std::unordered_map<const Inst*, Value> predicates;
 	std::function<Value(Value)> ClonePredicate = [&](Value value) -> Value {
 		value = value.Resolve();
 		if (value.IsEmpty()) return {};
+		const auto* inst = value.TryInstruction();
+		if (inst == nullptr)
+			return ValidateRuntimeValue(program, value, RuntimeValueType::Integer) ? value : Value {};
+		// Preserve shared EXEC predicates, including subgraphs with no uniform factor.
+		const auto [found, inserted] = predicates.try_emplace(inst);
+		if (!inserted) return found->second;
+		auto& result = found->second;
 		if (ValidateRuntimeValue(program, value, RuntimeValueType::Integer)) {
 			capture_control_reads |=
 			    !ValidateRuntimeValue(program, value, RuntimeValueType::ImmutableInteger);
-			return Clone(value);
+			return result = Clone(value);
 		}
-		const auto* inst = value.TryInstruction();
-		if (inst == nullptr) return {};
 		const auto op = inst->GetOpcode();
 		if (op == ValueOpcode::ConditionRef || op == ValueOpcode::LogicalNot) {
 			const auto operand = ClonePredicate(inst->Arg(0));
-			if (operand.IsEmpty() || op == ValueOpcode::ConditionRef) return operand;
+			if (operand.IsEmpty() || op == ValueOpcode::ConditionRef) return result = operand;
 			auto& node = plan.value_storage.emplace_back(op);
 			node.SetArg(0, operand);
-			return Value(&node);
+			return result = Value(&node);
 		}
 		if (op != ValueOpcode::LogicalAnd && op != ValueOpcode::LogicalOr) return {};
 		auto left = ClonePredicate(inst->Arg(0));
@@ -1019,7 +1025,7 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		auto& node = plan.value_storage.emplace_back(op);
 		node.SetArg(0, left);
 		node.SetArg(1, right);
-		return Value(&node);
+		return result = Value(&node);
 	};
 	plan.control_flow = ResourceControlFlow(program, ClonePredicate);
 	plan.uniform_fill = AnalyzeUniformFill(program);
@@ -1406,6 +1412,7 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 			}
 			EXIT_IF(image.indirect_root == memory.resource &&
 			        inst.GetOpcode() != ValueOpcode::ImageSampleRaw &&
+			        inst.GetOpcode() != ValueOpcode::ImageGatherRaw &&
 			        inst.GetOpcode() != ValueOpcode::ImageRead);
 		}
 	}

tests/ResourceMaterializationTests.cpp

@@ -640,6 +640,64 @@ void TestFailedMaterializationRejectsStage() {
         "missing runtime user data did not reject the cached stage");
 }
 
+void TestSharedControlPredicates() {
+  using namespace Libs::Graphics::ShaderRecompiler::IR;
+  using Libs::Graphics::ShaderRecompiler::CFG::TerminatorKind;
+  constexpr uint32_t depth = 8;
+  for (const auto operation : {ValueOpcode::LogicalAnd, ValueOpcode::LogicalOr}) {
+    Program program;
+    program.user_data_count = 1;
+    program.srt_plan_complete = program.resource_tracking_complete = true;
+    auto &block = AddValueBlock(program);
+    const auto emit = [&](ValueOpcode opcode, std::initializer_list<Value> args) {
+      return Value(&block.AppendNewInst(opcode, args));
+    };
+    const auto user = emit(ValueOpcode::GetUserData, {Value(ScalarReg(0))});
+    const auto uniform = emit(ValueOpcode::INotEqual32, {user, Value(0u)});
+    const auto lane = emit(ValueOpcode::GetBuiltin,
+        {Value(static_cast<uint32_t>(StageInputKind::LocalInvocationId)), Value(0u)});
+    auto varying = emit(ValueOpcode::INotEqual32, {lane, Value(0u)});
+    // Both failed and partially uniform subgraphs share their operands exponentially.
+    for (uint32_t i = 0; i < depth; ++i)
+      varying = emit(ValueOpcode::LogicalOr, {varying, varying});
+    auto condition = emit(operation, {varying, uniform});
+    for (uint32_t i = 0; i < depth; ++i)
+      condition = emit(ValueOpcode::LogicalOr, {condition, condition});
+    condition = emit(ValueOpcode::LogicalNot, {condition});
+    condition = emit(ValueOpcode::ConditionRef, {condition});
+    for (uint32_t i = 0; i < 3; ++i) AddValueBlock(program);
+    for (uint32_t i = 0; i < 2; ++i) {
+      program.blocks[i]->condition = condition;
+      program.blocks[i]->terminator.kind = TerminatorKind::ConditionalBranch;
+      program.blocks[i]->terminator.true_block = program.blocks[i == 0 ? 1 : 3];
+      program.blocks[i]->terminator.false_block = program.blocks[2];
+    }
+    auto plan = ExtractResourcePlan(program);
+    Check(plan.value_storage.size() == depth + 5 &&
+              plan.control_flow.size() == 4 &&
+              plan.control_flow[0].condition == plan.control_flow[1].condition &&
+              !plan.capture_specialization_reads,
+          "shared predicates expanded into trees or retained varying dependencies");
+    for (const uint32_t flag : {0u, 1u}) {
+      const std::array<uint32_t, 1> user_data{flag};
+      SrtWalker walker(plan, {.user_data = user_data});
+      uint32_t result = 99;
+      const bool conjunction = operation == ValueOpcode::LogicalAnd;
+      const bool known = conjunction ? flag == 0 : flag != 0;
+      Check(walker.Evaluate(plan.control_flow[0].condition, result) == known &&
+                (!known || result == uint32_t{conjunction}),
+            "shared predicate changed its uniform factor or cached runtime values");
+      std::vector<uint32_t> flat;
+      Check(walker.RefreshFlatBuffer(flat) &&
+                plan.visited_blocks == (known
+                    ? (conjunction ? std::vector<uint8_t>{1, 1, 0, 1}
+                                   : std::vector<uint8_t>{1, 0, 1, 0})
+                    : std::vector<uint8_t>{1, 1, 1, 1}),
+            "shared predicate skipped an unknown branch or followed a rejected branch");
+    }
+  }
+}
+
 void TestFiniteImageRefreshReusesScalarReads() {
   using namespace Libs::Graphics::ShaderRecompiler::IR;
   Program program;
@@ -793,6 +851,7 @@ int main() {
   for (const bool memory_condition : {false, true})
     TestWrittenDescriptorPredicateReads(memory_condition);
   TestFailedMaterializationRejectsStage();
+  TestSharedControlPredicates();
   TestFiniteImageRefreshReusesScalarReads();
   TestMixedSamplerVariantsShareRuntimeDescriptor();
   std::puts("ResourceMaterializationTests: all cases passed");

