shader: add missing 64-bit integer image atomics

Co-authored-by: shadowbeat070 <70843316+shadowbeat070@users.noreply.github.com>

CMakeLists.txt

@@ -259,7 +259,7 @@ set(gpu_tiler_shader_names
 	depth
 	promote_d16
 	demote_d16
-	swap_bgra16
+	color_transform
 )
 file(GLOB gpu_tiler_shader_includes CONFIGURE_DEPENDS "${gpu_tiler_shader_dir}/gpu_tiler_*.inc")
 foreach(shader_name IN LISTS gpu_tiler_shader_names)

src/graphics/guest_gpu/gpu_format.cpp

@@ -20,7 +20,8 @@ struct FormatInfo {
 
 constexpr FormatInfo kFormatInfo[] = {
 	{BufferFormat::k8UNorm, 1, 0, 1, true, false},
-	{BufferFormat::k8SNorm, 0, 0, 1, false, false},
+	{BufferFormat::k8UScaled, 1, 0, 0, true, false},
+	{BufferFormat::k8SNorm, 1, 0, 1, false, false},
 	{BufferFormat::k8UInt, 1, 0, 1, true, true},
 	{BufferFormat::k16UNorm, 2, 0, 2, true, false},
 	{BufferFormat::k16SNorm, 2, 0, 2, true, false},
@@ -28,6 +29,7 @@ constexpr FormatInfo kFormatInfo[] = {
 	{BufferFormat::k16SInt, 2, 0, 2, true, false, true},
 	{BufferFormat::k16Float, 2, 0, 2, true, false},
 	{BufferFormat::k8_8UNorm, 2, 0, 2, true, false},
+	{BufferFormat::k8_8UScaled, 2, 0, 0, true, false},
 	{BufferFormat::k8_8SNorm, 2, 0, 2, true, false},
 	{BufferFormat::k8_8UInt, 2, 0, 2, true, true},
 	{BufferFormat::k8_8SInt, 2, 0, 2, true, false, true},
@@ -41,6 +43,7 @@ constexpr FormatInfo kFormatInfo[] = {
 	{BufferFormat::k16_16Float, 4, 0, 4, true, false},
 	{BufferFormat::k11_11_10UInt, 4, 0, 4, true, true},
 	{BufferFormat::k11_11_10Float, 4, 0, 4, true, false},
+	{BufferFormat::k10_11_11Float, 4, 0, 4, true, false},
 	{BufferFormat::k10_10_10_2UNorm, 4, 0, 4, true, false},
 	{BufferFormat::k10_10_10_2UInt, 4, 0, 4, true, true},
 	{BufferFormat::k8_8_8_8UNorm, 4, 0, 4, true, false},
@@ -67,7 +70,7 @@ constexpr FormatInfo kFormatInfo[] = {
 	{BufferFormat::k9_9_9_5Float, 4, 0, 0, true, false},
 	{BufferFormat::k5_6_5UNorm, 2, 0, 2, true, false},
 	{BufferFormat::k5_5_5_1UNorm, 2, 0, 2, true, false},
-	{BufferFormat::k1_5_5_5UNorm, 0, 0, 2, false, false},
+	{BufferFormat::k1_5_5_5UNorm, 2, 0, 2, false, false},
 	{BufferFormat::k4_4_4_4UNorm, 2, 0, 2, true, false},
 	{BufferFormat::kFmask8_S2_F1, 1, 0, 1, true, false},
 	{BufferFormat::kFmask8_S4_F1, 1, 0, 1, true, false},
@@ -231,7 +234,12 @@ TextureNumericClass SampledTextureNumericClass(BufferFormat format) {
 }
 
 BufferFormat RemapTextureFormat(BufferFormat format) {
-	return format == BufferFormat::k11_11_10UInt ? BufferFormat::k32UInt : format;
+	switch (format) {
+		case BufferFormat::k8UScaled: return BufferFormat::k8UNorm;
+		case BufferFormat::k8_8UScaled: return BufferFormat::k8_8UNorm;
+		case BufferFormat::k11_11_10UInt: return BufferFormat::k32UInt;
+		default: return format;
+	}
 }
 
 } // namespace Libs::Graphics::Prospero

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

@@ -195,6 +195,9 @@ bool TextureCache::SameBacking(const ImageInfo& cached, const ImageInfo& request
 	if (cached.tile_mode != requested.tile_mode) {
 		return false;
 	}
+	if (cached.GetColorTransform() != requested.GetColorTransform()) {
+		return false;
+	}
 	if (!ImageViewOps::FormatsCompatible(cached.pixel_format, requested.pixel_format) ||
 	    cached.type != requested.type) {
 		return false;
@@ -634,6 +637,7 @@ void TextureCache::CopyImage(ImageId destination_id, ImageId source_id) {
 	const bool source_depth = source.info.IsDepth();
 	const bool dest_depth   = destination.info.IsDepth();
 	const bool direct_copy =
+	    source.info.GetColorTransform() == destination.info.GetColorTransform() &&
 	    (source.backing.image_type == destination.backing.image_type ||
 	     (source.backing.image_type != vk::ImageType::e1D &&
 	      destination.backing.image_type != vk::ImageType::e1D)) &&
@@ -647,7 +651,7 @@ void TextureCache::CopyImage(ImageId destination_id, ImageId source_id) {
 			EXIT("TextureCache: cross-format multisample image copy is unsupported\n");
 		}
 		auto& copy_buffer = m_buffer_cache.GetUtilityBuffer(MemoryUsage::DeviceLocal);
-		destination.CopyImageWithBuffer(source, copy_buffer);
+		destination.CopyImageWithBuffer(source, copy_buffer, m_tiler);
 	}
 	if (source.IsGpuModified()) {
 		destination.MarkGpuModified();
@@ -710,7 +714,8 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 	        cached.info.data.size / cached.info.resources.layers &&
 	    !requested.HasStencil() && !cached.info.HasStencil() && !requested.HasMetadata() &&
 	    !cached.info.HasMetadata();
-	bool recreate = cached.info.resources < requested.resources;
+	bool recreate = cached.info.resources < requested.resources ||
+	                requested.IsVolume() != cached.info.IsVolume();
 	switch (binding) {
 		case BindingType::Texture:
 			recreate |= requested.IsDepth() && !cached.info.IsDepth();
@@ -801,6 +806,9 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 			}
 			return {merged_id};
 		}
+		if (requested.GetColorTransform() != cached.info.GetColorTransform()) {
+			return {ExpandImage(requested, cached_id)};
+		}
 		if (requested.IsBlock() && !cached.info.IsBlock()) {
 			return {ExpandImage(requested, cached_id)};
 		}
@@ -836,6 +844,9 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 		      requested.resources.layers == cached.info.resources.layers))) {
 			return {ExpandImage(requested, cached_id)};
 		}
+		if (requested.type != cached.info.type) {
+			return {merged_id};
+		}
 		if (requested.pixel_format != cached.info.pixel_format ||
 		    requested.data.size <= cached.info.data.size) {
 			const auto result_id = merged_id ? merged_id : cached_id;
@@ -905,8 +916,8 @@ struct TextureCache::TextureTransfer {
 	TextureUploadLayout              layout;
 	std::vector<vk::BufferImageCopy> regions;
 	std::vector<GpuTileInfo>         tiles;
-	bool                             swap_bgra16 = false;
-	bool                             valid       = false;
+	ColorTransform                   color_transform = ColorTransform::None;
+	bool                             valid           = false;
 
 	[[nodiscard]] uint64_t LinearSize() const {
 		uint64_t size = 0;
@@ -935,7 +946,7 @@ TextureCache::BuildTextureTransfer(const Image& image, BindingType binding,
 	const char* owner            = "TextureCache readback";
 
 	TextureTransfer transfer;
-	transfer.swap_bgra16 = info.bgra16;
+	transfer.color_transform = info.GetColorTransform();
 	if (upload) {
 		if ((render_target || video_out) &&
 		    (info.resources.layers == 0 || info.data.size % info.resources.layers != 0 ||
@@ -1023,10 +1034,9 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 		TileManager::Result linear {source.Handle(), source_offset, info.data.size};
 		if (!transfer.tiles.empty()) {
 			linear = m_tiler.Detile(source.Handle(), source_offset, info.data.size,
-			                        transfer.LinearSize(), transfer.tiles);
-		}
-		if (transfer.swap_bgra16) {
-			linear = m_tiler.SwapBgra16(linear);
+			                        transfer.LinearSize(), transfer.tiles, transfer.color_transform);
+		} else if (transfer.color_transform != ColorTransform::None) {
+			linear = m_tiler.TransformColor(linear, transfer.color_transform, true);
 		}
 		upload(transfer.regions, linear);
 		return;
@@ -1741,14 +1751,14 @@ void TextureCache::DownloadImage(Image& image, Buffer& destination, uint64_t des
 	}
 
 	auto&      texture   = transfer.texture;
-	const auto transform = texture.swap_bgra16 ? TileManager::ColorTransform::SwapBgra16
-	                                           : TileManager::ColorTransform::None;
+	const auto transform = texture.color_transform;
 	if (texture.tiles.empty()) {
-		if (transform == TileManager::ColorTransform::SwapBgra16) {
+		if (transform != ColorTransform::None) {
 			auto linear = m_tiler.GetScratchBuffer(destination_size);
 			image.Download(texture.regions, linear.buffer, 0, linear.size);
-			m_tiler.SwapBgra16(linear,
-			                   {destination.Handle(), destination_offset, destination_size});
+			m_tiler.TransformColor(linear,
+			                       {destination.Handle(), destination_offset, destination_size},
+			                       transform, false);
 			return;
 		}
 		for (auto& copy: texture.regions) {

src/graphics/host_gpu/renderer/colorRenderTarget.cpp

@@ -264,7 +264,7 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 	desc.type              = TextureCache::BindingType::RenderTarget;
 	desc.info.data         = {rt.base.addr, backing_size};
 	desc.info.pixel_format = target_format.format;
-	desc.info.guest_format = transfer_format;
+	desc.info.guest_format = target_format.guest_format;
 	desc.info.type         = image_type;
 	desc.info.extent       = {width, height, depth};
 	desc.info.resources    = {levels, volume ? 1u : view.image_layers};

src/graphics/host_gpu/renderer/depthRenderTarget.cpp

@@ -160,8 +160,7 @@ static TextureCache::ImageDesc MakeDepthTargetDesc(const CommandBuffer& buffer,
 			           z.depth_view.slice_max);
 	}
 	// EXPCLEAR permits an HTile acceleration state; the host attachment is already expanded.
-	if (z.z_info.partially_resident ||
-	    z.stencil_info.partially_resident || z.z_info.max_mip_level != 0 ||
+	if (z.z_info.max_mip_level != 0 ||
 	    z.depth_view.current_mip_level != 0 || unsupported_shading_rate_encoding ||
 	    depth_address == 0 || (depth_address & 0xffffu) != 0) {
 		DepthFatal("unsupported depth register state");

src/graphics/host_gpu/renderer/image/image.cpp

@@ -6,6 +6,7 @@
 #include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
 #include "graphics/host_gpu/renderer/image/imageView.h"
+#include "graphics/host_gpu/renderer/image/tiler.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "kernel/memory.h"
 
@@ -456,7 +457,7 @@ uint32_t Image::CopyRows(uint64_t row_size, uint32_t rows, uint64_t capacity) no
 	return static_cast<uint32_t>(std::min<uint64_t>(rows, capacity / row_size));
 }
 
-void Image::CopyImageWithBuffer(Image& source, Buffer& buffer) {
+void Image::CopyImageWithBuffer(Image& source, Buffer& buffer, TileManager& tiler) {
 	EXIT_IF(buffer.Handle() == nullptr || source.backing.samples != 1 || backing.samples != 1);
 	m_scheduler.EndRendering();
 	const uint32_t levels = std::min(source.backing.mip_levels, backing.mip_levels);
@@ -474,6 +475,8 @@ void Image::CopyImageWithBuffer(Image& source, Buffer& buffer) {
 	const uint32_t destination_block = info.IsBlock() ? 4u : 1u;
 	EXIT_IF(levels == 0 || source_bytes == 0 || source_bytes != destination_bytes ||
 	        source_block != destination_block);
+	const auto source_transform = source.info.GetColorTransform();
+	const auto target_transform = info.GetColorTransform();
 
 	vk::BufferMemoryBarrier2 barrier {};
 	barrier.srcStageMask        = vk::PipelineStageFlagBits2::eTransfer;
@@ -535,9 +538,19 @@ void Image::CopyImageWithBuffer(Image& source, Buffer& buffer) {
 				command.copyImageToBuffer(source.backing.image,
 				                          vk::ImageLayout::eTransferSrcOptimal, buffer.Handle(),
 				                          source_copy);
-				barrier.srcAccessMask = vk::AccessFlagBits2::eTransferWrite;
-				barrier.dstAccessMask = vk::AccessFlagBits2::eTransferRead;
-				command.pipelineBarrier2(dependency);
+				if (source_transform != target_transform) {
+					const TileManager::Result bytes {buffer.Handle(), 0, copy_size};
+					if (source_transform != ColorTransform::None) {
+						tiler.TransformColor(bytes, bytes, source_transform, false);
+					}
+					if (target_transform != ColorTransform::None) {
+						tiler.TransformColor(bytes, bytes, target_transform, true);
+					}
+				} else {
+					barrier.srcAccessMask = vk::AccessFlagBits2::eTransferWrite;
+					barrier.dstAccessMask = vk::AccessFlagBits2::eTransferRead;
+					command.pipelineBarrier2(dependency);
+				}
 				command.copyBufferToImage(buffer.Handle(), backing.image,
 				                          vk::ImageLayout::eTransferDstOptimal, destination_copy);
 			}

src/graphics/host_gpu/renderer/image/image.h

@@ -18,6 +18,7 @@ namespace Libs::Graphics {
 
 class Buffer;
 class CommandScheduler;
+class TileManager;
 struct ImageTestAccess;
 
 using ImageId = Common::SlotId;
@@ -66,7 +67,7 @@ public:
 	void CopyImage(Image& source);
 	void Resolve(Image& source, const ImageSubresourceRange& source_range,
 	             const ImageSubresourceRange& destination_range);
-	void CopyImageWithBuffer(Image& source, Buffer& buffer);
+	void CopyImageWithBuffer(Image& source, Buffer& buffer, TileManager& tiler);
 	void CopyMip(Image& source, uint32_t mip, uint32_t layer);
 
 	void InvalidateCpuWrite(uint64_t vaddr, uint64_t size) {

src/graphics/host_gpu/renderer/image/imageInfo.h

@@ -15,6 +15,8 @@
 
 namespace Libs::Graphics {
 
+enum class ColorTransform : uint32_t { None, SwapBgra16, Reverse10_11_11 };
+
 enum class VideoOutCompression : uint8_t { Uncompressed, Dcc256_256_0, Dcc256_64_64, Unsupported };
 
 enum class ImageMetadataKind : uint8_t { None, Htile, Dcc, Cmask };
@@ -70,6 +72,14 @@ struct ImageInfo {
 	bool                         bgra16          = false;
 	std::array<ImageMipInfo, 16> mip_layout {};
 
+	[[nodiscard]] ColorTransform GetColorTransform() const noexcept {
+		if (bgra16) return ColorTransform::SwapBgra16;
+		switch (guest_format) {
+			case Prospero::BufferFormat::k10_11_11Float:
+				return ColorTransform::Reverse10_11_11;
+			default: return ColorTransform::None;
+		}
+	}
 	[[nodiscard]] constexpr bool HasStencil() const noexcept { return !stencil.Empty(); }
 	[[nodiscard]] constexpr bool HasMetadata() const noexcept {
 		return metadata.kind != ImageMetadataKind::None;
@@ -91,7 +101,8 @@ struct ImageInfo {
 	}
 	[[nodiscard]] bool IsCompatible(const ImageInfo& other) const noexcept {
 		return pixel_format == other.pixel_format && samples == other.samples &&
-		       bytes_per_block == other.bytes_per_block;
+		       bytes_per_block == other.bytes_per_block &&
+		       GetColorTransform() == other.GetColorTransform();
 	}
 	[[nodiscard]] int32_t MipOf(const ImageInfo& container) const noexcept {
 		if (container.resources.levels > container.mip_layout.size()) {

src/graphics/host_gpu/renderer/image/textureCommon.cpp

@@ -35,6 +35,9 @@ struct HostFormatInfo {
 
 HostFormatInfo ResolveHostFormat(Prospero::BufferFormat guest_format,
                                  Prospero::ChannelOrder order) {
+	if (guest_format == Prospero::BufferFormat::k10_11_11Float) {
+		return {vk::Format::eB10G11R11UfloatPack32, Prospero::ColorMappingBgra};
+	}
 	if (order == Prospero::ChannelOrder::kAlt) {
 		switch (guest_format) {
 			case Prospero::BufferFormat::k8_8_8_8UNorm:
@@ -131,7 +134,8 @@ RenderTargetFormatInfo TextureGetRenderTargetFormat(Prospero::ChannelLayout layo
 		if (host_format.format != vk::Format::eUndefined && bytes != 0) {
 			const auto order_mapping =
 			    kRenderTargetColorMappings[static_cast<size_t>(order)][encoding.components - 1u];
-			return {host_format.format, bytes, host_format.host_to_storage.Then(order_mapping)};
+			return {host_format.format, bytes, host_format.host_to_storage.Then(order_mapping),
+			        encoding.buffer_format};
 		}
 	}
 	EXIT("unsupported render-target format combination: layout=%u type=%u order=%u\n",

src/graphics/host_gpu/renderer/image/textureCommon.h

@@ -15,6 +15,7 @@ struct RenderTargetFormatInfo {
 	vk::Format                      format            = vk::Format::eUndefined;
 	uint32_t                        bytes_per_element = 0;
 	Prospero::ColorComponentMapping export_mapping;
+	Prospero::BufferFormat          guest_format = Prospero::BufferFormat::kInvalid;
 };
 
 struct SurfaceFormatInfo {

src/graphics/host_gpu/renderer/image/tiler.cpp

@@ -13,7 +13,7 @@
 #include "gpu_tiler_shaders/gpu_tiler_standard4_spv.h"
 #include "gpu_tiler_shaders/gpu_tiler_standard64_3d_spv.h"
 #include "gpu_tiler_shaders/gpu_tiler_standard64_spv.h"
-#include "gpu_tiler_shaders/gpu_tiler_swap_bgra16_spv.h"
+#include "gpu_tiler_shaders/gpu_tiler_color_transform_spv.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
@@ -31,7 +31,7 @@ TileManager::TileManager(GraphicContext& graphics, CommandScheduler& scheduler,
                          StreamBuffer& stream_buffer)
     : m_graphics(graphics), m_scheduler(scheduler), m_stream_buffer(stream_buffer) {
 	static_assert(FamilyCount == 9);
-	static_assert(sizeof(Push) == 52);
+	static_assert(sizeof(Push) == 60);
 	std::array<vk::DescriptorSetLayoutBinding, 3> bindings {};
 	for (uint32_t index = 0; index < 2; index++) {
 		bindings[index] = {index, vk::DescriptorType::eStorageBuffer, 1,
@@ -77,8 +77,8 @@ TileManager::~TileManager() {
 	if (m_d32_to_d16 != nullptr) {
 		m_graphics.device.destroyPipeline(m_d32_to_d16, nullptr);
 	}
-	if (m_swap_bgra16 != nullptr) {
-		m_graphics.device.destroyPipeline(m_swap_bgra16, nullptr);
+	if (m_color_transform != nullptr) {
+		m_graphics.device.destroyPipeline(m_color_transform, nullptr);
 	}
 	if (m_pipeline_layout != nullptr) {
 		m_graphics.device.destroyPipelineLayout(m_pipeline_layout, nullptr);
@@ -90,7 +90,8 @@ TileManager::~TileManager() {
 
 void TileManager::Prepare(bool tile, uint64_t tiled_capacity, uint64_t linear_capacity,
                           std::span<const GpuTileInfo> infos, uint64_t source_base,
-                          uint64_t target_base, std::vector<Dispatch>& dispatches) {
+                          uint64_t target_base, std::vector<Dispatch>& dispatches,
+                          ColorTransform transform) {
 	EXIT_IF(infos.empty() || tiled_capacity == 0 || linear_capacity == 0);
 	const auto& limits = m_graphics.GetPhysicalDeviceProperties().limits;
 	EXIT_NOT_IMPLEMENTED(tiled_capacity > UINT32_MAX || linear_capacity > UINT32_MAX);
@@ -108,6 +109,8 @@ void TileManager::Prepare(bool tile, uint64_t tiled_capacity, uint64_t linear_ca
 	dispatches.clear();
 	dispatches.reserve(infos.size());
 	for (const auto& info: infos) {
+		EXIT_IF((transform == ColorTransform::SwapBgra16 && info.bytes_per_element != 8u) ||
+		        (transform == ColorTransform::Reverse10_11_11 && info.bytes_per_element != 4u));
 		TileBlockLayout block {};
 		const uint32_t  tiled_width  = info.tiled_width != 0 ? info.tiled_width : info.pitch;
 		const uint32_t  tiled_height = info.tiled_height != 0 ? info.tiled_height : info.height;
@@ -193,6 +196,7 @@ void TileManager::Prepare(bool tile, uint64_t tiled_capacity, uint64_t linear_ca
 		dispatch.push.tail_x           = info.tail_x;
 		dispatch.push.tail_y           = info.tail_y;
 		dispatch.push.tail             = info.tail;
+		dispatch.push.color_transform  = static_cast<uint32_t>(transform);
 		dispatches.push_back(dispatch);
 	}
 
@@ -338,13 +342,14 @@ void TileManager::Record(vk::Buffer source, uint64_t source_offset,
 
 TileManager::Result TileManager::Detile(vk::Buffer tiled, uint64_t tiled_offset,
                                         uint64_t tiled_capacity, uint64_t linear_capacity,
-                                        std::span<const GpuTileInfo> infos) {
+                                        std::span<const GpuTileInfo> infos,
+                                        ColorTransform transform) {
 	const auto&    limits = m_graphics.GetPhysicalDeviceProperties().limits;
 	const uint64_t descriptor_alignment =
 	    std::max<uint64_t>(limits.minStorageBufferOffsetAlignment, 4);
 	const uint64_t        source_base = tiled_offset & (descriptor_alignment - 1);
 	std::vector<Dispatch> dispatches;
-	Prepare(false, tiled_capacity, linear_capacity, infos, source_base, 0, dispatches);
+	Prepare(false, tiled_capacity, linear_capacity, infos, source_base, 0, dispatches, transform);
 	auto scratch = GetScratchBuffer(linear_capacity, tiled);
 	Record(tiled, tiled_offset, tiled_capacity, scratch.buffer, 0, scratch.size, dispatches,
 	       true);
@@ -375,15 +380,11 @@ void TileManager::TileImage(Image& image, std::span<const vk::BufferImageCopy> r
 	    std::max<uint64_t>(limits.minStorageBufferOffsetAlignment, 4);
 	const uint64_t        target_base = tiled_offset & (descriptor_alignment - 1);
 	std::vector<Dispatch> dispatches;
-	// Reserve stream parameters before recording the image download and conversion.
-	Prepare(true, tiled_capacity, linear_capacity, infos, 0, target_base, dispatches);
+	// Reserve stream parameters before recording the image download.
+	Prepare(true, tiled_capacity, linear_capacity, infos, 0, target_base, dispatches, transform);
 	auto linear = GetScratchBuffer(linear_capacity, tiled);
 	image.Download(regions, linear.buffer, 0, linear.size);
-	Result source {linear.buffer, 0, linear.size};
-	if (transform == ColorTransform::SwapBgra16) {
-		source = SwapBgra16(source);
-	}
-	Record(source.buffer, source.offset, linear_capacity, tiled, tiled_offset, tiled_capacity,
+	Record(linear.buffer, 0, linear_capacity, tiled, tiled_offset, tiled_capacity,
 	       dispatches, false);
 }
 
@@ -582,9 +583,20 @@ void TileManager::ConvertD16(Result source, Result target, D16Direction directio
 	                        &barriers[1], 0, nullptr);
 }
 
-void TileManager::SwapBgra16(Result input, Result output, uint32_t pixels) {
-	if (m_swap_bgra16 == nullptr) {
-		const auto module = CompileSPV(GPU_TILER_SWAP_BGRA16_SPV, m_graphics.device);
+void TileManager::TransformColor(Result input, Result output, ColorTransform transform,
+                                 bool to_host) {
+	uint32_t element_bytes = 0;
+	switch (transform) {
+		case ColorTransform::SwapBgra16: element_bytes = 8; break;
+		case ColorTransform::Reverse10_11_11: element_bytes = 4; break;
+		default: EXIT("invalid color transform\n");
+	}
+	EXIT_IF(input.size == 0 || input.size % element_bytes != 0 ||
+	        input.size / element_bytes > UINT32_MAX ||
+	        output.size < input.size);
+	const auto pixels = static_cast<uint32_t>(input.size / element_bytes);
+	if (m_color_transform == nullptr) {
+		const auto module = CompileSPV(GPU_TILER_COLOR_TRANSFORM_SPV, m_graphics.device);
 		vk::PipelineShaderStageCreateInfo stage {};
 		stage.stage  = vk::ShaderStageFlagBits::eCompute;
 		stage.module = module;
@@ -593,14 +605,12 @@ void TileManager::SwapBgra16(Result input, Result output, uint32_t pixels) {
 		create.stage  = stage;
 		create.layout = m_pipeline_layout;
 		const auto result =
-		    m_graphics.device.createComputePipelines(nullptr, 1, &create, nullptr, &m_swap_bgra16);
+		    m_graphics.device.createComputePipelines(nullptr, 1, &create, nullptr, &m_color_transform);
 		m_graphics.device.destroyShaderModule(module, nullptr);
-		RequireVulkanSuccess(result, "create BGRA16 swap pipeline");
+		RequireVulkanSuccess(result, "create color transform pipeline");
 	}
-	const uint64_t bytes = static_cast<uint64_t>(pixels) * 8u;
-	EXIT_IF(pixels == 0);
-	const auto input_binding  = BindStorage(input, bytes);
-	const auto output_binding = BindStorage(output, bytes);
+	const auto input_binding  = BindStorage(input, input.size);
+	const auto output_binding = BindStorage(output, input.size);
 
 	const vk::DescriptorBufferInfo infos[] {
 	    input_binding.info,
@@ -631,13 +641,15 @@ void TileManager::SwapBgra16(Result input, Result output, uint32_t pixels) {
 	command.pipelineBarrier(
 	    vk::PipelineStageFlagBits::eAllCommands | vk::PipelineStageFlagBits::eHost,
 	    vk::PipelineStageFlagBits::eComputeShader, {}, 0, nullptr, 2, barriers, 0, nullptr);
-	command.bindPipeline(vk::PipelineBindPoint::eCompute, m_swap_bgra16);
+	command.bindPipeline(vk::PipelineBindPoint::eCompute, m_color_transform);
 	command.pushDescriptorSetKHR(vk::PipelineBindPoint::eCompute, m_pipeline_layout, 0,
 	                             static_cast<uint32_t>(writes.size()), writes.data());
 	Push push {};
-	push.src_base = input_binding.base;
-	push.dst_base = output_binding.base;
-	push.width    = pixels;
+	push.src_base        = input_binding.base;
+	push.dst_base        = output_binding.base;
+	push.width           = pixels;
+	push.color_transform = static_cast<uint32_t>(transform);
+	push.to_host         = to_host;
 	command.pushConstants(m_pipeline_layout, vk::ShaderStageFlagBits::eCompute, 0, sizeof(push),
 	                      &push);
 	command.dispatch((pixels + 63u) / 64u, 1, 1);
@@ -648,17 +660,11 @@ void TileManager::SwapBgra16(Result input, Result output, uint32_t pixels) {
 	                        0, nullptr);
 }
 
