graphics: share signed storage conversion and support R16 image writes

Co-authored-by: fxpw <fxpw@outlook.com>

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

@@ -308,9 +308,8 @@ void ValidateStorageTexture(const ShaderRecompiler::IR::ImageResource& resource,
 	const bool descriptor_ok = IsSupportedStorageTextureDescriptor(resource, descriptor);
 	const bool encoding_ok   = IsSupportedStorageTextureEncoding(resource, descriptor);
 	const bool uint_resource    = resource.numeric_class == Prospero::TextureNumericClass::Uint;
-	const bool raw_sint_storage = format == Prospero::BufferFormat::k32SInt && uint_resource &&
+	const bool raw_sint_storage = uint_resource &&
-	                              resource.written &&
+	    ShaderRecompiler::IR::RawSignedStorageFormat(resource, format) != Prospero::BufferFormat::kInvalid;
-	                              (resource.atomic ? !resource.atomic64 : !resource.read);
 	const auto numeric_class =
 	    Prospero::SampledTextureNumericClass(Prospero::RemapTextureFormat(format));
 	const bool raw_float_atomic = format == Prospero::BufferFormat::k32Float && uint_resource &&
@@ -588,10 +587,11 @@ TextureBinding RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageR
 			pixel_format = depth_format->depth_attachment_format;
 		}
 	}
+	const auto raw_sint_format = ShaderRecompiler::IR::RawSignedStorageFormat(resource, format);
 	const auto storage_view_format = resource.atomic64 ? vk::Format::eR64Uint
-	                                 : storage && (resource.atomic ||
+	                                 : storage && resource.atomic ? vk::Format::eR32Uint
-	                                               format == Prospero::BufferFormat::k32SInt)
+	                                 : raw_sint_format != Prospero::BufferFormat::kInvalid
-	                                     ? vk::Format::eR32Uint
+	                                     ? VulkanFormat(raw_sint_format)
 	                                     : SrgbStorageViewFormat(pixel_format);
 	const auto view_format         = storage && storage_view_format != vk::Format::eUndefined
 	                                     ? storage_view_format

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

@@ -385,9 +385,10 @@ Format::BufferFormatInfo ImageConversionFormat(const EmitterState&   state,
 	const auto format = image.conversion_format;
 	if (format == Prospero::BufferFormat::kInvalid) return {};
 	const auto info = Format::GetFormatInfo(format);
-	EXIT_IF(Prospero::RemapTextureFormat(format) == format || info.component_count == 0u ||
+	const bool raw_signed = IR::RawSignedStorageFormat(image, format) != Prospero::BufferFormat::kInvalid;
+	EXIT_IF((Prospero::RemapTextureFormat(format) == format && !raw_signed) || info.component_count == 0u ||
 	        info.component_count > 4u);
-	EXIT_IF(info.type != Format::ComponentType::Uscaled &&
+	EXIT_IF(!raw_signed && info.type != Format::ComponentType::Uscaled &&
 	        ((info.type != Format::ComponentType::Uint &&
 	          info.type != Format::ComponentType::Unorm) || !info.packed_bitfield ||
 	         info.byte_size != sizeof(uint32_t)));
@@ -610,15 +611,16 @@ uint32_t EmitOneDimensionalGatherLz(ValueEmitContext& ctx, const IR::MemoryInfo&
 uint32_t StoreTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t data, bool integer) {
 	const auto swizzle = ctx.state.program.info.images[mem.resource].shader_swizzle;
 	const auto info = ImageConversionFormat(ctx.state, mem);
-	const bool packed = info.format != Prospero::BufferFormat::kInvalid;
+	const bool converted = info.format != Prospero::BufferFormat::kInvalid;
-	EXIT_IF(packed && (!info.packed_bitfield ||
+	const bool packed = info.packed_bitfield;
-	                  (info.type != Format::ComponentType::Uint &&
+	EXIT_IF(converted && info.type != Format::ComponentType::Sint &&
-	                   info.type != Format::ComponentType::Unorm)));
+	        (!packed || (info.type != Format::ComponentType::Uint &&
+	                     info.type != Format::ComponentType::Unorm)));
 	const auto zero = ConstantU32(ctx.state, 0);
 	uint32_t values[4] {zero, zero, zero, zero};
 	auto packed_value = zero;
 	const auto dmask = mem.dmask != 0u ? mem.dmask : 1u;
-	for (uint32_t component = 0; component < (packed ? info.component_count : 4u); component++) {
+	for (uint32_t component = 0; component < (converted ? info.component_count : 4u); component++) {
 		const auto source = InverseSwizzle(swizzle, component);
 		const bool selected = source < 4u && ((dmask >> source) & 1u) != 0u;
 		uint32_t raw = zero;
@@ -638,9 +640,13 @@ uint32_t StoreTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t d
 				}
 			}
 		}
-		if (packed) {
+		if (converted) {
 			if (!selected) continue;
 			auto value = EncodeFormattedStoreComponent(ctx.state, info, component, raw, mem.data_bits);
+			if (!packed) {
+				values[component] = value;
+				continue;
+			}
 			if (info.type == Format::ComponentType::Uint) {
 				value = EmitUMin32(ctx.state, value,
 				                  ConstantU32(ctx.state, (1u << info.component_bits[component]) - 1u));