tests/ResourceTrackingTests.cpp

@@ -270,7 +270,7 @@ MakeIndirectImageFixture(bool malformed, uint32_t table_stride = 32, uint32_t ta
   sample.kind = ResourceKind::Image;
   sample.image_dimension = Decoder::ImageDimension::Dim2D;
   const auto sampled = fixture->Emit(ValueOpcode::ImageSampleRaw,
-                                     {image, sampler, fixture->ImageAddress()},
+                                     {image, sampler, fixture->ImageAddress(), Value(true)},
                                      fixture->AddMemory(sample, 0x10f0));
   const auto sampled_x =
       fixture->Emit(ValueOpcode::CompositeExtractU32x4, {sampled, Value(0u)});
@@ -486,7 +486,8 @@ void TestWaterfallImageTable() {
       memory.kind = ResourceKind::Image;
       memory.image_dimension = Decoder::ImageDimension::Dim2D;
       const auto value = fixture->Emit(ValueOpcode::ImageSampleRaw,
-          {image, sampler, fixture->ImageAddress()}, fixture->AddMemory(memory, 0x92cu), block);
+          {image, sampler, fixture->ImageAddress(), block == entry ? active : guard},
+          fixture->AddMemory(memory, 0x92cu), block);
       fixture->Emit(ValueOpcode::ReferenceU32,
           {fixture->Emit(ValueOpcode::CompositeExtractU32x4, {value, Value(0u)}, 0, block)}, 0, block);
     };
@@ -641,7 +642,7 @@ void TestGuardedDirectImageTable() {
     MemoryInfo memory;
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
-    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress()},
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress(), Value(true)},
                  fixture.AddMemory(memory, 0x128));
     fixture.PlanAndTrack();
     const auto source = fixture.program.info.images[0].source;
@@ -989,7 +990,7 @@ void TestBoundedComputeImageLoop() {
     sample.kind = ResourceKind::Image;
     sample.image_dimension = Decoder::ImageDimension::Dim2D;
     fixture.Emit(ValueOpcode::ImageSampleRaw,
-                 {image, sampler, fixture.ImageAddress()},
+                 {image, sampler, fixture.ImageAddress(), Value(true)},
                  fixture.AddMemory(sample, 0x29c));
     fixture.PlanAndTrack();
     const auto source = fixture.program.info.images[0].source;
@@ -1325,7 +1326,9 @@ void TestUniformizedMaterialImageKeys() {
     MemoryInfo image_memory;
     image_memory.kind = ResourceKind::Image;
     image_memory.image_dimension = Decoder::ImageDimension::Dim2D;
-    const auto sampled = fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, address},
+    const auto sampled = fixture.Emit(ValueOpcode::ImageSampleRaw,
+        {image, sampler, address,
+         variant == Variant::EmptyFirstEntry ? material_guard : sample_guard},
         fixture.AddMemory(image_memory, 0x1aecu), sample);
     if (first_lane) {
       const auto component = fixture.Emit(
@@ -1477,7 +1480,7 @@ void TestImageDescriptorFields() {
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
     memory.image_r128 = r128;
-    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress()},
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress(), Value(true)},
                  fixture.AddMemory(memory, 0x80));
     fixture.PlanAndTrack();
     auto plan = ExtractResourcePlan(fixture.program);
@@ -1571,7 +1574,7 @@ void TestUniformScalarBufferImage() {
   sample.kind = ResourceKind::Image;
   sample.image_dimension = Decoder::ImageDimension::Dim2D;
   fixture.Emit(ValueOpcode::ImageSampleRaw,
-                {image, sampler, fixture.ImageAddress()}, fixture.AddMemory(sample, 0x228));
+                {image, sampler, fixture.ImageAddress(), Value(true)}, fixture.AddMemory(sample, 0x228));
   fixture.PlanAndTrack();
   const auto plan = ExtractResourcePlan(fixture.program);
 