-TileManager::Result TileManager::SwapBgra16(Result input) {
-	EXIT_NOT_IMPLEMENTED(input.size == 0 || input.size % 8u != 0 || input.size / 8u > UINT32_MAX);
+TileManager::Result TileManager::TransformColor(Result input, ColorTransform transform,
+                                                bool to_host) {
 	auto result = GetScratchBuffer(input.size, input.buffer);
-	SwapBgra16(input, result, static_cast<uint32_t>(input.size / 8u));
+	TransformColor(input, result, transform, to_host);
 	return result;
 }
 
-void TileManager::SwapBgra16(Result input, Result output) {
-	EXIT_NOT_IMPLEMENTED(input.size == 0 || input.size % 8u != 0 || input.size / 8u > UINT32_MAX ||
-	                     output.size < input.size);
-	SwapBgra16(input, output, static_cast<uint32_t>(input.size / 8u));
-}
-
 } // namespace Libs::Graphics

src/graphics/host_gpu/renderer/image/tiler.h

@@ -3,6 +3,7 @@
 
 #include "common/common.h"
 #include "graphics/guest_gpu/tile.h"
+#include "graphics/host_gpu/renderer/image/imageInfo.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 
 #include <array>
@@ -42,7 +43,6 @@ struct GpuTileInfo {
 class TileManager final {
 public:
 	enum class D16Direction { Promote, Demote };
-	enum class ColorTransform { None, SwapBgra16 };
 
 	struct Result {
 		vk::Buffer buffer = nullptr;
@@ -65,7 +65,8 @@ public:
 
 	// Consume scratch results before the next acquisition, or pass their buffer as input.
 	[[nodiscard]] Result Detile(vk::Buffer tiled, uint64_t tiled_offset, uint64_t tiled_capacity,
-	                            uint64_t linear_capacity, std::span<const GpuTileInfo> infos);
+	                            uint64_t linear_capacity, std::span<const GpuTileInfo> infos,
+	                            ColorTransform transform = ColorTransform::None);
 	void Tile(vk::Buffer linear, uint64_t linear_offset, uint64_t linear_capacity, vk::Buffer tiled,
 	          uint64_t tiled_offset, uint64_t tiled_capacity, std::span<const GpuTileInfo> infos);
 	void TileImage(Image& image, std::span<const vk::BufferImageCopy> regions, vk::Buffer tiled,
@@ -75,8 +76,8 @@ public:
 	[[nodiscard]] Result GetScratchBuffer(uint64_t size, vk::Buffer input = nullptr);
 	void                 ConvertD16(Result source, Result target, D16Direction direction, bool d32,
 	                                const D16Layout& layout);
-	[[nodiscard]] Result SwapBgra16(Result input);
-	void                 SwapBgra16(Result input, Result output);
+	[[nodiscard]] Result TransformColor(Result input, ColorTransform transform, bool to_host);
+	void TransformColor(Result input, Result output, ColorTransform transform, bool to_host);
 
 private:
 	friend struct TileManagerTestAccess;
@@ -100,6 +101,8 @@ private:
 		uint32_t tail_x;
 		uint32_t tail_y;
 		uint32_t tail;
+		uint32_t color_transform;
+		uint32_t to_host;
 	};
 	struct Dispatch {
 		Push     push {};
@@ -118,12 +121,11 @@ private:
 	                                             uint32_t max_groups) noexcept;
 	void Prepare(bool tile, uint64_t tiled_capacity, uint64_t linear_capacity,
 	             std::span<const GpuTileInfo> infos, uint64_t source_base, uint64_t target_base,
-	             std::vector<Dispatch>& dispatches);
+	             std::vector<Dispatch>& dispatches, ColorTransform transform = ColorTransform::None);
 	void Record(vk::Buffer source, uint64_t source_offset, uint64_t source_capacity,
 	            vk::Buffer target, uint64_t target_offset, uint64_t target_capacity,
 	            std::span<Dispatch> dispatches, bool clear_target);
 	[[nodiscard]] vk::Pipeline GetPipeline(uint32_t slot);
-	void                       SwapBgra16(Result input, Result output, uint32_t pixels);
 
 	GraphicContext&                         m_graphics;
 	CommandScheduler&                       m_scheduler;
@@ -132,11 +134,11 @@ private:
 	vk::DescriptorSetLayout                 m_descriptor_layout = nullptr;
 	vk::PipelineLayout                      m_pipeline_layout   = nullptr;
 	std::array<vk::Pipeline, PipelineCount> m_pipelines {};
-	vk::Pipeline                            m_d16_to_d24  = nullptr;
-	vk::Pipeline                            m_d16_to_d32  = nullptr;
-	vk::Pipeline                            m_d24_to_d16  = nullptr;
-	vk::Pipeline                            m_d32_to_d16  = nullptr;
-	vk::Pipeline                            m_swap_bgra16 = nullptr;
+	vk::Pipeline                            m_d16_to_d24      = nullptr;
+	vk::Pipeline                            m_d16_to_d32      = nullptr;
+	vk::Pipeline                            m_d24_to_d16      = nullptr;
+	vk::Pipeline                            m_d32_to_d16      = nullptr;
+	vk::Pipeline                            m_color_transform = nullptr;
 };
 
 } // namespace Libs::Graphics

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

@@ -30,6 +30,18 @@ BlendMappingSupport ClassifyBlendMapping(const HW::BlendControl&
 	if (mapping.Map(3) == 3u) {
 		return BlendMappingSupport::Direct;
 	}
+	if (blend.separate_alpha_blend &&
+	    blend.color_srcblend == static_cast<uint8_t>(Prospero::BlendFactor::kSrcAlpha) &&
+	    blend.color_destblend == static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrcAlpha) &&
+	    blend.alpha_destblend == blend.color_destblend &&
+	    blend.color_comb_fcn == static_cast<uint8_t>(Prospero::BlendOp::kAdd) &&
+	    blend.alpha_comb_fcn == blend.color_comb_fcn) {
+		switch (static_cast<Prospero::BlendFactor>(blend.alpha_srcblend)) {
+			case Prospero::BlendFactor::kZero: return BlendMappingSupport::SourceAlphaZero;
+			case Prospero::BlendFactor::kOne: return BlendMappingSupport::SourceAlphaOne;
+			default: break;
+		}
+	}
 	// Moving alpha requires the same equation for all channels.
 	if (blend.separate_alpha_blend && (blend.alpha_srcblend != blend.color_srcblend ||
 	                                   blend.alpha_destblend != blend.color_destblend ||

src/graphics/host_gpu/renderer/pipeline/blendMapping.h

@@ -15,6 +15,8 @@ struct ColorComponentMapping;
 enum class BlendMappingSupport {
 	Direct,
 	SourceAlpha, // Requires logical alpha in the second blend source.
+	SourceAlphaOne, // Source alpha for RGB, one for alpha, with a shared destination equation.
+	SourceAlphaZero, // Source alpha for RGB, zero for alpha, with a shared destination equation.
 	Unsupported,
 };
 

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

@@ -68,7 +68,8 @@ vk::DescriptorType NativeDescriptorType(BindingKind kind) {
 		case BindingKind::BdaPagetable:
 		case BindingKind::FaultBuffer:
 		case BindingKind::FlattenedSrt:
-		case BindingKind::ShaderData: return vk::DescriptorType::eStorageBuffer;
+		case BindingKind::ShaderData:
+		case BindingKind::SharedMemory: return vk::DescriptorType::eStorageBuffer;
 		case BindingKind::Count: EXIT("invalid native descriptor binding kind");
 	}
 	EXIT("invalid native descriptor binding kind");
@@ -779,6 +780,7 @@ void RenderExecutor::PrepareBindings(const ShaderStageRuntime& runtime,
 	prepared.gds = {nullptr, 0, VK_WHOLE_SIZE};
 	prepared.flattened_srt = {};
 	prepared.shader_data_buffer = {};
+	prepared.shared_memory = {};
 	prepared.images.resize(program.info.images.size());
 	prepared.samplers.clear();
 	prepared.shader_data.clear();
@@ -876,6 +878,7 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 	auto&       images   = prepared.images;
 	EXIT_IF(images.size() != program.info.images.size());
 	auto& texture_cache = m_context.GetTextureCache();
+	// Acquire each view before a later overlapping descriptor can replace its image.
 	for (uint32_t i = 0; i < program.info.images.size(); i++) {
 		const auto old_image = texture_cache.m_slot_images.try_get(images[i].image_id);
 		if (old_image == nullptr || (!old_image->registered && !old_image->info.data.Empty()) ||
@@ -887,8 +890,6 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
 			BindImage(images[i].image_id,
 			          images[i].desc.type == TextureCache::BindingType::Storage);
 		}
-	}
-	for (uint32_t i = 0; i < program.info.images.size(); i++) {
 		auto& binding = images[i];
 		binding.mip_views.clear();
 		const auto& resource = program.info.images[i];
@@ -1121,12 +1122,15 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 					}
 					case BindingKind::FlattenedSrt:
 					case BindingKind::ShaderData:
+					case BindingKind::SharedMemory:
 					case BindingKind::Gds: {
 						const vk::DescriptorBufferInfo* view = &descriptors.gds;
 						if (binding.kind == BindingKind::FlattenedSrt) {
 							view = &descriptors.flattened_srt;
 						} else if (binding.kind == BindingKind::ShaderData) {
 							view = &descriptors.shader_data_buffer;
+						} else if (binding.kind == BindingKind::SharedMemory) {
+							view = &descriptors.shared_memory;
 						}
 						EXIT_IF(view->buffer == nullptr);
 						m_descriptor_buffers.push_back(*view);

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

@@ -44,6 +44,7 @@ struct PreparedBindings {
 	vk::DescriptorBufferInfo              gds {nullptr, 0, VK_WHOLE_SIZE};
 	vk::DescriptorBufferInfo              flattened_srt;
 	vk::DescriptorBufferInfo              shader_data_buffer;
+	vk::DescriptorBufferInfo              shared_memory;
 	std::vector<uint32_t>                 shader_data;
 };
 

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

@@ -39,6 +39,16 @@ namespace Libs::Graphics {
 
 namespace {
 
+uint8_t RemapSourceAlphaFactor(uint8_t factor) {
+	switch (static_cast<Prospero::BlendFactor>(factor)) {
+		case Prospero::BlendFactor::kSrcAlpha:
+			return static_cast<uint8_t>(Prospero::BlendFactor::kSrc1Alpha);
+		case Prospero::BlendFactor::kOneMinusSrcAlpha:
+			return static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrc1Alpha);
+		default: return factor;
+	}
+}
+
 vk::PolygonMode ResolvePolygonMode(const HW::ModeControl& mode, bool cull_front, bool cull_back) {
 	// CxPrimitiveSetup::PolygonMode disables both per-face modes when it is zero.
 	if (mode.poly_mode == 0) {
@@ -289,11 +299,14 @@ struct PipelineCache::ProgramCache {
 		lookup_key.code_size       = static_cast<uint32_t>(params.code.size());
 		BuildStageStaticKey(input_info, lookup_key.static_state);
 		auto                                         entry = programs.find(lookup_key);
-		const ShaderRecompiler::IR::SrtRuntime       runtime {
+		ShaderRecompiler::IR::SrtRuntime             runtime {
 		    .user_data                  = user_data,
 		    .shader_base                = params.Base(),
 		    .read_specialization_memory = ReadShaderGuestMemory,
 		};
+		if constexpr (std::is_same_v<InputInfo, ShaderComputeInputInfo>) {
+			runtime.workgroup_counts = input_info.workgroup_counts;
+		}
 		if (entry != programs.end()) {
 			EXIT_IF(!ShaderRecompiler::IR::MaterializeResources(
 			    entry->second.resource_plan, runtime, entry->second.resources,
@@ -619,14 +632,24 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 		           pixel_info.target_output_mode[0] != 0 && pixel_info.target_output_mode[0] != 7 &&
 		           std::all_of(std::begin(pixel_info.target_output_mode) + 1,
 		                       std::end(pixel_info.target_output_mode),
-		                       [](uint8_t mode) { return mode == 0; }) &&
-		           ClassifyBlendMapping(blend, pixel_info.target_export_mapping[0]) ==
-		               BlendMappingSupport::SourceAlpha) {
-			// Preserve logical alpha when the export mapping moves it.
-			pixel_info.alpha_blend_source_remap = true;
-			pixel_info.dual_source_blending     = true;
-			pixel_info.target_output_mode[1]    = pixel_info.target_output_mode[0];
-			pixel_info.target_export_mapping[1] = {};
+		                       [](uint8_t mode) { return mode == 0; })) {
+			switch (ClassifyBlendMapping(blend, pixel_info.target_export_mapping[0])) {
+				case BlendMappingSupport::SourceAlpha:
+					pixel_info.alpha_blend_source = ShaderAlphaBlendSource::SourceAlpha;
+					break;
+				case BlendMappingSupport::SourceAlphaOne:
+					pixel_info.alpha_blend_source = ShaderAlphaBlendSource::SourceAlphaOne;
+					break;
+				case BlendMappingSupport::SourceAlphaZero:
+					pixel_info.alpha_blend_source = ShaderAlphaBlendSource::SourceAlphaZero;
+					break;
+				default: break;
+			}
+			if (pixel_info.alpha_blend_source != ShaderAlphaBlendSource::None) {
+				pixel_info.dual_source_blending     = true;
+				pixel_info.target_output_mode[1]    = pixel_info.target_output_mode[0];
+				pixel_info.target_export_mapping[1] = {};
+			}
 		}
 	}
 	if (context.GetClipControl().clip_disable) {
@@ -660,19 +683,8 @@ ShaderProgram PipelineCache::GetComputeProgram(const HW::ComputeShaderInfo& regs
                                                ShaderComputeInputInfo&      input_info) {
 	input_info.host_subgroup_size = m_graphics.SupportsComputeWave64() ? 64u : 32u;
 	const auto        params      = PrepareProgram(regs, sh, input_info);
-	// Use one effective size for the cache key, LDS declaration, and access bounds.
-	const auto max_lds_dwords =
-	    m_graphics.GetPhysicalDeviceProperties().limits.maxComputeSharedMemorySize / 4u;
-	if (input_info.lds_size_dwords > max_lds_dwords) {
-		static std::atomic_bool warned = false;
-		if (!warned.exchange(true, std::memory_order_relaxed)) {
-			PipelineCacheLog("GPU warning: game compute shader requests {} bytes of LDS, but "
-			                 "the Vulkan device limit is {} bytes. Clamping LDS; rendering may "
-			                 "be incorrect.",
-			                 input_info.lds_size_dwords * 4u, max_lds_dwords * 4u);
-		}
-	}
-	input_info.lds_size_dwords = std::min(input_info.lds_size_dwords, max_lds_dwords);
+	input_info.lds_storage = input_info.lds_size_dwords * 4u >
+	    m_graphics.GetPhysicalDeviceProperties().limits.maxComputeSharedMemorySize;
 	uint32_t          push_data_cursor = 0;
 	return m_program_cache->Get(params, input_info, push_data_cursor);
 }
@@ -729,10 +741,12 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 		}
 		const auto& rt                        = ctx.GetRenderTarget(colors[i].target_slot);
 		const auto& bc                        = ctx.GetBlendControl(colors[i].target_slot);
-		const bool alpha_remap =
-		    slot == 0 && ps_input_info != nullptr && ps_input_info->alpha_blend_source_remap;
+		auto alpha_source = ShaderAlphaBlendSource::None;
+		if (slot == 0 && ps_input_info != nullptr) {
+			alpha_source = ps_input_info->alpha_blend_source;
+		}
 		static_params.blend_enable[slot] = bc.enable && !rt.info.blend_bypass;
-		if (static_params.blend_enable[slot] && !alpha_remap &&
+		if (static_params.blend_enable[slot] && alpha_source == ShaderAlphaBlendSource::None &&
 		    ClassifyBlendMapping(bc, colors[i].export_mapping) != BlendMappingSupport::Direct) {
 			static_params.blend_enable[slot] = false;
 			static std::atomic_bool warned = false;
@@ -744,18 +758,32 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 				    bc.alpha_srcblend, bc.alpha_destblend, bc.separate_alpha_blend ? 1 : 0));
 			}
 		}
-		if (alpha_remap) {
-			static_params.blend_alpha_source_remap = true;
-		}
 		if (static_params.blend_enable[slot]) {
-			static_params.color_srcblend[slot]       = bc.color_srcblend;
-			static_params.color_comb_fcn[slot]       = bc.color_comb_fcn;
-			static_params.color_destblend[slot]      = bc.color_destblend;
-			static_params.separate_alpha_blend[slot] = bc.separate_alpha_blend;
-			if (bc.separate_alpha_blend) {
-				static_params.alpha_srcblend[slot]  = bc.alpha_srcblend;
-				static_params.alpha_comb_fcn[slot]  = bc.alpha_comb_fcn;
-				static_params.alpha_destblend[slot] = bc.alpha_destblend;
+			auto blend = bc;
+			switch (alpha_source) {
+				case ShaderAlphaBlendSource::SourceAlpha:
+					blend.color_srcblend  = RemapSourceAlphaFactor(blend.color_srcblend);
+					blend.color_destblend = RemapSourceAlphaFactor(blend.color_destblend);
+					blend.separate_alpha_blend = false;
+					break;
+				case ShaderAlphaBlendSource::SourceAlphaOne:
+				case ShaderAlphaBlendSource::SourceAlphaZero:
+					// The second source carries the mapped source factor; its alpha stays logical Sa.
+					blend.color_srcblend = static_cast<uint8_t>(Prospero::BlendFactor::kSrc1Color);
+					blend.color_destblend =
+					    static_cast<uint8_t>(Prospero::BlendFactor::kOneMinusSrc1Alpha);
+					blend.separate_alpha_blend = false;
+					break;
+				case ShaderAlphaBlendSource::None: break;
+			}
+			static_params.color_srcblend[slot]       = blend.color_srcblend;
+			static_params.color_comb_fcn[slot]       = blend.color_comb_fcn;
+			static_params.color_destblend[slot]      = blend.color_destblend;
+			static_params.separate_alpha_blend[slot] = blend.separate_alpha_blend;
+			if (blend.separate_alpha_blend) {
+				static_params.alpha_srcblend[slot]  = blend.alpha_srcblend;
+				static_params.alpha_comb_fcn[slot]  = blend.alpha_comb_fcn;
+				static_params.alpha_destblend[slot] = blend.alpha_destblend;
 			}
 		}
 	}

src/graphics/host_gpu/renderer/pipeline/pipelineCache.h

@@ -55,7 +55,6 @@ struct PipelineStaticParameters {
 	uint8_t                    alpha_destblend[RENDER_COLOR_ATTACHMENTS_MAX]      = {};
 	bool                       separate_alpha_blend[RENDER_COLOR_ATTACHMENTS_MAX] = {};
 	bool                       blend_enable[RENDER_COLOR_ATTACHMENTS_MAX]         = {};
-	bool                       blend_alpha_source_remap                           = false;
 
 	bool operator==(const PipelineStaticParameters& other) const noexcept;
 };
@@ -65,7 +64,7 @@ struct PipelineStaticParameters {
 static_assert(std::is_trivially_copyable_v<PipelineStaticParameters>);
 static_assert(std::is_standard_layout_v<PipelineStaticParameters>);
 static_assert(alignof(PipelineStaticParameters) == 1);
-static_assert(sizeof(PipelineStaticParameters) == 126);
+static_assert(sizeof(PipelineStaticParameters) == 125);
 
 struct PipelineRenderingState {
 	std::array<vk::Format, RENDER_COLOR_ATTACHMENTS_MAX> color_formats {};

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

@@ -137,17 +137,14 @@ static void GetInputFormat(const ShaderBufferResource& res, vk::Format& format,
 	}
 }
 
-static vk::BlendFactor GetBlendFactor(uint32_t factor, bool remap_source_alpha) {
+static vk::BlendFactor GetBlendFactor(uint32_t factor) {
 	switch (static_cast<Prospero::BlendFactor>(factor)) {
 		case Prospero::BlendFactor::kZero: return vk::BlendFactor::eZero;
 		case Prospero::BlendFactor::kOne: return vk::BlendFactor::eOne;
 		case Prospero::BlendFactor::kSrcColor: return vk::BlendFactor::eSrcColor;
 		case Prospero::BlendFactor::kOneMinusSrcColor: return vk::BlendFactor::eOneMinusSrcColor;
-		case Prospero::BlendFactor::kSrcAlpha:
-			return remap_source_alpha ? vk::BlendFactor::eSrc1Color : vk::BlendFactor::eSrcAlpha;
-		case Prospero::BlendFactor::kOneMinusSrcAlpha:
-			return remap_source_alpha ? vk::BlendFactor::eOneMinusSrc1Color
-			                          : vk::BlendFactor::eOneMinusSrcAlpha;
+		case Prospero::BlendFactor::kSrcAlpha: return vk::BlendFactor::eSrcAlpha;
+		case Prospero::BlendFactor::kOneMinusSrcAlpha: return vk::BlendFactor::eOneMinusSrcAlpha;
 		case Prospero::BlendFactor::kDstAlpha: return vk::BlendFactor::eDstAlpha;
 		case Prospero::BlendFactor::kOneMinusDstAlpha: return vk::BlendFactor::eOneMinusDstAlpha;
 		case Prospero::BlendFactor::kDstColor: return vk::BlendFactor::eDstColor;
@@ -397,20 +394,18 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 		color_blend_attachment[i].colorWriteMask =
 		    vk::ColorComponentFlags {static_params.color_mask[i]};
 		color_blend_attachment[i].blendEnable = static_params.blend_enable[i] ? VK_TRUE : VK_FALSE;
-		// Only target 0 can use the second blend source carrying logical alpha.
-		const bool remap_source_alpha         = i == 0 && static_params.blend_alpha_source_remap;
 		color_blend_attachment[i].srcColorBlendFactor =
-		    GetBlendFactor(static_params.color_srcblend[i], remap_source_alpha);
+		    GetBlendFactor(static_params.color_srcblend[i]);
 		color_blend_attachment[i].dstColorBlendFactor =
-		    GetBlendFactor(static_params.color_destblend[i], remap_source_alpha);
+		    GetBlendFactor(static_params.color_destblend[i]);
 		color_blend_attachment[i].colorBlendOp = GetBlendOp(static_params.color_comb_fcn[i]);
 		color_blend_attachment[i].srcAlphaBlendFactor =
 		    (static_params.separate_alpha_blend[i]
-		         ? GetBlendFactor(static_params.alpha_srcblend[i], remap_source_alpha)
+		         ? GetBlendFactor(static_params.alpha_srcblend[i])
 		         : color_blend_attachment[i].srcColorBlendFactor);
 		color_blend_attachment[i].dstAlphaBlendFactor =
 		    (static_params.separate_alpha_blend[i]
-		         ? GetBlendFactor(static_params.alpha_destblend[i], remap_source_alpha)
+		         ? GetBlendFactor(static_params.alpha_destblend[i])
 		         : color_blend_attachment[i].dstColorBlendFactor);
 		color_blend_attachment[i].alphaBlendOp =
 		    (static_params.separate_alpha_blend[i] ? GetBlendOp(static_params.alpha_comb_fcn[i])

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -20,6 +20,7 @@
 #include "graphics/host_gpu/vulkanCommon.h"
 #include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/recompiler/ir/passes/ResourceMaterialization.h"
+#include "graphics/shader/recompiler/ir/passes/BindingLayout.h"
 #include "graphics/shader/shader.h"
 #include "kernel/eventQueue.h"
 #include "kernel/pthread.h"
@@ -195,6 +196,35 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
 	return true;
 }
 
