shader: FLAT opcode=0x24

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

3rdparty/cpuinfo/src/arm/mach/init.c

@@ -14,7 +14,7 @@
 #include <cpuinfo/internal-api.h>
 #include <cpuinfo/log.h>
 
-/* Polyfill recent CPUFAMILY_ARM_* values for older SDKs */
+/* Polyfill recent CPUFAMILY_ARM_* values for older platform headers */
 #ifndef CPUFAMILY_ARM_MONSOON_MISTRAL
 	#define CPUFAMILY_ARM_MONSOON_MISTRAL   0xE81E7EF6
 #endif

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)

README.md

@@ -1,8 +1,8 @@
 # KytyPS5
 
-[![Build KytyPS5 (Windows)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28Windows%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
-[![Build KytyPS5 (Linux)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28Linux%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
-[![Build KytyPS5 (macOS)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28macOS%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Build KytyPS5 (Windows)](https://img.shields.io/github/check-runs/KytyPS5/KytyPS5/main?nameFilter=Build%20KytyPS5%20%28Windows%29&label=Build%20KytyPS5%20%28Windows%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Build KytyPS5 (Linux)](https://img.shields.io/github/check-runs/KytyPS5/KytyPS5/main?nameFilter=Build%20KytyPS5%20%28Linux%29&label=Build%20KytyPS5%20%28Linux%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Build KytyPS5 (macOS)](https://img.shields.io/github/check-runs/KytyPS5/KytyPS5/main?nameFilter=Build%20KytyPS5%20%28macOS%29&label=Build%20KytyPS5%20%28macOS%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
 [![Platform](https://img.shields.io/badge/platform-Windows%20x64%20%7C%20Linux%20x64%20%7C%20macOS%20x86__64-0078D4.svg)](#system-requirements)
 [![Status](https://img.shields.io/badge/status-active%20development-orange.svg)](#current-status)
 [![License](https://img.shields.io/badge/license-GPL--2.0-blue.svg)](LICENSE)
@@ -72,8 +72,8 @@ or graphical glitches, so please include the version you tested when reporting a
       <img src="docs/screenshots/ps5-02.png" width="300" alt="Demon's Souls running in KytyPS5">
     </td>
     <td align="center">
-      <strong>UFC 5</strong><br>
-      <img src="docs/screenshots/ps5-06.png" width="300" alt="UFC 5 running in KytyPS5">
+      <strong>UFC 6</strong><br>
+      <img src="docs/screenshots/ps5-06.jpg" width="300" alt="UFC 6 running in KytyPS5">
     </td>
   </tr>
 </table>

src/common/emulatorConfig.h

@@ -48,7 +48,7 @@ struct ConfigOptions {
 	int32_t                user_id                     = DEFAULT_USER_ID;
 	std::string            audio_input_device;
 	std::optional<ControllerColor> controller_color;
-	uint32_t               controller_speaker_volume      = 100;
+	uint32_t               controller_speaker_volume      = 50;
 	uint32_t               controller_vibration_intensity = 100;
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int32_t                gpu_index                   = -1;

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/bufferCache.cpp

@@ -258,15 +258,8 @@ void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size, bool is_write) {
 		}
 		auto& buffer = m_slot_buffers[FindBuffer(vaddr, size)];
 
-		// Widen nearby CPU reads so they share one GPU drain.
-		constexpr uint64_t WindowSize   = 512 * 1024;
-		const auto         buffer_begin = buffer.CpuAddress();
-		const auto         buffer_end   = buffer_begin + buffer.Size();
-		const auto window_begin = std::max(Common::AlignDown(vaddr, WindowSize), buffer_begin);
-		const auto window_end = std::min(std::max(window_begin + WindowSize, vaddr + size), buffer_end);
-
-		if (DownloadBufferMemory<false>(buffer, window_begin, window_end - window_begin)) {
-			m_memory_tracker.UnmarkRegionAsGpuModified(window_begin, window_end - window_begin);
+		if (DownloadBufferMemory<false>(buffer, vaddr, size)) {
+			m_memory_tracker.UnmarkRegionAsGpuModified(vaddr, size);
 		}
 		if (is_write) {
 			m_memory_tracker.MarkRegionAsCpuModified(vaddr, size);

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;
@@ -1037,6 +1047,8 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 	auto info = destination.info;
 	if (image.depth_id) {
 		info.data            = image.info.data;
+		info.resources       = {image.stencil_subresources.level_count,
+		                        image.stencil_subresources.layer_count};
 		info.guest_format    = Prospero::BufferFormat::k8UInt;
 		info.bytes_per_block = 1;
 		if (info.IsTiled()) info.pitch = TileGetDepthPitch(info.extent.width, 1, 0);
@@ -1051,6 +1063,11 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 	auto copies = BuildDepthCopies(info, full_slice_size, image.depth_id
 	                                                        ? vk::ImageAspectFlagBits::eStencil
 	                                                        : vk::ImageAspectFlagBits::eDepth);
+	if (image.depth_id) {
+		for (auto& copy: copies) {
+			copy.imageSubresource.baseArrayLayer += image.stencil_subresources.base_layer;
+		}
+	}
 	TileManager::Result linear {source.Handle(), source_offset, source.Size() - source_offset};
 	if (info.IsTiled()) {
 		const auto tiles = BuildDepthTiles(info);
@@ -1239,7 +1256,8 @@ ImageId TextureCache::AssociateStencil(ImageId depth_id, GuestRange stencil) {
 	}
 	auto& record = m_slot_images[association];
 	TouchImage(record);
-	record.depth_id = depth_id;
+	record.depth_id             = depth_id;
+	record.stencil_subresources = depth.stencil_subresources;
 	return association;
 }
 