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

@@ -435,6 +435,13 @@ uint32_t EncodeFormattedStoreComponent(EmitterState& state, const Format::Buffer
 			data = EmitBitCastU32F32(state, EmitF16BitsToF32(state, data));
 		}
 	}
+	if (info.type == Format::ComponentType::Sint && bits < 32u) {
+		const auto maximum = (1u << (bits - 1u)) - 1u;
+		data = EmitSMax32(state, data, ConstantU32(state, ~maximum));
+		data = EmitSMin32(state, data, ConstantU32(state, maximum));
+		return Binary(state, spv::OpBitwiseAnd, TypeU32(state), data,
+		              ConstantU32(state, (1u << bits) - 1u));
+	}
 	const bool normalized = info.type == Format::ComponentType::Unorm ||
 	                        info.type == Format::ComponentType::Snorm;
 	const bool scaled = info.type == Format::ComponentType::Uscaled ||

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

@@ -169,6 +169,17 @@ struct ImageResource {
 	bool operator==(const ImageResource& other) const = default;
 };
 
+[[nodiscard]] constexpr Prospero::BufferFormat
+RawSignedStorageFormat(const ImageResource& image, Prospero::BufferFormat format) {
+	if (image.resource_class != ImageResourceClass::Storage || !image.written)
+		return Prospero::BufferFormat::kInvalid;
+	if (format == Prospero::BufferFormat::k32SInt && (image.atomic ? !image.atomic64 : !image.read))
+		return Prospero::BufferFormat::k32UInt;
+	if (format == Prospero::BufferFormat::k16SInt && !image.read && !image.atomic)
+		return Prospero::BufferFormat::k16UInt;
+	return Prospero::BufferFormat::kInvalid;
+}
+
 struct SamplerResource {
 	uint32_t source                = 0;
 	uint32_t first_use_pc          = 0;

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

@@ -755,8 +755,9 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 		if (storage || image.conversion_format != Prospero::BufferFormat::kInvalid) {
 			image.shader_swizzle = DescriptorImageSwizzle(descriptor);
 		}
-		const bool raw_sint_storage = storage && format == Prospero::BufferFormat::k32SInt &&
+		const auto raw_sint_format = RawSignedStorageFormat(base, format);
-		                              base.written && (base.atomic ? !base.atomic64 : !base.read);
+		const bool raw_sint_storage = raw_sint_format != Prospero::BufferFormat::kInvalid;
+		if (raw_sint_format == Prospero::BufferFormat::k16UInt) image.conversion_format = format;
 		image.numeric_class = Prospero::SampledTextureNumericClass(
 		    Prospero::RemapTextureFormat(format));
 		if (!storage && format == Prospero::BufferFormat::k11_11_10UNorm) {

tests/ResourceMaterializationTests.cpp

@@ -1164,6 +1164,82 @@ void TestMixedSamplerVariantsShareRuntimeDescriptor() {
         "sampler variants retained stale or duplicated descriptors after refresh");
 }
 