+static void BindSharedMemory(RenderContext& context, ShaderComputeInputInfo& input,
+                             PreparedBindings& bindings, uint64_t indirect_args = 0) {
+	if (ShaderRecompiler::IR::FindBinding(input.stage.program->bindings,
+	        ShaderRecompiler::IR::DescriptorBindingKind::SharedMemory) == nullptr) {
+		return;
+	}
+	auto& cache = context.GetBufferCache();
+	if (indirect_args != 0) {
+		cache.ReadMemory(indirect_args, sizeof(vk::DispatchIndirectCommand));
+		std::memcpy(input.workgroup_counts, reinterpret_cast<const void*>(indirect_args),
+		            sizeof(input.workgroup_counts));
+	}
+	// LDS has no contents to preserve between dispatches. The existing shader hazard
+	// barriers also order other users of this GPU-only utility buffer.
+	auto& storage = cache.GetUtilityBuffer(MemoryUsage::DeviceLocal);
+	const auto limit = std::min<uint64_t>(storage.Size(),
+	    context.GetGraphics().GetPhysicalDeviceProperties().limits.maxStorageBufferRange);
+	uint64_t size = sizeof(uint32_t);
+	if (std::ranges::find(input.workgroup_counts, 0u) == std::end(input.workgroup_counts)) {
+		size = uint64_t {input.lds_size_dwords} * sizeof(uint32_t);
+		EXIT_IF(size == 0 || size > limit);
+		for (const auto count: input.workgroup_counts) {
+			EXIT_IF(size > limit / count);
+			size *= count;
+		}
+	}
+	bindings.shared_memory = {storage.Handle(), 0, size};
+}
+
 void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
                                     uint32_t thread_group_x, uint32_t thread_group_y,
                                     uint32_t thread_group_z, uint32_t mode) {
@@ -254,6 +284,18 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	ShaderComputeInputInfo input_info {};
 	const bool use_thread_dimensions = (mode & DISPATCH_INITIATOR_USE_THREAD_DIMENSIONS) != 0;
 	input_info.dispatch_thread_dimensions = use_thread_dimensions;
+	input_info.workgroup_counts[0] = thread_group_x;
+	input_info.workgroup_counts[1] = thread_group_y;
+	input_info.workgroup_counts[2] = thread_group_z;
+	if (use_thread_dimensions) {
+		const uint32_t group_sizes[] = {cs_regs.cs_regs.num_thread_x, cs_regs.cs_regs.num_thread_y,
+		                                cs_regs.cs_regs.num_thread_z};
+		for (uint32_t axis = 0; axis < 3u; ++axis) {
+			const auto size = std::max(group_sizes[axis], 1u);
+			const auto threads = input_info.workgroup_counts[axis];
+			input_info.workgroup_counts[axis] = threads / size + (threads % size != 0u);
+		}
+	}
 	const auto compute_program =
 	    m_context.GetPipelineCache().GetComputeProgram(cs_regs, sh_regs, input_info);
 	if (use_thread_dimensions) {
@@ -273,18 +315,12 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}
 
 	if (use_thread_dimensions) {
-		auto groups_from_threads = [](uint32_t threads, uint32_t group_size) {
-			return (threads == 0
-			            ? 0u
-			            : (threads + std::max(group_size, 1u) - 1u) / std::max(group_size, 1u));
-		};
-
 		const uint32_t old_x = thread_group_x;
 		const uint32_t old_y = thread_group_y;
 		const uint32_t old_z = thread_group_z;
-		thread_group_x       = groups_from_threads(thread_group_x, cs_regs.cs_regs.num_thread_x);
-		thread_group_y       = groups_from_threads(thread_group_y, cs_regs.cs_regs.num_thread_y);
-		thread_group_z       = groups_from_threads(thread_group_z, cs_regs.cs_regs.num_thread_z);
+		thread_group_x       = input_info.workgroup_counts[0];
+		thread_group_y       = input_info.workgroup_counts[1];
+		thread_group_z       = input_info.workgroup_counts[2];
 
 		static std::atomic<uint32_t> log_count {0};
 		if (log_count.fetch_add(1, std::memory_order_relaxed) < 32) {
@@ -311,13 +347,15 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		m_context.PrepareBda();
 	}
 	RebindImages(bindings);
+	BindSharedMemory(m_context, input_info, bindings);
 	RebindBuffers(bindings);
 
 	auto              vk_buffer        = buffer.Handle();
 	PreparedBindings* descriptor_stage = &bindings;
 	CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
 	               std::span {&descriptor_stage, 1u});
-	bool has_storage_writes = HasShaderBufferWrites(input_info.stage);
+	bool has_storage_writes = bindings.shared_memory.buffer != nullptr ||
+	    HasShaderBufferWrites(input_info.stage);
 	has_storage_writes =
 	    std::any_of(program.info.images.begin(), program.info.images.end(),
 	                [](const auto& image) {
@@ -424,6 +462,7 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	if (program.info.uses_dma) {
 		m_context.PrepareBda();
 	}
+	BindSharedMemory(m_context, input_info, bindings, args_addr);
 	RebindImages(bindings);
 	// Acquiring arguments can merge cache buffers; finalize shader bindings afterward.
 	const auto [args_buffer, args_offset] = m_context.GetBufferCache().ObtainBuffer(
@@ -434,7 +473,8 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	CommitBindings(buffer, vk::PipelineBindPoint::eCompute, pipeline,
 	               std::span {&descriptor_stage, 1u});
 	const auto vk_buffer = buffer.Handle();
-	const bool has_storage_writes = HasShaderBufferWrites(input_info.stage) ||
+	const bool has_storage_writes = bindings.shared_memory.buffer != nullptr ||
+	    HasShaderBufferWrites(input_info.stage) ||
 	    std::any_of(program.info.images.begin(), program.info.images.end(), [](const auto& image) {
 		    return image.written && image.resource_class ==
 		                                ShaderRecompiler::IR::ImageResourceClass::Storage;

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -784,6 +784,7 @@ static bool GetDrawTopology(const HW::UserConfig& ucfg, vk::PrimitiveTopology& t
 			topology = vk::PrimitiveTopology::eTriangleList;
 			break;
 		case Prospero::PrimitiveType::kTriFan:
+		case Prospero::PrimitiveType::kPolygon:
 			topology = vk::PrimitiveTopology::eTriangleFan;
 			break;
 		case Prospero::PrimitiveType::kTriStrip:
@@ -994,6 +995,7 @@ static void EmitDrawPrimitives(const HW::UserConfig& ucfg, vk::CommandBuffer vk_
 		case Prospero::PrimitiveType::kTriFan:
 		case Prospero::PrimitiveType::kTriStrip:
 		case Prospero::PrimitiveType::kRectList:
+		case Prospero::PrimitiveType::kPolygon:
 		case Prospero::PrimitiveType::kRectListLegacy:
 		case Prospero::PrimitiveType::kPatch:
 			if (draw.IsIndexed()) {

src/graphics/host_gpu/shaders/gpu_tiler_color.inc

@@ -0,0 +1,20 @@
+// ColorTransform values match imageInfo.h. Preserve the floating-point bit fields exactly.
+uvec2 transform_color(uvec2 value, uint transform, bool to_host) {
+	switch (transform) {
+		case 1u:
+			return uvec2((value.y & 0xffffu) | (value.x & 0xffff0000u),
+			             (value.x & 0xffffu) | (value.y & 0xffff0000u));
+		case 2u: {
+			uint packed = value.x;
+			if (to_host) {
+				value.x = (packed >> 21u) | (((packed >> 10u) & 0x7ffu) << 11u) |
+				          ((packed & 0x3ffu) << 22u);
+			} else {
+				value.x = (packed >> 22u) | (((packed >> 11u) & 0x7ffu) << 10u) |
+				          ((packed & 0x7ffu) << 21u);
+			}
+			break;
+		}
+	}
+	return value;
+}

src/graphics/host_gpu/shaders/gpu_tiler_color_transform.comp

@@ -1,4 +1,5 @@
 #version 450
+#extension GL_GOOGLE_include_directive : require
 
 layout(local_size_x = 64) in;
 
@@ -19,17 +20,23 @@ layout(push_constant, std430) uniform Push {
 	uint tail_x;
 	uint tail_y;
 	uint tail;
+	uint color_transform;
+	uint to_host;
 } params;
 
+#include "gpu_tiler_color.inc"
+
 void main() {
 	uint pixel = gl_GlobalInvocationID.x;
 	if (pixel >= params.width) {
 		return;
 	}
-	uint src = (params.src_base >> 2u) + pixel * 2u;
-	uint dst = (params.dst_base >> 2u) + pixel * 2u;
-	uint bg = input_buffer.data[src];
-	uint ra = input_buffer.data[src + 1u];
-	output_buffer.data[dst] = (ra & 0x0000ffffu) | (bg & 0xffff0000u);
-	output_buffer.data[dst + 1u] = (bg & 0x0000ffffu) | (ra & 0xffff0000u);
+	uint words = params.color_transform == 1u ? 2u : 1u;
+	uint src = (params.src_base >> 2u) + pixel * words;
+	uint dst = (params.dst_base >> 2u) + pixel * words;
+	uvec2 value = uvec2(input_buffer.data[src], 0u);
+	if (words == 2u) value.y = input_buffer.data[src + 1u];
+	value = transform_color(value, params.color_transform, params.to_host != 0u);
+	output_buffer.data[dst] = value.x;
+	if (words == 2u) output_buffer.data[dst + 1u] = value.y;
 }

src/graphics/host_gpu/shaders/gpu_tiler_common.inc

@@ -17,10 +17,19 @@ layout(set = 0, binding = 2, std140) uniform Push {
 	uint tail_x;
 	uint tail_y;
 	uint tail;
+	uint color_transform;
 } params;
 
+#include "gpu_tiler_color.inc"
+
 void copy_element(uint src, uint dst) {
-	if (ELEMENT_BYTES >= 4) {
+	if ((ELEMENT_BYTES == 4u || ELEMENT_BYTES == 8u) && params.color_transform != 0u) {
+		uvec2 value = uvec2(input_buffer.data[src >> 2], 0u);
+		if (ELEMENT_BYTES == 8u) value.y = input_buffer.data[(src >> 2) + 1u];
+		value = transform_color(value, params.color_transform, TILE == 0u);
+		output_buffer.data[dst >> 2] = value.x;
+		if (ELEMENT_BYTES == 8u) output_buffer.data[(dst >> 2) + 1u] = value.y;
+	} else if (ELEMENT_BYTES >= 4) {
 		for (uint i = 0; i < ELEMENT_BYTES; i += 4) {
 			output_buffer.data[(dst + i) >> 2] = input_buffer.data[(src + i) >> 2];
 		}

src/graphics/host_gpu/vulkanCommon.cpp

@@ -48,6 +48,7 @@ constexpr FormatMapping kFormatMappings[] = {
     {Prospero::BufferFormat::k11_11_10Float, vk::Format::eB10G11R11UfloatPack32},
     {Prospero::BufferFormat::k10_10_10_2UNorm, vk::Format::eA2B10G10R10UnormPack32},
     {Prospero::BufferFormat::k10_10_10_2SNorm, vk::Format::eA2B10G10R10SnormPack32},
+    {Prospero::BufferFormat::k10_10_10_2UScaled, vk::Format::eA2B10G10R10UscaledPack32},
     {Prospero::BufferFormat::k10_10_10_2UInt, vk::Format::eA2B10G10R10UintPack32},
     {Prospero::BufferFormat::k8_8_8_8UNorm, vk::Format::eR8G8B8A8Unorm},
     {Prospero::BufferFormat::k8_8_8_8SNorm, vk::Format::eR8G8B8A8Snorm},

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -670,7 +670,9 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
 	IR::EliminateDeadCode(ir.blocks);
 
 	IR::CollectShaderInfo(ir, options.input_info);
-	IR::AllocateBindings(ir, push_data_start_dword);
+	IR::AllocateBindings(ir, push_data_start_dword,
+	                     ir.stage == ShaderType::Compute && options.input_info.compute != nullptr &&
+	                         options.input_info.compute->lds_storage);
 	std::string ir_dump;
 	if (options.dump_ir) {
 		ir_dump = MakeIrDump(translated.cfg_dump, ir);

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

@@ -18,7 +18,7 @@ namespace {
 	std::abort();
 }
 
-void ValidateNativeProgram(const IR::Program& program) {
+void ValidateNativeProgram(const IR::Program& program, bool lds_storage) {
 	using Kind                                             = IR::DescriptorBindingKind;
 	constexpr auto                               KindCount = static_cast<size_t>(Kind::Count);
 	std::array<std::vector<uint32_t>, KindCount> expected;
@@ -53,11 +53,14 @@ void ValidateNativeProgram(const IR::Program& program) {
 		Expect(Kind::Samplers, Dense(program.info.samplers.size()));
 	}
 	auto& buffers = expected[static_cast<size_t>(Kind::Buffers)];
-	const bool uses_gds = IR::CollectMemoryResources(program, buffers);
+	const auto shared = IR::CollectMemoryResources(program, buffers);
 	present[static_cast<size_t>(Kind::Buffers)] = !buffers.empty();
-	if (uses_gds) {
+	if (shared.gds) {
 		Expect(Kind::Gds);
 	}
+	if (shared.lds && lds_storage) {
+		Expect(Kind::SharedMemory);
+	}
 	if (program.info.uses_dma) {
 		Expect(Kind::BdaPagetable);
 		Expect(Kind::FaultBuffer);
@@ -344,7 +347,8 @@ std::vector<uint32_t> EmitProgram(const IR::Program& program,
 	    !program.shader_info_complete || !program.binding_layout_complete) {
 		Fail(program, "SPIR-V emitter requires a fully planned native shader program");
 	}
-	ValidateNativeProgram(program);
+	ValidateNativeProgram(program, program.stage == ShaderType::Compute &&
+	                                   input_info.compute != nullptr && input_info.compute->lds_storage);
 	IR::ValidateProgram(program, true);
 	EmitterState state(program, input_info);
 	const auto* workgroup = ShaderWorkgroupInput(program.stage, input_info);

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

@@ -30,6 +30,28 @@ uint32_t CompareEqual64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_va
 	return Unary(state, not_equal ? spv::OpAny : spv::OpAll, TypeBool(state), compare);
 }
 
+uint32_t EmitMinMaxF64(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
+	const auto bits_type = TypeU64(state);
+	const auto lhs_bits = Unary(state, spv::OpBitcast, bits_type, lhs);
+	const auto rhs_bits = Unary(state, spv::OpBitcast, bits_type, rhs);
+	const auto ordered = Binary(state, max_value ? spv::OpFOrdGreaterThan : spv::OpFOrdLessThan,
+	                            TypeBool(state), lhs, rhs);
+	auto result = Select(state, TypeF64(state), ordered, lhs, rhs);
+
+	// Equal values have identical bits except signed zero: min chooses -0, max chooses +0.
+	const auto equal = Binary(state, spv::OpFOrdEqual, TypeBool(state), lhs, rhs);
+	const auto equal_bits = Binary(state, max_value ? spv::OpBitwiseAnd : spv::OpBitwiseOr,
+	                               bits_type, lhs_bits, rhs_bits);
+	result = Select(state, TypeF64(state), equal,
+	                Unary(state, spv::OpBitcast, TypeF64(state), equal_bits), result);
+
+	// Non-IEEE mode selects the other operand for NaN, including rhs when both are NaN.
+	result = Select(state, TypeF64(state), Unary(state, spv::OpIsNan, TypeBool(state), rhs),
+	                lhs, result);
+	return Select(state, TypeF64(state), Unary(state, spv::OpIsNan, TypeBool(state), lhs),
+	              rhs, result);
+}
+
 uint32_t CompareOrdered64(EmitterState& state, uint32_t lhs_value, uint32_t rhs_value,
                           spv::Op high_compare, spv::Op low_compare) {
 	const auto lhs         = ExtractPair(state, lhs_value);
@@ -201,6 +223,25 @@ uint32_t EmitF32ToU32(EmitterState& state, uint32_t src, bool signed_value) {
 }
 
 } // namespace
+
+uint32_t EmitFPFma32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
+	const auto result = EmitGlsl<GLSLstd450Fma, IR::Type::F32>(state, a, b, c);
+	state.builder.AddAnnotation(spv::OpDecorate, result, spv::DecorationNoContraction);
+	return result;
+}
+
+uint32_t EmitFPMad32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
+	// Legacy MAD/MAC round and flush between multiply and add, irrespective of SP_DENORM.
+	a = EmitFlushF32DenormToSignedZero(state, a);
+	b = EmitFlushF32DenormToSignedZero(state, b);
+	c = EmitFlushF32DenormToSignedZero(state, c);
+	const auto product = EmitFPMul32(state, a, b);
+	state.builder.AddAnnotation(spv::OpDecorate, product, spv::DecorationNoContraction);
+	const auto sum = EmitFPAdd32(state, EmitFlushF32DenormToSignedZero(state, product), c);
+	state.builder.AddAnnotation(spv::OpDecorate, sum, spv::DecorationNoContraction);
+	return EmitFlushF32DenormToSignedZero(state, sum);
+}
+
 uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
 	const auto min_ab   = EmitMinMaxF32Value(state, a, b, false);
 	const auto min3     = EmitMinMaxF32Value(state, min_ab, c, false);
@@ -468,6 +509,14 @@ uint32_t EmitFPMax32(EmitterState& state, uint32_t arg0, uint32_t arg1) {
 	return EmitMinMaxF32Value(state, arg0, arg1, true);
 }
 
+uint32_t EmitFPMin64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
+	return EmitMinMaxF64(state, arg0, arg1, false);
+}
+
+uint32_t EmitFPMax64(EmitterState& state, uint32_t arg0, uint32_t arg1) {
+	return EmitMinMaxF64(state, arg0, arg1, true);
+}
+
 uint32_t EmitFPMinTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2) {
 	return EmitFMinMax3(state, arg0, arg1, arg2, false);
 }

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

@@ -459,7 +459,7 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 	}
 	// Skip dormant color exports after their valid mask; MRT1 is reserved for logical alpha.
 	if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
-	    exp.index != 0 && state.input_info.pixel->alpha_blend_source_remap) {
+	    exp.index != 0 && state.input_info.pixel->alpha_blend_source != ShaderAlphaBlendSource::None) {
 		return;
 	}
 	EmitIfCondition(state, exec, [&]() {
@@ -503,14 +503,36 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const auto vector_type = uint_output ? TypeU32Vector(state, 4) : TypeF32Vector(state, 4);
 		auto       value       = ExportVector(ctx, data, exp, uint_output);
 		if (state.program.stage == ShaderType::Pixel && exp.kind == IR::ExportTargetKind::Mrt &&
-		    exp.index == 0 && !uint_output && state.input_info.pixel->alpha_blend_source_remap) {
+		    exp.index == 0 && !uint_output &&
+		    state.input_info.pixel->alpha_blend_source != ShaderAlphaBlendSource::None) {
 			// Broadcast logical alpha before swizzling the primary output.
 			const auto blend_output =
 			    OutputVariableForExport(state, {.kind = IR::ExportTargetKind::Mrt, .index = 1});
 			if (blend_output != 0) {
-				const auto alpha = state.builder.AllocateId();
+				auto alpha = state.builder.AllocateId();
 				state.builder.AddFunction(spv::OpVectorShuffle, vector_type, alpha, value, value,
 				                          3u, 3u, 3u, 3u);
+				uint32_t alpha_factor = 0;
+				switch (state.input_info.pixel->alpha_blend_source) {
+					case ShaderAlphaBlendSource::SourceAlphaOne:
+						alpha_factor = ConstantF32Value(state, 1.0f);
+						break;
+					case ShaderAlphaBlendSource::SourceAlphaZero:
+						alpha_factor = ConstantF32Value(state, 0.0f);
+						break;
+					default: break;
+				}
+				if (alpha_factor != 0) {
+					const auto mapping = state.input_info.pixel->target_export_mapping[0];
+					for (uint32_t component = 0; component < 4; component++) {
+						if (mapping.Map(component) != 3u) continue;
+						const auto factors = state.builder.AllocateId();
+						state.builder.AddFunction(spv::OpCompositeInsert, vector_type, factors,
+						                          alpha_factor, alpha, component);
+						alpha = factors;
+						break;
+					}
+				}
 				state.builder.AddFunction(spv::OpStore, blend_output, alpha);
 			}
 		}
@@ -594,9 +616,11 @@ void EmitBarrier(EmitterState& state) {
 		return;
 	}
 	const auto memory_scope = tessellation ? spv::ScopeInvocation : spv::ScopeWorkgroup;
-	const auto semantics    = tessellation ? spv::MemorySemanticsMaskNone
-	                                       : spv::MemorySemanticsAcquireReleaseMask |
-	                                             spv::MemorySemanticsWorkgroupMemoryMask;
+	const auto memory = state.lds_storage_class == spv::StorageClassStorageBuffer
+	                        ? spv::MemorySemanticsUniformMemoryMask
+	                        : spv::MemorySemanticsWorkgroupMemoryMask;
+	const auto semantics = tessellation ? spv::MemorySemanticsMaskNone
+	                                    : spv::MemorySemanticsAcquireReleaseMask | memory;
 	state.builder.AddFunction(spv::OpControlBarrier, ConstantU32(state, spv::ScopeWorkgroup),
 	                          ConstantU32(state, memory_scope), ConstantU32(state, semantics));
 }

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

@@ -391,28 +391,50 @@ Format::BufferFormatInfo ImageConversionFormat(const EmitterState&   state,
 	const auto format = state.program.info.images[mem.resource].conversion_format;
 	if (format == Prospero::BufferFormat::kInvalid) return {};
 	const auto info = Format::GetFormatInfo(format);
-	EXIT_IF(Prospero::SampledTextureNumericClass(format) != Prospero::TextureNumericClass::Uint ||
-	        Prospero::RemapTextureFormat(format) == format ||
-	        info.type != Format::ComponentType::Uint || !info.packed_bitfield ||
-	        info.byte_size != sizeof(uint32_t) || info.component_count == 0u ||
+	EXIT_IF(Prospero::RemapTextureFormat(format) == format || info.component_count == 0u ||
 	        info.component_count > 4u);
+	EXIT_IF(info.type != Format::ComponentType::Uscaled &&
+	        (info.type != Format::ComponentType::Uint || !info.packed_bitfield ||
+	         info.byte_size != sizeof(uint32_t)));
 	return info;
 }
 
+uint32_t ImageGatherSource(const EmitterState& state, const IR::MemoryInfo& mem) {
+	const auto component = ImageGatherComponent(mem.dmask);
+	const auto info      = ImageConversionFormat(state, mem);
+	if (info.format == Prospero::BufferFormat::kInvalid) return component;
+	if (info.packed_bitfield) return 0u;
+	const auto selector =
+	    (state.program.info.images[mem.resource].shader_swizzle >> (component * 3u)) & 7u;
+	return selector >= 4u ? (selector - 4u) % info.component_count : 0u;
+}
+
 uint32_t UnpackImageTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t texel) {
 	const auto info = ImageConversionFormat(ctx.state, mem);
 	if (info.format == Prospero::BufferFormat::kInvalid) return texel;
 
-	const auto packed = ctx.state.builder.AllocateId();
-	ctx.state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(ctx.state), packed, texel, 0u);
-	uint32_t components[4] = {ConstantU32(ctx.state, 0), ConstantU32(ctx.state, 0),
-	                          ConstantU32(ctx.state, 0), ConstantU32(ctx.state, 0)};
+	const auto numeric_class = ctx.state.program.info.images[mem.resource].numeric_class;
+	const auto scalar_type   = ImageScalarType(ctx.state, numeric_class);
+	uint32_t   packed        = 0;
+	if (info.packed_bitfield) {
+		packed = ctx.state.builder.AllocateId();
+		ctx.state.builder.AddFunction(spv::OpCompositeExtract, scalar_type, packed, texel, 0u);
+	}
+	uint32_t components[4] {};
 	for (uint32_t component = 0; component < info.component_count; component++) {
 		components[component] = ctx.state.builder.AllocateId();
-		ctx.state.builder.AddFunction(spv::OpBitFieldUExtract, TypeU32(ctx.state),
-		                              components[component], packed,
-		                              ConstantU32(ctx.state, info.component_bit_offset[component]),
-		                              ConstantU32(ctx.state, info.component_bits[component]));
+		if (info.packed_bitfield) {
+			ctx.state.builder.AddFunction(spv::OpBitFieldUExtract, scalar_type,
+			                              components[component], packed,
+			                              ConstantU32(ctx.state, info.component_bit_offset[component]),
+			                              ConstantU32(ctx.state, info.component_bits[component]));
+		} else {
+			ctx.state.builder.AddFunction(spv::OpCompositeExtract, scalar_type,
+			                              components[component], texel, component);
+			// UNorm backing preserves the guest's filtering; scaling precedes swizzle constants.
+			components[component] = Binary(ctx.state, spv::OpFMul, scalar_type, components[component],
+			                               ConstantF32Value(ctx.state, 255.0f));
+		}
 	}
 	for (uint32_t component = info.component_count; component < 4u; component++) {
 		components[component] = components[component % info.component_count];
@@ -423,16 +445,19 @@ uint32_t UnpackImageTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint
 	for (uint32_t component = 0; component < 4u; component++) {
 		const auto selector = (swizzle >> (component * 3u)) & 7u;
 		if (selector == 1u) {
-			selected[component] = ConstantU32(ctx.state, 1u);
+			selected[component] = info.packed_bitfield ? ConstantU32(ctx.state, 1u)
+			                                          : ConstantF32Value(ctx.state, 1.0f);
 		} else if (selector >= 4u) {
 			selected[component] = components[selector - 4u];
 		} else {
-			selected[component] = ConstantU32(ctx.state, 0u);
+			selected[component] = info.packed_bitfield ? ConstantU32(ctx.state, 0u)
+			                                          : ZeroF32(ctx.state);
 		}
 	}
 	const auto result = ctx.state.builder.AllocateId();
-	ctx.state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(ctx.state, 4), result,
-	                              selected[0], selected[1], selected[2], selected[3]);
+	ctx.state.builder.AddFunction(spv::OpCompositeConstruct,
+	                              ImageVectorType(ctx.state, numeric_class, 4), result, selected[0],
+	                              selected[1], selected[2], selected[3]);
 	return result;
 }
 
@@ -443,10 +468,24 @@ uint32_t UnpackImageGather(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uin
 	const auto component = ImageGatherComponent(mem.dmask);
 	const auto selector =
 	    (ctx.state.program.info.images[mem.resource].shader_swizzle >> (component * 3u)) & 7u;
+	const auto numeric_class = ctx.state.program.info.images[mem.resource].numeric_class;
+	const auto vector_type   = ImageVectorType(ctx.state, numeric_class, 4);
 	if (selector < 4u) {
-		const auto value = ConstantU32(ctx.state, selector == 1u ? 1u : 0u);
-		return ctx.state.builder.Constant(spv::OpConstantComposite, TypeU32Vector(ctx.state, 4),
-		                                  value, value, value, value);
+		uint32_t value;
+		switch (info.type) {
+			case Format::ComponentType::Uscaled:
+				value = ConstantF32Value(ctx.state, selector == 1u ? 1.0f : 0.0f);
+				break;
+			default: value = ConstantU32(ctx.state, selector == 1u ? 1u : 0u); break;
+		}
+		return ctx.state.builder.Constant(spv::OpConstantComposite, vector_type, value, value,
+		                                  value, value);
+	}
+	if (!info.packed_bitfield) {
+		const auto scale  = ConstantF32Value(ctx.state, 255.0f);
+		const auto scales = ctx.state.builder.Constant(spv::OpConstantComposite, vector_type,
+		                                                scale, scale, scale, scale);
+		return Binary(ctx.state, spv::OpFMul, vector_type, gathered, scales);
 	}
 
 	uint32_t values[4] {};
@@ -488,10 +527,7 @@ uint32_t EmitOneDimensionalGatherLz(ValueEmitContext& ctx, const IR::MemoryInfo&
 	                                     LoadSamplerDescriptor(state, mem.sampler));
 	const auto vector_type = ImageVectorType(state, numeric_class, 4);
 	const auto scalar_type = ImageScalarType(state, numeric_class);
-	const auto component =
-	    ImageConversionFormat(state, mem).format == Prospero::BufferFormat::kInvalid
-	        ? ImageGatherComponent(mem.dmask)
-	        : 0u;
+	const auto component = ImageGatherSource(state, mem);
 	uint32_t values[2] {};
 	for (uint32_t index = 0; index < 2u; index++) {
 		const auto sample_coord =
@@ -515,6 +551,7 @@ uint32_t EmitOneDimensionalGatherLz(ValueEmitContext& ctx, const IR::MemoryInfo&
 uint32_t PackImageTexel(ValueEmitContext& ctx, const IR::MemoryInfo& mem, uint32_t texel) {
 	const auto info = ImageConversionFormat(ctx.state, mem);
 	if (info.format == Prospero::BufferFormat::kInvalid) return texel;
+	EXIT_IF(info.type != Format::ComponentType::Uint || !info.packed_bitfield);
 
 	auto packed = ConstantU32(ctx.state, 0u);
 	for (uint32_t component = 0; component < info.component_count; component++) {
@@ -669,7 +706,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const bool  dref           = HasFlag(mem, Decoder::ImageSampleFlagCompare);
 		if (dref && state.program.info.images[mem.resource].conversion_format !=
 		                Prospero::BufferFormat::kInvalid) {
-			ctx.Fail(inst, "uses depth comparison with a packed integer image");
+			ctx.Fail(inst, "uses depth comparison with a converted image");
 			return;
 		}
 		if (op == IR::ValueOpcode::ImageGatherRaw) {
@@ -700,9 +737,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				                        ? AddressF32(ctx, mem, *address, layout.dref)
 				                        : ZeroF32(state);
 			} else {
-				component_or_dref = ConstantU32(
-				    state, ImageConversionFormat(state, mem).format == Prospero::BufferFormat::kInvalid
-				               ? ImageGatherComponent(mem.dmask) : 0u);
+				component_or_dref = ConstantU32(state, ImageGatherSource(state, mem));
 			}
 			uint32_t operand_mask = 0;
 			uint32_t offset = 0;