@@ -1445,6 +1463,18 @@ vk::ImageView TextureCache::FindDepthTarget(ImageId id, const ImageDesc& desc) {
 	TouchImage(image);
 	image.MarkGpuModified();
 	image.usage.depth_target = true;
+	if (desc.info.HasStencil()) {
+		const auto layers = image.info.resources.layers;
+		EXIT_IF(layers == 0 || image.info.data.size % layers != 0 ||
+		        desc.info.data.address < image.info.data.address);
+		const auto slice_size = image.info.data.size / layers;
+		const auto offset     = desc.info.data.address - image.info.data.address;
+		EXIT_IF(slice_size == 0 || offset % slice_size != 0 || offset / slice_size >= layers ||
+		        desc.info.resources.layers > layers - offset / slice_size);
+		image.stencil_subresources = {0, desc.info.resources.levels,
+		                             static_cast<uint32_t>(offset / slice_size),
+		                             desc.info.resources.layers};
+	}
 	image.info.stencil = desc.info.stencil;
 	image.info.metadata = desc.info.metadata;
 	if (desc.info.HasMetadata()) {
@@ -1491,9 +1521,11 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 	if (command.IsInvalid() || !GuestRange {address, size}.Valid()) {
 		EXIT("TextureCache: invalid image clear\n");
 	}
-	std::scoped_lock     lock {m_lock};
-	ImageId              selected {};
-	vk::ImageAspectFlags aspect {};
+	std::scoped_lock      lock {m_lock};
+	ImageId               selected {};
+	ImageId               stencil_id {};
+	vk::ImageAspectFlags  aspect {};
+	ImageSubresourceRange subresources;
 	for (const auto id: FindImagesInRegion(address, size, false)) {
 		auto owner = m_slot_images.try_get(id);
 		if (owner == nullptr) {
@@ -1501,10 +1533,13 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 		}
 		vk::ImageAspectFlags candidate {};
 		ImageId              candidate_id = id;
+		ImageSubresourceRange candidate_subresources {
+		    0, owner->info.resources.levels, 0, owner->info.TransferLayers()};
 		if (owner->depth_id && owner->info.data.address == address &&
 		    owner->info.data.size == size) {
-			candidate    = vk::ImageAspectFlagBits::eStencil;
-			candidate_id = owner->depth_id;
+			candidate              = vk::ImageAspectFlagBits::eStencil;
+			candidate_id           = owner->depth_id;
+			candidate_subresources = owner->stencil_subresources;
 			owner        = m_slot_images.try_get(candidate_id);
 			if (owner == nullptr || owner->backing.image == nullptr || !owner->info.HasStencil()) {
 				continue;
@@ -1517,11 +1552,13 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 		if (!candidate) {
 			continue;
 		}
-		if (selected && selected != candidate_id) {
+		if (selected && (selected != candidate_id || subresources != candidate_subresources)) {
 			return false;
 		}
-		selected = candidate_id;
-		aspect   = candidate;
+		selected     = candidate_id;
+		stencil_id   = candidate == vk::ImageAspectFlagBits::eStencil ? id : ImageId {};
+		aspect       = candidate;
+		subresources = candidate_subresources;
 	}
 	if (!selected) {
 		return false;
@@ -1543,7 +1580,12 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 		clear.depthStencil.stencil = stencil_clear;
 	}
 	ClearImage(command, selected, image.backing.format,
-	           {aspect, 0, image.info.resources.levels, 0, image.info.TransferLayers()}, clear);
+	           {aspect, subresources.base_level, subresources.level_count,
+	            subresources.base_layer, subresources.layer_count}, clear);
+	if (stencil_id) {
+		TrackImage(stencil_id);
+		CommitGpuWrite(m_slot_images[stencil_id]);
+	}
 	return true;
 }
 
@@ -1709,14 +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) {
@@ -1856,23 +1898,6 @@ void TextureCache::InvalidateMemoryFromGPU(uint64_t address, uint64_t size) {
 	}
 }
 
-bool TextureCache::IsRegionGpuModified(uint64_t address, uint64_t size) {
-	if (!GuestRange {address, size}.Valid()) {
-		return false;
-	}
-	std::scoped_lock lock {m_lock};
-	for (const auto id: FindImagesInRegion(address, size, false)) {
-		const auto& image = m_slot_images[id];
-		// PPSA17168: S_LOAD_DWORD reads shader data at an address overlapping an old
-		// render target whose memory the CPU has reused. The cached image still retains
-		// its earlier GPU-modified flag.
-		if (!image.depth_id && image.IsGpuModified() && !image.IsDefinitelyCpuDirty()) {
-			return true;
-		}
-	}
-	return false;
-}
-
 void TextureCache::InvalidateCpuAliases(uint64_t address, uint64_t size) {
 	const auto page_begin = Common::AlignDown(address, TRACKER_PAGE_SIZE);
 	const auto page_end   = Common::AlignUp(address + size, TRACKER_PAGE_SIZE);

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

@@ -61,7 +61,6 @@ public:
 	                                        uint32_t packed_clear);
 	void               InvalidateMemory(uint64_t address, uint64_t size);
 	void               InvalidateMemoryFromGPU(uint64_t address, uint64_t size);
-	[[nodiscard]] bool IsRegionGpuModified(uint64_t address, uint64_t size);
 
 	[[nodiscard]] bool IsMeta(uint64_t address);
 	[[nodiscard]] bool IsMetaCleared(uint64_t address, uint32_t slice);

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);
 			}
@@ -673,7 +686,9 @@ Prospero::BufferFormat RenderTargetTransferFormat(uint32_t bytes_per_element) {
 } // namespace ImageOps
 
 Image::Image(GraphicContext& graphics, CommandScheduler& scheduler, const ImageInfo& image_info)
-    : info(image_info), m_graphics(graphics), m_scheduler(scheduler) {
+    : info(image_info),
+      stencil_subresources {0, image_info.resources.levels, 0, image_info.resources.layers},
+      m_graphics(graphics), m_scheduler(scheduler) {
 	KYTY_PROFILER_FUNCTION();
 	ImageOps::Validate(info);
 	m_cpu_dirty =

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) {
@@ -148,6 +149,8 @@ public:
 	uint64_t         track_addr     = 0;
 	uint64_t         track_addr_end = 0;
 	ImageId          depth_id {};
+	// The current stencil plane's mapping into the depth image, also retained by its association.
+	ImageSubresourceRange stencil_subresources;
 	uint64_t         tick_accessed_last = 0;
 	size_t           lru_id             = 0;
 

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");
@@ -130,7 +131,7 @@ NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource
 	const auto aligned_offset = Common::AlignDown(offset, alignment);
 	const auto adjustment     = offset - aligned_offset;
 	const auto max_range      = graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange;
-	if (adjustment % sizeof(uint32_t) != 0 || adjustment >= 256 || size > max_range - adjustment) {
+	if (adjustment >= 256 || size > max_range - adjustment) {
 		EXIT("storage buffer offset adjustment is unsupported\n");
 	}
 	buffer_offset = static_cast<uint32_t>(adjustment);
@@ -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) {
@@ -94,8 +104,10 @@ void PipelineCacheLog(fmt::format_string<Args...> format, Args&&... args) {
 }
 
 bool ReadShaderGuestMemory(void*, uint64_t address, std::span<uint32_t> values) {
+	// Scalar and unformatted buffer dependencies use the same backing as native raw loads.
+	// Image synchronization belongs to formatted buffer bindings, not these reads.
 	return !values.empty() &&
-	       Libs::LibKernel::Memory::TryReadGpuCleanBacking(address, values.data(), values.size_bytes());
+	       Libs::LibKernel::Memory::TryReadBufferBacking(address, values.data(), values.size_bytes());
 }
 
 void DumpShaderSpirv(const char* stage_name, uint64_t shader_hash,
@@ -287,11 +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,
@@ -617,14 +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) {
@@ -658,19 +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);
 }
@@ -727,10 +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;
@@ -742,18 +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) {
@@ -272,22 +314,75 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		return;
 	}
 