+void TestSignedStorageSpecialization() {
+  using namespace Libs::Graphics::ShaderRecompiler::IR;
+  using F = Libs::Graphics::Prospero::BufferFormat;
+  using N = Libs::Graphics::Prospero::TextureNumericClass;
+  Program program;
+  program.stage = Libs::Graphics::ShaderType::Compute;
+  program.srt_plan_complete = program.resource_tracking_complete = true;
+  program.user_data_count = 1u;
+  auto &block = AddValueBlock(program);
+  DescriptorSource source;
+  source.dword_count = 8u;
+  source.dwords.fill(Value(0u));
+  source.dwords[0] = Value(0x100u);
+  source.dwords[1] = Value(&block.AppendNewInst(ValueOpcode::GetUserData, {Value(ScalarReg(0))}));
+  constexpr auto swizzle = Libs::Graphics::DstSel(5, 4, 1, 0);
+  source.dwords[3] = Value(swizzle |
+      (uint32_t(Libs::Graphics::Prospero::ImageType::kColor2D) << 28u));
+  program.descriptor_sources.push_back(source);
+  program.info.images.push_back({.source = 0u, .resource_class = ImageResourceClass::Storage,
+      .dimension = Libs::Graphics::ShaderRecompiler::Decoder::ImageDimension::Dim2D, .written = true});
+  auto plan = ExtractResourcePlan(program);
+  std::array<uint32_t, 1> data{};
+  const SrtRuntime runtime{.user_data = data};
+  ResourceSnapshot snapshot;
+  ResourceSpecialization specialization, signed16;
+  for (const auto format : {F::k16SInt, F::k32SInt, F::k16SInt}) {
+    data[0] = uint32_t(format) << 20u;
+    Check(MaterializeResources(plan, runtime, snapshot, specialization),
+          "write-only signed storage image failed materialization");
+    const auto conversion = format == F::k16SInt ? format : F::kInvalid;
+    const auto view = format == F::k16SInt ? F::k16UInt : F::k32UInt;
+    Check(specialization.images[0].numeric_class == N::Uint &&
+              specialization.images[0].conversion_format == conversion &&
+              specialization.images[0].shader_swizzle == swizzle &&
+              snapshot.images[0].dwords[1] == data[0] &&
+              RawSignedStorageFormat(plan.info.images[0], format) == view,
+          "signed storage specialization lost guest format, conversion, swizzle or raw view width");
+    if (format == F::k16SInt) {
+      if (signed16.images.empty()) signed16 = specialization;
+      else Check(specialization == signed16, "signed storage format refresh retained a stale specialization");
+    } else {
+      Check(specialization != signed16, "signed 16/32-bit stores shared incompatible conversion state");
+    }
+    ApplyResourceSpecialization(program, specialization);
+    const auto binding = DescriptorBindingForImage(program.info.images[0]);
+    Check(binding && ImageBindingResourceClass(*binding) == ImageResourceClass::Storage,
+          "signed storage conversion did not reuse the unsigned storage binding");
+  }
+  for (const auto format : {F::k16SInt, F::k32SInt}) {
+    data[0] = uint32_t(format) << 20u;
+    auto &image = plan.info.images[0];
+    image.read = true;
+    Check(RawSignedStorageFormat(image, format) == F::kInvalid &&
+              !MaterializeResources(plan, runtime, snapshot, specialization),
+          "signed raw storage view accepted a non-atomic read/write image");
+    image.atomic = true;
+    Check((RawSignedStorageFormat(image, format) != F::kInvalid) == (format == F::k32SInt) &&
+              MaterializeResources(plan, runtime, snapshot, specialization) == (format == F::k32SInt),
+          "signed storage conversion changed atomic format eligibility");
+    image.atomic64 = true;
+    Check(RawSignedStorageFormat(image, format) == F::kInvalid &&
+              !MaterializeResources(plan, runtime, snapshot, specialization),
+          "signed 16/32-bit image acquired a 64-bit atomic raw view");
+    image.resource_class = ImageResourceClass::Sampled;
+    image.written = image.atomic = image.atomic64 = false;
+    Check(MaterializeResources(plan, runtime, snapshot, specialization) &&
+              specialization.images[0].numeric_class == N::Sint &&
+              specialization.images[0].conversion_format == F::kInvalid &&
+              RawSignedStorageFormat(image, format) == F::kInvalid,
+          "storage conversion changed ordinary signed image reads");
+    image.resource_class = ImageResourceClass::Storage;
+    image.written = true;
+    image.read = false;
+  }
+}
+
 void TestIndirectImageSamplerClasses(bool mixed) {
   using namespace Libs::Graphics::ShaderRecompiler::IR;
   using Libs::Graphics::Prospero::BufferFormat;
@@ -1256,6 +1332,7 @@ int main() {
   TestCyclicControlPredicatesRemainUnknown();
   TestEmptyScalarBufferSpecialization();
   TestMixedSamplerVariantsShareRuntimeDescriptor();
+  TestSignedStorageSpecialization();
   for (const bool mixed : {false, true}) TestIndirectImageSamplerClasses(mixed);
   std::puts("ResourceMaterializationTests: all cases passed");
   return 0;