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

@@ -160,10 +160,12 @@ EMIT_NATIVE(FPMul64, OpFMul, F64, uint32_t, uint32_t)
 inline constexpr auto EmitFPFma64 =
     EmitGlsl<GLSLstd450Fma, IR::Type::F64, uint32_t, uint32_t, uint32_t>;
 uint32_t              EmitFPRecip64(EmitterState& state, uint32_t arg0);
-inline constexpr auto EmitFPFma32 =
-    EmitGlsl<GLSLstd450Fma, IR::Type::F32, uint32_t, uint32_t, uint32_t>;
+uint32_t EmitFPFma32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
+uint32_t EmitFPMad32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
 uint32_t EmitFPMin32(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitFPMax32(EmitterState& state, uint32_t arg0, uint32_t arg1);
+uint32_t EmitFPMin64(EmitterState& state, uint32_t arg0, uint32_t arg1);
+uint32_t EmitFPMax64(EmitterState& state, uint32_t arg0, uint32_t arg1);
 uint32_t EmitFPMinTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
 uint32_t EmitFPMaxTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
 uint32_t EmitFPMedTri32(EmitterState& state, uint32_t arg0, uint32_t arg1, uint32_t arg2);
@@ -178,6 +180,9 @@ inline constexpr auto EmitFPRoundEven32 = EmitGlsl<GLSLstd450RoundEven, IR::Type
 inline constexpr auto EmitFPFloor32     = EmitGlsl<GLSLstd450Floor, IR::Type::F32, uint32_t>;
 inline constexpr auto EmitFPCeil32      = EmitGlsl<GLSLstd450Ceil, IR::Type::F32, uint32_t>;
 inline constexpr auto EmitFPTrunc32     = EmitGlsl<GLSLstd450Trunc, IR::Type::F32, uint32_t>;
+inline constexpr auto EmitFPFloor64     = EmitGlsl<GLSLstd450Floor, IR::Type::F64, uint32_t>;
+inline constexpr auto EmitFPCeil64      = EmitGlsl<GLSLstd450Ceil, IR::Type::F64, uint32_t>;
+inline constexpr auto EmitFPTrunc64     = EmitGlsl<GLSLstd450Trunc, IR::Type::F64, uint32_t>;
 inline constexpr auto EmitFPFract32     = EmitGlsl<GLSLstd450Fract, IR::Type::F32, uint32_t>;
 inline constexpr auto EmitFPFract64     = EmitGlsl<GLSLstd450Fract, IR::Type::F64, uint32_t>;
 uint32_t              EmitFPSin(EmitterState& state, uint32_t arg0);
@@ -316,15 +321,21 @@ inline constexpr auto EmitImageSampleRaw       = EmitImage;
 inline constexpr auto EmitImageGatherRaw       = EmitImage;
 inline constexpr auto EmitImageAtomicCompareSwap32 = EmitImage;
 inline constexpr auto EmitImageAtomicSwap32    = EmitImage;
+inline constexpr auto EmitImageAtomicSwap64    = EmitImage;
 inline constexpr auto EmitImageAtomicIAdd32    = EmitImage;
+inline constexpr auto EmitImageAtomicIAdd64    = EmitImage;
 inline constexpr auto EmitImageAtomicSMin32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMin32    = EmitImage;
+inline constexpr auto EmitImageAtomicUMin64    = EmitImage;
 inline constexpr auto EmitImageAtomicSMax32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMax32    = EmitImage;
 inline constexpr auto EmitImageAtomicUMax64    = EmitImage;
 inline constexpr auto EmitImageAtomicAnd32     = EmitImage;
+inline constexpr auto EmitImageAtomicAnd64     = EmitImage;
 inline constexpr auto EmitImageAtomicOr32      = EmitImage;
+inline constexpr auto EmitImageAtomicOr64      = EmitImage;
 inline constexpr auto EmitImageAtomicXor32     = EmitImage;
+inline constexpr auto EmitImageAtomicXor64     = EmitImage;
 inline constexpr auto EmitImageAtomicFMin32    = EmitImage;
 inline constexpr auto EmitImageAtomicFMax32    = EmitImage;
 void                  EmitUnreachable(ValueEmitContext& ctx, const IR::Inst& inst);

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

@@ -83,7 +83,14 @@ struct EmitterState {
 	EmitterState(const IR::Program& program_, ShaderStageInputInfo input_info_)
 	    : builder(program_.stage == ShaderType::Mesh ? 0x00010400u : 0x00010300u),
 	      program(program_), input_info(input_info_),
-	      requirements(AnalyzeProgramRequirements(program_)) {}
+	      requirements(AnalyzeProgramRequirements(program_)) {
+		if (ShaderWorkgroupInput(program.stage, input_info) != nullptr) {
+			lds_storage_class = spv::StorageClassWorkgroup;
+		}
+		if (IR::FindBinding(program.bindings, IR::DescriptorBindingKind::SharedMemory) != nullptr) {
+			lds_storage_class = spv::StorageClassStorageBuffer;
+		}
+	}
 
 	Builder                                          builder;
 	const IR::Program&                               program;
@@ -123,6 +130,8 @@ struct EmitterState {
 	uint32_t                                         push_constant_variable  = 0;
 	uint32_t                                         shader_data_storage_variable = 0;
 	uint32_t                                         flattened_srt_variable  = 0;
+	spv::StorageClass                                lds_storage_class = spv::StorageClassFunction;
+	uint32_t                                         lds_base_dwords         = 0;
 	uint32_t                                         lds_variable            = 0;
 	uint32_t                                         lds_u64_variable        = 0;
 	std::array<uint32_t, 2>                          scratch_variable {};
@@ -454,6 +463,9 @@ void EmitAtomicMemoryBarrier(EmitterState& state, IR::ResourceKind kind);
 uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
                                     bool max_value);
 
+uint32_t EmitDsFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
+                                      bool max_value);
+
 uint32_t EmitDsSwizzleTargetLane(EmitterState& state, uint32_t subid, uint32_t control);
 
 inline constexpr auto EmitSelectValueU32 =

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

@@ -710,24 +710,37 @@ uint32_t FormattedLoad(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Me
 uint32_t EncodeFormattedStoreComponent(ValueEmitContext& ctx,
                                        const Format::BufferFormatInfo& info,
                                        uint32_t component, uint32_t data) {
+	auto& state = ctx.state;
 	const auto bits = info.component_bits[component];
-	if (info.type == Format::ComponentType::Unorm) {
-		const auto value = EmitBitCastF32U32(ctx.state, data);
-		const auto clamped = EmitFPMin32(
-		    ctx.state, EmitFPMax32(ctx.state, value, ConstantF32Value(ctx.state, 0.0f)),
-		    ConstantF32Value(ctx.state, 1.0f));
-		const auto scaled = EmitFPMul32(
-		    ctx.state, clamped, ConstantF32Value(ctx.state, float((1u << bits) - 1u)));
-		return Unary(ctx.state, spv::OpConvertFToU, TypeU32(ctx.state),
-		             EmitFPRoundEven32(ctx.state, scaled));
+	const bool normalized = info.type == Format::ComponentType::Unorm ||
+	                        info.type == Format::ComponentType::Snorm;
+	const bool scaled = info.type == Format::ComponentType::Uscaled ||
+	                    info.type == Format::ComponentType::Sscaled;
+	if (normalized || scaled) {
+		// Two-bit packed channels are unsigned even in signed formats.
+		const bool is_signed = bits != 2u && (info.type == Format::ComponentType::Snorm ||
+		                                      info.type == Format::ComponentType::Sscaled);
+		const auto max_value = static_cast<float>((1u << (bits - (is_signed ? 1u : 0u))) - 1u);
+		const auto lower = normalized ? (is_signed ? -1.0f : 0.0f)
+		                              : (is_signed ? -max_value - 1.0f : 0.0f);
+		const auto upper = normalized ? 1.0f : max_value;
+		auto value = Select(state, TypeF32(state), EmitClassifyF32Bits(state, data).nan,
+		                    ConstantF32Value(state, 0.0f), EmitBitCastF32U32(state, data));
+		value = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(
+		    state, value, ConstantF32Value(state, lower), ConstantF32Value(state, upper));
+		if (normalized) {
+			value = EmitFPRoundEven32(
+			    state, EmitFPMul32(state, value, ConstantF32Value(state, max_value)));
+		}
+		// Float-to-integer conversion truncates scaled values toward zero.
+		value = Unary(state, is_signed ? spv::OpConvertFToS : spv::OpConvertFToU,
+		              is_signed ? TypeI32(state) : TypeU32(state), value);
+		return is_signed ? Unary(state, spv::OpBitcast, TypeU32(state), value) : value;
 	}
-	if (bits == 16u && (info.type == Format::ComponentType::Snorm ||
-	                    info.type == Format::ComponentType::Float)) {
-		const auto value = EmitBitCastF32U32(ctx.state, data);
-		const auto pair = EmitCompositeConstructF32x2(ctx.state, value,
-		                                               ConstantF32Value(ctx.state, 0.0f));
-		return info.type == Format::ComponentType::Float ? EmitPackHalf2x16(ctx.state, pair)
-		                                               : EmitPackSnorm2x16(ctx.state, pair);
+	if (bits == 16u && info.type == Format::ComponentType::Float) {
+		const auto pair = EmitCompositeConstructF32x2(
+		    state, EmitBitCastF32U32(state, data), ConstantF32Value(state, 0.0f));
+		return EmitPackHalf2x16(state, pair);
 	}
 	return data;
 }
@@ -1195,16 +1208,22 @@ void EmitSharedAtomic64(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const auto in_bounds = Binary(state, spv::OpULessThan, TypeBool(state), index,
 		                              ConstantU32(state, LdsDwordCount(state) / 2u));
 		EmitIfCondition(state, in_bounds, [&]() {
+			auto native_index = index;
+			auto semantics = spv::MemorySemanticsWorkgroupMemoryMask;
+			if (state.lds_storage_class == spv::StorageClassStorageBuffer) {
+				native_index = EmitAddU32(state, index, EmitBinaryU32(
+				    state, spv::OpShiftRightLogical, state.lds_base_dwords, ConstantU32(state, 1)));
+				semantics = spv::MemorySemanticsUniformMemoryMask;
+			}
 			const auto pointer = state.builder.AllocateId();
 			state.builder.AddFunction(spv::OpAccessChain,
-			                          TypePointer(state, spv::StorageClassWorkgroup, TypeScalarU64(state)),
-			                          pointer, state.lds_u64_variable, ConstantU32(state, 0), index);
+			                          TypePointer(state, state.lds_storage_class, TypeScalarU64(state)),
+			                          pointer, state.lds_u64_variable, ConstantU32(state, 0), native_index);
 			const auto value = Unary(state, spv::OpBitcast, TypeScalarU64(state), ctx.Arg(inst, 1));
 			state.builder.AddFunction(
 			    SpirvAtomicOpcode(inst.GetOpcode()), TypeScalarU64(state), state.builder.AllocateId(),
 			    pointer, ConstantU32(state, spv::ScopeWorkgroup),
-			    ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask |
-			                           spv::MemorySemanticsWorkgroupMemoryMask), value);
+			    ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask | semantics), value);
 		});
 	});
 }
@@ -1221,34 +1240,10 @@ uint32_t EmitBufferFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst) {
 void EmitSharedFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto& mem       = ctx.Memory(inst);
 	const bool  max_value = inst.GetOpcode() == IR::ValueOpcode::SharedAtomicFMax32;
-	EmitIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
-		const auto index = DwordIndex(ctx, inst, mem);
-		const auto resource = PrepareMemoryResourceAccess(ctx.state, mem);
-		EmitIfCondition(
-		    ctx.state, EmitMemoryElementInBounds(ctx.state, resource, index), [&]() {
-			    ctx.state.builder.AddFunction(spv::OpStore, ctx.scratch_u32_variable,
-			                                  ctx.Arg(inst, 1));
-			    const auto data = ctx.state.builder.AllocateId();
-			    ctx.state.builder.AddFunction(spv::OpLoad, TypeU32(ctx.state), data,
-			                                  ctx.scratch_u32_variable);
-			    AtomicUpdate(
-			        ctx.state, EmitMemoryElementPointer(ctx.state, resource, index),
-			        mem.kind, [&](uint32_t old) {
-				        const auto old_f =
-				            Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), old);
-				        const auto compare_f =
-				            Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), ctx.Arg(inst, 2));
-				        const auto data_f =
-				            Unary(ctx.state, spv::OpBitcast, TypeF32(ctx.state), data);
-				        const auto compare = Binary(
-				            ctx.state, max_value ? spv::OpFOrdGreaterThan : spv::OpFOrdLessThan,
-				            TypeBool(ctx.state), max_value ? old_f : compare_f,
-				            max_value ? compare_f : old_f);
-				        return Unary(ctx.state, spv::OpBitcast, TypeU32(ctx.state),
-				                     Select(ctx.state, TypeF32(ctx.state), compare, data_f, old_f));
-			        });
-		    });
-	});
+	EmitAtomicUpdate(ctx, inst, mem,
+	                 [max_value](EmitterState& state, uint32_t old, uint32_t value) {
+		                 return EmitDsFloatAtomicReplacement(state, old, value, max_value);
+	                 });
 }
 
 uint32_t EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst) {

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

@@ -121,8 +121,8 @@ void EnsureLdsStorage(EmitterState& state) {
 	if (state.lds_variable != 0) {
 		return;
 	}
-	if (ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {
-		EXIT("function LDS was not prepared before SPIR-V function emission\n");
+	if (state.lds_storage_class != spv::StorageClassWorkgroup) {
+		EXIT("LDS backing was not prepared before SPIR-V function emission\n");
 	}
 	const auto define = [&](uint32_t type, uint32_t bytes) {
 		const auto array = state.builder.DecoratedType(
@@ -227,12 +227,14 @@ uint32_t EmitMemoryElementInBounds(EmitterState& state, const MemoryResourceAcce
 uint32_t EmitMemoryElementPointer(EmitterState& state, const MemoryResourceAccess& access,
                                   uint32_t index) {
 	if (access.kind == IR::ResourceKind::Lds || access.kind == IR::ResourceKind::Scratch) {
+		const auto storage_class = access.kind == IR::ResourceKind::Scratch
+		                               ? spv::StorageClassFunction : state.lds_storage_class;
+		if (storage_class == spv::StorageClassStorageBuffer) {
+			return EmitStorageBufferElementPointer(
+			    state, access, EmitAddU32(state, state.lds_base_dwords, index),
+			    TypeStorageBufferElementPointer(state, 32));
+		}
 		const auto pointer = state.builder.AllocateId();
-		const auto storage_class =
-		    access.kind == IR::ResourceKind::Scratch ? spv::StorageClassFunction
-		    : ShaderWorkgroupInput(state.program.stage, state.input_info) != nullptr
-		        ? spv::StorageClassWorkgroup
-		        : spv::StorageClassFunction;
 		if (access.kind == IR::ResourceKind::Lds && state.requirements.shared_int64_atomics) {
 			state.builder.AddFunction(spv::OpAccessChain, TypeU32ElementPointer(state, storage_class),
 			                          pointer, access.object_pointer, ConstantU32(state, 0), index);
@@ -359,10 +361,12 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::ImageAtomicCompareSwap32:
 		case IR::ValueOpcode::BufferAtomicCmpSwap32: return spv::OpAtomicCompareExchange;
 		case IR::ValueOpcode::ImageAtomicSwap32:
+		case IR::ValueOpcode::ImageAtomicSwap64:
 		case IR::ValueOpcode::BufferAtomicSwap32:
 		case IR::ValueOpcode::BufferAtomicSwap64:
 		case IR::ValueOpcode::SharedAtomicSwap32: return spv::OpAtomicExchange;
 		case IR::ValueOpcode::ImageAtomicIAdd32:
+		case IR::ValueOpcode::ImageAtomicIAdd64:
 		case IR::ValueOpcode::BufferAtomicIAdd32:
 		case IR::ValueOpcode::SharedAtomicIAdd64:
 		case IR::ValueOpcode::SharedAtomicIAdd32: return spv::OpAtomicIAdd;
@@ -372,6 +376,7 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::BufferAtomicSMin32:
 		case IR::ValueOpcode::SharedAtomicSMin32: return spv::OpAtomicSMin;
 		case IR::ValueOpcode::ImageAtomicUMin32:
+		case IR::ValueOpcode::ImageAtomicUMin64:
 		case IR::ValueOpcode::BufferAtomicUMin32:
 		case IR::ValueOpcode::SharedAtomicUMin32: return spv::OpAtomicUMin;
 		case IR::ValueOpcode::ImageAtomicSMax32:
@@ -382,15 +387,18 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::BufferAtomicUMax32:
 		case IR::ValueOpcode::SharedAtomicUMax32: return spv::OpAtomicUMax;
 		case IR::ValueOpcode::ImageAtomicAnd32:
+		case IR::ValueOpcode::ImageAtomicAnd64:
 		case IR::ValueOpcode::BufferAtomicAnd32:
 		case IR::ValueOpcode::BufferAtomicAnd64:
 		case IR::ValueOpcode::SharedAtomicAnd32: return spv::OpAtomicAnd;
 		case IR::ValueOpcode::ImageAtomicOr32:
+		case IR::ValueOpcode::ImageAtomicOr64:
 		case IR::ValueOpcode::BufferAtomicOr32:
 		case IR::ValueOpcode::BufferAtomicOr64:
 		case IR::ValueOpcode::SharedAtomicOr64:
 		case IR::ValueOpcode::SharedAtomicOr32: return spv::OpAtomicOr;
 		case IR::ValueOpcode::ImageAtomicXor32:
+		case IR::ValueOpcode::ImageAtomicXor64:
 		case IR::ValueOpcode::BufferAtomicXor32:
 		case IR::ValueOpcode::SharedAtomicXor32: return spv::OpAtomicXor;
 		default: return spv::OpNop;
@@ -424,7 +432,10 @@ void EmitAtomicMemoryBarrier(EmitterState& state, IR::ResourceKind kind) {
 	const auto scope  = kind == IR::ResourceKind::Lds ? spv::ScopeWorkgroup : spv::ScopeDevice;
 	const auto memory = [&] {
 		switch (kind) {
-			case IR::ResourceKind::Lds: return spv::MemorySemanticsWorkgroupMemoryMask;
+			case IR::ResourceKind::Lds:
+				return state.lds_storage_class == spv::StorageClassStorageBuffer
+				           ? spv::MemorySemanticsUniformMemoryMask
+				           : spv::MemorySemanticsWorkgroupMemoryMask;
 			case IR::ResourceKind::Image: return spv::MemorySemanticsImageMemoryMask;
 			default: return spv::MemorySemanticsUniformMemoryMask;
 		}
@@ -433,6 +444,20 @@ void EmitAtomicMemoryBarrier(EmitterState& state, IR::ResourceKind kind) {
 	                          ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask | memory));
 }
 
+namespace {
+
+uint32_t FloatOrderKey(EmitterState& state, uint32_t bits) {
+	const auto negative = EmitCompareU32Constant(state, spv::OpINotEqual,
+	                                             EmitAndConstant(state, bits, 0x80000000u), 0u);
+	const auto negative_key = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpNot, TypeU32(state), negative_key, bits);
+	const auto positive_key =
+	    EmitBinaryU32(state, spv::OpBitwiseXor, bits, ConstantU32(state, 0x80000000u));
+	return EmitSelectValueU32(state, negative, negative_key, positive_key);
+}
+
+} // namespace
+
 uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
                                     bool max_value) {
 	struct OrderedBits {
@@ -442,15 +467,7 @@ uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t
 	};
 	const auto classify = [&](uint32_t bits) {
 		const auto cls = EmitClassifyF32Bits(state, bits);
-		const auto negative = EmitCompareU32Constant(state, spv::OpINotEqual,
-		                                             EmitAndConstant(state, bits, 0x80000000u), 0u);
-		const auto negative_key = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpNot, TypeU32(state), negative_key, bits);
-		const auto positive_key =
-		    EmitBinaryU32(state, spv::OpBitwiseXor, bits, ConstantU32(state, 0x80000000u));
-		return OrderedBits {
-		    cls.nan, cls.zero,
-		    EmitSelectValueU32(state, negative, negative_key, positive_key)};
+		return OrderedBits {cls.nan, cls.zero, FloatOrderKey(state, bits)};
 	};
 	const auto source_class = classify(source);
 	const auto old_class    = classify(old);
@@ -465,6 +482,45 @@ uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t
 	    old);
 }
 
+uint32_t EmitDsFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
+                                      bool max_value) {
+	struct OrderedBits {
+		uint32_t nan;
+		uint32_t signaling_nan;
+		uint32_t key;
+	};
+	const auto classify = [&](uint32_t bits) {
+		const auto abs = EmitAndConstant(state, bits, 0x7fffffffu);
+		const auto nan = EmitCompareU32Constant(state, spv::OpUGreaterThan, abs, 0x7f800000u);
+		const auto signaling_nan = EmitLogicalAndBool(
+		    state, nan, EmitCompareU32Constant(state, spv::OpIEqual,
+		                                      EmitAndConstant(state, bits, 0x00400000u), 0u));
+		const auto small = EmitCompareU32Constant(state, spv::OpULessThan, abs, 0x00800000u);
+		const auto sign = EmitAndConstant(state, bits, 0x80000000u);
+		const auto positive_denorm = EmitLogicalAndBool(
+		    state, EmitCompareU32Constant(state, spv::OpIEqual, sign, 0u),
+		    EmitCompareU32Constant(state, spv::OpINotEqual, abs, 0u));
+		// Negative denorms tie with -0; positive denorms tie only with each other.
+		const auto small_bits = EmitSelectValueU32(
+		    state, positive_denorm, ConstantU32(state, 1u), sign);
+		const auto key_bits = EmitSelectValueU32(state, small, small_bits, bits);
+		return OrderedBits {nan, signaling_nan, FloatOrderKey(state, key_bits)};
+	};
+	const auto lhs = classify(old);
+	const auto rhs = classify(source);
+	const auto choose_source = state.builder.AllocateId();
+	state.builder.AddFunction(max_value ? spv::OpUGreaterThan : spv::OpULessThanEqual,
+	                          TypeBool(state), choose_source, rhs.key, lhs.key);
+	// Keep original bits: ordering denorms does not flush the selected operand.
+	auto result = EmitSelectValueU32(state, choose_source, source, old);
+	result = EmitSelectValueU32(state, lhs.nan, source, result);
+	result = EmitSelectValueU32(state, rhs.nan, old, result);
+	result = EmitSelectValueU32(state, rhs.signaling_nan,
+	                            EmitOrU32(state, source, ConstantU32(state, 0x00400000u)), result);
+	return EmitSelectValueU32(state, lhs.signaling_nan,
+	                          EmitOrU32(state, old, ConstantU32(state, 0x00400000u)), result);
+}
+
 uint32_t EmitDsSwizzleTargetLane(EmitterState& state, uint32_t subid, uint32_t control) {
 	if ((control & 0xc000u) == 0xc000u) {
 		const uint32_t mask         = control & 0x1fu;

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

@@ -275,6 +275,17 @@ void DefineDescriptors(EmitterState& state) {
 			case IR::DescriptorBindingKind::Gds:
 				state.gds_variable = Define(StorageBufferType(state), "gds");
 				break;
+			case IR::DescriptorBindingKind::SharedMemory:
+				state.lds_variable = Define(StorageBufferType(state), "lds_dwords");
+				state.builder.AddAnnotation(spv::OpDecorate, state.lds_variable, spv::DecorationCoherent);
+				if (state.requirements.shared_int64_atomics) {
+					state.lds_u64_variable = Define(StorageBufferType(state, 64), "lds_qwords");
+					state.builder.AddAnnotation(spv::OpDecorate, state.lds_u64_variable, spv::DecorationCoherent);
+					for (const auto variable: {state.lds_variable, state.lds_u64_variable}) {
+						state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationAliased);
+					}
+				}
+				break;
 			default: {
 				EXIT_IF(IR::ImageBindingResourceClass(binding.kind) ==
 				        IR::ImageResourceClass::None);
@@ -408,6 +419,7 @@ uint32_t BuiltInForInput(IR::StageInputKind kind) {
 		case IR::StageInputKind::BaryCoordSmooth: return spv::BuiltInBaryCoordKHR;
 		case IR::StageInputKind::BaryCoordNoPerspective: return spv::BuiltInBaryCoordNoPerspKHR;
 		case IR::StageInputKind::WorkgroupId: return spv::BuiltInWorkgroupId;
+		case IR::StageInputKind::NumWorkgroups: return spv::BuiltInNumWorkgroups;
 		case IR::StageInputKind::LocalInvocationId: return spv::BuiltInLocalInvocationId;
 		case IR::StageInputKind::LocalInvocationIndex: return spv::BuiltInLocalInvocationIndex;
 		case IR::StageInputKind::GlobalInvocationId: return spv::BuiltInGlobalInvocationId;
@@ -420,15 +432,19 @@ void DefineInputs(EmitterState& state) {
 	for (const auto& input: state.program.info.inputs) {
 		state.inputs.push_back({input});
 	}
+	const auto add_builtin = [&](IR::StageInputKind kind, uint32_t components,
+	                             const char* name) {
+		if (std::ranges::none_of(state.inputs, [kind](const InputBinding& input) {
+			    return input.kind == kind;
+		    })) {
+			state.inputs.push_back({{kind, 0, components, name}});
+		}
+	};
+	if (IR::FindBinding(state.program.bindings, IR::DescriptorBindingKind::SharedMemory) != nullptr) {
+		add_builtin(IR::StageInputKind::WorkgroupId, 3, "gl_WorkGroupID");
+		add_builtin(IR::StageInputKind::NumWorkgroups, 3, "gl_NumWorkGroups");
+	}
 	if (state.lane_count == 2) {
-		const auto add_builtin = [&](IR::StageInputKind kind, uint32_t components,
-		                             const char* name) {
-			if (std::ranges::none_of(state.inputs, [kind](const InputBinding& input) {
-				    return input.kind == kind;
-			    })) {
-				state.inputs.push_back({{kind, 0, components, name}});
-			}
-		};
 		add_builtin(IR::StageInputKind::LocalInvocationIndex, 1, "gl_LocalInvocationIndex");
 		if (std::ranges::any_of(state.inputs, [](const InputBinding& input) {
 			    return input.kind == IR::StageInputKind::GlobalInvocationId;
@@ -459,6 +475,7 @@ void DefineInputs(EmitterState& state) {
 			case IR::StageInputKind::Layer:
 			case IR::StageInputKind::SampleId: type = TypeI32(state); break;
 			case IR::StageInputKind::WorkgroupId:
+			case IR::StageInputKind::NumWorkgroups:
 			case IR::StageInputKind::LocalInvocationId:
 			case IR::StageInputKind::GlobalInvocationId: type = TypeU32Vector(state, 3); break;
 			case IR::StageInputKind::FragCoord: type = TypeF32Vector(state, 4); break;
@@ -677,7 +694,8 @@ void DefineModule(EmitterState& state) {
 		state.builder.RequireCapability(spv::CapabilityInt64);
 		state.builder.RequireCapability(spv::CapabilityInt64Atomics);
 	}
-	if (state.requirements.shared_int64_atomics) {
+	if (state.requirements.shared_int64_atomics &&
+	    state.lds_storage_class == spv::StorageClassWorkgroup) {
 		state.builder.RequireVersion(0x00010400u);
 		state.builder.RequireExtension("SPV_KHR_workgroup_memory_explicit_layout");
 		state.builder.RequireCapability(spv::CapabilityWorkgroupMemoryExplicitLayoutKHR);