-	const bool large_workgroup =
-	    (input_info.threads_num[0] * input_info.threads_num[1] * input_info.threads_num[2] >= 512);
-	const bool                   has_sampler = !program.info.samplers.empty();
-	static std::atomic<uint32_t> dispatch_log_count {0};
-	if ((large_workgroup || has_sampler) &&
-	    dispatch_log_count.fetch_add(1, std::memory_order_relaxed) < 512) {
+	if (use_thread_dimensions) {
+		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       = 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) {
+			LOGF("GraphicsRenderDispatchDirect: use-thread-dimensions %ux%ux%u / %ux%ux%u -> "
+			     "groups %ux%ux%u\n",
+			     old_x, old_y, old_z, std::max(cs_regs.cs_regs.num_thread_x, 1u),
+			     std::max(cs_regs.cs_regs.num_thread_y, 1u),
+			     std::max(cs_regs.cs_regs.num_thread_z, 1u), thread_group_x, thread_group_y,
+			     thread_group_z);
+		}
+	}
+
+	buffer.EndRendering();
+	auto& pipeline =
+	    m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
+	auto& bindings = m_compute_bindings;
+	PrepareBindings(input_info.stage, bindings);
+	if (program.bindings.dispatch_thread_dword != ShaderRecompiler::IR::PushData::NoStart) {
+		std::copy(std::begin(input_info.dispatch_threads_num), std::end(input_info.dispatch_threads_num),
+		          bindings.shader_data.begin() + program.bindings.dispatch_thread_dword);
+	}
+	FindBuffers(bindings);
+	if (program.info.uses_dma) {
+		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 = 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) {
+		                return image.written &&
+		                       image.resource_class ==
+		                           ShaderRecompiler::IR::ImageResourceClass::Storage;
+	                }) ||
+	    has_storage_writes;
+	if (has_storage_writes) {
+		// A host fence used to serialize every dispatch. Preserve its read-before-write ordering
+		// while allowing the queue to execute asynchronously.
+		ShaderWriteHazardBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
+	}
+	if (Config::GraphicsDebugDumpEnabled()) {
 		const auto sampled_images = std::count_if(
 		    program.info.images.begin(), program.info.images.end(), [](const auto& image) {
 			    return image.resource_class == ShaderRecompiler::IR::ImageResourceClass::Sampled;
 		    });
 		const uint32_t frame_num = static_cast<uint32_t>(m_context.GetGpu().GetFrameNum());
 		LOGF("GraphicsRenderDispatchDirect: frame=%u shader=0x%016" PRIx64
+		     " hash=0x%016" PRIx64 " tick=%" PRIu64
 		     " groups=%ux%ux%u mode=0x%08" PRIx32 " local=%ux%ux%u "
 		     "buffers=%zu textures=%zu sampled=%zu storage=%zu samplers=%zu push=%u\n",
-		     frame_num, sh_ctx.GetCs().cs_regs.data_addr, thread_group_x, thread_group_y,
-		     thread_group_z, mode, input_info.threads_num[0], input_info.threads_num[1],
+		     frame_num, sh_ctx.GetCs().cs_regs.data_addr, program.shader_hash,
+		     m_context.GetCommandScheduler().CurrentTick(),
+		     thread_group_x, thread_group_y, thread_group_z, mode,
+		     input_info.threads_num[0], input_info.threads_num[1],
 		     input_info.threads_num[2], program.info.buffers.size(), program.info.images.size(),
 		     sampled_images, program.info.images.size() - sampled_images,
 		     program.info.samplers.size(),
@@ -334,61 +429,6 @@ 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);
-
-		static std::atomic<uint32_t> log_count {0};
-		if (log_count.fetch_add(1, std::memory_order_relaxed) < 32) {
-			LOGF("GraphicsRenderDispatchDirect: use-thread-dimensions %ux%ux%u / %ux%ux%u -> "
-			     "groups %ux%ux%u\n",
-			     old_x, old_y, old_z, std::max(cs_regs.cs_regs.num_thread_x, 1u),
-			     std::max(cs_regs.cs_regs.num_thread_y, 1u),
-			     std::max(cs_regs.cs_regs.num_thread_z, 1u), thread_group_x, thread_group_y,
-			     thread_group_z);
-		}
-	}
-
-	buffer.EndRendering();
-	auto& pipeline =
-	    m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
-	auto& bindings = m_compute_bindings;
-	PrepareBindings(input_info.stage, bindings);
-	FindBuffers(bindings);
-	if (program.info.uses_dma) {
-		m_context.PrepareBda();
-	}
-	RebindImages(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);
-	has_storage_writes =
-	    std::any_of(program.info.images.begin(), program.info.images.end(),
-	                [](const auto& image) {
-		                return image.written &&
-		                       image.resource_class ==
-		                           ShaderRecompiler::IR::ImageResourceClass::Storage;
-	                }) ||
-	    has_storage_writes;
-	if (has_storage_writes) {
-		// A host fence used to serialize every dispatch. Preserve its read-before-write ordering
-		// while allowing the queue to execute asynchronously.
-		ShaderWriteHazardBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
-	}
 	vk_buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline.pipeline);
 	vk_buffer.dispatch(thread_group_x, thread_group_y, thread_group_z);
 
@@ -422,6 +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(
@@ -432,7 +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/presentation/window/vulkanWindow.cpp

@@ -47,6 +47,12 @@ struct VulkanExtensions {
 	std::vector<vk::LayerProperties>     available_layers;
 };
 
+vk::PhysicalDeviceVulkan11Features WindowContext::RequiredVulkan11Features() noexcept {
+	vk::PhysicalDeviceVulkan11Features features {};
+	features.storageBuffer16BitAccess = VK_TRUE;
+	return features;
+}
+
 vk::PhysicalDeviceVulkan12Features WindowContext::RequiredVulkan12Features() noexcept {
 	vk::PhysicalDeviceVulkan12Features features {};
 	features.samplerMirrorClampToEdge  = VK_TRUE;
@@ -55,6 +61,8 @@ vk::PhysicalDeviceVulkan12Features WindowContext::RequiredVulkan12Features() noe
 	features.shaderOutputViewportIndex = VK_TRUE;
 	features.bufferDeviceAddress       = VK_TRUE;
 	features.shaderBufferInt64Atomics  = VK_TRUE;
+	features.storageBuffer8BitAccess   = VK_TRUE;
+	features.shaderSampledImageArrayNonUniformIndexing = VK_TRUE;
 	return features;
 }
 
@@ -206,6 +214,7 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		depth_clip_control.pNext = &depth_clip_enable;
 
 		vk::PhysicalDeviceVulkan12Features features12 {};
+		vk::PhysicalDeviceVulkan11Features features11 {};
 #if defined(__APPLE__)
 		features12.pNext = &depth_clip_control;
 #else
@@ -213,6 +222,8 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		fragment_barycentric.pNext = &depth_clip_control;
 		features12.pNext           = &fragment_barycentric;
 #endif
+		features11.pNext       = features12.pNext;
+		features12.pNext       = &features11;
 		features13.pNext       = &features12;
 		device_features2.pNext = &features13;
 
@@ -256,6 +267,10 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 
 		check_feature(features12.samplerMirrorClampToEdge, "samplerMirrorClampToEdge",
 		              required_features12.samplerMirrorClampToEdge);
+		check_feature(features11.storageBuffer16BitAccess, "storageBuffer16BitAccess",
+		              WindowContext::RequiredVulkan11Features().storageBuffer16BitAccess);
+		check_feature(features12.storageBuffer8BitAccess, "storageBuffer8BitAccess",
+		              required_features12.storageBuffer8BitAccess);
 		check_feature(features12.timelineSemaphore, "timelineSemaphore",
 		              required_features12.timelineSemaphore);
 		check_feature(features12.shaderOutputLayer, "shaderOutputLayer",
@@ -266,6 +281,9 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		              required_features12.bufferDeviceAddress);
 		check_feature(features12.shaderBufferInt64Atomics, "shaderBufferInt64Atomics",
 		              required_features12.shaderBufferInt64Atomics);
+		check_feature(features12.shaderSampledImageArrayNonUniformIndexing,
+		              "shaderSampledImageArrayNonUniformIndexing",
+		              required_features12.shaderSampledImageArrayNonUniformIndexing);
 		check_feature(features13.robustImageAccess, "robustImageAccess");
 		check_feature(features13.dynamicRendering, "dynamicRendering",
 		              required_features13.dynamicRendering);
@@ -478,6 +496,9 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	workgroup_layout.pNext = &depth_clip_control;
 	features12.pNext = workgroup_layout_extension ? static_cast<void*>(&workgroup_layout)
 	                                             : static_cast<void*>(&depth_clip_control);
+	auto features11 = WindowContext::RequiredVulkan11Features();
+	features11.pNext = features12.pNext;
+	features12.pNext = &features11;
 	if (!features12.shaderSharedInt64Atomics || !workgroup_layout.workgroupMemoryExplicitLayout) {
 		Log::WriteToConsoleAndLog(fmt::format(
 		    "WARNING: Native 64-bit LDS atomics are unavailable: shaderSharedInt64Atomics={}, "
@@ -560,8 +581,7 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	device_features.shaderInt64 = VK_TRUE;
 	device_features.shaderFloat64 =
 	    supported_features2.features.shaderFloat64 &&
-	    float_controls.shaderSignedZeroInfNanPreserveFloat64 &&
-	    float_controls.shaderRoundingModeRTEFloat32;
+	    float_controls.shaderSignedZeroInfNanPreserveFloat64;
 	// if (device_features.shaderFloat64 && !float_controls.shaderDenormPreserveFloat64) {
 	// 	Log::WriteToConsoleAndLog(
 	// 	    "WARNING: Vulkan device does not guarantee FP64 denormal preservation; "