tests/ShaderRecompilerComputeTests.cpp

@@ -37,6 +37,7 @@
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "graphics/host_gpu/renderer/sync.h"
 #include "graphics/host_gpu/vulkanCommon.h"
+#include <vulkan/vulkan_format_traits.hpp>
 #include "graphics/presentation/window/windowInternal.h"
 #include "graphics/shader/recompiler/BufferFormat.h"
 #include "graphics/shader/recompiler/ShaderRecompiler.h"
@@ -2009,7 +2010,6 @@ public:
     u32 height = 0;
     u32 layers = 1;
     u32 mip_levels = 1;
-    u32 dwords_per_pixel = 0;
   };
 
   [[nodiscard]] vk::Device Device() const { return m_device; }
@@ -15749,11 +15749,12 @@ public:
     return std::max(value, 1u);
   }
 
-  static size_t ImageMipDwordCount(u32 width, u32 height, u32 dwords_per_pixel,
+  static size_t ImageMipDwordCount(u32 width, u32 height, vk::Format format,
                                    u32 level, u32 layers = 1) {
-    return static_cast<size_t>(MipExtent(width, level)) *
+    const auto block = vk::blockExtent(format);
-           static_cast<size_t>(MipExtent(height, level)) * layers *
+    const auto blocks_x = (MipExtent(width, level) + block[0] - 1u) / block[0];
-           dwords_per_pixel;
+    const auto blocks_y = (MipExtent(height, level) + block[1] - 1u) / block[1];
+    return (size_t{blocks_x} * blocks_y * layers * vk::blockSize(format) + 3u) / 4u;
   }
 