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

@@ -689,9 +689,7 @@ void EmitProgram(EmitterState& state) {
 	DefineBvhIntersect(state);
 	for (const auto* block: program.blocks) {
 		if (std::ranges::any_of(*block, [](const IR::Inst& inst) {
-			    return inst.GetOpcode() == IR::ValueOpcode::SwizzleU32 ||
-			           inst.GetOpcode() == IR::ValueOpcode::SharedAtomicFMin32 ||
-			           inst.GetOpcode() == IR::ValueOpcode::SharedAtomicFMax32;
+			    return inst.GetOpcode() == IR::ValueOpcode::SwizzleU32;
 		    })) {
 			ctx.scratch_u32_variable = state.builder.AllocateId();
 			if (state.lane_count == 2) {
@@ -752,6 +750,17 @@ void EmitProgram(EmitterState& state) {
 		state.builder.AddFunction(spv::OpStore, state.pixel_valid_mask_variable,
 		                          ConstantU32(state, 1));
 	}
+	if (state.lds_storage_class == spv::StorageClassStorageBuffer && state.lds_variable != 0) {
+		const auto group_x = EmitInputComponentU32(state, IR::StageInputKind::WorkgroupId, 0);
+		const auto group_y = EmitInputComponentU32(state, IR::StageInputKind::WorkgroupId, 1);
+		const auto group_z = EmitInputComponentU32(state, IR::StageInputKind::WorkgroupId, 2);
+		const auto count_x = EmitInputComponentU32(state, IR::StageInputKind::NumWorkgroups, 0);
+		const auto count_y = EmitInputComponentU32(state, IR::StageInputKind::NumWorkgroups, 1);
+		const auto row = EmitAddU32(state, group_y, EmitBinaryU32(state, spv::OpIMul, group_z, count_y));
+		const auto index = EmitAddU32(state, group_x, EmitBinaryU32(state, spv::OpIMul, row, count_x));
+		state.lds_base_dwords = EmitBinaryU32(state, spv::OpIMul, index,
+		                                     ConstantU32(state, LdsDwordCount(state)));
+	}
 	EmitMemoryOffsets(state);
 	if (program.blocks.empty()) {
 		EmitReturn(ctx);

src/graphics/shader/recompiler/frontend/decode/ImageOps.cpp

@@ -20,6 +20,12 @@ struct MimgGatherInfo {
 	uint32_t flags    = 0;
 };
 
+struct MimgAtomicInfo {
+	uint32_t encoding    = 0;
+	Opcode   decoded     = Opcode::UNSUPPORTED;
+	bool     supports_64 = false;
+};
+
 constexpr ImageDimension DecodeImageDimension(uint32_t dim) {
 	switch (dim) {
 		case 0u: return ImageDimension::Dim1D;
@@ -190,17 +196,17 @@ constexpr MimgGatherInfo MIMG_GATHER_OPCODE_LIST[] = {
     {0x61u, Opcode::IMAGE_GATHER4H, ImageSampleFlagGatherHorizontal},
 };
 
-constexpr Detail::OpcodeMap MIMG_ATOMIC_OPCODE_LIST[] = {
-    {0x0fu, Opcode::IMAGE_ATOMIC_SWAP},
+constexpr MimgAtomicInfo MIMG_ATOMIC_OPCODE_LIST[] = {
+    {0x0fu, Opcode::IMAGE_ATOMIC_SWAP, true},
     {0x10u, Opcode::IMAGE_ATOMIC_CMPSWAP},
-    {0x11u, Opcode::IMAGE_ATOMIC_ADD},
+    {0x11u, Opcode::IMAGE_ATOMIC_ADD, true},
     {0x14u, Opcode::IMAGE_ATOMIC_SMIN},
-    {0x15u, Opcode::IMAGE_ATOMIC_UMIN},
+    {0x15u, Opcode::IMAGE_ATOMIC_UMIN, true},
     {0x16u, Opcode::IMAGE_ATOMIC_SMAX},
-    {0x17u, Opcode::IMAGE_ATOMIC_UMAX},
-    {0x18u, Opcode::IMAGE_ATOMIC_AND},
-    {0x19u, Opcode::IMAGE_ATOMIC_OR},
-    {0x1au, Opcode::IMAGE_ATOMIC_XOR},
+    {0x17u, Opcode::IMAGE_ATOMIC_UMAX, true},
+    {0x18u, Opcode::IMAGE_ATOMIC_AND, true},
+    {0x19u, Opcode::IMAGE_ATOMIC_OR, true},
+    {0x1au, Opcode::IMAGE_ATOMIC_XOR, true},
     {0x1eu, Opcode::IMAGE_ATOMIC_FMIN},
     {0x1fu, Opcode::IMAGE_ATOMIC_FMAX},
 };
@@ -210,7 +216,7 @@ constexpr auto MIMG_GATHER_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_GATHER_OPCO
 constexpr auto MIMG_ATOMIC_OPS = Detail::MakeOpcodeTable<0x100>(MIMG_ATOMIC_OPCODE_LIST);
 
 Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const MimgGatherInfo* gather,
-                        const Detail::OpcodeMap* atomic) {
+                        const MimgAtomicInfo* atomic) {
 	if (sample != nullptr) {
 		return Opcode::IMAGE_SAMPLE;
 	}
@@ -246,7 +252,7 @@ uint32_t DecodeMimgSampleFlags(const MimgSampleInfo* sample, const MimgGatherInf
 
 uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
                                      const MimgSampleInfo* sample, const MimgGatherInfo* gather,
-                                     const Detail::OpcodeMap* atomic) {
+                                     const MimgAtomicInfo* atomic) {
 	if (sample != nullptr) {
 		return ImageSampleAddressComponents(sample->flags, dimension);
 	}
@@ -370,7 +376,7 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 			SetUnsupported(inst, Family::MIMG, opcode, "MIMG image atomic has invalid DMASK");
 		} else if (inst.dmask == mask64) {
 			inst.data_bits = 64u;
-			if (inst.opcode != Opcode::IMAGE_ATOMIC_UMAX) {
+			if (!atomic->supports_64) {
 				SetUnsupported(inst, Family::MIMG, opcode,
 				               "MIMG 64-bit image atomic opcode is not implemented");
 			}

src/graphics/shader/recompiler/frontend/decode/MemoryOps.cpp

@@ -205,9 +205,9 @@ uint32_t DsSourceCount(Opcode opcode) {
 		case Opcode::DS_WRITE2ST64_B32:
 		case Opcode::DS_WRITE2_B64:
 		case Opcode::DS_WRITE2ST64_B64:
-		case Opcode::DS_MSKOR_B32:
+		case Opcode::DS_MSKOR_B32: return 3u;
 		case Opcode::DS_MIN_F32:
-		case Opcode::DS_MAX_F32: return 3u;
+		case Opcode::DS_MAX_F32:
 		case Opcode::DS_PERMUTE_B32:
 		case Opcode::DS_BPERMUTE_B32: return 2u;
 		case Opcode::DS_READ_ADDTID_B32:

src/graphics/shader/recompiler/frontend/decode/ShaderDecoder.h

@@ -188,9 +188,12 @@ enum class Opcode {
 	V_FRACT_F32,
 	V_FRACT_F64,
 	V_TRUNC_F32,
+	V_TRUNC_F64,
 	V_CEIL_F32,
+	V_CEIL_F64,
 	V_RNDNE_F32,
 	V_FLOOR_F32,
+	V_FLOOR_F64,
 	V_EXP_F32,
 	V_LOG_F32,
 	V_RSQ_F32,
@@ -221,6 +224,9 @@ enum class Opcode {
 	V_MAC_F32,
 	V_MADMK_F32,
 	V_MADAK_F32,
+	V_FMAC_F32,
+	V_FMAMK_F32,
+	V_FMAAK_F32,
 	V_MIN_F32,
 	V_MAX_F32,
 	V_ADD_F16,
@@ -248,6 +254,8 @@ enum class Opcode {
 	V_FMA_F64,
 	V_ADD_F64,
 	V_MUL_F64,
+	V_MIN_F64,
+	V_MAX_F64,
 	V_FMA_F32,
 	V_FMA_F16,
 	V_PACK_B32_F16,

src/graphics/shader/recompiler/frontend/decode/VectorAluOps.cpp

@@ -65,9 +65,9 @@ constexpr Vop2OpcodeInfo VOP2_OPCODE_LIST[] = {
     {0x2au, Opcode::V_SUBREV_CO_CI_U32},
     {0x26u, Opcode::V_SUB_NC_U32, Vop2SdwaProfile::IntegerPartialDestination},
     {0x27u, Opcode::V_SUBREV_NC_U32, Vop2SdwaProfile::IntegerFullDestination},
-    {0x2bu, Opcode::V_MAC_F32},
-    {0x2cu, Opcode::V_MADMK_F32},
-    {0x2du, Opcode::V_MADAK_F32},
+    {0x2bu, Opcode::V_FMAC_F32},
+    {0x2cu, Opcode::V_FMAMK_F32},
+    {0x2du, Opcode::V_FMAAK_F32},
     {0x2fu, Opcode::V_CVT_PKRTZ_F16_F32, Vop2SdwaProfile::PackedFloat16},
     {0x32u, Opcode::V_ADD_F16, Vop2SdwaProfile::Float16},
     {0x33u, Opcode::V_SUB_F16, Vop2SdwaProfile::Float16},
@@ -91,6 +91,9 @@ constexpr OpcodeMap VOP1_OPCODE_LIST[] = {
     {0x0fu, Opcode::V_CVT_F32_F64},
     {0x10u, Opcode::V_CVT_F64_F32},
     {0x16u, Opcode::V_CVT_F64_U32},
+    {0x17u, Opcode::V_TRUNC_F64},
+    {0x18u, Opcode::V_CEIL_F64},
+    {0x1au, Opcode::V_FLOOR_F64},
     {0x2fu, Opcode::V_RCP_F64},
     {0x05u, Opcode::V_CVT_F32_I32},
     {0x06u, Opcode::V_CVT_F32_U32},
@@ -156,6 +159,9 @@ constexpr OpcodeMap VOP3_ENCODED_VOP1_OPCODE_LIST[] = {
     {0x0fu, Opcode::V_CVT_F32_F64},
     {0x10u, Opcode::V_CVT_F64_F32},
     {0x16u, Opcode::V_CVT_F64_U32},
+    {0x17u, Opcode::V_TRUNC_F64},
+    {0x18u, Opcode::V_CEIL_F64},
+    {0x1au, Opcode::V_FLOOR_F64},
     {0x2fu, Opcode::V_RCP_F64},
     {0x05u, Opcode::V_CVT_F32_I32},
     {0x06u, Opcode::V_CVT_F32_U32},
@@ -292,6 +298,8 @@ constexpr OpcodeMap VOP3_OPCODE_LIST[] = {
     {0x14cu, Opcode::V_FMA_F64},
     {0x164u, Opcode::V_ADD_F64},
     {0x165u, Opcode::V_MUL_F64},
+    {0x166u, Opcode::V_MIN_F64},
+    {0x167u, Opcode::V_MAX_F64},
     {0x148u, Opcode::V_BFE_U32},
     {0x149u, Opcode::V_BFE_I32},
     {0x14au, Opcode::V_BFI_B32},
@@ -478,7 +486,9 @@ bool IsVopcCompareExec(Opcode opcode);
 bool IsVop1Float64Opcode(Opcode opcode) {
 	return opcode == Opcode::V_CVT_F64_I32 || opcode == Opcode::V_CVT_F32_F64 ||
 	       opcode == Opcode::V_CVT_F64_F32 || opcode == Opcode::V_CVT_F64_U32 ||
-	       opcode == Opcode::V_RCP_F64 || opcode == Opcode::V_FRACT_F64;
+	       opcode == Opcode::V_RCP_F64 || opcode == Opcode::V_FRACT_F64 ||
+	       opcode == Opcode::V_TRUNC_F64 || opcode == Opcode::V_CEIL_F64 ||
+	       opcode == Opcode::V_FLOOR_F64;
 }
 
 bool IsVop1FloatSourceOpcode(Opcode opcode) {
@@ -487,6 +497,9 @@ bool IsVop1FloatSourceOpcode(Opcode opcode) {
 		case Opcode::V_CVT_F64_F32:
 		case Opcode::V_RCP_F64:
 		case Opcode::V_FRACT_F64:
+		case Opcode::V_TRUNC_F64:
+		case Opcode::V_CEIL_F64:
+		case Opcode::V_FLOOR_F64:
 		case Opcode::V_MOV_B32:
 		case Opcode::V_CVT_F32_F16:
 		case Opcode::V_CVT_U32_F32:
@@ -787,6 +800,9 @@ bool IsVop2FloatOpcode(Opcode opcode) {
 		case Opcode::V_MAC_F32:
 		case Opcode::V_MADMK_F32:
 		case Opcode::V_MADAK_F32:
+		case Opcode::V_FMAC_F32:
+		case Opcode::V_FMAMK_F32:
+		case Opcode::V_FMAAK_F32:
 		case Opcode::V_CVT_PKRTZ_F16_F32:
 		case Opcode::V_ADD_F16:
 		case Opcode::V_SUB_F16:
@@ -1044,6 +1060,7 @@ void FinalizeVop2Instruction(std::span<const uint32_t> code, uint32_t word_index
                              Instruction& inst) {
 	switch (inst.opcode) {
 		case Opcode::V_MADMK_F32:
+		case Opcode::V_FMAMK_F32:
 		case Opcode::V_FMAMK_F16:
 			inst.src2      = inst.src1;
 			inst.src1      = {};
@@ -1051,6 +1068,7 @@ void FinalizeVop2Instruction(std::span<const uint32_t> code, uint32_t word_index
 			inst.src_count = 3;
 			break;
 		case Opcode::V_MADAK_F32:
+		case Opcode::V_FMAAK_F32:
 		case Opcode::V_FMAAK_F16:
 			inst.src2      = {};
 			inst.src2.kind = OperandKind::LiteralConstant;
@@ -1113,8 +1131,10 @@ void DecodeVop2Dpp(uint32_t pc, std::span<const uint32_t> code, uint32_t word_in
 	DecodeVectorGpr(vsrc1, inst.src1);
 	DecodeScalarSource(src0 + 256u, pc, inst.src0);
 	ApplyDppModifier(inst.src0, modifier, code[word_index] & 0x1ffu);
-	inst.src1.negate       = ((modifier >> 22u) & 0x1u) != 0u;
-	inst.src1.absolute     = ((modifier >> 23u) & 0x1u) != 0u;
+	if (!inst.src0.dpp8) {
+		inst.src1.negate   = ((modifier >> 22u) & 0x1u) != 0u;
+		inst.src1.absolute = ((modifier >> 23u) & 0x1u) != 0u;
+	}
 	const bool packed_fmac = inst.opcode == Opcode::V_PK_FMAC_F16;
 	if (packed_fmac) {
 		inst.src0.negate_hi = inst.src0.negate;
@@ -1235,6 +1255,8 @@ uint32_t NativeVop3SourceCount(Opcode opcode) {
 	switch (opcode) {
 		case Opcode::V_ADD_F64:
 		case Opcode::V_MUL_F64:
+		case Opcode::V_MIN_F64:
+		case Opcode::V_MAX_F64:
 		case Opcode::V_MUL_LO_U32:
 		case Opcode::V_MUL_HI_U32:
 		case Opcode::V_MUL_LO_I32:
@@ -1384,9 +1406,12 @@ bool SupportsNativeVop3SourceModifiers(Opcode opcode) {
 		case Opcode::V_MIN_F32:
 		case Opcode::V_MAX_F32:
 		case Opcode::V_MAC_F32:
+		case Opcode::V_FMAC_F32:
 		case Opcode::V_MAD_F32:
 		case Opcode::V_ADD_F64:
 		case Opcode::V_MUL_F64:
+		case Opcode::V_MIN_F64:
+		case Opcode::V_MAX_F64:
 		case Opcode::V_FMA_F64:
 		case Opcode::V_FMA_F32:
 		case Opcode::V_PACK_B32_F16:
@@ -1417,6 +1442,7 @@ bool SupportsNativeVop3ResultModifiers(Opcode opcode) {
 		case Opcode::V_MIN_F32:
 		case Opcode::V_MAX_F32:
 		case Opcode::V_MAC_F32:
+		case Opcode::V_FMAC_F32:
 		case Opcode::V_MAD_F32:
 		case Opcode::V_FMA_F32:
 		case Opcode::V_FMA_F16:
@@ -1565,6 +1591,7 @@ void DecodeVop2(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 		return;
 	}
 	switch (src0) {
+		case 233u: DecodeVop2Dpp(pc, code, word_index, opcode, vdst, vsrc1, inst); return;
 		case 249u: DecodeVop2Sdwa(pc, code, word_index, opcode, vdst, vsrc1, inst); return;
 		case 250u: DecodeVop2Dpp(pc, code, word_index, opcode, vdst, vsrc1, inst); return;
 		default: break;

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

@@ -546,7 +546,10 @@ void Translator::BUFFER_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode
 	}
 }
 
-void Translator::IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode) {
+void Translator::IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode32,
+                               IR::ValueOpcode opcode64) {
+	const auto opcode = inst.data_bits == 64u ? opcode64 : opcode32;
+	EXIT_IF(opcode == IR::ValueOpcode::Count);
 	const auto memory   = MemoryInfoFromDecoded(inst);
 	const auto resource = GetImageResource(memory);
 	const auto address  = MakeImageAddress(inst, MemorySourceAt(inst, 1));
@@ -1041,24 +1044,30 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::IMAGE_ATOMIC_CMPSWAP:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicCompareSwap32);
 		case Decoder::Opcode::IMAGE_ATOMIC_SWAP:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSwap32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSwap32,
+			                    IR::ValueOpcode::ImageAtomicSwap64);
 		case Decoder::Opcode::IMAGE_ATOMIC_ADD:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicIAdd32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicIAdd32,
+			                    IR::ValueOpcode::ImageAtomicIAdd64);
 		case Decoder::Opcode::IMAGE_ATOMIC_SMIN:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSMin32);
 		case Decoder::Opcode::IMAGE_ATOMIC_UMIN:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMin32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMin32,
+			                    IR::ValueOpcode::ImageAtomicUMin64);
 		case Decoder::Opcode::IMAGE_ATOMIC_SMAX:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicSMax32);
 		case Decoder::Opcode::IMAGE_ATOMIC_UMAX:
-			return IMAGE_ATOMIC(inst, inst.data_bits == 64u ? IR::ValueOpcode::ImageAtomicUMax64
-			                                                : IR::ValueOpcode::ImageAtomicUMax32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicUMax32,
+			                    IR::ValueOpcode::ImageAtomicUMax64);
 		case Decoder::Opcode::IMAGE_ATOMIC_AND:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicAnd32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicAnd32,
+			                    IR::ValueOpcode::ImageAtomicAnd64);
 		case Decoder::Opcode::IMAGE_ATOMIC_OR:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicOr32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicOr32,
+			                    IR::ValueOpcode::ImageAtomicOr64);
 		case Decoder::Opcode::IMAGE_ATOMIC_XOR:
-			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicXor32);
+			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicXor32,
+			                    IR::ValueOpcode::ImageAtomicXor64);
 		case Decoder::Opcode::IMAGE_ATOMIC_FMIN:
 			return IMAGE_ATOMIC(inst, IR::ValueOpcode::ImageAtomicFMin32);
 		case Decoder::Opcode::IMAGE_ATOMIC_FMAX:

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

@@ -895,8 +895,13 @@ void IncludeInstructionVectorRegisters(const Decoder::Instruction& inst, uint32_
 		case Decoder::Opcode::V_CVT_F64_U32: include_vector(inst.dst, 2u); break;
 		case Decoder::Opcode::V_FMA_F64: include_vector(inst.src2, 2u); [[fallthrough]];
 		case Decoder::Opcode::V_ADD_F64:
+		case Decoder::Opcode::V_MIN_F64:
+		case Decoder::Opcode::V_MAX_F64:
 		case Decoder::Opcode::V_MUL_F64: include_vector(inst.src1, 2u); [[fallthrough]];
 		case Decoder::Opcode::V_RCP_F64:
+		case Decoder::Opcode::V_TRUNC_F64:
+		case Decoder::Opcode::V_CEIL_F64:
+		case Decoder::Opcode::V_FLOOR_F64:
 		case Decoder::Opcode::V_FRACT_F64: include_vector(inst.dst, 2u); [[fallthrough]];
 		case Decoder::Opcode::V_CVT_F32_F64: include_vector(inst.src0, 2u); break;
 		case Decoder::Opcode::V_CMP_EQ_F64:

src/graphics/shader/recompiler/frontend/translate/Translator.h

@@ -86,7 +86,8 @@ private:
 	void          BUFFER_LOAD(const Decoder::Instruction& inst);
 	void          BUFFER_STORE(const Decoder::Instruction& inst);
 	void          BUFFER_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
-	void          IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
+	void          IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode32,
+	                           IR::ValueOpcode opcode64 = IR::ValueOpcode::Count);
 	void DS_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode, bool returns_value);
 	void FLAT_LOAD(const Decoder::Instruction& inst);
 	void FLAT_STORE(const Decoder::Instruction& inst);

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

@@ -456,9 +456,12 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_FRACT_F32: return FloatUnary(inst, IR::ValueOpcode::FPFract32);
 		case O::V_FRACT_F64: return FloatUnary(inst, IR::ValueOpcode::FPFract64);
 		case O::V_TRUNC_F32: return FloatUnary(inst, IR::ValueOpcode::FPTrunc32);
+		case O::V_TRUNC_F64: return FloatUnary(inst, IR::ValueOpcode::FPTrunc64);
 		case O::V_CEIL_F32: return FloatUnary(inst, IR::ValueOpcode::FPCeil32);
+		case O::V_CEIL_F64: return FloatUnary(inst, IR::ValueOpcode::FPCeil64);
 		case O::V_RNDNE_F32: return FloatUnary(inst, IR::ValueOpcode::FPRoundEven32);
 		case O::V_FLOOR_F32: return FloatUnary(inst, IR::ValueOpcode::FPFloor32);
+		case O::V_FLOOR_F64: return FloatUnary(inst, IR::ValueOpcode::FPFloor64);
 		case O::V_EXP_F32: return FloatUnary(inst, IR::ValueOpcode::FPExp2);
 		case O::V_LOG_F32: return FloatUnary(inst, IR::ValueOpcode::FPLog2);
 		case O::V_RSQ_F32: return FloatUnary(inst, IR::ValueOpcode::FPRecipSqrt32);
@@ -471,13 +474,18 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_MUL_F32: return FloatBinary(inst, IR::ValueOpcode::FPMul32, false);
 		case O::V_ADD_F64: return FloatBinary(inst, IR::ValueOpcode::FPAdd64, false);
 		case O::V_MUL_F64: return FloatBinary(inst, IR::ValueOpcode::FPMul64, false);
+		case O::V_MIN_F64: return FloatBinary(inst, IR::ValueOpcode::FPMin64, false);
+		case O::V_MAX_F64: return FloatBinary(inst, IR::ValueOpcode::FPMax64, false);
 		case O::V_MIN_F32: return FloatBinary(inst, IR::ValueOpcode::FPMin32, false);
 		case O::V_MAX_F32: return FloatBinary(inst, IR::ValueOpcode::FPMax32, false);
 		case O::V_LDEXP_F32: return FloatBinary(inst, IR::ValueOpcode::FPLdexp, false);
-		case O::V_MAC_F32: return FloatTernary(inst, IR::ValueOpcode::FPFma32, true, true);
+		case O::V_MAC_F32: return FloatTernary(inst, IR::ValueOpcode::FPMad32, true, true);
 		case O::V_MADMK_F32:
 		case O::V_MADAK_F32:
-		case O::V_MAD_F32:
+		case O::V_MAD_F32: return FloatTernary(inst, IR::ValueOpcode::FPMad32, false, true);
+		case O::V_FMAC_F32: return FloatTernary(inst, IR::ValueOpcode::FPFma32, true, true);
+		case O::V_FMAMK_F32:
+		case O::V_FMAAK_F32:
 		case O::V_FMA_F32: return FloatTernary(inst, IR::ValueOpcode::FPFma32, false, true);
 		case O::V_FMA_F64: return FloatTernary(inst, IR::ValueOpcode::FPFma64, false, false);
 		case O::V_MIN3_F32: return FloatTernary(inst, IR::ValueOpcode::FPMinTri32, false, false);