@@ -1866,7 +1869,7 @@ void TestImagesSamplersAndAliases() {
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
     memory.image_sample_flags = sample_flags;
-    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, image_address},
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, image_address, Value(true)},
                  fixture.AddMemory(memory, pc));
     return std::pair{image, sampler};
   };
@@ -1938,14 +1941,14 @@ void TestSampleAdjustSamplerScratch() {
   memory.image_dimension = Decoder::ImageDimension::Dim2D;
   memory.image_sample_flags = Decoder::ImageSampleFlagAdjust;
   fixture.Emit(ValueOpcode::ImageSampleRaw,
-               {image, sampler, fixture.ImageAddress()},
+               {image, sampler, fixture.ImageAddress(), Value(true)},
                fixture.AddMemory(memory, 0x1ec));
   // PPSA24156's zero border payload leaves the reserved scratch shift as
   // the entire DWORD after the frontend folds its OR with zero.
   const auto shift_sampler = fixture.Sampler(
       {Value(0x36u), Value(0xfff000u), Value(0x02500000u), scratch}, 0x200);
   fixture.Emit(ValueOpcode::ImageSampleRaw,
-               {image, shift_sampler, fixture.ImageAddress()},
+               {image, shift_sampler, fixture.ImageAddress(), Value(true)},
                fixture.AddMemory(memory, 0x200));
   fixture.PlanAndTrack();
   auto plan = ExtractResourcePlan(fixture.program);
@@ -1994,7 +1997,7 @@ void TestSampleAdjustSamplerScratch() {
     rejected_memory.image_dimension = Decoder::ImageDimension::Dim2D;
     rejected_memory.image_sample_flags = flags;
     rejected.Emit(ValueOpcode::ImageSampleRaw,
-                  {rejected_image, rejected_sampler, rejected.ImageAddress()},
+                  {rejected_image, rejected_sampler, rejected.ImageAddress(), Value(true)},
                   rejected.AddMemory(rejected_memory, 0x200));
 