src/graphics/presentation/window/window.cpp

@@ -362,11 +362,13 @@ static void GameEventDidEnterForeground(WindowLoopState& game) {
 	SetPause(game, false);
 }
 
-void WindowContext::Resize(uint32_t new_width, uint32_t new_height) {
-	EXIT_IF(new_width == 0 || new_height == 0);
+void WindowContext::Resize(int new_width, int new_height) {
+	if (new_width <= 0 || new_height <= 0) {
+		return;
+	}
 	Common::LockGuard lock(mutex);
-	graphic_ctx.screen_width  = new_width;
-	graphic_ctx.screen_height = new_height;
+	graphic_ctx.screen_width  = static_cast<uint32_t>(new_width);
+	graphic_ctx.screen_height = static_cast<uint32_t>(new_height);
 }
 
 void WindowContext::ProcessWindowEvent(const SDL_WindowEvent& event) {

src/graphics/presentation/window/windowInternal.h

@@ -34,6 +34,7 @@ struct WindowContext {
 	~WindowContext();
 	KYTY_CLASS_NO_COPY(WindowContext);
 
+	[[nodiscard]] static vk::PhysicalDeviceVulkan11Features RequiredVulkan11Features() noexcept;
 	[[nodiscard]] static vk::PhysicalDeviceVulkan12Features RequiredVulkan12Features() noexcept;
 	[[nodiscard]] static vk::PhysicalDeviceVulkan13Features RequiredVulkan13Features() noexcept;
 	[[nodiscard]] static uint32_t InitialWindowFlags(bool fullscreen) noexcept;
@@ -42,7 +43,7 @@ struct WindowContext {
 	void                                                    RefreshSurfaceCapabilities();
 	void                                                    UpdateIcon();
 	void                                                    UpdateTitle();
-	void                                                    Resize(uint32_t width, uint32_t height);
+	void                                                    Resize(int width, int height);
 	void ProcessWindowEvent(const SDL_WindowEvent& event);
 	void ProcessDisplayEvent(const SDL_DisplayEvent& event);
 	void ProcessEvent(double time_seconds);

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -648,6 +648,11 @@ CompileResult CompileProgram(TranslateResult translated, const CompileOptions& o
 				})) {
 					continue;
 				}
+				const auto resource = inst.Flags<uint32_t>();
+				first               = resource < ir.info.buffers.size() &&
+				                              ir.info.buffers[resource].indirect_root == resource
+				                          ? 1u
+				                          : 0u;
 			} else if (op == IR::ValueOpcode::GetImageResource) {
 				const auto resource = inst.Flags<uint32_t>();
 				first = resource < ir.info.images.size() &&
@@ -665,7 +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);
@@ -85,9 +88,13 @@ void ValidateNativeProgram(const IR::Program& program) {
 	}
 	const auto has_shader_data_storage = present[static_cast<size_t>(Kind::ShaderData)];
 	const auto shader_data_dwords = program.bindings.ShaderDataDwords();
+	const auto user_data_dwords = program.bindings.user_data_registers.size();
+	const bool has_dispatch_threads = program.bindings.dispatch_thread_dword != IR::PushData::NoStart;
 	if ((program.bindings.UsesPushData() &&
 	     !IR::PushData::CanFit(program.bindings.push_data_start_dword, shader_data_dwords)) ||
-	    program.bindings.memory_offset_dword != program.bindings.user_data_registers.size() ||
+	    (has_dispatch_threads && (program.stage != ShaderType::Compute ||
+	                              program.bindings.dispatch_thread_dword != user_data_dwords)) ||
+	    program.bindings.memory_offset_dword != user_data_dwords + (has_dispatch_threads ? 3u : 0u) ||
 	    program.bindings.memory_offset_count != buffers.size() ||
 	    has_shader_data_storage != (shader_data_dwords != 0 && !program.bindings.UsesPushData()) ||
 	    !std::is_sorted(program.bindings.user_data_registers.begin(),
@@ -218,7 +225,8 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 					Fail(program, "buffer operation has invalid memory metadata");
 				}
 				const auto& memory = program.memory_info[memory_index];
-				if (memory.kind == IR::ResourceKind::IndirectBuffer) {
+				if (memory.kind == IR::ResourceKind::IndirectBuffer &&
+				    inst.GetOpcode() != IR::ValueOpcode::ReadConstBuffer) {
 					requirements.subgroup_local_invocation_id = true;
 				}
 				if (memory.kind == IR::ResourceKind::Buffer) {
@@ -226,6 +234,9 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 					if (memory.resource >= program.info.buffers.size()) {
 						Fail(program, "buffer operation has invalid resource metadata");
 					}
+					const auto bits = StorageBufferElementBits(program, memory);
+					requirements.buffer_u8 |= bits == 8u;
+					requirements.buffer_u16 |= bits == 16u;
 					if ((program.info.buffers[memory.resource].packed_stride & (1u << 20u)) != 0u) {
 						if (program.stage != ShaderType::Compute) {
 							Fail(program, "buffer ADD_TID is only valid for compute shaders");
@@ -263,6 +274,7 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 				}
 			}
 			switch (inst.GetOpcode()) {
+				case IR::ValueOpcode::StoreCompletion: requirements.subgroup_barrier = true; break;
 				case IR::ValueOpcode::BvhIntersect: requirements.bvh = true; break;
 				case IR::ValueOpcode::Ballot: requirements.subgroup_ballot = true; break;
 				case IR::ValueOpcode::DppMoveU32:
@@ -288,6 +300,7 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
 					break;
 				}
 				case IR::ValueOpcode::SwizzleU32:
+				case IR::ValueOpcode::PermuteU32:
 				case IR::ValueOpcode::BpermuteU32: {
 					requirements.subgroup_ballot              = true;
 					requirements.subgroup_shuffle             = true;
@@ -334,7 +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/spirvEmitterAnalysis.cpp

@@ -83,7 +83,7 @@ uint32_t DescriptorElementPointer(EmitterState& state, uint32_t result_ptr_type,
 	}
 	const auto pointer = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpAccessChain, result_ptr_type, pointer, variable_id,
-	                          ConstantU32(state, array_index));
+	                          array_index);
 	return pointer;
 }
 
@@ -152,8 +152,9 @@ uint32_t ImageViewSizeType(EmitterState& state, ImageDimension dimension) {
 	}
 }
 
-uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mip) {
-	const auto pointer = ImageDescriptorPointer(state, resource, mip);
+uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mip,
+                             uint32_t array_index) {
+	const auto pointer = ImageDescriptorPointer(state, resource, mip, array_index);
 	const auto image = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpLoad, ImageType(state, state.program.info.images.at(resource)),
 	                          image, pointer);
@@ -167,30 +168,40 @@ uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler) {
 	const auto pointer_type =
 	    state.builder.Type(spv::OpTypePointer, spv::StorageClassUniformConstant, sampler_type);
 	const auto pointer = DescriptorElementPointer(
-	    state, pointer_type, state.sampler_variable, array_index,
+	    state, pointer_type, state.sampler_variable, ConstantU32(state, array_index),
 	    IR::DescriptorBindingKind::Samplers, sampler, "sampler descriptor array was not emitted");
 	const auto sampler_id = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpLoad, sampler_type, sampler_id, pointer);
 	return sampler_id;
 }
 
-uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler_id, uint32_t mip) {
+uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler_id, uint32_t mip,
+                          uint32_t array_index) {
 	const auto& image_resource = state.program.info.images.at(resource);
 	EXIT_IF(image_resource.resource_class != IR::ImageResourceClass::Sampled);
-	const auto  image          = LoadImageDescriptor(state, resource, mip);
+	const auto  image          = LoadImageDescriptor(state, resource, mip, array_index);
 	const auto  sampled_image = state.builder.AllocateId();
 	const auto  sampled_type =
 	    state.builder.Type(spv::OpTypeSampledImage, ImageType(state, image_resource));
 	state.builder.AddFunction(spv::OpSampledImage, sampled_type, sampled_image, image, sampler_id);
+	if (array_index != 0u) {
+		state.builder.RequireExtension("SPV_EXT_descriptor_indexing");
+		state.builder.RequireCapability(spv::CapabilityShaderNonUniform);
+		state.builder.RequireCapability(spv::CapabilitySampledImageArrayNonUniformIndexing);
+		state.builder.AddAnnotation(spv::OpDecorate, sampled_image, spv::DecorationNonUniform);
+	}
 	return sampled_image;
 }
 
-uint32_t ImageDescriptorPointer(EmitterState& state, uint32_t resource, uint32_t mip) {
+uint32_t ImageDescriptorPointer(EmitterState& state, uint32_t resource, uint32_t mip,
+                                uint32_t array_index) {
 	const auto& image = state.program.info.images.at(resource);
 	EXIT_IF(mip >= image.mip_count);
 	const auto kind = IR::DescriptorBindingForImage(image);
 	EXIT_IF(!kind.has_value());
-	const auto array_index  = ResourceForDescriptor(state, *kind, resource) + mip;
+	if (array_index == 0u) {
+		array_index = ConstantU32(state, ResourceForDescriptor(state, *kind, resource) + mip);
+	}
 	const auto pointer_type = state.builder.Type(
 	    spv::OpTypePointer, spv::StorageClassUniformConstant, ImageType(state, image));
 	const auto variable = state.image_variables[IR::ImageBindingIndex(*kind)];
@@ -213,7 +224,7 @@ void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_
 		EmitWrite(0);
 		return;
 	}
-	EmitImageMipSwitch(state, mip_lod, image.mip_count, 0, EmitWrite);
+	EmitIndexSwitch(state, mip_lod, image.mip_count, 0, EmitWrite);
 }
 
 spv::ExecutionModel ExecutionModelForStage(ShaderType stage) {

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

@@ -226,13 +226,13 @@ void DefineBvhIntersect(EmitterState& s) {
 	if (!s.requirements.bvh) return;
 	const auto u = TypeU32(s), f = TypeF32(s), b = TypeBool(s), wide = TypeScalarU64(s);
 	const auto vec4 = TypeU32Vector(s, 4), vec3 = TypeF32Vector(s, 3);
-	const auto signature = s.builder.Type(spv::OpTypeFunction, vec4, vec4, u, f, vec3, vec3, vec3);
+	const auto signature = s.builder.Type(spv::OpTypeFunction, vec4, vec4, wide, f, vec3, vec3, vec3);
 	s.bvh_intersect_function = s.builder.AllocateId();
 	s.builder.AddName(s.bvh_intersect_function, "bvh_intersect");
 	s.builder.AddFunction(spv::OpFunction, vec4, s.bvh_intersect_function,
 	                      spv::FunctionControlMaskNone, signature);
 	std::array<uint32_t, 6> args;
-	const std::array types {vec4, u, f, vec3, vec3, vec3};
+	const std::array types {vec4, wide, f, vec3, vec3, vec3};
 	for (uint32_t i = 0; i < args.size(); ++i) {
 		args[i] = s.builder.AllocateId();
 		s.builder.AddFunction(spv::OpFunctionParameter, types[i], args[i]);
@@ -253,7 +253,7 @@ void DefineBvhIntersect(EmitterState& s) {
 	const auto shr = [&](uint32_t value, uint32_t shift) {
 		return Binary(s, spv::OpShiftRightLogical, u, value, ConstantU32(s, shift));
 	};
-	const auto kind = and_bits(node, 7);
+	const auto kind = and_bits(Unary(s, spv::OpUConvert, u, node), 7);
 	const auto triangle = Binary(s, spv::OpULessThan, b, kind, ConstantU32(s, 4));
 	const auto full = Binary(s, spv::OpIEqual, b, kind, ConstantU32(s, 5));
 	const auto half = Binary(s, spv::OpIEqual, b, kind, ConstantU32(s, 4));
@@ -262,7 +262,7 @@ void DefineBvhIntersect(EmitterState& s) {
 	const auto grow = and_bits(shr(words[1], 23), 0xff);
 	const auto base = Binary(s, spv::OpShiftLeftLogical, wide,
 	    DeviceAddressFromWords(s, words[0], and_bits(words[1], 0xff)), ConstantDeviceAddress(s, 8));
-	const auto index = Unary(s, spv::OpUConvert, wide, shr(node, 3));
+	const auto index = Binary(s, spv::OpShiftRightLogical, wide, node, ConstantDeviceAddress(s, 3));
 	const auto last_index = Binary(s, spv::OpIAdd, wide, index,
 	    Select(s, wide, full, ConstantDeviceAddress(s, 1), ConstantDeviceAddress(s, 0)));
 	const auto address = Binary(s, spv::OpIAdd, wide, base,
@@ -325,17 +325,23 @@ void DefineBvhIntersect(EmitterState& s) {
 uint32_t EmitBvhIntersect(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto& s = ctx.state;
 	const auto* ray = ctx.ImageAddress(inst.Arg(1));
-	if (ray == nullptr || ray->NumArgs() < 11 || s.bvh_intersect_function == 0)
-		ctx.Fail(inst, "BVH intersection requires eleven address components and its shared function");
+	const auto node_words = inst.Flags<uint32_t>();
+	if (ray == nullptr || (node_words != 1u && node_words != 2u) ||
+	    ray->NumArgs() < node_words + 10u || s.bvh_intersect_function == 0)
+		ctx.Fail(inst, "BVH intersection requires its node pointer, ray components and shared function");
 	const auto vector = [&](uint32_t first) {
 		Vec3 values;
 		for (uint32_t i = 0; i < 3; ++i)
 			values[i] = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, first + i));
 		return Vector(s, TypeF32Vector(s, 3), values);
 	};
-	const auto descriptor = ctx.Arg(inst, 0), node = ctx.Arg(*ray, 0);
-	const auto extent = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, 1));
-	const auto origin = vector(2), direction = vector(5), inverse = vector(8);
+	const auto descriptor = ctx.Arg(inst, 0);
+	const auto node = node_words == 2u
+	                      ? DeviceAddressFromWords(s, ctx.Arg(*ray, 0), ctx.Arg(*ray, 1))
+	                      : Unary(s, spv::OpUConvert, TypeScalarU64(s), ctx.Arg(*ray, 0));
+	const auto extent = Unary(s, spv::OpBitcast, TypeF32(s), ctx.Arg(*ray, node_words));
+	const auto origin = vector(node_words + 1u), direction = vector(node_words + 4u),
+	           inverse = vector(node_words + 7u);
 	return EmitValueOrDefaultIfCondition(s, ctx.Arg(inst, 2), TypeU32Vector(s, 4),
 	    ConstantU32CompositeZero(s, 4), [&] {
 		const auto result = s.builder.AllocateId();