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

@@ -199,6 +199,7 @@ enum class StageInputKind {
 	BaryCoordSmoothCentroid,
 	BaryCoordNoPerspective,
 	WorkgroupId,
+	NumWorkgroups,
 	LocalInvocationId,
 	LocalInvocationIndex,
 	GlobalInvocationId,
@@ -301,11 +302,12 @@ enum class DescriptorBindingKind : uint32_t {
 	FaultBuffer,
 	FlattenedSrt,
 	ShaderData,
+	SharedMemory,
 	Count,
 };
 
 static_assert(static_cast<uint32_t>(DescriptorBindingKind::Samplers) == 49u);
-static_assert(static_cast<uint32_t>(DescriptorBindingKind::Count) == 55u);
+static_assert(static_cast<uint32_t>(DescriptorBindingKind::Count) == 56u);
 
 struct PushData {
 	static constexpr uint32_t DwordCount = 32;
@@ -488,6 +490,8 @@ struct DescriptorSource {
 		uint32_t selector_stride = 0;
 		uint32_t selector_offset = 0;
 		uint32_t table_offset    = 0;
+		uint32_t table_stride    = 0;
+		uint32_t workgroup_axis  = UINT32_MAX;
 		uint32_t selector_shift  = 0;
 		Value    key_count;
 		Value                 selector_first;

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

@@ -224,15 +224,21 @@ ImageOpcodeInfo ImageOpcodeInfoOf(ValueOpcode opcode) {
 			return {ImageAccess::Write, ImageResourceClass::Storage, false};
 		case ValueOpcode::ImageAtomicCompareSwap32:
 		case ValueOpcode::ImageAtomicSwap32:
+		case ValueOpcode::ImageAtomicSwap64:
 		case ValueOpcode::ImageAtomicIAdd32:
+		case ValueOpcode::ImageAtomicIAdd64:
 		case ValueOpcode::ImageAtomicSMin32:
 		case ValueOpcode::ImageAtomicUMin32:
+		case ValueOpcode::ImageAtomicUMin64:
 		case ValueOpcode::ImageAtomicSMax32:
 		case ValueOpcode::ImageAtomicUMax32:
 		case ValueOpcode::ImageAtomicUMax64:
 		case ValueOpcode::ImageAtomicAnd32:
+		case ValueOpcode::ImageAtomicAnd64:
 		case ValueOpcode::ImageAtomicOr32:
+		case ValueOpcode::ImageAtomicOr64:
 		case ValueOpcode::ImageAtomicXor32:
+		case ValueOpcode::ImageAtomicXor64:
 		case ValueOpcode::ImageAtomicFMin32:
 		case ValueOpcode::ImageAtomicFMax32:
 			return {ImageAccess::Atomic, ImageResourceClass::Storage, false};

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

@@ -179,6 +179,7 @@ VALUE_OPCODE(FPCmpClass32, U1, F32, U32)
 VALUE_OPCODE(FPAdd32, F32, F32, F32)
 VALUE_OPCODE(FPSub32, F32, F32, F32)
 VALUE_OPCODE(FPFma32, F32, F32, F32, F32)
+VALUE_OPCODE(FPMad32, F32, F32, F32, F32)
 VALUE_OPCODE(FPMul32, F32, F32, F32)
 VALUE_OPCODE(FPAdd64, F64, F64, F64)
 VALUE_OPCODE(FPMul64, F64, F64, F64)
@@ -186,6 +187,8 @@ VALUE_OPCODE(FPFma64, F64, F64, F64, F64)
 VALUE_OPCODE(FPRecip64, F64, F64)
 VALUE_OPCODE(FPMin32, F32, F32, F32)
 VALUE_OPCODE(FPMax32, F32, F32, F32)
+VALUE_OPCODE(FPMin64, F64, F64, F64)
+VALUE_OPCODE(FPMax64, F64, F64, F64)
 VALUE_OPCODE(FPMinTri32, F32, F32, F32, F32)
 VALUE_OPCODE(FPMaxTri32, F32, F32, F32, F32)
 VALUE_OPCODE(FPMedTri32, F32, F32, F32, F32)
@@ -202,6 +205,9 @@ VALUE_OPCODE(FPRoundEven32, F32, F32)
 VALUE_OPCODE(FPFloor32, F32, F32)
 VALUE_OPCODE(FPCeil32, F32, F32)
 VALUE_OPCODE(FPTrunc32, F32, F32)
+VALUE_OPCODE(FPFloor64, F64, F64)
+VALUE_OPCODE(FPCeil64, F64, F64)
+VALUE_OPCODE(FPTrunc64, F64, F64)
 VALUE_OPCODE(FPFract32, F32, F32)
 VALUE_OPCODE(FPFract64, F64, F64)
 
@@ -274,8 +280,8 @@ VALUE_OPCODE(WriteSharedU32, Void, U32, U32, U1)
 VALUE_OPCODE(WriteSharedU32x2, Void, U32, U32, U32, U1)
 VALUE_OPCODE(WriteSharedU32x3, Void, U32, U32, U32, U32, U1)
 VALUE_OPCODE(WriteSharedU32x4, Void, U32, U32, U32, U32, U32, U1)
-VALUE_OPCODE(SharedAtomicFMin32, Void, U32, U32, U32, U1)
-VALUE_OPCODE(SharedAtomicFMax32, Void, U32, U32, U32, U1)
+VALUE_OPCODE(SharedAtomicFMin32, Void, U32, U32, U1)
+VALUE_OPCODE(SharedAtomicFMax32, Void, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicSwap32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicIAdd32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicIAdd64, Void, U32, U64, U1)
@@ -304,15 +310,21 @@ VALUE_OPCODE(ImageSampleRaw, U32x4, ImageResource, SamplerResource, ImageAddress
 VALUE_OPCODE(ImageGatherRaw, U32x4, ImageResource, SamplerResource, ImageAddress, U32, U32, U32, U32)
 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)
 VALUE_OPCODE(ImageAtomicIAdd32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicIAdd64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicSMin32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicUMin32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicUMin64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicSMax32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicUMax32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicUMax64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicAnd32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicAnd64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicOr32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicOr64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicXor32, U32, ImageResource, ImageAddress, U32, U1)
+VALUE_OPCODE(ImageAtomicXor64, U64, ImageResource, ImageAddress, U64, U1)
 VALUE_OPCODE(ImageAtomicFMin32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(ImageAtomicFMax32, U32, ImageResource, ImageAddress, U32, U1)
 VALUE_OPCODE(GetAttribute, U32, U32, U32)

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

@@ -56,14 +56,15 @@ void AddBinding(BindingLayout& layout, DescriptorBindingKind kind,
 
 } // namespace
 
-bool CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffers) {
+SharedMemoryResources CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffers) {
 	std::array<bool, ShaderInfo::MaxBuffers> live_buffers {};
-	bool uses_gds = false;
+	SharedMemoryResources shared;
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
 			const auto op = inst.GetOpcode();
 			if (BufferAccessOf(op) == BufferAccess::None &&
-			    SharedAccessOf(op) == SharedAccess::None) {
+			    SharedAccessOf(op) == SharedAccess::None &&
+			    AddressOpcodeInfoOf(op).access == AddressAccess::None) {
 				continue;
 			}
 			const auto index = inst.Flags<MemoryFlags>().index;
@@ -74,11 +75,13 @@ bool CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffe
 			if (memory.planning_only) {
 				continue;
 			}
+			shared.lds |= memory.kind == ResourceKind::FlatLocal;
 			if (SharedAccessOf(op) != SharedAccess::None) {
 				if (memory.kind != ResourceKind::Lds && memory.kind != ResourceKind::Gds) {
 					BindingFail("typed shader contains invalid shared-memory metadata");
 				}
-				uses_gds |= memory.kind == ResourceKind::Gds;
+				shared.gds |= memory.kind == ResourceKind::Gds;
+				shared.lds |= memory.kind == ResourceKind::Lds;
 			} else if (memory.kind == ResourceKind::Buffer || memory.kind == ResourceKind::ScalarBuffer) {
 				EXIT_IF(memory.resource >= program.info.buffers.size());
 				live_buffers.at(memory.resource) = true;
@@ -93,7 +96,7 @@ bool CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffe
 			buffers.push_back(i);
 		}
 	}
-	return uses_gds;
+	return shared;
 }
 
 bool UsesFlattenedSrt(const Program& program) {
@@ -108,7 +111,7 @@ bool UsesFlattenedSrt(const Program& program) {
 	       std::ranges::any_of(program.info.images, uses_mapping);
 }
 
-void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
+void AllocateBindings(Program& program, uint32_t push_data_start_dword, bool lds_storage) {
 	if (!program.shader_info_complete || program.binding_layout_complete) {
 		EXIT("shader binding layout failed: %s", !program.shader_info_complete
 		                                             ? "shader info is not ready"
@@ -116,7 +119,7 @@ void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
 	}
 	BindingLayout next;
 	std::vector<uint32_t> buffers;
-	const bool            uses_gds = CollectMemoryResources(program, buffers);
+	const auto shared = CollectMemoryResources(program, buffers);
 	CollectShaderData(program, next);
 	next.memory_offset_count       = static_cast<uint32_t>(buffers.size());
 	next.push_data_start_dword =
@@ -160,9 +163,12 @@ void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
 		}
 		AddBinding(next, DescriptorBindingKind::Samplers, std::move(resources));
 	}
-	if (uses_gds) {
+	if (shared.gds) {
 		AddBinding(next, DescriptorBindingKind::Gds);
 	}
+	if (shared.lds && lds_storage) {
+		AddBinding(next, DescriptorBindingKind::SharedMemory);
+	}
 	if (program.info.uses_dma) {
 		AddBinding(next, DescriptorBindingKind::BdaPagetable);
 		AddBinding(next, DescriptorBindingKind::FaultBuffer);

src/graphics/shader/recompiler/ir/passes/BindingLayout.h

@@ -5,10 +5,15 @@
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 
-void AllocateBindings(Program& program, uint32_t push_data_start_dword = 0);
+void AllocateBindings(Program& program, uint32_t push_data_start_dword = 0,
+                      bool lds_storage = false);
 
-// Collect live direct buffer indices and report whether the shader accesses GDS.
-bool CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffers);
+struct SharedMemoryResources {
+	bool lds = false;
+	bool gds = false;
+};
+
+SharedMemoryResources CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffers);
 bool UsesFlattenedSrt(const Program& program);
 
 const DescriptorBinding* FindBinding(const BindingLayout& layout, DescriptorBindingKind kind);

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

@@ -111,7 +111,8 @@ enum class SamplerClass : uint8_t { Float, Integer, PointInteger };
 
 SamplerClass ClassifySampler(const ImageResource& image) {
 	if (image.numeric_class == Prospero::TextureNumericClass::Sint ||
-	    image.conversion_format != Prospero::BufferFormat::kInvalid) {
+	    (image.numeric_class == Prospero::TextureNumericClass::Uint &&
+	     image.conversion_format != Prospero::BufferFormat::kInvalid)) {
 		return SamplerClass::PointInteger;
 	}
 	return image.numeric_class == Prospero::TextureNumericClass::Uint ? SamplerClass::Integer
@@ -284,12 +285,19 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		}
 		if (indirect.material_source == UINT32_MAX) {
 			uint32_t key_count = 0;
-			const bool evaluated = clean.Evaluate(indirect.key_count, key_count);
-			if (std::bit_cast<int32_t>(key_count) <= 0) key_count = 0;
-			if (table_value.dword_count != 2u || !evaluated ||
-			    key_count > MaxIndirectDescriptorProbes ||
-			    uint64_t {indirect.table_offset} + uint64_t {key_count} * descriptor_bytes >
-			        UINT32_MAX + 1ull) {
+			if (indirect.workgroup_axis != UINT32_MAX) {
+				if (indirect.workgroup_axis >= runtime.workgroup_counts.size()) return false;
+				key_count = runtime.workgroup_counts[indirect.workgroup_axis];
+				if (key_count == 0u) return false;
+			} else {
+				if (table_value.dword_count != 2u || !clean.Evaluate(indirect.key_count, key_count))
+					return false;
+				if (std::bit_cast<int32_t>(key_count) <= 0) key_count = 0;
+			}
+			if (indirect.table_stride == 0u || key_count > MaxIndirectDescriptorProbes ||
+			    (key_count != 0u && uint64_t {indirect.table_offset} +
+			         uint64_t {key_count - 1u} * indirect.table_stride + descriptor_bytes >
+			             UINT32_MAX + 1ull)) {
 				return false;
 			}
 			keys.resize(key_count);
@@ -375,7 +383,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		candidate.dword_count = dword_count;
 		if (sources.empty()) {
 			const auto table_offset =
-			    static_cast<uint32_t>(key * descriptor_bytes) + indirect.table_offset;
+			    static_cast<uint32_t>(key * indirect.table_stride) + indirect.table_offset;
 			if (!ReadScalarTable(table_base, table_size, table_offset, runtime,
 			                     std::span(candidate.dwords).first(dword_count))) {
 				return false;

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

@@ -77,6 +77,8 @@ Value CanonicalizeSampleAdjustDword3(Value value) {
 		value            = value.Resolve();
 		const auto* inst = value.TryInstruction();
 		if (inst == nullptr || inst->GetOpcode() != ValueOpcode::BitwiseOr32) {
+			if ((PossibleU32Bits(value) & ~SamplerDword3ReservedMask) == 0)
+				return Value(0u);
 			return value;
 		}
 		const auto left           = inst->Arg(0).Resolve();
@@ -834,7 +836,8 @@ private:
 				if (a.material_source != b.material_source || a.table_source != b.table_source ||
 				    a.selector_stride != b.selector_stride || a.selector_offset != b.selector_offset ||
 				    a.selector_shift != b.selector_shift ||
-				    a.table_offset != b.table_offset || a.sources != b.sources ||
+				    a.table_offset != b.table_offset || a.table_stride != b.table_stride ||
+				    a.workgroup_axis != b.workgroup_axis || a.sources != b.sources ||
 				    !EquivalentValue(m_program, a.key_count, b.key_count) ||
 				    a.selector_first.IsEmpty() != b.selector_first.IsEmpty() ||
 				    (!a.selector_first.IsEmpty() &&
@@ -1010,7 +1013,17 @@ private:
 		return false;
 	}
 
-	bool MatchTableOffset(Value value, Value& key, uint32_t& offset, uint32_t shift = 5u) const {
+	uint32_t WorkgroupAxis(Value key) const {
+		const auto* builtin = key.Resolve().TryInstruction();
+		uint32_t axis = UINT32_MAX;
+		if (m_program.stage != ShaderType::Compute || builtin == nullptr ||
+		    builtin->GetOpcode() != ValueOpcode::GetBuiltin || builtin->NumArgs() != 2u ||
+		    builtin->Arg(0) != Value(static_cast<uint32_t>(StageInputKind::WorkgroupId)) ||
+		    !ImmediateU32(builtin->Arg(1), axis) || axis >= 3u) return UINT32_MAX;
+		return axis;
+	}
+
+	bool MatchTableOffset(Value value, Value& key, uint32_t& offset, uint32_t& stride) const {
 		offset = 0;
 		for (;;) {
 			const auto* inst = value.Resolve().TryInstruction();
@@ -1019,10 +1032,17 @@ private:
 			}
 			uint32_t immediate;
 			if (inst->GetOpcode() == ValueOpcode::ShiftLeftLogical32 &&
-			    ImmediateU32(inst->Arg(1), immediate) && immediate == shift) {
+			    ImmediateU32(inst->Arg(1), immediate) && immediate < 32u) {
 				key = inst->Arg(0).Resolve();
+				stride = 1u << immediate;
 				return key.GetType() == Type::U32;
 			}
+			if (inst->GetOpcode() == ValueOpcode::IMul32) {
+				if (ImmediateU32(inst->Arg(0), stride)) key = inst->Arg(1).Resolve();
+				else if (ImmediateU32(inst->Arg(1), stride)) key = inst->Arg(0).Resolve();
+				else return false;
+				return stride != 0u && key.GetType() == Type::U32;
+			}
 			if (inst->GetOpcode() != ValueOpcode::IAdd32) {
 				return false;
 			}
@@ -1563,7 +1583,8 @@ private:
 
 	bool MatchDescriptorTable(Inst& handle, const DescriptorSource& descriptor,
 	                          IndirectDescriptorPlan& plan,
-	                          DescriptorSource& table_source, Value& key, uint32_t& table_offset) {
+	                          DescriptorSource& table_source, Value& key, uint32_t& table_offset,
+	                          uint32_t& table_stride) {
 		Inst* table_handle = nullptr;
 		for (uint32_t dword = 0; dword < handle.NumArgs(); ++dword) {
 			auto* read = descriptor.dwords[dword].Resolve().TryInstruction();
@@ -1579,6 +1600,7 @@ private:
 			auto* current_handle = read->Arg(0).Resolve().TryInstruction();
 			Value current_key;
 			uint32_t offset = 0;
+			uint32_t stride = 0;
 			if (current_handle == nullptr ||
 			    current_handle->GetOpcode() != (memory->kind == ResourceKind::ScalarAddress
 			                                      ? ValueOpcode::GetAddressResource
@@ -1586,7 +1608,7 @@ private:
 			    (memory->kind == ResourceKind::ScalarAddress && read->Parent() != handle.Parent()) ||
 			    (table_handle != nullptr &&
 			     !EquivalentValue(m_program, Value(table_handle), Value(current_handle))) ||
-			    !MatchTableOffset(read->Arg(1), current_key, offset, handle.NumArgs() == 4u ? 4u : 5u) ||
+			    !MatchTableOffset(read->Arg(1), current_key, offset, stride) ||
 			    memory->offset > UINT32_MAX - offset) {
 				return false;
 			}
@@ -1594,12 +1616,14 @@ private:
 			if (dword == 0u) {
 				key = current_key;
 				table_offset = offset;
-			} else if (!EquivalentValue(m_program, key, current_key) ||
+				table_stride = stride;
+			} else if (table_stride != stride || !EquivalentValue(m_program, key, current_key) ||
 			           static_cast<uint64_t>(table_offset) + dword * sizeof(uint32_t) != offset) {
 				return false;
 			}
 			table_handle = current_handle;
-			if (handle.NumArgs() == 8u ? !UsesOnlyImageDescriptors(*read) :
+			if (handle.NumArgs() == 8u
+			        ? WorkgroupAxis(key) == UINT32_MAX && !UsesOnlyImageDescriptors(*read) :
 			    std::ranges::any_of(read->Uses(), [](const Use& use) {
 				return use.user->GetOpcode() != ValueOpcode::GetBufferResource ||
 				       std::ranges::any_of(use.user->Uses(), [](const Use& consumer) {
@@ -1654,11 +1678,21 @@ private:
 		DescriptorSource table_source;
 		Value key;
 		uint32_t table_offset = 0;
-		if (!MatchDescriptorTable(handle, descriptor, plan, table_source, key, table_offset)) return false;
+		uint32_t table_stride = 0;
+		if (!MatchDescriptorTable(handle, descriptor, plan, table_source, key, table_offset,
+		                          table_stride)) return false;
 		DescriptorSource material_source;
 		DescriptorSource::IndirectDescriptor indirect;
 		indirect.table_offset = table_offset;
-		if (table_source.dword_count == 2u) {
+		indirect.table_stride = table_stride;
+		indirect.workgroup_axis = WorkgroupAxis(key);
+		if (indirect.workgroup_axis != UINT32_MAX) {
+			// The descriptor also supplies dimensions to shader arithmetic. Keep its reads;
+			// only the image handle is projected onto the bounded workgroup key.
+			plan.reads.fill(nullptr);
+		} else if (table_stride != 32u) {
+			return false;
+		} else if (table_source.dword_count == 2u) {
 			const auto* selector = key.Resolve().TryInstruction();
 			const bool bitscan = selector != nullptr && selector->GetOpcode() == ValueOpcode::FindILsb32 &&
 			    selector->NumArgs() == 1u && !m_shader_writes &&
@@ -1733,7 +1767,9 @@ private:
 		DescriptorSource table_source;
 		Value key;
 		uint32_t table_offset = 0;
-		if (!MatchDescriptorTable(handle, descriptor, plan, table_source, key, table_offset) ||
+		uint32_t table_stride = 0;
+		if (!MatchDescriptorTable(handle, descriptor, plan, table_source, key, table_offset,
+		                          table_stride) || table_stride != 16u ||
 		    table_source.dword_count != 4u || table_offset != 0u) return false;
 		const auto* lane_read = key.Resolve().TryInstruction();
 		if (lane_read == nullptr || lane_read->GetOpcode() != ValueOpcode::ReadFirstLane)
@@ -1763,6 +1799,7 @@ private:
 		    !EquivalentValue(m_program, selected->Arg(0), read->Arg(4))) return false;
 		DescriptorSource material_source;
 		DescriptorSource::IndirectDescriptor indirect;
+		indirect.table_stride = table_stride;
 		if (!MakeRuntimeTableSource(*read, material_source) ||
 		    !BoundedSelectedIndex(read->Arg(1), read->Arg(4), indirect)) return false;
 		indirect.material_source = InternSource(material_source);

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

@@ -327,7 +327,7 @@ void CollectBuiltinInputs(const Program& program, ShaderInfo& info) {
 
 void CollectOutputs(const Program& program, ShaderStageInputInfo input_info, ShaderInfo& info) {
 	const bool alpha_remap = program.stage == ShaderType::Pixel && input_info.pixel != nullptr &&
-	                         input_info.pixel->alpha_blend_source_remap;
+	                         input_info.pixel->alpha_blend_source != ShaderAlphaBlendSource::None;
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
 			if (inst.GetOpcode() != ValueOpcode::SetAttribute) {

src/graphics/shader/recompiler/ir/passes/SrtWalker.h

@@ -17,6 +17,7 @@ struct SrtRuntime {
 	SrtMemoryReader           read_memory                = nullptr;
 	void*                     userdata                   = nullptr;
 	SrtMemoryReader           read_specialization_memory = nullptr;
+	std::span<const uint32_t> workgroup_counts;
 };
 
 enum class RuntimeValueType { Any, Integer };

src/graphics/shader/shader.cpp

@@ -571,9 +571,11 @@ static void ShaderGetStaticInputInfoCS(const HW::ComputeShaderInfo& regs,
                                        const ShaderMappedData& data, ShaderComputeInputInfo& info) {
 	const bool dispatch_thread_dimensions = info.dispatch_thread_dimensions;
 	const auto host_subgroup_size         = info.host_subgroup_size;
+	const auto workgroup_counts           = std::to_array(info.workgroup_counts);
 	info                                  = {};
 	info.dispatch_thread_dimensions       = dispatch_thread_dimensions;
 	info.host_subgroup_size               = host_subgroup_size;
+	std::ranges::copy(workgroup_counts, info.workgroup_counts);
 	info.threads_num[0]                   = regs.cs_regs.num_thread_x;
 	info.threads_num[1]                   = regs.cs_regs.num_thread_y;
 	info.threads_num[2]                   = regs.cs_regs.num_thread_z;
@@ -660,7 +662,7 @@ void BuildStageStaticKey(const ShaderPixelInputInfo& info, std::vector<uint32_t>
 	key.push_back(static_cast<uint32_t>(info.ps_sample_mask_export_enable));
 	key.push_back(static_cast<uint32_t>(info.ps_early_z));
 	key.push_back(static_cast<uint32_t>(info.dual_source_blending));
-	key.push_back(static_cast<uint32_t>(info.alpha_blend_source_remap));
+	key.push_back(static_cast<uint32_t>(info.alpha_blend_source));
 	key.insert(key.end(), std::begin(info.target_output_mode), std::end(info.target_output_mode));
 	for (uint32_t base = 0; base < info.target_export_mapping.size(); base += 4u) {
 		uint32_t packed = 0;

src/graphics/shader/shader.h

@@ -148,14 +148,18 @@ struct ShaderVertexInputInfo {
 struct ShaderComputeInputInfo: ShaderWorkgroupInputInfo {
 	uint8_t            float_mode                 = 0xc0;
 	uint32_t           dispatch_threads_num[3]    = {0, 0, 0};
+	uint32_t           workgroup_counts[3]        = {0, 0, 0};
 	bool               group_id[3]                = {false, false, false};
 	bool               dispatch_thread_dimensions = false;
+	bool               lds_storage                = false;
 	int                thread_ids_num             = 0;
 	int                workgroup_register         = 0;
 	bool               tg_size_en                 = false;
 	ShaderStageRuntime stage;
 };
 