     CheckFatal([&] { rejected.PlanAndTrack(); },
@@ -2112,7 +2115,7 @@ void TestGatherLodSamplerValidation() {
         fixture.Emit(ValueOpcode::ImageGatherRaw,
                      {fixture.Image(image_words), fixture.Sampler(sampler_words),
                       fixture.ImageAddress(), image_words[1], image_words[3],
-                      sampler_words[1], sampler_words[2]},
+                      sampler_words[1], sampler_words[2], Value(true)},
                      fixture.AddMemory(memory, i * 4));
       }
       fixture.PlanAndTrack();
@@ -2514,7 +2517,7 @@ ResourcePlan ConditionalSamplerPlan(bool diamond, bool reverse, bool reverse_phi
     sample.kind = ResourceKind::Image;
     sample.image_dimension = Decoder::ImageDimension::Dim2D;
     const auto result =
-        fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, address},
+        fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, address, Value(true)},
                      fixture.AddMemory(sample, 0x4d4));
     fixture.Emit(ValueOpcode::ReferenceU32,
                  {fixture.Emit(ValueOpcode::CompositeExtractU32x4,
@@ -2619,7 +2622,7 @@ void TestFiniteImagePhiCycle() {
   MemoryInfo memory;
   memory.kind = ResourceKind::Image;
   memory.image_dimension = Decoder::ImageDimension::Dim2D;
-  fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress()},
+  fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress(), Value(true)},
                fixture.AddMemory(memory, 4));
   fixture.PlanAndTrack();
   auto plan = ExtractResourcePlan(fixture.program);
@@ -2685,12 +2688,12 @@ void TestFiniteInlineSampler() {
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
     if (format == F::k32Float)
-      fixture.Emit(ValueOpcode::ImageSampleRaw, {image, previous, fixture.ImageAddress()},
+      fixture.Emit(ValueOpcode::ImageSampleRaw, {image, previous, fixture.ImageAddress(), Value(true)},
                    fixture.AddMemory(memory, 0x440u));
-    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress()},
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress(), Value(true)},
                  fixture.AddMemory(memory, 0x1dd0u));
     if (format == F::k32Float)
-      fixture.Emit(ValueOpcode::ImageSampleRaw, {image, direct, fixture.ImageAddress()},
+      fixture.Emit(ValueOpcode::ImageSampleRaw, {image, direct, fixture.ImageAddress(), Value(true)},
                    fixture.AddMemory(memory, 0x1dd8u));
   }
   fixture.PlanAndTrack();
@@ -2764,7 +2767,7 @@ void TestFiniteInlineSampler() {
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
     memory.image_sample_flags = flags;
-    adjusted.Emit(ValueOpcode::ImageSampleRaw, {image, shared, adjusted.ImageAddress()},
+    adjusted.Emit(ValueOpcode::ImageSampleRaw, {image, shared, adjusted.ImageAddress(), Value(true)},
                   adjusted.AddMemory(memory, 0x1dd0u));
   }
   CheckFatal([&] { adjusted.PlanAndTrack(); }, "not a valid runtime value",
@@ -2817,7 +2820,7 @@ void TestFiniteImageBitScanSentinel() {
     MemoryInfo memory;
     memory.kind = ResourceKind::Image;
     memory.image_dimension = Decoder::ImageDimension::Dim2D;
-    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress()},
+    fixture.Emit(ValueOpcode::ImageSampleRaw, {image, sampler, fixture.ImageAddress(), Value(true)},
                  fixture.AddMemory(memory, 4));
     if (!nonzero) {
       CheckFatal([&] { fixture.PlanAndTrack(); }, "not a valid runtime value",

tests/ShaderRecompilerComputeTests.cpp

@@ -11083,6 +11083,78 @@ public:
       context.ShutdownGpu();
       LibKernel::Memory::InstallGpuResources(nullptr);
     }
+    {
+      // PPSA01593 samples a DCC R32 color surface through a comparison sampler.
+      constexpr uint64_t comparison_dcc = base + 0x10000;
+      std::memset(mapped, 0, 0x10000);
+      std::memset(reinterpret_cast<void *>(comparison_dcc), 0xff, 0x1000);
+      RenderContext context(m_runtime_context);
+      context.InitializeGpu(nullptr);
+      LibKernel::Memory::InstallGpuResources(&context);
+      auto &scheduler = context.GetCommandScheduler();
+      HW::Context registers{};
+      HW::UserConfig user_config{};
+      HW::Shader shaders{};
+      scheduler.Begin(registers, user_config, shaders);
+      context.MapMemory(base, allocation_size);
+      auto &cache = context.GetTextureCache();
+      auto &executor = context.GetRenderExecutor();
+      const ShaderTextureResource descriptor{{static_cast<uint32_t>(base >> 8u),
+          0xc1600000u, 0x000fc00fu, 0x91b00204u, 0, 0x00700000u,
+          0x002b0000u | (static_cast<uint32_t>(comparison_dcc >> 8u) << 24u),
+          static_cast<uint32_t>(comparison_dcc >> 16u)}};
+      ShaderRecompiler::IR::DescriptorValue value{};
+      value.dword_count = 8;
+      std::copy_n(descriptor.fields, 8, value.dwords.begin());
+      ShaderRecompiler::IR::ImageResource resource{};
+      resource.resource_class = ShaderRecompiler::IR::ImageResourceClass::Storage;
+      resource.numeric_class = Prospero::TextureNumericClass::Float;
+      resource.dimension = ShaderRecompiler::Decoder::ImageDimension::Dim2D;
+      resource.written = true;
+      const auto color = RenderExecutorTestAccess::ResolveTexture(executor, resource, value);
+      (void)cache.FindTexture(color.image_id, color.desc);
+      vk::ClearValue paint{};
+      paint.color.float32 = std::array{0.375f, 0.f, 0.f, 0.f};
+      TextureCacheTestAccess::ClearImage(cache, scheduler.Current(), color.image_id,
+          {vk::ImageAspectFlagBits::eColor, 0, 1, 0, 1}, paint);
+      resource.resource_class = ShaderRecompiler::IR::ImageResourceClass::Sampled;
+      resource.written = false;
+      resource.read = true;
+      resource.depth_compare = true;
+      auto comparison = RenderExecutorTestAccess::ResolveTexture(executor, resource, value);
+      (void)cache.FindTexture(comparison.image_id, comparison.desc);
+      const auto &depth = cache.GetImage(comparison.image_id);
+      Require(name, "expanded DCC comparison alias",
+              comparison.image_id != color.image_id &&
+                  depth.info.pixel_format == vk::Format::eD32Sfloat &&
+                  depth.info.guest_format == Prospero::BufferFormat::k32Float &&
+                  depth.info.tile_mode == Prospero::TileMode::kRenderTarget &&
+                  depth.info.extent == vk::Extent3D{64, 64, 1} &&
+                  depth.info.metadata.kind == ImageMetadataKind::Dcc &&
+                  depth.info.metadata.range == GuestRange{comparison_dcc, 0x1000} &&
+                  ReadCachedTexel(name, context, comparison.image_id) ==
+                      std::vector<u32>{std::bit_cast<u32>(0.375f)},
+              "comparison promotion lost native color data or guest DCC layout");
+      for (const u32 code : {0x40404040u, 0x80808080u}) {
+        WriteMetadata(context, comparison_dcc, 0x1000, code);
+        comparison = RenderExecutorTestAccess::ResolveTexture(executor, resource, value);
+        (void)cache.FindTexture(comparison.image_id, comparison.desc);
+        Require(name, "DCC clear on a comparison backing",
+                ReadCachedTexel(name, context, comparison.image_id, {63, 63, 0}) ==
+                    std::vector<u32>{code == 0x40404040u ? 0u : 0x3f800000u},
+                "DCC clear used the native depth format or color aspect");
+        std::array<uint8_t, 0x1000> metadata{};
+        Require(name, "comparison clear expands metadata",
+                LibKernel::Memory::TryReadBacking(comparison_dcc, metadata.data(),
+                                                 metadata.size()) &&
+                    std::ranges::all_of(metadata, [](uint8_t byte) { return byte == 0xff; }),
+                "comparison clear retained the native DCC clear key");
+      }
+      context.UnmapMemory(base, allocation_size);
+      scheduler.Finish();
+      context.ShutdownGpu();
+      LibKernel::Memory::InstallGpuResources(nullptr);
+    }
     Require(name, "unmap",
             Libs::LibKernel::Memory::KernelMunmap(base, allocation_size) == 0,
             "captured scene mapping release failed");
@@ -34888,12 +34960,11 @@ TestCase ImageSampleOpcodeAliasUsesNormalCoords() {
   return test;
 }
 