   Image CreateImageMips(const char *shader_name, u32 width, u32 height,
@@ -15772,7 +15773,8 @@ public:
     ret.height = height;
     ret.layers = layers;
     ret.mip_levels = std::max<u32>(static_cast<u32>(initial_mips.size()), 1u);
-    ret.dwords_per_pixel = dwords_per_pixel;
+    Require(shader_name, "image format size", dwords_per_pixel == (vk::blockSize(format) + 3u) / 4u,
+            "image fixture word size disagrees with its native format");
 
     vk::ImageCreateInfo image_info{};
     image_info.sType = vk::StructureType::eImageCreateInfo;
@@ -15845,13 +15847,13 @@ public:
       size_t total_dwords = 0;
       for (u32 level = 0; level < ret.mip_levels; level++) {
         total_dwords +=
-            ImageMipDwordCount(width, height, dwords_per_pixel, level, layers);
+            ImageMipDwordCount(width, height, format, level, layers);
       }
       std::vector<u32> contents(total_dwords, 0);
       size_t offset = 0;
       for (u32 level = 0; level < ret.mip_levels; level++) {
         const auto level_dwords =
-            ImageMipDwordCount(width, height, dwords_per_pixel, level, layers);
+            ImageMipDwordCount(width, height, format, level, layers);
         const auto &src = initial_mips[level];
         for (size_t i = 0; i < src.size() && i < level_dwords; i++) {
           contents[offset + i] = src[i];
@@ -15891,9 +15893,8 @@ public:
   }
 
   std::vector<u32> ReadImage(const char *shader_name, Image *image) {
-    const auto dword_count = static_cast<size_t>(image->width) *
+    const auto dword_count = ImageMipDwordCount(image->width, image->height,
-                             static_cast<size_t>(image->height) *
+                                               image->format, 0, image->layers);
-                             image->layers * image->dwords_per_pixel;
     auto staging = CreateHostBuffer(shader_name, dword_count * sizeof(u32),
                                    vk::BufferUsageFlagBits::eTransferDst, {});
 
@@ -19549,7 +19550,7 @@ private:
       copies.push_back(copy);
       offset += static_cast<vk::DeviceSize>(
           ImageMipDwordCount(image->width, image->height,
-                             image->dwords_per_pixel, level, image->layers) *
+                             image->format, level, image->layers) *
           sizeof(u32));
     }
     cmd.copyBufferToImage(staging, image->image,
@@ -31882,23 +31883,23 @@ TestCase TBufferLoadFormatXSintSignExtendsSubDword() {
 
 TestCase TBufferStoreFormatXSintWritesSubDword() {
   using O = ShaderOpcode;
-
+  TestCase test;
-  std::vector<u32> code;
+  test.name = "TBufferStoreFormatXSintWritesSubDword";
-  AppendVMovU32(&code, 20, 0);
+  constexpr std::array<u32, 6> inputs{0xffffff80u, 0xffffff7fu, 128u,
-  AppendVMovLiteral(&code, 0, 0xffffff80u);
+                                     0xffff8001u, 0xffff7fffu, 32768u};
-  code.push_back(EncodeMtbuf0(0x04, 6, 0));
+  constexpr std::array<u32, 6> encoded{0x80u, 0x80u, 0x7fu, 0x8001u, 0x8000u, 0x7fffu};
-  code.push_back(EncodeMtbuf1(0x04, 0, 0, 20));
+  test.initial.assign(inputs.size(), 0x11223344u);
-  AppendVMovU32(&code, 20, 4);
+  test.expected = test.initial;
-  AppendVMovLiteral(&code, 1, 0xffff8001u);
+  for (u32 index = 0; index < inputs.size(); ++index) {
-  code.push_back(EncodeMtbuf0(0x04, 12, 0));
+    AppendVMovU32(&test.code, 20, index * 4u);
-  code.push_back(EncodeMtbuf1(0x04, 1, 0, 20));
+    AppendVMovLiteral(&test.code, 0, inputs[index]);
-  AppendEnd(&code);
+    test.code.push_back(EncodeMtbuf0(0x04, index < 3u ? 6u : 12u, 0));
-
+    test.code.push_back(EncodeMtbuf1(0x04, 0, 0, 20));
-  return {"TBufferStoreFormatXSintWritesSubDword",
+    test.expected[index] = (test.expected[index] & (index < 3u ? 0xffffff00u : 0xffff0000u)) | encoded[index];
-          code,
+  }
-          {0x11223344u, 0x55667788u},
+  AppendEnd(&test.code);
-          {0x11223380u, 0x55668001u},
+  test.opcodes = {O::V_MOV_B32, O::TBUFFER_STORE_FORMAT_X, O::S_ENDPGM};
-          {O::V_MOV_B32, O::TBUFFER_STORE_FORMAT_X, O::S_ENDPGM}};
+  return test;
 }
 