+enum class ShaderAlphaBlendSource : uint8_t { None, SourceAlpha, SourceAlphaOne, SourceAlphaZero };
+
 struct ShaderPixelInputInfo {
 	uint32_t                                       interpolator_settings[32]    = {0};
 	uint32_t                                       input_num                    = 0;
@@ -179,8 +183,8 @@ struct ShaderPixelInputInfo {
 	bool                                           ps_sample_mask_export_enable = false;
 	bool                                           ps_sample_shading            = false;
 	bool                                           dual_source_blending         = false;
-	// Export logical alpha through MRT1 for blending after channel swizzling.
-	bool                                           alpha_blend_source_remap     = false;
+	// Export logical alpha or per-channel source factors through MRT1 after channel swizzling.
+	ShaderAlphaBlendSource                         alpha_blend_source = ShaderAlphaBlendSource::None;
 	bool                                           ps_early_z                   = false;
 	bool                                           ps_execute_on_noop           = false;
 	ShaderStageRuntime                             stage;

src/kernel/eventQueue.cpp

@@ -564,7 +564,7 @@ int KYTY_SYSV_ABI KernelAddAmprEvent(KernelEqueue eq, int id, void* udata) {
 	if (eq != KERNEL_EQUEUE_INVALID) {
 		KernelEqueueEvent event {};
 		event.event.ident         = static_cast<uintptr_t>(id);
-		event.event.filter        = KERNEL_EVFILT_USER;
+		event.event.filter        = KERNEL_EVFILT_AMPR;
 		event.event.flags         = EV_ADD | EV_CLEAR;
 		event.event.fflags        = 0;
 		event.event.data          = 0;
@@ -592,7 +592,7 @@ int KYTY_SYSV_ABI KernelDeleteAmprEvent(KernelEqueue eq, int id) {
 	LOGF("\t AMPR event delete: eq = 0x%016" PRIx64 ", id = %d\n", static_cast<uint64_t>(eq), id);
 
 	if (eq != KERNEL_EQUEUE_INVALID) {
-		(void)KernelDeleteEvent(eq, static_cast<uintptr_t>(id), KERNEL_EVFILT_USER);
+		(void)KernelDeleteEvent(eq, static_cast<uintptr_t>(id), KERNEL_EVFILT_AMPR);
 	}
 
 	return OK;

src/kernel/eventQueue.h

@@ -18,6 +18,7 @@ constexpr int16_t KERNEL_EVFILT_FILE      = -4;
 constexpr int16_t KERNEL_EVFILT_GRAPHICS  = -14;
 constexpr int16_t KERNEL_EVFILT_VIDEO_OUT = -13;
 constexpr int16_t KERNEL_EVFILT_HRTIMER   = -15;
+constexpr int16_t KERNEL_EVFILT_AMPR      = -25;
 
 class KernelEqueuePrivate;
 struct KernelEqueueEvent;

src/libs/libAmpr.cpp

@@ -1193,8 +1193,9 @@ static int ExecuteCommand(const CommandBufferState::Command& entry) {
 	               std::get_if<CommandBufferState::KernelEventCommand>(&entry.data)) {
 		const auto& command = *payload;
 		const auto  eq      = static_cast<LibKernel::EventQueue::KernelEqueue>(command.eq);
-		auto        result  = LibKernel::EventQueue::KernelTriggerUserEvent(
-		    eq, command.id, reinterpret_cast<void*>(command.data));
+		auto        result  = LibKernel::EventQueue::KernelTriggerEvent(
+		    eq, static_cast<uintptr_t>(command.id), LibKernel::EventQueue::KERNEL_EVFILT_AMPR,
+		    reinterpret_cast<void*>(command.data));
 		if (result != OK) {
 			LOGF("\tAPR submit event failed: eq=0x%016" PRIx64 ", id=%" PRId32
 			     ", result=0x%08" PRIx32 "\n",

src/libs/libBluetoothHid.cpp

@@ -0,0 +1,42 @@
+#include "common/abi.h"
+#include "common/logging/log.h"
+#include "libs/errno.h"
+#include "libs/libs.h"
+#include "loader/symbolDatabase.h"
+
+namespace Libs {
+
+LIB_VERSION("BluetoothHid", 1, "BluetoothHid", 1, 1);
+
+namespace BluetoothHid {
+
+static int KYTY_SYSV_ABI BluetoothHidInit() {
+	PRINT_NAME();
+
+	return OK;
+}
+
+static int KYTY_SYSV_ABI BluetoothHidRegisterDevice() {
+	PRINT_NAME();
+
+	return OK;
+}
+
+// Callback signature is unverified; the stub never invokes it.
+static int KYTY_SYSV_ABI BluetoothHidRegisterCallback(uint64_t callback_func) {
+	PRINT_NAME();
+
+	LOGF("\t callback_func = 0x%016" PRIx64 "\n", callback_func);
+
+	return OK;
+}
+
+} // namespace BluetoothHid
+
+LIB_DEFINE(InitBluetoothHid_1) {
+	LIB_FUNC("tul3-GzejQc", BluetoothHid::BluetoothHidInit);
+	LIB_FUNC("4FUZ+c52d2k", BluetoothHid::BluetoothHidRegisterDevice);
+	LIB_FUNC("4Ypfo9RIwfM", BluetoothHid::BluetoothHidRegisterCallback);
+}
+
+} // namespace Libs

src/libs/libDeviceService.cpp

@@ -0,0 +1,36 @@
+#include "common/abi.h"
+#include "common/logging/log.h"
+#include "libs/errno.h"
+#include "libs/libs.h"
+#include "loader/symbolDatabase.h"
+
+namespace Libs {
+
+LIB_VERSION("DeviceService", 1, "Mbus", 1, 1);
+
+namespace DeviceService {
+
+static int KYTY_SYSV_ABI DeviceServiceInitialize() {
+	PRINT_NAME();
+
+	return OK;
+}
+
+static int KYTY_SYSV_ABI DeviceServiceQueryDeviceInfo(uint32_t device_handle, void* out_info) {
+	PRINT_NAME();
+
+	LOGF("\t device_handle = 0x%08" PRIx32 "\n"
+	     "\t out_info      = 0x%016" PRIx64 "\n",
+	     device_handle, reinterpret_cast<uint64_t>(out_info));
+
+	return OK;
+}
+
+} // namespace DeviceService
+
+LIB_DEFINE(InitDeviceService_1) {
+	LIB_FUNC("84fDxStrG44", DeviceService::DeviceServiceInitialize);
+	LIB_FUNC("UNMEa+5lrUA", DeviceService::DeviceServiceQueryDeviceInfo);
+}
+
+} // namespace Libs

src/libs/libFont.cpp

@@ -78,6 +78,9 @@ constexpr int SCE_FONT_TEXT_PARSER_RESULT_TERMINATE = 0;
 constexpr int SCE_FONT_TEXT_PARSER_RESULT_ERROR     = -1;
 constexpr int SCE_FONT_WRITING_FORM_HORIZONTAL      = 0x10;
 constexpr int FONT_BITMAP_MAX_DIM                   = 128;
+constexpr int FONT_ERROR_INVALID_PARAMETER          = static_cast<int>(0x80460002u);
+constexpr int FONT_ERROR_INVALID_FONT_HANDLE        = static_cast<int>(0x80460005u);
+constexpr int FONT_ERROR_NO_SUPPORT_CODE            = static_cast<int>(0x80460041u);
 
 struct FontMemory {
 	uint16_t                   type;
@@ -118,6 +121,14 @@ struct FontGlyphMetrics {
 	} vertical;
 };
 
+struct FontKerning {
+	float offset_x;
+	float offset_y;
+	float position_x;
+	float position_y;
+};
+static_assert(sizeof(FontKerning) == 16);
+
 struct FontRenderSurface {
 	void*   buffer;
 	int32_t width_byte;
@@ -801,15 +812,6 @@ static bool ensure_stb_font(FontState* font) {
 	return font->has_ttf;
 }
 
-static float stb_font_pixel_height(const FontState* font) {
-	const float scale = (font != nullptr && font->scale_h > 1.0f ? font->scale_h : 16.0f);
-	return static_cast<float>(std::clamp(static_cast<int>(scale + 0.5f), 8, FONT_BITMAP_MAX_DIM));
-}
-
-static float stb_font_scale(FontState* font) {
-	return stbtt_ScaleForPixelHeight(&font->font_info, stb_font_pixel_height(font));
-}
-
 static uint32_t stb_supported_codepoint(FontState* font, uint32_t code) {
 	if (font == nullptr || !ensure_stb_font(font)) {
 		return static_cast<uint32_t>('?');
@@ -825,31 +827,26 @@ static bool get_stb_metrics(FontState* font, uint32_t code, FontGlyphMetrics* me
 		return false;
 	}
 
-	code              = stb_supported_codepoint(font, code);
-	const float scale = stb_font_scale(font);
+	code                = stb_supported_codepoint(font, code);
+	const float scale_x = stbtt_ScaleForMappingEmToPixels(&font->font_info, font->scale_w);
+	const float scale_y = stbtt_ScaleForMappingEmToPixels(&font->font_info, font->scale_h);
 
 	int advance = 0;
-	int lsb     = 0;
 	int x0      = 0;
 	int y0      = 0;
 	int x1      = 0;
 	int y1      = 0;
-	stbtt_GetCodepointHMetrics(&font->font_info, static_cast<int>(code), &advance, &lsb);
-	stbtt_GetCodepointBitmapBox(&font->font_info, static_cast<int>(code), scale, scale, &x0, &y0,
-	                            &x1, &y1);
+	stbtt_GetCodepointHMetrics(&font->font_info, static_cast<int>(code), &advance, nullptr);
+	stbtt_GetCodepointBox(&font->font_info, static_cast<int>(code), &x0, &y0, &x1, &y1);
 
-	const float width     = static_cast<float>(std::max(x1 - x0, 0));
-	const float height    = static_cast<float>(std::max(y1 - y0, 0));
-	const float advance_f = std::max(static_cast<float>(advance) * scale, width);
-
-	metrics->width                = width;
-	metrics->height               = height;
-	metrics->horizontal.bearing_x = static_cast<float>(x0);
-	metrics->horizontal.bearing_y = static_cast<float>(-y0);
-	metrics->horizontal.advance   = advance_f;
+	metrics->width                = static_cast<float>(x1 - x0) * scale_x;
+	metrics->height               = static_cast<float>(y1 - y0) * scale_y;
+	metrics->horizontal.bearing_x = static_cast<float>(x0) * scale_x;
+	metrics->horizontal.bearing_y = static_cast<float>(y1) * scale_y;
+	metrics->horizontal.advance   = static_cast<float>(advance) * scale_x;
 	metrics->vertical.bearing_x   = 0.0f;
 	metrics->vertical.bearing_y   = 0.0f;
-	metrics->vertical.advance     = stb_font_pixel_height(font);
+	metrics->vertical.advance     = font->scale_h;
 
 	return true;
 }
@@ -860,14 +857,15 @@ static bool init_stb_image(std::array<uint8_t, FONT_BITMAP_MAX_DIM * FONT_BITMAP
 		return false;
 	}
 
-	code              = stb_supported_codepoint(font, code);
-	const float scale = stb_font_scale(font);
+	code                = stb_supported_codepoint(font, code);
+	const float scale_x = stbtt_ScaleForMappingEmToPixels(&font->font_info, font->scale_w);
+	const float scale_y = stbtt_ScaleForMappingEmToPixels(&font->font_info, font->scale_h);
 
 	int x0 = 0;
 	int y0 = 0;
 	int x1 = 0;
 	int y1 = 0;
-	stbtt_GetCodepointBitmapBox(&font->font_info, static_cast<int>(code), scale, scale, &x0, &y0,
+	stbtt_GetCodepointBitmapBox(&font->font_info, static_cast<int>(code), scale_x, scale_y, &x0, &y0,
 	                            &x1, &y1);
 
 	image->fill(0);
@@ -884,7 +882,7 @@ static bool init_stb_image(std::array<uint8_t, FONT_BITMAP_MAX_DIM * FONT_BITMAP
 	}
 
 	stbtt_MakeCodepointBitmap(&font->font_info, image->data(), static_cast<int>(width),
-	                          static_cast<int>(height), FONT_BITMAP_MAX_DIM, scale, scale,
+	                          static_cast<int>(height), FONT_BITMAP_MAX_DIM, scale_x, scale_y,
 	                          static_cast<int>(code));
 
 	return true;
@@ -1425,7 +1423,7 @@ int KYTY_SYSV_ABI FontGetHorizontalLayout(FontHandle font_handle, FontHorizontal
 
 	auto* font = static_cast<FontState*>(font_handle);
 	if (ensure_stb_font(font)) {
-		const float scale    = stb_font_scale(font);
+		const float scale = stbtt_ScaleForMappingEmToPixels(&font->font_info, font->scale_h);
 		int         ascent   = 0;
 		int         descent  = 0;
 		int         line_gap = 0;
@@ -2081,6 +2079,40 @@ int KYTY_SYSV_ABI FontGetRenderCharGlyphMetrics(FontHandle font_handle, uint32_t
 	return OK;
 }
 
+int KYTY_SYSV_ABI FontGetKerning(FontHandle font_handle, uint32_t pre_code, uint32_t code,
+                                FontKerning* kerning) {
+	PRINT_NAME();
+	if (kerning == nullptr) {
+		return FONT_ERROR_INVALID_PARAMETER;
+	}
+	*kerning = {};
+	auto* font = static_cast<FontState*>(font_handle);
+	if (font == nullptr) {
+		return FONT_ERROR_INVALID_FONT_HANDLE;
+	}
+	if (pre_code == 0 || !ensure_stb_font(font)) {
+		return OK;
+	}
+	const auto glyph_index = [&](uint32_t value) {
+		if ((value & 0x80000000u) != 0) {
+			return static_cast<int>(value & 0xffffu);
+		}
+		return stbtt_FindGlyphIndex(&font->font_info, static_cast<int>(value));
+	};
+	const int left  = glyph_index(pre_code);
+	const int right = glyph_index(code);
+	if (left <= 0 || left >= font->font_info.numGlyphs ||
+	    right <= 0 || right >= font->font_info.numGlyphs) {
+		return FONT_ERROR_NO_SUPPORT_CODE;
+	}
+	kerning->offset_x = static_cast<float>(stbtt_GetGlyphKernAdvance(&font->font_info, left, right)) *
+	                    stbtt_ScaleForMappingEmToPixels(&font->font_info, font->scale_w);
+	LOGF("\t handle = 0x%016" PRIx64 ", pre_code = 0x%08" PRIx32 ", code = 0x%08" PRIx32
+	     ", offset_x = %f\n", reinterpret_cast<uint64_t>(font_handle), pre_code, code,
+	     static_cast<double>(kerning->offset_x));
+	return OK;
+}
+
 int KYTY_SYSV_ABI FontGetCharGlyphMetrics(FontHandle font_handle, uint32_t code,
                                           FontGlyphMetrics* metrics) {
 	PRINT_NAME();
@@ -2253,6 +2285,7 @@ LIB_DEFINE(InitFont_1) {
 	LIB_FUNC("FXP359ygujs", Font::FontDestroyLibrary);
 	LIB_FUNC("C-4Qw5Srlyw", Font::FontGenerateCharGlyph);
 	LIB_FUNC("L97d+3OgMlE", Font::FontGetCharGlyphMetrics);
+	LIB_FUNC("sDuhHGNhHvE", Font::FontGetKerning);
 	LIB_FUNC("IQtleGLL5pQ", Font::FontGetRenderCharGlyphMetrics);
 	LIB_FUNC("8-zmgsxkBek", Font::FontGlyphDefineAttribute);
 	LIB_FUNC("whrS4oksXc4", Font::FontMemoryInit);

src/libs/libUsbd.cpp

@@ -0,0 +1,44 @@
+#include "common/abi.h"
+#include "common/threads.h"
+#include "libs/errno.h"
+#include "libs/libs.h"
+#include "loader/symbolDatabase.h"
+
+namespace Libs {
+
+LIB_VERSION("Usbd", 1, "Usbd", 1, 1);
+
+namespace Usbd {
+
+constexpr uint32_t POLL_INTERVAL_MS = 16;
+
+static int KYTY_SYSV_ABI UsbdInit() {
+	PRINT_NAME();
+
+	return OK;
+}
+
+static int KYTY_SYSV_ABI UsbdExit() {
+	PRINT_NAME();
+
+	return OK;
+}
+
+static int KYTY_SYSV_ABI UsbdHandleEventsTimeout() {
+	PRINT_NAME();
+
+	// Avoid a busy loop while USB events are unimplemented.
+	Common::Thread::SleepMicro(POLL_INTERVAL_MS * 1000);
+
+	return OK;
+}
+
+} // namespace Usbd
+
+LIB_DEFINE(InitUsbd_1) {
+	LIB_FUNC("TOhg7P6kTH4", Usbd::UsbdInit);
+	LIB_FUNC("Fq6+0Fm55xU", Usbd::UsbdExit);
+	LIB_FUNC("+wU6CGuZcWk", Usbd::UsbdHandleEventsTimeout);
+}
+
+} // namespace Libs

src/libs/libs.cpp

@@ -79,8 +79,10 @@ LIB_DEFINE(InitVideoOutVrrStatus_1);
 
 LIB_DEFINE(InitAppContent_1);
 LIB_DEFINE(InitAudio_1);
+LIB_DEFINE(InitBluetoothHid_1);
 LIB_DEFINE(InitConvertKeycode_1);
 LIB_DEFINE(InitDbgAddressSanitizer_1);
+LIB_DEFINE(InitDeviceService_1);
 LIB_DEFINE(InitDialog_1);
 LIB_DEFINE(InitFont_1);
 LIB_DEFINE(InitFontFt_1);
@@ -99,6 +101,7 @@ LIB_DEFINE(InitShare_1);
 LIB_DEFINE(InitSysmodule_1);
 LIB_DEFINE(InitSystemService_1);
 LIB_DEFINE(InitTextToSpeech2_1);
+LIB_DEFINE(InitUsbd_1);
 LIB_DEFINE(InitUserService_1);
 LIB_DEFINE(InitWebBrowserDialog_1);
 
@@ -120,6 +123,9 @@ void InitAll(Loader::SymbolDatabase* s) {
 	InitFont_1(s);
 	InitFontFt_1(s);
 	InitAgcDriver_1(s);
+	InitBluetoothHid_1(s);
+	InitDeviceService_1(s);
+	InitUsbd_1(s);
 	InitHmd2_1(s);
 	InitLibKernel_1(s);
 	LibMouse::InitMouse_1(s);

src/loader/runtimeLinker.cpp

@@ -13,6 +13,7 @@
 #include "common/threads.h"
 #include "common/virtualMemory.h"
 #include "graphics/host_gpu/pageManager.h"
+#include "kernel/fileSystem.h"
 #include "kernel/memory.h"
 #include "kernel/pthread.h"
 #include "loader/elf.h"
@@ -1264,7 +1265,9 @@ void RuntimeLinker::Execute(const std::filesystem::path& game_patch) {
 	}
 #endif
 
-	PreloadAdjacentPrograms();
+	// The runtime automatically loads libc; other PRXs are requested by the application.
+	const auto libc_path = Libs::LibKernel::FileSystem::GetRealFilename("/app0/sce_module/libc.prx");
+	auto* libc = Common::File::IsFileExisting(libc_path) ? LoadProgram(libc_path) : nullptr;
 	RelocateAll();
 
 	if (!game_patch.empty()) {
@@ -1273,7 +1276,9 @@ void RuntimeLinker::Execute(const std::filesystem::path& game_patch) {
 			EXIT("Failed to apply game cheat\n");
 		}
 	}
-	StartAllModules();
+	if (libc != nullptr && libc->dynamic_info->init_vaddr != 0) {
+		StartModule(libc, 0, nullptr, nullptr);
+	}
 
 	LOGF_COLOR(Log::Color::BrightYellow, "---\n--- Execute: %s\n---\n", "Main");
 
@@ -1526,84 +1531,6 @@ void RuntimeLinker::StackTrace(uint64_t frame_ptr, uint64_t stack_ptr) {
 	}
 }
 
-static std::string GetProgramModuleName(const Program* program) {
-	EXIT_IF(program == nullptr);
-
-	if (program->dynamic_info != nullptr && program->dynamic_info->so_name != nullptr &&
-	    program->dynamic_info->so_name[0] != '\0') {
-		return std::string(program->dynamic_info->so_name);
-	}
-
-	return Common::FilenameWithoutDirectory(Common::PathToGenericString(program->file_name));
-}
-
-static bool ModuleStartDependenciesSatisfied(const Program*               program,
-                                             const std::vector<Program*>& programs,
-                                             const std::vector<Program*>& started) {
-	EXIT_IF(program == nullptr);
-	EXIT_IF(program->dynamic_info == nullptr);
-
-	for (const auto* needed: program->dynamic_info->needed) {
-		if (needed == nullptr || needed[0] == '\0') {
-			continue;
-		}
-
-		const auto needed_name = std::string(needed);
-
-		for (auto* dependency: programs) {
-			if (dependency == nullptr || dependency == program || dependency->elf == nullptr ||
-			    !dependency->elf->IsShared()) {
-				continue;
-			}
-
-			const auto dependency_name = GetProgramModuleName(dependency);
-			if (Common::EqualNoCase(dependency_name, needed_name) ||
-			    Common::EqualNoCase(Common::FilenameWithoutDirectory(
-			                            Common::PathToGenericString(dependency->file_name)),
-			                        needed_name)) {
-				if (std::find(started.begin(), started.end(), dependency) == started.end()) {
-					return false;
-				}
-				break;
-			}
-		}
-	}
-
-	return true;
-}
-
-void RuntimeLinker::StartAllModules() {
-	Common::LockGuard lock(m_mutex);
-
-	std::vector<Program*> started;
-
-	for (;;) {
-		bool progressed = false;
-
-		for (auto* p: m_programs) {
-			if (p->elf->IsShared() && p->dynamic_info->init_vaddr != 0 &&
-			    std::find(started.begin(), started.end(), p) == started.end() &&
-			    ModuleStartDependenciesSatisfied(p, m_programs, started)) {
-				StartModule(p, 0, nullptr, nullptr);
-				started.push_back(p);
-				progressed = true;
-			}
-		}
-
-		if (!progressed) {
-			break;
-		}
-	}
-
-	for (auto* p: m_programs) {
-		if (p->elf->IsShared() && p->dynamic_info->init_vaddr != 0 &&
-		    std::find(started.begin(), started.end(), p) == started.end()) {
-			StartModule(p, 0, nullptr, nullptr);
-			started.push_back(p);
-		}
-	}
-}
-
 void RuntimeLinker::StopAllModules() {
 	Common::LockGuard lock(m_mutex);
 
@@ -1614,75 +1541,6 @@ void RuntimeLinker::StopAllModules() {
 	}
 }
 
-static bool IsAdjacentModuleFile(const std::string& name) {
-	auto lower = Common::ToLower(name);
-	return lower.ends_with(".prx") || lower.ends_with(".sprx");
-}
-
-static bool SkipAdjacentModuleFile(const std::string& name) {
-	auto lower = Common::ToLower(name);
-	return lower == "eboot.bin" || lower == "libkernel.prx" || lower == "libkernel_sys.prx";
-}
-
-void RuntimeLinker::PreloadAdjacentPrograms() {
-	if (m_programs.empty()) {
-		return;
-	}
-
-	std::vector<std::filesystem::path> module_paths;
-
-	auto is_loaded = [this](const std::filesystem::path& path) {
-		auto fixed_path = Common::FixFilenameSlash(Common::PathToGenericString(path));
-		for (auto* program: m_programs) {
-			if (Common::EqualNoCase(
-			        Common::FixFilenameSlash(Common::PathToGenericString(program->file_name)),
-			        fixed_path)) {
-				return true;
-			}
-		}
-		return false;
-	};
-
-	auto add_path = [&module_paths, &is_loaded](const std::filesystem::path& path) {
-		if (is_loaded(path)) {
-			return;
-		}
-		for (const auto& p: module_paths) {
-			if (Common::EqualNoCase(Common::PathToGenericString(p),
-			                        Common::PathToGenericString(path))) {
-				return;
-			}
-		}
-		module_paths.push_back(path);
-	};
-
-	auto add_dir = [&add_path](const std::filesystem::path& dir) {
-		if (!Common::File::IsDirectoryExisting(dir)) {
-			return;
-		}
-		for (const auto& entry: Common::File::GetDirEntries(dir)) {
-			if (entry.is_file && IsAdjacentModuleFile(entry.name) &&
-			    !SkipAdjacentModuleFile(entry.name)) {
-				add_path(dir / entry.name);
-			}
-		}
-	};
-
-	auto root = m_programs.at(0)->file_name.parent_path();
-	if (root.empty()) {
-		return;
-	}
-
-	add_dir(root);
-	add_dir(root / "sce_module");
-	add_dir(root / "sce_modules");
-
-	for (const auto& path: module_paths) {
-		auto* program                        = LoadProgram(path);
-		program->fail_if_global_not_resolved = false;
-	}
-}
-
 int RuntimeLinker::StartModule(Program* program, size_t args, const void* argp,
                                module_func_t func) {
 	EXIT_IF(program == nullptr);
@@ -2099,12 +1957,6 @@ void RuntimeLinker::ParseProgramDynamicInfo(Program* program) {
 	EXIT_NOT_IMPLEMENTED(program->dynamic_info->debug != 0);
 	EXIT_NOT_IMPLEMENTED(program->dynamic_info->textrel != 0);
 
-	std::vector<uint64_t> needed;
-	GetDynValues(elf, &needed, DT_NEEDED);
-	for (auto need: needed) {
-		program->dynamic_info->needed.push_back(program->dynamic_info->str_table + need);
-	}
-
 	uint64_t so_name = 0;
 	GetDynValue(elf, &so_name, DT_SONAME);
 	program->dynamic_info->so_name = program->dynamic_info->str_table + so_name;

src/loader/runtimeLinker.h

@@ -102,11 +102,10 @@ struct DynamicInfo {
 
 	const char* so_name = nullptr;
 