-TestCase ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu() {
+template <u32 wave_size> TestCase ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu() {
   using O = ShaderOpcode;
 
   std::vector<u32> code;
-  AppendVMovU32(&code, 20,
-                1); // Non-constant +1 X offset is not a SPIR-V ConstOffset.
+  AppendVMovU32(&code, 20, 1); // +1 X remains a full dword with A16 coordinates.
   AppendVMovLiteral(&code, 21, 0x36003900u); // x=0.625, y=0.375 packed as f16.
   AppendVMovU32(&code, 22, 0);
   code.push_back(EncodeMimg0(0x30, 0xf));
@@ -34904,18 +34975,63 @@ TestCase ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu() {
   AppendEnd(&code);
 
   auto image = MakeRgbaImage(4, 4);
-  SetRgbaPixel(&image, 4, 2, 1, 0x3f800000u, 0x40000000u, 0x40400000u,
+  SetRgbaPixel(&image, 4, 2, 1, 0x3f000000u, 0, 0, 0);
+  SetRgbaPixel(&image, 4, 3, 1, 0x3f800000u, 0x40000000u, 0x40400000u,
                0x40800000u);
 
   TestCase test;
-  test.name = "ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu";
+  test.name = wave_size == 32 ? "ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu32"
+                              : "ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu64";
   test.code = code;
+  test.compute_info.wave_size = wave_size;
+  test.has_compute_info = true;
   test.expected = {0x3f800000u, 0x40000000u, 0x40400000u, 0x40800000u};
   test.opcodes = {O::V_MOV_B32, O::IMAGE_SAMPLE, O::BUFFER_STORE_DWORD,
                   O::S_ENDPGM};
   test.sampled_image_rgba = image;
-  test.required_spirv = {"UnpackHalf2x16"};
-  test.forbidden_spirv = {"OpBitFieldSExtract"};
+  test.required_spirv = {"UnpackHalf2x16", "Lod|ConstOffset"};
+  test.forbidden_spirv = {"OpBitFieldSExtract", "OpImageQuerySizeLod", "OpFDiv"};
+  return test;
+}
+
+template <u32 wave_size> TestCase ImageSampleOffsetsReadFourNeighbors() {
+  using O = ShaderOpcode;
+  TestCase test;
+  test.name = wave_size == 32 ? "ImageSampleOffsetsReadFourNeighbors32"
+                              : "ImageSampleOffsetsReadFourNeighbors64";
+  test.compute_info.wave_size = wave_size;
+  test.has_compute_info = true;
+  constexpr std::array offsets{0x01u, 0x3fu, 0x3f00u, 0x0100u};
+  constexpr std::array opcodes{0x37u, 0x37u, 0x34u, 0x32u};
+  for (u32 sample = 0; sample < offsets.size(); ++sample) {
+    AppendVMovLiteral(&test.code, 20, offsets[sample]);
+    u32 coordinate = 21;
+    if (opcodes[sample] == 0x32u) {
+      for (float gradient : {0.25f, 0.f, 0.f, 0.25f}) {
+        AppendVMovLiteral(&test.code, coordinate++, std::bit_cast<u32>(gradient));
+      }
+    }
+    // All instructions address the center of texel (2, 1) before applying the offset.
+    AppendVMovLiteral(&test.code, coordinate, std::bit_cast<u32>(0.625f));
+    AppendVMovLiteral(&test.code, coordinate + 1, std::bit_cast<u32>(0.375f));
+    if (opcodes[sample] == 0x34u) AppendVMovU32(&test.code, coordinate + 2, 0);
+    test.code.push_back(EncodeMimg0(opcodes[sample], 0x1));
+    test.code.push_back(EncodeMimg1(sample, 20));
+    AppendStoreVgpr(&test.code, sample, sample);
+  }
+  AppendEnd(&test.code);
+  test.sampled_image_rgba = MakeRgbaImage(4, 4);
+  SetRgbaPixel(&test.sampled_image_rgba, 4, 2, 1, std::bit_cast<u32>(16.f), 0, 0, 0);
+  constexpr std::array<std::array<u32, 2>, 4> neighbors{{{3, 1}, {1, 1}, {2, 0}, {2, 2}}};
+  for (u32 sample = 0; sample < neighbors.size(); ++sample) {
+    const auto value = std::bit_cast<u32>(float(sample + 1));
+    SetRgbaPixel(&test.sampled_image_rgba, 4, neighbors[sample][0], neighbors[sample][1],
+                 value, 0, 0, 0);
+    test.expected.push_back(value);
+  }
+  test.opcodes = {O::V_MOV_B32, O::IMAGE_SAMPLE, O::BUFFER_STORE_DWORD, O::S_ENDPGM};
+  test.required_spirv = {"Lod|ConstOffset", "Grad|ConstOffset"};
+  test.forbidden_spirv = {"OpImageQuerySizeLod", "OpFDiv", "OpBitFieldSExtract", "OpSwitch"};
   return test;
 }
 