 TestCase TBufferLoadFormatXy88IntegerComponents() {
@@ -38292,6 +38293,82 @@ TestCase ImageStoreR32FloatUsesFormatlessStorageImage() {
   return test;
 }
 
+void CheckSigned16ImageStore(VulkanHarness &vulkan) {
+  constexpr std::array<int32_t, 8> values{INT32_MIN, -32769, -32768, -1, 0, 32767, 32768, INT32_MAX};
+  const u32 alignment = vulkan.RuntimeContext().GetPhysicalDeviceProperties()
+                            .limits.minStorageBufferOffsetAlignment;
+  for (const u32 wave : {32u, 64u}) {
+    for (const bool d16 : {false, true}) {
+      for (const u32 component : {0u, 2u}) {
+        TestCase test;
+        test.name = "Signed16ImageAndBufferStore";
+        const u32 width = wave - 2u;
+        const u32 output = (wave * 4u + alignment - 1u) / alignment * alignment;
+        test.user_data = MakeStorageTextureData(Prospero::BufferFormat::k16SInt);
+        test.user_data[1] |= ((width - 1u) & 3u) << 30u;
+        test.user_data[2] = ((width - 1u) >> 2u) | (1u << 14u);
+        test.has_user_data = test.has_compute_info = true;
+        test.image_descriptor_swizzle = component == 0u ? DstSel(4, 0, 0, 1) : DstSel(0, 0, 4, 1);
+        const std::array<u32, 8> descriptors{0x1000u, 4u << 16u, wave, 0x5204u,
+            0x2000u, 2u << 16u, wave, (u32(Prospero::BufferFormat::k16SInt) << 12u) | 4u};
+        std::copy(descriptors.begin(), descriptors.end(), test.user_data.begin() + 8u);
+        test.storage_buffer_ranges = {{0u, wave * 4u}, {output, wave * 2u}};
+        test.compute_info.wave_size = wave;
+        test.compute_info.threads_num[0] = wave;
+        test.compute_info.threads_num[1] = test.compute_info.threads_num[2] = 1u;
+        test.compute_info.thread_ids_num = 1u;
+        test.compute_info.lds_size_dwords = 0u;
+        test.initial.assign(output / 4u + wave / 2u, 0xdeadbeefu);
+        for (u32 lane = 0; lane < wave; ++lane)
+          test.initial[lane] = d16 ? 0x7a5b0000u | (u32(values[lane % values.size()]) & 0xffffu)
+                                   : u32(values[lane % values.size()]);
+        test.expected = test.initial;
+        std::vector<u32> pixels(width, 0x13571357u), expected = pixels;
+        for (u32 lane = 0; lane < wave; ++lane) {
+          const auto input = values[lane % values.size()];
+          const auto value = d16 ? int32_t(std::bit_cast<int16_t>(uint16_t(input))) : input;
+          const u32 encoded = uint16_t(std::clamp(value, -32768, 32767));
+          const auto set_half = [&](std::vector<u32> &words, u32 index) {
+            const auto shift = (index & 1u) * 16u;
+            words[index / 2u] = (words[index / 2u] & ~(0xffffu << shift)) | (encoded << shift);
+          };
+          set_half(test.expected, output / 2u + lane);
+          if (lane < width) {
+            set_half(expected, lane);
+            if (lane & 1u) set_half(expected, width + lane);
+          }
+        }
+        auto &code = test.code;
+        code = {EncodeMubuf0(0x0c, 0, true, false), EncodeMubuf1(4, 2, 0),
+                EncodeSopp(0x0c, 0), EncodeMubuf0(d16 ? 0x84 : 0x04, 0, true, false),
+                EncodeMubuf1(4, 3, 0), EncodeVop1(0x01, 20, Vgpr(0))};
+        AppendVMovU32(&code, 21, 0u);
+        code.insert(code.end(), {EncodeMimg0(0x08, 1u << component),
+                                 EncodeMimg1(4, 20, 0, 0, false, d16)});
+        AppendVMovU32(&code, 21, 1u);
+        AppendSMovLiteral(&code, 126, 0xaaaaaaaau);
+        AppendSMovLiteral(&code, 127, wave == 64u ? 0xaaaaaaaau : 0u);
+        code.insert(code.end(), {EncodeMimg0(0x08, 1u << component),
+                                 EncodeMimg1(4, 20, 0, 0, false, d16)});
+        AppendEnd(&code);
+        test.opcodes = {ShaderOpcode::IMAGE_STORE};
+        const auto compiled = CompileCase(test, vulkan.SubgroupSize());
+        auto buffer = vulkan.CreateStorageBuffer(test.name, test.initial, test.initial.size());
+        auto image = vulkan.CreateImage2D(test.name, width, 2u, vk::Format::eR16Sint,
+            vk::ImageUsageFlagBits::eStorage, pixels, 1u, vk::ImageLayout::eGeneral, vk::Format::eR16Uint);
+        vulkan.Dispatch(test, compiled, buffer, nullptr, nullptr, nullptr, &image);
+        CompareWords(test, "signed image saturation, sparse channels, masked lanes and OOB", expected,
+                     vulkan.ReadImage(test.name, &image));
+        CompareWords(test, "shared buffer conversion and adjacent halfwords", test.expected,
+                     vulkan.ReadBuffer(test.name, buffer, test.expected.size()));
+        vulkan.DestroyImage(&image);
+        vulkan.DestroyBuffer(&buffer);
+      }
+    }
+  }
+  std::printf("[compute] %-32s ok\n", "Signed16ImageAndBufferStore");
+}
+
 TestCase ImageStoreR32SintUsesRawUintView() {
   using O = ShaderOpcode;
 