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

@@ -366,18 +366,23 @@ void EmitAuxPositionExport(ValueEmitContext& ctx, uint32_t data, const IR::Expor
 			state.builder.AddFunction(spv::OpStore, state.point_size_variable, f32);
 			continue;
 		}
-		auto StoreDistance = [&](uint32_t variable, uint32_t index) {
+		auto StoreDistance = [&](IR::StageOutputKind kind, uint32_t variable, uint32_t index) {
 			if (index == UINT32_MAX) {
 				return;
 			}
-			const auto pointer = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpAccessChain,
-			                          TypePointer(state, spv::StorageClassOutput, TypeF32(state)),
-			                          pointer, variable, ConstantU32(state, index));
+			uint32_t pointer;
+			if (state.program.stage == ShaderType::Mesh) {
+				pointer = MeshOutputPointer(state, kind, index);
+			} else {
+				pointer = state.builder.AllocateId();
+				state.builder.AddFunction(spv::OpAccessChain,
+				                          TypePointer(state, spv::StorageClassOutput, TypeF32(state)),
+				                          pointer, variable, ConstantU32(state, index));
+			}
 			state.builder.AddFunction(spv::OpStore, pointer, f32);
 		};
-		StoreDistance(state.clip_distance_variable, output.clip_distance);
-		StoreDistance(state.cull_distance_variable, output.cull_distance);
+		StoreDistance(IR::StageOutputKind::ClipDistance, state.clip_distance_variable, output.clip_distance);
+		StoreDistance(IR::StageOutputKind::CullDistance, state.cull_distance_variable, output.cull_distance);
 	}
 }
 
@@ -454,7 +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, [&]() {
@@ -498,14 +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);
 			}
 		}
@@ -574,6 +601,14 @@ uint32_t EmitIdentity(ValueEmitContext&, uint32_t value) {
 
 void EmitVoid(ValueEmitContext&) {}
 
+void EmitStoreCompletion(EmitterState& state) {
+	state.builder.AddFunction(
+	    spv::OpControlBarrier, ConstantU32(state, spv::ScopeSubgroup),
+	    ConstantU32(state, spv::ScopeDevice),
+	    ConstantU32(state, spv::MemorySemanticsReleaseMask | spv::MemorySemanticsUniformMemoryMask |
+	                          spv::MemorySemanticsImageMemoryMask));
+}
+
 void EmitBarrier(EmitterState& state) {
 	const auto tessellation = state.program.stage == ShaderType::TessellationControl;
 	if (!tessellation && ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {
@@ -581,9 +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));
 }
@@ -623,6 +660,14 @@ uint32_t EmitGetBuiltin(ValueEmitContext& ctx, IR::Value kind, IR::Value index)
 	return EmitBuiltinU32(ctx.state, static_cast<IR::StageInputKind>(kind.U32()), index.U32());
 }
 
+uint32_t EmitGetDispatchThreadExtent(ValueEmitContext& ctx, const IR::Inst& inst) {
+	const auto start = ctx.state.program.bindings.dispatch_thread_dword;
+	if (start == IR::PushData::NoStart || !inst.Arg(0).IsImmediate() || inst.Arg(0).U32() >= 3u) {
+		ctx.Fail(inst, "invalid dispatch thread extent");
+	}
+	return EmitShaderDataDwordLoad(ctx.state, start + inst.Arg(0).U32());
+}
+
 uint32_t EmitUndefU1(EmitterState& state, const IR::Inst& inst) {
 	const auto result = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpUndef, TypeId(state, inst.GetType()), result);