@@ -34986,7 +35102,8 @@ TestCase ImageGatherCompareOpcodes() {
       O::V_MOV_B32,         O::IMAGE_GATHER4_C,      O::IMAGE_GATHER4_C_LZ,
       O::IMAGE_GATHER4_C_O, O::IMAGE_GATHER4_C_LZ_O, O::BUFFER_STORE_DWORD,
       O::S_ENDPGM};
-  test.required_spirv = {"OpImageDrefGather", "OpBitFieldSExtract"};
+  test.required_spirv = {"OpImageDrefGather", "ConstOffset"};
+  test.forbidden_spirv = {"OpBitFieldSExtract"};
   test.compile_only = true;
   return test;
 }
@@ -35538,7 +35655,7 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
   std::memcpy(&memory_flag_bits, &memory_flags, sizeof(memory_flags));
   auto &sample = block->AppendNewInst(
       ValueOpcode::ImageSampleRaw,
-      {Value(&image), Value(&sampler), Value(&address)}, memory_flag_bits);
+      {Value(&image), Value(&sampler), Value(&address), Value(true)}, memory_flag_bits);
   auto &sample_x = block->AppendNewInst(ValueOpcode::CompositeExtractU32x4,
                                         {Value(&sample), Value(0u)});
   block->AppendNewInst(ValueOpcode::ReferenceU32, {Value(&sample_x)});
@@ -35749,29 +35866,63 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
         resource == 4u ? vk::ImageViewType::e2DArray : vk::ImageViewType::e2D, layers));
   }
   const auto native_sampler = vulkan.CreateSampler(test.name);
-  for (const bool fetch : {false, true}) {
+  enum class Access { Sample, Gather, GatherHomogeneous, Fetch };
+  auto *access = &sample;
+  for (const auto mode : {Access::Sample, Access::Gather,
+                          Access::GatherHomogeneous, Access::Fetch}) {
+    const bool fetch = mode == Access::Fetch;
+    const bool gather = mode == Access::Gather || mode == Access::GatherHomogeneous;
+    const bool homogeneous = mode == Access::GatherHomogeneous;
+    if (mode == Access::Gather) {
+      test.name = "IndirectImageGatherContiguousRuns";
+      const auto position = std::ranges::find_if(*block, [&](const Inst &inst) {
+        return &inst == access;
+      });
+      auto &read = *block->PrependNewInst(
+          position, ValueOpcode::ImageGatherRaw,
+          {Value(&image), Value(&sampler), Value(&address), Value(0u), Value(0u),
+           Value(0u), Value(0u), Value(true)}, memory_flag_bits);
+      access->ReplaceUsesWith(Value(&read), false);
+      access = &read;
+      compiled.program.info.images[4].cube = false;
+      auto &gather_memory = compiled.program.memory_info[0];
+      gather_memory.dmask = 1;
+      gather_memory.image_sample_flags = ShaderRecompiler::Decoder::ImageSampleFlagLevelZero |
+          ShaderRecompiler::Decoder::ImageSampleFlagOffset;
+      // GATHER4_LZ_O places its packed signed offset before the coordinates.
+      address.SetArg(0, Value(0x3fu));
+      address.SetArg(1, Value(std::bit_cast<u32>(0.75f)));
+      address.SetArg(2, Value(std::bit_cast<u32>(0.25f)));
+      address.SetArg(3, Value(0u));
+    }
+    auto &selected_resources = compiled.program.info.images[0].indirect_resources;
+    selected_resources = homogeneous ? std::vector<u32>{0u, 2u, 3u, 5u, 6u}
+                                     : root.indirect_resources;
+    if (homogeneous) test.name = "IndirectImageGatherHomogeneous";
+    compiled.resources.flattened_srt[6] = homogeneous ? 4u : ordinals[5];
     if (fetch) {
       test.name = "IndirectImageFetchContiguousRuns";
       const auto position = std::ranges::find_if(*block, [&](const Inst &inst) {
-        return &inst == &sample;
+        return &inst == access;
       });
       auto &read = *block->PrependNewInst(
           position, ValueOpcode::ImageRead,
           {Value(&image), Value(&address), Value(true)}, memory_flag_bits);
-      sample.ReplaceUsesWith(Value(&read), false);
+      access->ReplaceUsesWith(Value(&read), false);
       sampler.Invalidate();
       compiled.program.info.samplers.clear();
       compiled.program.info.sampled_pairs.clear();
       compiled.program.info.images[4].cube = false;
+      compiled.program.memory_info[0].dmask = 0xf;
       compiled.program.memory_info[0].image_sample_flags = 0;
       compiled.program.memory_info[0].image_has_mip = true;
       for (u32 component = 0; component < 4u; ++component)
         address.SetArg(component, Value(0u));
       // Native DIM=2DArray places LOD after the layer, even for a 2D candidate.
       address.SetArg(3, Value(1u));
-      compiled.program.shader_info_complete = false;
-      CollectShaderInfo(compiled.program, {.compute = &compute});
     }
+    compiled.program.shader_info_complete = false;
+    CollectShaderInfo(compiled.program, {.compute = &compute});
     for (const u32 wave_size : {32u, 64u}) {