@@ -45517,6 +45594,7 @@ int main(int argc, char **argv) {
     RunCase(&vulkan, BufferStoreFormatXRejectsPartialRecord());
     RunCase(&vulkan, BufferStoreFormatXyRejectsPartialRecord());
     RunCase(&vulkan, BufferFormatStoreVariants());
+    RunCase(&vulkan, TBufferStoreFormatXSintWritesSubDword());
     RunCase(&vulkan, BufferStoreByteD16Hi(false));
     RunCase(&vulkan, BufferStoreByteD16Hi(true));
     return 0;
@@ -45634,6 +45712,7 @@ int main(int argc, char **argv) {
   if (argc == 2 && std::strcmp(argv[1], "--packed-texture-only") == 0) {
     VulkanHarness vulkan;
     vulkan.CheckPackedTextureComponents();
+    CheckSigned16ImageStore(vulkan);
     vulkan.CheckBgra16Readback();
     RunCase(&vulkan, ImageLoadPackedUintUnpacksAndSwizzles());
     RunCase(&vulkan, ImageLoadPackedUnorm<false, false>());
@@ -46414,6 +46493,7 @@ int main(int argc, char **argv) {
   CheckTessellationPrograms();
   CheckPixelParameterAliases();
   CheckRectListShaders();
+  CheckSigned16ImageStore(vulkan);
   CheckBoundedFormattedBufferCopy(vulkan);
   CheckIndirectBufferStore(vulkan);
   CheckBoundedRecordBufferAccess(vulkan);