shader: emit native texel offsets using sample execution context

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

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,
+uint32_t PackedOffset(ValueEmitContext& ctx, IR::Value value, ImageDimension dimension,
-                      const ImageSampleLayout& layout, ImageDimension dimension) {
+                      uint32_t& mask) {
 	const auto components = ImageDimensionInfoFor(dimension).spatial_components;
-	const auto zero       = ConstantI32(ctx.state, 0);
+	uint32_t values[3] {};
-	if (layout.offset == NoImageComponent || mem.image_address_components <= layout.offset) {
+	if (value.IsImmediate()) {
-		if (components == 1u) return zero;
+		// RDNA2 packs signed six-bit texel offsets into the low bits of each byte.
-		const auto result = ctx.state.builder.AllocateId();
+		for (uint32_t index = 0; index < components; index++) {
-		if (components == 3u) {
+			const auto bits = (value.U32() >> (index * 8u)) & 0x3fu;
-			ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 3),
+			values[index] = ConstantI32(ctx.state, static_cast<int32_t>(bits ^ 0x20u) - 0x20);
-			                              result, zero, zero, zero);
-		} else {
-			ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 2),
-			                              result, zero, zero);
 		}
-		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),
+	mask |= spv::ImageOperandsOffsetMask;
-	                             AddressU32(ctx, mem, address, layout.offset));
+	const auto packed = Unary(ctx.state, spv::OpBitcast, TypeI32(ctx.state), ctx.Def(value));
-	uint32_t   values[3] = {zero, zero, zero};
 	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, components),
-		ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 3),
+	                              result, std::span(values).first(components));
-		                              result, values[0], values[1], values[2]);
-	} else {
-		ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeI32Vector(ctx.state, 2),
-		                              result, values[0], values[1]);
-	}
 	return result;
 }
 
@@ -794,6 +789,12 @@ 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");
@@ -843,8 +844,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 					operand_mask = spv::ImageOperandsConstOffsetsMask;
 					offset = HorizontalOffsets(state, dimension);
 				} else if (layout.offset != NoImageComponent) {
-					operand_mask = spv::ImageOperandsOffsetMask;
+					offset = PackedOffset(ctx, offset_value, dimension, operand_mask);
-					offset = PackedOffset(ctx, mem, *address, layout, dimension);
 				}
 				const auto EmitMip = [&](uint32_t mip) {
 					const auto sampled = MakeSampledImage(state, resource, sampler_id, mip, array_index);
@@ -886,7 +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;
@@ -916,9 +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,
@@ -928,7 +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(ImageSampleRaw, U32x4, ImageResource, SamplerResource, ImageAddress, U1)
-VALUE_OPCODE(ImageGatherRaw, U32x4, ImageResource, SamplerResource, ImageAddress, U32, U32, U32, U32)
+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)

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

@@ -34960,12 +34960,11 @@ TestCase ImageSampleOpcodeAliasUsesNormalCoords() {
   return test;
 }
 
-TestCase ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu() {
+template <u32 wave_size> TestCase ImageSampleA16OffsetKeepsTexelOffset32BitOnGpu() {
   using O = ShaderOpcode;
 
   std::vector<u32> code;
-  AppendVMovU32(&code, 20,
+  AppendVMovU32(&code, 20, 1); // +1 X remains a full dword with A16 coordinates.
-                1); // Non-constant +1 X offset is not a SPIR-V ConstOffset.
   AppendVMovLiteral(&code, 21, 0x36003900u); // x=0.625, y=0.375 packed as f16.
   AppendVMovU32(&code, 22, 0);
   code.push_back(EncodeMimg0(0x30, 0xf));
@@ -34976,18 +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.required_spirv = {"UnpackHalf2x16", "Lod|ConstOffset"};
-  test.forbidden_spirv = {"OpBitFieldSExtract"};
+  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;
 }
 
@@ -35058,7 +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;
 }
@@ -35610,7 +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)});
@@ -35836,7 +35881,7 @@ void CheckIndirectImageKeySwitch(VulkanHarness &vulkan) {
       auto &read = *block->PrependNewInst(
           position, ValueOpcode::ImageGatherRaw,
           {Value(&image), Value(&sampler), Value(&address), Value(0u), Value(0u),
-           Value(0u), Value(0u)}, memory_flag_bits);
+           Value(0u), Value(0u), Value(true)}, memory_flag_bits);
       access->ReplaceUsesWith(Value(&read), false);
       access = &read;
       compiled.program.info.images[4].cube = false;
@@ -37772,7 +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);
@@ -40183,7 +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{};
@@ -43440,6 +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});