       compute.wave_size = wave_size;
       compute.host_subgroup_size = vulkan.SubgroupSize();
@@ -35782,31 +35933,38 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
       compiled.packed_user_data.resize(compiled.program.bindings.ShaderDataDwords());
       ValidateSpirv(test.name, compiled.spirv);
       const auto halves = wave_size > compute.host_subgroup_size ? 2u : 1u;
+      const auto image_opcode = fetch ? "OpImageFetch" : gather ? "OpImageGather"
+                                                          : "OpImageSampleExplicitLod";
       Require(test.name, "native array image runs", tools.Disassemble(compiled.spirv, &text) &&
-                  CountText(text, fetch ? "OpImageFetch" : "OpImageSampleExplicitLod") == 3u * halves,
+                  CountText(text, image_opcode) == (homogeneous ? 1u : 3u) * halves,
               "mixed runs expanded into per-image accesses");
-      if (fetch) {
+      if (homogeneous) {
+        Require(test.name, "homogeneous gather array", text.find("OpSwitch") == std::string::npos,
+                "homogeneous gather emitted a descriptor switch");
+      }
+      if (fetch || gather) {
         std::vector<u32> decorated;
-        std::vector<u32> fetched_images;
+        std::vector<u32> image_operands;
         for (size_t offset = 5; offset < compiled.spirv.size();) {
           const auto words = std::span<const u32>(compiled.spirv).subspan(
               offset, compiled.spirv[offset] >> 16u);
           const auto opcode = static_cast<spv::Op>(words[0] & 0xffffu);
           if (opcode == spv::OpDecorate && words[2] == spv::DecorationNonUniform)
             decorated.push_back(words[1]);
-          if (opcode == spv::OpImageFetch) fetched_images.push_back(words[3]);
+          if (opcode == (fetch ? spv::OpImageFetch : spv::OpImageGather))
+            image_operands.push_back(words[3]);
           offset += words.size();
         }
-        Require(test.name, "nonuniform fetch operands",
-                std::ranges::count_if(fetched_images, [&](u32 operand) {
+        Require(test.name, "nonuniform image operands",
+                std::ranges::count_if(image_operands, [&](u32 operand) {
                   return std::ranges::find(decorated, operand) != decorated.end();
-                }) == 2u * halves,
-                "a dynamically selected fetch image lacks its nonuniform decoration");
+                }) == (homogeneous ? 1u : 2u) * halves,
+                "a dynamically selected image operand lacks its nonuniform decoration");
       }
       test.initial.assign(wave_size, 0xdeadbeefu);
       test.expected.assign(wave_size, 0u);
       for (u32 lane = 0; lane < ordinals.size(); ++lane) {
-        const auto resource = root.indirect_resources[ordinals[lane]];
+        const auto resource = selected_resources[compiled.resources.flattened_srt[1u + lane]];
         test.expected[lane] = std::bit_cast<u32>(float(resource + (fetch && resource != 0u ? 10u : 0u)));
       }
       auto buffer = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size());
@@ -35814,7 +35972,7 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
                       fetch ? vk::Sampler{} : native_sampler, textures);
       const auto actual = vulkan.ReadBuffer(test.name, buffer, test.expected.size());
       vulkan.DestroyBuffer(&buffer);
-      CompareWords(test, "root, native slots, cube and mapping bounds", test.expected, actual);
+      CompareWords(test, "root, native slots, array runs and mapping bounds", test.expected, actual);
     }
     std::printf("[compute] %-32s ok\n", test.name);
   }
@@ -37659,7 +37817,10 @@ std::vector<TestCase> MakeCases() {
   AddCase(ImageD16GatherPacksHalfPairs);
   AddCase(ImageSampleA16SamplerCoordsOnGpu);
   AddCase(ImageSampleOpcodeAliasUsesNormalCoords);
-  AddCase(ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu);
+  AddCase(ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu<32>);
+  AddCase(ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu<64>);
+  AddCase(ImageSampleOffsetsReadFourNeighbors<32>);
+  AddCase(ImageSampleOffsetsReadFourNeighbors<64>);
   AddCase(ImageSampleA16CompareBiasRdna2AddressOrder);
   AddCase(ImageGatherCompareOpcodes);
   AddCase(ImageStoreVariants);
@@ -40070,7 +40231,7 @@ void CheckImageSamplerSpecialization() {
               sizeof(signed_memory_flags));
   mixed_sampler_block->AppendNewInst(
       ValueOpcode::ImageSampleRaw,
-      {Value(&image_handle), Value(&sampler_handle), Value(&image_address)},
+      {Value(&image_handle), Value(&sampler_handle), Value(&image_address), Value(true)},
       signed_memory_flag_bits);
 