-	std::vector<const char*> needed;
-	std::vector<ModuleId>    export_modules;
-	std::vector<ModuleId>    import_modules;
-	std::vector<LibraryId>   export_libs;
-	std::vector<LibraryId>   import_libs;
+	std::vector<ModuleId>  export_modules;
+	std::vector<ModuleId>  import_modules;
+	std::vector<LibraryId> export_libs;
+	std::vector<LibraryId> import_libs;
 };
 
 struct Program {
@@ -157,7 +156,6 @@ public:
 	void  Execute(const std::filesystem::path& game_patch = {});
 	int   StartModule(Program* program, size_t args, const void* argp, module_func_t func);
 	int   StopModule(Program* program, size_t args, const void* argp, module_func_t func);
-	void  StartAllModules();
 	void  StopAllModules();
 	void  DeleteTlss(int thread_id);
 	void  SetApplicationHeapApi(void* const api[10]);
@@ -186,7 +184,6 @@ private:
 	static void Relocate(Program* program);
 	static void DeleteProgram(Program* program);
 	static void SetupTlsHandler(Program* program);
-	void        PreloadAdjacentPrograms();
 
 	static const ModuleId*  FindModule(const Program& program, const std::string& id);
 	static const LibraryId* FindLibrary(const Program& program, const std::string& id);

tests/KernelFileSystemTests.cpp

@@ -10,6 +10,7 @@
 #include "common/threads.h"
 #include "common/stringUtils.h"
 #include "graphics/presentation/window/windowInternal.h"
+#include "kernel/eventQueue.h"
 #include "kernel/fileSystem.h"
 #include "libs/errno.h"
 #include "libs/network.h"
@@ -640,6 +641,8 @@ void CheckAmprOrdering(Loader::SymbolDatabase &symbols, uint32_t file_id) {
   using WriteAddress = int (KYTY_SYSV_ABI *)(void *, volatile uint64_t *, uint64_t);
   using ReadFile = int (KYTY_SYSV_ABI *)(void *, uint64_t, uint64_t, uint32_t,
                                        void *, uint64_t, uint64_t);
+  using WriteKernelEvent = int (KYTY_SYSV_ABI *)(void *, uint64_t, uint64_t, uint64_t,
+                                               uint64_t, uint64_t);
   struct Result { int32_t result; uint32_t error_offset; };
   using SubmitApr = int (KYTY_SYSV_ABI *)(void *, uint32_t, Result *, uint32_t *);
   using SubmitAmm = int (KYTY_SYSV_ABI *)(void *, uint32_t, uint32_t, uint32_t *);
@@ -653,6 +656,7 @@ void CheckAmprOrdering(Loader::SymbolDatabase &symbols, uint32_t file_id) {
   const auto wait_address = reinterpret_cast<WaitAddress>(find("DLfoNxTFNVk"));
   const auto write_address = reinterpret_cast<WriteAddress>(find("sJXyWHjP-F8"));
   const auto read_file = reinterpret_cast<ReadFile>(find("mQ16-QdKv7k"));
+  const auto write_event = reinterpret_cast<WriteKernelEvent>(find("H896Pt-yB4I"));
   const auto submit_apr = reinterpret_cast<SubmitApr>(find("ASoW5WE-UPo"));
   const auto submit_amm = reinterpret_cast<SubmitAmm>(find("NnKhlMJtIsI"));
   const auto wait_apr = reinterpret_cast<WaitSubmission>(find("rqwFKI4PAiM"));
@@ -723,6 +727,42 @@ void CheckAmprOrdering(Loader::SymbolDatabase &symbols, uint32_t file_id) {
             "submission wait observes the completed result");
     }
   }
+  reset(1);
+  namespace EventQueue = Libs::LibKernel::EventQueue;
+  EventQueue::KernelEqueue queue = EventQueue::KERNEL_EQUEUE_INVALID;
+  uint64_t ampr_user_data = 1, user_data = 2;
+  Check(EventQueue::KernelCreateEqueue(&queue, "apr-completion") == OK &&
+            EventQueue::KernelAddAmprEvent(queue, 42, &ampr_user_data) == OK &&
+            EventQueue::KernelAddUserEventEdge(queue, 42) == OK,
+        "AMPR and user events can share an identifier");
+  std::array<char, 3> output {};
+  Result event_result {1234, 5678};
+  uint32_t event_submission = 0;
+  auto *reader = buffers[0].header.data();
+  Check(read_file(reader, reinterpret_cast<uint64_t>(&buffers[0].header[3]),
+                  reinterpret_cast<uint64_t>(&buffers[0].header[4]), file_id,
+                  output.data(), output.size(), 0) == OK &&
+            write_event(reader, queue, 42, 0x123456789abc, 0, 0) == OK &&
+            submit_apr(reader, 3, &event_result, &event_submission) == OK &&
+            wait_apr(event_submission) == OK && event_result.result == OK,
+        "APR read submits its completion event");
+  std::array<EventQueue::KernelEvent, 2> events {};
+  int event_count = 0;
+  const Libs::LibKernel::KernelUseconds poll = 0;
+  Check(EventQueue::KernelWaitEqueue(queue, events.data(), 2, &event_count, &poll) == OK &&
+            event_count == 1 && events[0].ident == 42 && events[0].filter == -25 &&
+            events[0].data == 0x123456789abc && events[0].udata == &ampr_user_data &&
+            std::memcmp(output.data(), "APR", 3) == 0,
+        "APR completion reports the AMPR filter, payload and registration user data");
+  Check(EventQueue::KernelDeleteAmprEvent(queue, 42) == OK &&
+            EventQueue::KernelTriggerEvent(queue, 42, -25, nullptr) ==
+                Libs::LibKernel::KERNEL_ERROR_ENOENT &&
+            EventQueue::KernelTriggerUserEvent(queue, 42, &user_data) == OK &&
+            EventQueue::KernelWaitEqueue(queue, events.data(), 2, &event_count, &poll) == OK &&
+            event_count == 1 && events[0].filter == EventQueue::KERNEL_EVFILT_USER &&
+            events[0].udata == &user_data,
+        "deleting AMPR leaves the user event with the same identifier intact");
+  Check(EventQueue::KernelDeleteEqueue(queue) == OK, "delete APR completion queue");
   reset(0);
   uint64_t cpu_fence = 0, amm_done = 0, lower_done = 0;
   std::array<uint32_t, 2> ids {};

tests/ResourceMaterializationTests.cpp

@@ -126,6 +126,11 @@ Libs::Graphics::ShaderRecompiler::IR::Program MixedSamplerProgram() {
   const auto image1 = AddSource(8);
   const auto sampler0 = AddSource(4);
   const auto sampler1 = AddSource(4);
+  program.descriptor_sources[image1].dwords[0] = Value(1u);
+  program.descriptor_sources[image1].dwords[1] = Value(static_cast<uint32_t>(
+      Libs::Graphics::Prospero::BufferFormat::k11_11_10UInt) << 20u);
+  program.descriptor_sources[image1].dwords[3] = Value(static_cast<uint32_t>(
+      Libs::Graphics::Prospero::ImageType::kColor2D) << 28u);
   for (uint32_t index = 0; index < 2; ++index) {
     auto &value = block.AppendNewInst(
         ValueOpcode::GetUserData, {Value(static_cast<ScalarReg>(index))});
@@ -142,9 +147,7 @@ Libs::Graphics::ShaderRecompiler::IR::Program MixedSamplerProgram() {
        .resource_class = ImageResourceClass::Sampled,
        .numeric_class = Libs::Graphics::Prospero::TextureNumericClass::Float,
        .dimension =
-           Libs::Graphics::ShaderRecompiler::Decoder::ImageDimension::Dim2D,
-       .conversion_format =
-           Libs::Graphics::Prospero::BufferFormat::k8_8_8_8UNorm});
+           Libs::Graphics::ShaderRecompiler::Decoder::ImageDimension::Dim2D});
   program.info.samplers.push_back({.source = sampler0});
   program.info.samplers.push_back({.source = sampler1});
   program.info.sampled_pairs.push_back({.image = 0, .sampler = 0});

tests/ResourceTrackingTests.cpp

@@ -1890,6 +1890,13 @@ void TestSampleAdjustSamplerScratch() {
   fixture.Emit(ValueOpcode::ImageSampleRaw,
                {image, sampler, fixture.ImageAddress()},
                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()},
+               fixture.AddMemory(memory, 0x200));
   fixture.PlanAndTrack();
 
   const auto source = fixture.program.info.samplers[0].source;
@@ -1905,6 +1912,12 @@ void TestSampleAdjustSamplerScratch() {
   Check(SrtWalker(fixture.program, runtime).EvaluateDescriptor(source, descriptor) &&
             descriptor.dwords[3] == 0x80000abcu,
         "SampleAdjust canonicalization lost sampler border fields");
+  const auto shift_source = fixture.program.info.samplers[
+      shift_sampler.Instruction()->Flags<uint32_t>()].source;
+  Check(SrtWalker(fixture.program, runtime).EvaluateDescriptor(shift_source, descriptor) &&
+            std::ranges::equal(std::span(descriptor.dwords).first(4),
+                               std::array{0x36u, 0xfff000u, 0x02500000u, 0u}),
+        "SampleAdjust shift-only scratch was not reduced to its zero border payload");
 
   const auto CheckRejected = [](uint32_t flags, uint32_t shift,
                                 const char *message) {
@@ -3385,7 +3398,8 @@ void TestImageBindingAbi() {
             static_cast<uint32_t>(DescriptorBindingKind::FaultBuffer) == 52u &&
             static_cast<uint32_t>(DescriptorBindingKind::FlattenedSrt) == 53u &&
             static_cast<uint32_t>(DescriptorBindingKind::ShaderData) == 54u &&
-            static_cast<uint32_t>(DescriptorBindingKind::Count) == 55u,
+            static_cast<uint32_t>(DescriptorBindingKind::SharedMemory) == 55u &&
+            static_cast<uint32_t>(DescriptorBindingKind::Count) == 56u,
         "native descriptor binding anchors changed");
 
   const std::array sampled_dimensions{

tests/shaderCfgTests.cpp

@@ -3600,6 +3600,53 @@ void TestNewShaderRecompilerCapturedFractF64() {
   }
 }
 
+void TestNewShaderRecompilerF64AluEncodings() {
+  using namespace ShaderRecompiler::Decoder;
+  for (const bool maximum : {false, true}) {
+    const std::array<uint32_t, 2> captured = maximum
+        ? std::array<uint32_t, 2>{0xd5670006u, 0x00020480u}
+        : std::array<uint32_t, 2>{0xd5660006u, 0x00020cf2u};
+    Instruction decoded;
+    DecodeInstruction(captured, 0, decoded);
+    Check(decoded.family == Family::VOP3 &&
+              decoded.opcode == (maximum ? Opcode::V_MAX_F64 : Opcode::V_MIN_F64) &&
+              decoded.opcode_id == (maximum ? 0x167u : 0x166u) && decoded.word_count == 2 &&
+              decoded.dst.kind == OperandKind::Vgpr && decoded.dst.reg == 6 &&
+              decoded.src0.kind == (maximum ? OperandKind::IntegerInlineConstant
+                                            : OperandKind::FloatInlineConstant) &&
+              decoded.src0.value == (maximum ? 0u : 0x3f800000u) &&
+              decoded.src1.kind == OperandKind::Vgpr && decoded.src1.reg == (maximum ? 2u : 6u) &&
+              !decoded.src0.absolute && !decoded.src0.negate &&
+              !decoded.src1.absolute && !decoded.src1.negate &&
+              !decoded.dst.clamp && decoded.dst.omod == 0,
+          "captured V_MIN/MAX_F64 lost its inline source, register pair, or modifiers");
+  }
+  for (const auto [opcode, expected] : {
+      std::pair{0x17u, Opcode::V_TRUNC_F64}, std::pair{0x18u, Opcode::V_CEIL_F64},
+      std::pair{0x1au, Opcode::V_FLOOR_F64}}) {
+    for (const bool vop3 : {false, true}) {
+      const std::array<uint32_t, 2> words = vop3
+          ? std::array<uint32_t, 2>{EncodeVop3Word0(0x180 + opcode, 1, 0, 1),
+                                   EncodeVop3Word1(8, 0, 0) | (1u << 29)}
+          : std::array<uint32_t, 2>{EncodeVop1(opcode, 1, 258), 0};
+      Instruction decoded;
+      DecodeInstruction(words, 0, decoded);
+      Check(decoded.opcode == expected && decoded.word_count == (vop3 ? 2u : 1u) &&
+                decoded.dst.reg == 1 && decoded.src0.reg == (vop3 ? 8u : 2u) &&
+                decoded.src0.absolute == vop3 && decoded.src0.negate == vop3,
+            "FP64 rounding lost its VOP1/VOP3 encoding, odd pair, or source modifiers");
+    }
+    for (const uint32_t source : {249u, 250u}) {
+      const uint32_t invalid[] = {EncodeVop1(opcode, 1, source),
+          source == 249u ? 0x00060604u : EncodeVop1Dpp(4, 0xe4)};
+      Instruction decoded;
+      DecodeInstruction(invalid, 0, decoded);
+      Check(decoded.opcode == Opcode::UNSUPPORTED,
+            "FP64 rounding accepted unsupported SDWA or DPP encoding");
+    }
+  }
+}
+
 void TestNewShaderRecompilerVop1SdwaBfrev() {
   using namespace ShaderRecompiler::Decoder;
   const uint32_t captured[] = {0x7e0070f9u, 0x00040600u};
@@ -4366,25 +4413,32 @@ void TestImageAtomicWidthDecoder() {
             decoded.dst.reg == 2u && decoded.dmask == 3u && !decoded.glc,
         "captured image atomic lost its 64-bit data width");
 
-  for (const auto opcode : {0x10u, 0x17u}) {
+  for (const auto opcode : {0x0fu, 0x10u, 0x11u, 0x14u, 0x15u, 0x16u,
+                            0x17u, 0x18u, 0x19u, 0x1au, 0x1eu, 0x1fu}) {
+    const bool compare_swap = opcode == 0x10u;
+    const bool supports_64 = opcode != 0x10u && opcode != 0x14u &&
+                             opcode != 0x16u && opcode != 0x1eu && opcode != 0x1fu;
     for (uint32_t mask = 0; mask < 16u; ++mask) {
-      const uint32_t words[] = {EncodeMimg0(opcode, mask), captured[1]};
-      DecodeInstruction(words, 0, decoded);
-      const bool compare_swap = opcode == 0x10u;
-      const bool supported = mask == (compare_swap ? 3u : 1u) ||
-                             (!compare_swap && mask == 3u);
-      Check((decoded.opcode != Opcode::UNSUPPORTED) == supported,
-            "image atomic accepted an invalid or unsupported width mask");
-      if (supported) {
-        Check(decoded.data_bits == (!compare_swap && mask == 3u ? 64u : 32u),
-              "image atomic DMASK selected the wrong data width");
+      for (const bool glc : {false, true}) {
+        const uint32_t words[] = {EncodeMimg0(opcode, mask, glc), captured[1]};
+        DecodeInstruction(words, 0, decoded);
+        const bool supported = mask == (compare_swap ? 3u : 1u) ||
+                               (supports_64 && mask == 3u);
+        Check((decoded.opcode != Opcode::UNSUPPORTED) == supported,
+              "image atomic accepted an invalid or unsupported width mask");
+        if (supported) {
+          const auto bits = !compare_swap && mask == 3u ? 64u : 32u;
+          Check(decoded.data_bits == bits && decoded.glc == glc &&
+                    decoded.data_dwords == (compare_swap ? 2u : bits / 32u),
+                "image atomic DMASK selected the wrong data width or return mode");
+        }
+        const uint32_t d16_words[] = {words[0], words[1] | (1u << 31u)};
+        DecodeInstruction(d16_words, 0, decoded);
+        Check(decoded.opcode == Opcode::UNSUPPORTED,
+              "image atomic incorrectly accepted D16 data");
       }
     }
   }
-  const uint32_t unsupported[] = {EncodeMimg0(0x11u, 3u), captured[1]};
-  DecodeInstruction(unsupported, 0, decoded);
-  Check(decoded.opcode == Opcode::UNSUPPORTED && decoded.data_bits == 64u,
-        "unsupported 64-bit image atomic silently decoded as 32-bit");
 }
 
 void TestNewShaderDecoderArchitecture() {
@@ -7694,40 +7748,18 @@ void TestNewShaderRecompilerDsAddtidTranslation() {
   CheckSpirvBinaryValidates(result.spirv);
 }
 
-void TestNewShaderRecompilerDsFloatMinMaxTranslation() {
-  const uint32_t shader[] = {
-      EncodeDs0(0x12, 4), EncodeDs1Ex(0, 9, 7, 1),  // ds_min_f32 v7, v9, v1
-      EncodeDs0(0x13, 8), EncodeDs1Ex(0, 10, 8, 1), // ds_max_f32 v8, v10, v1
-      0xbf810000u,
-  };
-
-  auto options = MakeCompileOptions(ShaderType::Compute);
-  options.dump_ir = true;
-
-  auto result = RecompileForTest(shader, options);
-  Check((result.decoded_dump.find("ds_min_f32") != std::string::npos),
-        "new decoder did not decode DS float min");
-  Check((result.decoded_dump.find("ds_max_f32") != std::string::npos),
-        "new decoder did not decode DS float max");
-  Check((result.ir_dump.find("DsMinF32 null, v7, v1") != std::string::npos),
-        "DS float min did not lower to explicit IR");
-  Check((result.ir_dump.find("DsMaxF32 null, v8, v1") != std::string::npos),
-        "DS float max did not lower to explicit IR");
-  Check((result.ir_dump.find("v9") != std::string::npos),
-        "DS float min did not retain DATA1 compare operand");
-  Check((result.ir_dump.find("v10") != std::string::npos),
-        "DS float max did not retain DATA1 compare operand");
-  Check(SpirvContainsOpcode(result.spirv, 12),
-        "SPIR-V binary does not contain OpExtInst");
-  Check(SpirvContainsOpcode(result.spirv, 61),
-        "SPIR-V binary does not contain OpLoad");
-  Check(SpirvContainsOpcode(result.spirv, 62),
-        "SPIR-V binary does not contain OpStore");
-  Check(SpirvContainsOpcode(result.spirv, 65),
-        "SPIR-V binary does not contain OpAccessChain");
-  Check(SpirvContainsOpcode(result.spirv, 124),
-        "SPIR-V binary does not contain OpBitcast");
-  CheckSpirvBinaryValidates(result.spirv);
+void TestDsFloatMinMaxDecoder() {
+  using namespace ShaderRecompiler::Decoder;
+  for (bool max_value : {false, true}) {
+    const uint32_t shader[] = {
+        EncodeDs0(max_value ? 0x13 : 0x12, 4), EncodeDs1Ex(0, 250, 7, 1)};
+    Instruction decoded;
+    DecodeInstruction(shader, 0u, decoded);
+    Check(decoded.opcode == (max_value ? Opcode::DS_MAX_F32 : Opcode::DS_MIN_F32) &&
+              decoded.src_count == 2u && decoded.src0.reg == 1u &&
+              decoded.src1.reg == 7u && decoded.offset == 4u,
+          "DS float min/max must read only address and DATA0");
+  }
 }
 
 void TestNewShaderRecompilerCfgStraightLine() {
@@ -12851,6 +12883,35 @@ void TestBlendMappingClassification() {
   blend.color_srcblend = static_cast<uint8_t>(Factor::kDstAlpha);
   Check(classify(Prospero::ColorMappingAbgr) == Support::Unsupported,
         "destination alpha incorrectly used the physical alpha channel");
+
+  blend.separate_alpha_blend = true;
+  blend.color_srcblend = static_cast<uint8_t>(Factor::kSrcAlpha);
+  blend.alpha_srcblend = static_cast<uint8_t>(Factor::kOne);
+  blend.color_destblend = blend.alpha_destblend =
+      static_cast<uint8_t>(Factor::kOneMinusSrcAlpha);
+  Check(classify(Prospero::ColorMappingAbgr) == Support::SourceAlphaOne,
+        "separate alpha accumulation did not preserve its unit source factor");
+  blend.alpha_srcblend = static_cast<uint8_t>(Factor::kZero);
+  Check(classify(Prospero::ColorMappingAbgr) == Support::SourceAlphaZero,
+        "separate alpha attenuation did not preserve its zero source factor");
+  blend.alpha_srcblend = blend.color_srcblend;
+  Check(classify(Prospero::ColorMappingAbgr) == Support::SourceAlpha,
+        "equal separate alpha equations did not reuse the broadcast alpha mode");
+  blend.alpha_srcblend = static_cast<uint8_t>(Factor::kOne);
+  const auto exact = blend;
+  for (uint32_t changed = 0; changed < 6; changed++) {
+    blend = exact;
+    switch (changed) {
+      case 0: blend.color_srcblend = static_cast<uint8_t>(Factor::kZero); break;
+      case 1: blend.color_destblend = static_cast<uint8_t>(Factor::kOne); break;
+      case 2: blend.alpha_srcblend = static_cast<uint8_t>(Factor::kDstAlpha); break;
+      case 3: blend.alpha_destblend = static_cast<uint8_t>(Factor::kOne); break;
+      case 4: blend.color_comb_fcn = static_cast<uint8_t>(Prospero::BlendOp::kSubtract); break;
+      case 5: blend.alpha_comb_fcn = static_cast<uint8_t>(Prospero::BlendOp::kSubtract); break;
+    }
+    Check(classify(Prospero::ColorMappingAbgr) == Support::Unsupported,
+          "separate alpha lowering accepted a different blend equation");
+  }
 }
 
 void TestLogicalAlphaBlendExport() {
@@ -12860,58 +12921,85 @@ void TestLogicalAlphaBlendExport() {
   const auto ordinary_key = MakeStageStaticKey(pixel);
   pixel.dual_source_blending = true;
   const auto guest_key = MakeStageStaticKey(pixel);
-  pixel.alpha_blend_source_remap = true;
+  pixel.alpha_blend_source = ShaderAlphaBlendSource::SourceAlpha;
   const auto remapped_key = MakeStageStaticKey(pixel);
+  pixel.alpha_blend_source = ShaderAlphaBlendSource::SourceAlphaOne;
+  const auto separate_key = MakeStageStaticKey(pixel);
+  pixel.alpha_blend_source = ShaderAlphaBlendSource::SourceAlphaZero;
+  const auto zero_key = MakeStageStaticKey(pixel);
   Check(ordinary_key != guest_key && guest_key != remapped_key &&
-            ordinary_key != remapped_key,
-        "ordinary, guest dual-source, and logical-alpha shaders share a cache key");
+            ordinary_key != remapped_key && separate_key != ordinary_key &&
+            separate_key != guest_key && separate_key != remapped_key &&
+            zero_key != ordinary_key && zero_key != guest_key &&
+            zero_key != remapped_key && zero_key != separate_key,
+        "ordinary, guest dual-source, and logical-alpha modes share a cache key");
 
   auto options = MakeCompileOptions(ShaderType::Pixel);
   options.input_info.pixel = &pixel;
-  for (const bool compressed : {false, true}) {
-    const uint32_t shader[] = {
-        // RGBA = (1000, 2, 3, 0.25), packed into two half pairs when compressed.
-        EncodeVop1(0x01, 0, 255), compressed ? 0x400063d0u : 0x447a0000u,
-        EncodeVop1(0x01, 1, 255), compressed ? 0x34004200u : 0x40000000u,
-        EncodeVop1(0x01, 2, 255), 0x40400000u,
-        EncodeVop1(0x01, 3, 255), 0x3e800000u,
-        EncodeExp0(0, 0xf, false, compressed), EncodeExp1(0, 1, 2, 3),
-        EncodeExp0(1, 0xf, false), EncodeExp1(0, 0, 0, 0),
-        EncodeExp0(2, 0xf, true, false, true), EncodeExp1(0, 0, 0, 0),
-        EncodeSopp(0x01),
-    };
-    const auto result = RecompileForTest(shader, options);
-    CheckSpirvBinaryValidates(result.spirv);
-    const auto source = DisassembleSpirvBinary(result.spirv);
-    Check(source.find("OpDecorate %out_mrt_1 Location 0") != std::string::npos &&
-              source.find("OpDecorate %out_mrt_1 Index 1") != std::string::npos &&
-              CountSourceOccurrences(source, "OpStore %out_mrt_1 ") == 1 &&
-              source.find("out_mrt_2") == std::string::npos,
-          "inactive MRT exports overwrote the logical-alpha output");
-    Check(SpirvInstructionOpcodeCount(result.spirv, 252u) != 0,
-          "ignoring inactive MRT stores discarded their valid-mask export");
-    uint32_t alpha_input = 0;
-    uint32_t color_input = 0;
-    for (size_t i = 5; i < result.spirv.size(); i += result.spirv[i] >> 16u) {
-      if ((result.spirv[i] & 0xffffu) != 79u || (result.spirv[i] >> 16u) != 9u) {
-        continue;
-      }
-      const auto selectors = std::span(result.spirv).subspan(i + 5, 4);
-      if (std::ranges::equal(selectors, std::array{3u, 3u, 3u, 3u})) {
-        alpha_input = result.spirv[i + 3];
-      } else if (std::ranges::equal(selectors, std::array{3u, 2u, 1u, 0u})) {
-        color_input = result.spirv[i + 3];
+  for (const auto mode : {ShaderAlphaBlendSource::SourceAlpha,
+                          ShaderAlphaBlendSource::SourceAlphaOne}) {
+    pixel.alpha_blend_source = mode;
+    for (const bool compressed : {false, true}) {
+      const uint32_t shader[] = {
+          // RGBA = (1000, 2, 3, 0.25), packed into two half pairs when compressed.
+          EncodeVop1(0x01, 0, 255), compressed ? 0x400063d0u : 0x447a0000u,
+          EncodeVop1(0x01, 1, 255), compressed ? 0x34004200u : 0x40000000u,
+          EncodeVop1(0x01, 2, 255), 0x40400000u,
+          EncodeVop1(0x01, 3, 255), 0x3e800000u,
+          EncodeExp0(0, 0xf, false, compressed), EncodeExp1(0, 1, 2, 3),
+          EncodeExp0(1, 0xf, false), EncodeExp1(0, 0, 0, 0),
+          EncodeExp0(2, 0xf, true, false, true), EncodeExp1(0, 0, 0, 0),
+          EncodeSopp(0x01),
+      };
+      const auto result = RecompileForTest(shader, options);
+      CheckSpirvBinaryValidates(result.spirv);
+      const auto source = DisassembleSpirvBinary(result.spirv);
+      Check(source.find("OpDecorate %out_mrt_1 Location 0") != std::string::npos &&
+                source.find("OpDecorate %out_mrt_1 Index 1") != std::string::npos &&
+                CountSourceOccurrences(source, "OpStore %out_mrt_1 ") == 1 &&
+                source.find("out_mrt_2") == std::string::npos,
+            "inactive MRT exports overwrote the logical-alpha output");
+      Check(SpirvInstructionOpcodeCount(result.spirv, 252u) != 0,
+            "ignoring inactive MRT stores discarded their valid-mask export");
+      uint32_t alpha_input = 0;
+      uint32_t alpha_broadcast = 0;
+      uint32_t color_input = 0;
+      uint32_t one = 0;
+      bool unit_alpha_factor = false;
+      for (size_t i = 5; i < result.spirv.size(); i += result.spirv[i] >> 16u) {
+        const auto opcode = result.spirv[i] & 0xffffu;
+        if (opcode == 43u && (result.spirv[i] >> 16u) == 4u &&
+            result.spirv[i + 3] == 0x3f800000u) {
+          one = result.spirv[i + 2];
+        }
+        if (opcode == 82u && (result.spirv[i] >> 16u) == 6u &&
+            result.spirv[i + 3] == one && result.spirv[i + 4] == alpha_broadcast &&
+            result.spirv[i + 5] == 0u) {
+          unit_alpha_factor = true;
+        }
+        if ((result.spirv[i] & 0xffffu) != 79u || (result.spirv[i] >> 16u) != 9u) {
+          continue;
+        }
+        const auto selectors = std::span(result.spirv).subspan(i + 5, 4);
+        if (std::ranges::equal(selectors, std::array{3u, 3u, 3u, 3u})) {
+          alpha_input = result.spirv[i + 3];
+          alpha_broadcast = result.spirv[i + 2];
+        } else if (std::ranges::equal(selectors, std::array{3u, 2u, 1u, 0u})) {
+          color_input = result.spirv[i + 3];
+        }
       }
+      Check(alpha_input != 0 && alpha_input == color_input,
+            "blend source did not broadcast logical alpha before the physical export swizzle");
+      Check(unit_alpha_factor == (mode == ShaderAlphaBlendSource::SourceAlphaOne),
+            "secondary ABGR output must be (1, Sa, Sa, Sa) only for separate alpha accumulation");
     }
-    Check(alpha_input != 0 && alpha_input == color_input,
-          "blend source did not broadcast logical alpha before the physical export swizzle");
   }
 
   const uint32_t guest_shader[] = {
       EncodeExp0(0, 0xf, false), EncodeExp1(0, 1, 2, 3),
       EncodeExp0(1, 0xf), EncodeExp1(4, 5, 6, 7), EncodeSopp(0x01),
   };
-  pixel.alpha_blend_source_remap = false;
+  pixel.alpha_blend_source = ShaderAlphaBlendSource::None;
   pixel.target_output_mode[1] = pixel.target_output_mode[0];
   pixel.target_export_mapping = {};
   const auto guest = RecompileForTest(guest_shader, options);
@@ -14693,6 +14781,7 @@ int main() {
   TestNewShaderRecompilerDsWideAndAtomicTranslation();
   TestNewShaderRecompilerCapturedVop1SdwaByteConvert();
   TestNewShaderRecompilerCapturedFractF64();
+  TestNewShaderRecompilerF64AluEncodings();
   TestNewShaderRecompilerVop1SdwaBfrev();
   TestNewShaderRecompilerVop1SdwaNotDestination();
   TestNewShaderRecompilerScalarMemoryBindingDomains();
@@ -14701,6 +14790,7 @@ int main() {
   // here.
   TestScalarAshrI64Decoder();
   TestImageAtomicWidthDecoder();
+  TestDsFloatMinMaxDecoder();
   TestNewShaderDecoderArchitecture();
   TestImageAddressOperands();
   TestSopkCompareImmediateExtension();