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;
@@ -728,7 +763,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 				return sample;
 			};
 			const auto sample = image.mip_mode == IR::ImageMipMode::Dynamic && image.mip_count > 1u
-			                        ? EmitImageMipSwitch(
+			                        ? EmitIndexSwitch(
 			                              state, GatherMip(ctx, inst, mem, *address, layout,
 			                                               image.mip_count),
 			                              image.mip_count, result_type, EmitGather)
@@ -776,13 +811,13 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			operands.push_back(AddressF32(ctx, mem, *address, layout.bias));
 		}
 		const auto sampler_id = LoadSamplerDescriptor(state, mem.sampler);
-		const auto EmitSample = [&](uint32_t resource) {
+		const auto EmitSample = [&](uint32_t resource, uint32_t array_index = 0u) {
 			const auto& candidate = state.program.info.images[resource];
 			const auto coord =
 			    CoordF32(ctx, mem, *address, layout.coord,
 			             ImageDimensionInfoFor(candidate.dimension).coordinate_components,
 			             candidate.cube);
-			const auto            sampled = MakeSampledImage(state, resource, sampler_id);
+			const auto sampled = MakeSampledImage(state, resource, sampler_id, 0u, array_index);
 			const auto            sample  = state.builder.AllocateId();
 			std::vector<uint32_t> sample_operands {result_type, sample, sampled, coord};
 			if (dref) {
@@ -808,7 +843,7 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 		const auto* source = image.source < state.program.descriptor_sources.size()
 		                         ? &state.program.descriptor_sources[image.source]
 		                         : nullptr;
-		if (handle == nullptr || source == nullptr || !source->indirect_image.has_value() ||
+		if (handle == nullptr || source == nullptr || !source->indirect_descriptor.has_value() ||
 		    handle->NumArgs() == 0u) {
 			ctx.Fail(inst, "has invalid indirect image key provenance");
 			return;
@@ -819,55 +854,42 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 			ctx.Fail(inst, "has no indirect image runtime mapping");
 			return;
 		}
-		const auto LoadMapping = [&](uint32_t index) {
-			const auto pointer = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpAccessChain, TypeStorageBufferElementPointer(state),
-			                          pointer, state.flattened_srt_variable, ConstantU32(state, 0),
-			                          index);
-			const auto value = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpLoad, TypeU32(state), value, pointer);
-			return value;
-		};
-		const auto mapping  = ConstantU32(state, image.indirect_mapping_offset);
-		auto       low      = ConstantU32(state, 0u);
-		auto       high     = LoadMapping(mapping);
-		auto       selected = ConstantU32(state, 0u);
-		for (uint32_t iteration = 0; iteration < image.indirect_search_iterations;
-		     iteration++) {
-			const auto active = Binary(state, spv::OpULessThan, TypeBool(state), low, high);
-			const auto mid    = Binary(state, spv::OpShiftRightLogical, TypeU32(state),
-			                           Binary(state, spv::OpIAdd, TypeU32(state), low, high),
-			                           ConstantU32(state, 1u));
-			const auto probe  = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpSelect, TypeU32(state), probe, active, mid,
-			                          ConstantU32(state, 0u));
-			const auto entry = Binary(state, spv::OpIAdd, TypeU32(state), mapping,
-			                          Binary(state, spv::OpIAdd, TypeU32(state),
-			                                 Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
-			                                        probe, ConstantU32(state, 1u)),
-			                                 ConstantU32(state, 1u)));
-			const auto mapped_key = LoadMapping(entry);
-			const auto candidate  = LoadMapping(
-			    Binary(state, spv::OpIAdd, TypeU32(state), entry, ConstantU32(state, 1u)));
-			const auto equal = Binary(state, spv::OpIEqual, TypeBool(state), mapped_key, key);
-			const auto match = Binary(state, spv::OpLogicalAnd, TypeBool(state), active, equal);
-			const auto next_selected = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpSelect, TypeU32(state), next_selected, match,
-			                          candidate, selected);
-			selected              = next_selected;
-			const auto less = Binary(state, spv::OpULessThan, TypeBool(state), mapped_key, key);
-			const auto take_upper = Binary(state, spv::OpLogicalAnd, TypeBool(state), active, less);
-			const auto take_lower = Binary(state, spv::OpLogicalAnd, TypeBool(state), active,
-			                               Unary(state, spv::OpLogicalNot, TypeBool(state), less));
-			const auto next_low   = state.builder.AllocateId();
-			state.builder.AddFunction(
-			    spv::OpSelect, TypeU32(state), next_low, take_upper,
-			    Binary(state, spv::OpIAdd, TypeU32(state), mid, ConstantU32(state, 1u)), low);
-			low                  = next_low;
-			const auto next_high = state.builder.AllocateId();
-			state.builder.AddFunction(spv::OpSelect, TypeU32(state), next_high, take_lower, mid,
-			                          high);
-			high = next_high;
+		const auto selected = EmitIndirectResourceIndex(
+		    state, key, image.indirect_mapping_offset, image.indirect_search_iterations, 0u);
+		const auto kind = *IR::DescriptorBindingForImage(image);
+		const auto& binding = *IR::FindBinding(state.program.bindings, kind);
+		uint32_t first_child = 0;
+		bool homogeneous = true;
+		for (uint32_t ordinal = 0; ordinal < image.indirect_resources.size(); ++ordinal) {
+			const auto resource = image.indirect_resources[ordinal];
+			const auto& candidate = state.program.info.images[resource];
+			if (candidate.dimension != image.dimension || candidate.cube != image.cube ||
+			    IR::DescriptorBindingForImage(candidate) != kind ||
+			    candidate.mip_count != 1u) {
+				homogeneous = false;
+				break;
+			}
+			if (ordinal == 1u) first_child = ResourceForDescriptor(state, kind, resource);
+			const auto child_slot = first_child + ordinal - 1u;
+			if (ordinal > 1u && (child_slot >= binding.resources.size() ||
+			                     binding.resources[child_slot] != resource)) {
+				homogeneous = false;
+				break;
+			}
+		}
+		if (homogeneous) {
+			const auto child_index = Binary(state, spv::OpIAdd, TypeU32(state), selected,
+			                                ConstantU32(state, first_child - 1u));
+			const auto is_root = Binary(state, spv::OpIEqual, TypeBool(state), selected,
+			                            ConstantU32(state, 0u));
+			const auto array_index = state.builder.AllocateId();
+			state.builder.AddFunction(spv::OpSelect, TypeU32(state), array_index, is_root,
+			                          ConstantU32(state, ResourceForDescriptor(state, kind, mem.resource)),
+			                          child_index);
+			const auto sample = EmitSample(mem.resource, array_index);
+			const auto result = dref ? sample : UnpackImageTexel(ctx, mem, sample);
+			ctx.Define(inst, ResultVector(ctx, result, numeric_class, dref, mem));
+			return;
 		}
 		const auto            default_label = state.builder.AllocateId();
 		const auto            merge_label   = state.builder.AllocateId();

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);
@@ -178,7 +180,11 @@ 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);
 uint32_t              EmitFPCos(EmitterState& state, uint32_t arg0);
 #undef EMIT_NATIVE
@@ -188,16 +194,17 @@ void                  EmitVoid(ValueEmitContext&);
 inline constexpr auto EmitReference    = EmitVoid;
 inline constexpr auto EmitReferenceU32 = EmitVoid;
 inline constexpr auto EmitControlNop   = EmitVoid;
-inline constexpr auto EmitWaitcnt      = EmitVoid;
 inline constexpr auto EmitSendmsg      = EmitVoid;
 inline constexpr auto EmitTtraceData   = EmitVoid;
 inline constexpr auto EmitInstPrefetch = EmitVoid;
 void                  EmitBarrier(EmitterState& state);
+void                  EmitStoreCompletion(EmitterState& state);
 void                  EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitMeshDrawParameter(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitSetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetUserData(EmitterState& state, IR::ScalarReg reg);
+uint32_t              EmitGetDispatchThreadExtent(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetBuiltin(ValueEmitContext& ctx, IR::Value kind, IR::Value index);
 uint32_t              EmitUndefU1(EmitterState& state, const IR::Inst& inst);
 inline constexpr auto EmitUndefU8  = EmitUndefU1;
@@ -230,10 +237,12 @@ void                  EmitStoreMemory(ValueEmitContext& ctx, const IR::Inst& ins
 uint32_t              EmitAtomic32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBufferAtomic64(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitSharedAtomic64(ValueEmitContext& ctx, const IR::Inst& inst);
+void                  EmitSharedAtomicMaskedOr32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBufferFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitSharedFloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitSharedIncDec(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst);
+uint32_t              EmitPermuteU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitBpermuteU32(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitReadConst(ValueEmitContext& ctx, const IR::Inst& inst);
 void                  EmitReadConstBuffer(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -312,15 +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

@@ -61,6 +61,7 @@ const ImageDimensionInfo& ImageDimensionInfoFor(ImageDimension dimension);
 struct SpirvRequirements {
 	bool bvh                          = false;
 	bool subgroup_ballot              = false;
+	bool subgroup_barrier             = false;
 	bool subgroup_shuffle             = false;
 	bool subgroup_local_invocation_id = false;
 	bool compute_derivatives          = false;
@@ -69,6 +70,8 @@ struct SpirvRequirements {
 	bool function_scratch             = false;
 	bool pixel_valid_mask             = false;
 	bool buffer_int64_atomics         = false;
+	bool buffer_u8                    = false;
+	bool buffer_u16                   = false;
 	bool shared_int64_atomics         = false;
 	bool coherent_buffers             = false;
 	bool float64                      = false;
@@ -80,7 +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;
@@ -107,6 +117,8 @@ struct EmitterState {
 	uint32_t                                         lane_count              = 1;
 	uint32_t                                         lane_half               = 0;
 	uint32_t                                         storage_buffer_variable = 0;
+	uint32_t                                         storage_buffer_u8_variable = 0;
+	uint32_t                                         storage_buffer_u16_variable = 0;
 	uint32_t                                         storage_buffer_u64_variable = 0;
 	std::array<uint32_t, IR::ShaderInfo::MaxBuffers> memory_byte_offsets {};
 	uint32_t                                         bda_pagetable_variable  = 0;
@@ -118,6 +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 {};
@@ -167,10 +181,9 @@ uint32_t TypeI32Vector(EmitterState& state, uint32_t components);
 uint32_t TypeF32Vector(EmitterState& state, uint32_t components);
 uint32_t TypePointer(EmitterState& state, spv::StorageClass storage_class, uint32_t pointee);
 uint32_t TypeFunction(EmitterState& state);
-uint32_t TypeStorageBufferPointer(EmitterState& state);
-uint32_t TypeStorageBufferElementPointer(EmitterState& state);
-uint32_t TypeStorageBufferU64Pointer(EmitterState& state);
-uint32_t TypeStorageBufferU64ElementPointer(EmitterState& state);
+uint32_t TypeStorageBufferElement(EmitterState& state, uint32_t bits);
+uint32_t TypeStorageBufferPointer(EmitterState& state, uint32_t bits = 32);
+uint32_t TypeStorageBufferElementPointer(EmitterState& state, uint32_t bits = 32);
 uint32_t TypePhysicalU32Pointer(EmitterState& state);
 uint32_t TypePushConstantElementPointer(EmitterState& state);
 uint32_t TypeU32ArrayPointer(EmitterState& state, spv::StorageClass storage_class, uint32_t dwords);
@@ -315,14 +328,16 @@ uint32_t ImageType(EmitterState& state, const IR::ImageResource& image);
 