   ShaderRecompiler::IR::DescriptorValue native_image_descriptor{};
@@ -43327,6 +43488,16 @@ int main(int argc, char **argv) {
     RunCase(&vulkan, FlatVirtualAddressRebasesGuestAllocation());
     return 0;
   }
+  if (argc == 2 && std::strcmp(argv[1], "--image-sample-offset-only") == 0) {
+    VulkanHarness vulkan;
+    RunCase(&vulkan, ImageSampleOffsetsReadFourNeighbors<32>());
+    RunCase(&vulkan, ImageSampleOffsetsReadFourNeighbors<64>());
+    RunCase(&vulkan, ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu<32>());
+    RunCase(&vulkan, ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu<64>());
+    RunCase(&vulkan, ImageSampleAndGather());
+    RunCase(nullptr, ImageGatherCompareOpcodes());
+    return 0;
+  }
   if (argc == 2 && std::strcmp(argv[1], "--image-gather-lod-only") == 0) {
     VulkanHarness vulkan;
     RunCase(&vulkan, ImageGatherExplicitLod());

tests/shaderCfgTests.cpp

@@ -1316,7 +1316,7 @@ void TestShaderFeatureCollection() {
                        {Value(0u), Value(0u), Value(0u)});
   block->AppendNewInst(ValueOpcode::ImageGatherRaw,
                        {Value(0u), Value(0u), Value(0u), Value(0u),
-                        Value(0u), Value(0u), Value(0u)});
+                        Value(0u), Value(0u), Value(0u), Value(true)});
   auto &shared = block->AppendNewInst(ValueOpcode::LoadSharedU32,
                                       {Value(0u), Value(true)});
   shared.SetFlags(MemoryFlags{.index = 0});