 uint32_t ImageViewSizeType(EmitterState& state, ImageDimension dimension);
 
-uint32_t ImageDescriptorPointer(EmitterState& state, uint32_t resource, uint32_t mip = 0);
+uint32_t ImageDescriptorPointer(EmitterState& state, uint32_t resource, uint32_t mip = 0,
+                                uint32_t array_index = 0);
 
-uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mip = 0);
+uint32_t LoadImageDescriptor(EmitterState& state, uint32_t resource, uint32_t mip = 0,
+                             uint32_t array_index = 0);
 
 uint32_t LoadSamplerDescriptor(EmitterState& state, uint32_t sampler);
 
 uint32_t MakeSampledImage(EmitterState& state, uint32_t resource, uint32_t sampler_id,
-                          uint32_t mip = 0);
+                          uint32_t mip = 0, uint32_t array_index = 0);
 
 void EmitStorageImageWrite(EmitterState& state, uint32_t resource, uint32_t mip_lod, uint32_t coord,
                            uint32_t texel);
@@ -348,7 +363,8 @@ uint32_t DefineInterfaceVariable(EmitterState& state, uint32_t type, spv::Storag
 void     DefineModule(EmitterState& state);
 void     DefineTessellationInterfaces(EmitterState& state);
 void     DefineTessellationExecutionModes(EmitterState& state);
-void     DefineMeshOutputs(EmitterState& state);
+void     DefineMeshOutputs(EmitterState& state, uint32_t clip_distance_count,
+                           uint32_t cull_distance_count);
 void     EmitMeshEntryPoint(EmitterState& state);
 void     EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t MeshOutputPointer(EmitterState& state, IR::StageOutputKind kind, uint32_t index = 0);
@@ -392,6 +408,7 @@ uint32_t EmitShaderDataDwordLoad(EmitterState& state, uint32_t dword_index);
 uint32_t StorageBufferPackedStride(const EmitterState& state, const IR::MemoryInfo& mem);
 
 Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::MemoryInfo& mem);
+uint32_t StorageBufferElementBits(const IR::Program& program, const IR::MemoryInfo& mem);
 
 void EmitMemoryOffsets(EmitterState& state);
 
@@ -402,21 +419,19 @@ struct MemoryResourceAccess {
 	IR::ResourceKind      kind             = IR::ResourceKind::None;
 	uint32_t              object_pointer   = 0;
 	uint32_t              length           = 0;
-	uint32_t              index_offset     = 0;
 	uint32_t              byte_offset      = 0;
-	bool                  add_index_offset = false;
+	uint32_t              element_bits     = 32;
 	spv::MemoryAccessMask memory_access    = spv::MemoryAccessMaskNone;
 };
 
+uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t mapping_offset,
+                                   uint32_t search_iterations, uint32_t default_resource);
 MemoryResourceAccess PrepareMemoryResourceAccess(EmitterState& state, const IR::MemoryInfo& mem);
 
 MemoryResourceAccess PrepareStorageBufferResourceAccess(EmitterState&         state,
                                                         const IR::MemoryInfo& mem,
                                                         uint32_t variable, uint32_t pointer_type);
 
-uint32_t EmitMemoryElementIndex(EmitterState& state, const MemoryResourceAccess& access,
-                                uint32_t raw_index);
-
 uint32_t EmitMemoryElementInBounds(EmitterState& state, const MemoryResourceAccess& access,
                                    uint32_t index);
 
@@ -448,6 +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 =
@@ -517,17 +535,17 @@ uint32_t DeviceAddressFromWords(EmitterState& state, uint32_t low, uint32_t high
 
 // These templates accept local lambdas from several emitter translation units.
 template <typename Fn>
-auto EmitImageMipSwitch(EmitterState& state, uint32_t mip_lod, uint32_t mip_count,
+auto EmitIndexSwitch(EmitterState& state, uint32_t index, uint32_t count,
                         uint32_t result_type, Fn&& emit) {
 	constexpr bool has_result = !std::is_void_v<std::invoke_result_t<Fn, uint32_t>>;
-	EXIT_IF(mip_count == 0u);
+	EXIT_IF(count == 0u);
 	const auto merge_label = state.builder.AllocateId();
-	std::vector<uint32_t> labels(mip_count);
-	std::vector<uint32_t> words {spv::OpSwitch, mip_lod, merge_label};
-	for (uint32_t mip = 0; mip < mip_count; mip++) {
-		labels[mip] = state.builder.AllocateId();
-		words.push_back(mip);
-		words.push_back(labels[mip]);
+	std::vector<uint32_t> labels(count);
+	std::vector<uint32_t> words {spv::OpSwitch, index, merge_label};
+	for (uint32_t item = 0; item < count; item++) {
+		labels[item] = state.builder.AllocateId();
+		words.push_back(item);
+		words.push_back(labels[item]);
 	}
 	if constexpr (has_result) {
 		words[2] = labels.front();
@@ -538,13 +556,13 @@ auto EmitImageMipSwitch(EmitterState& state, uint32_t mip_lod, uint32_t mip_coun
 	if constexpr (has_result) {
 		phi = {spv::OpPhi, result_type, state.builder.AllocateId()};
 	}
-	for (uint32_t mip = 0; mip < mip_count; mip++) {
-		EmitLabel(state, labels[mip]);
+	for (uint32_t item = 0; item < count; item++) {
+		EmitLabel(state, labels[item]);
 		if constexpr (has_result) {
-			phi.push_back(emit(mip));
+			phi.push_back(emit(item));
 			phi.push_back(state.current_label);
 		} else {
-			emit(mip);
+			emit(item);
 		}
 		state.builder.AddFunction(spv::OpBranch, merge_label);
 	}

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

@@ -31,6 +31,51 @@ uint32_t EmitBinaryU32(EmitterState& state, spv::Op opcode, uint32_t lhs, uint32
 	return ret;
 }
 
+uint32_t EmitIndirectResourceIndex(EmitterState& state, uint32_t key, uint32_t mapping_offset,
+                                   uint32_t search_iterations, uint32_t default_resource) {
+	const auto LoadMapping = [&](uint32_t index) {
+		const auto pointer = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpAccessChain, TypeStorageBufferElementPointer(state),
+		                          pointer, state.flattened_srt_variable, ConstantU32(state, 0),
+		                          index);
+		const auto value = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpLoad, TypeU32(state), value, pointer);
+		return value;
+	};
+	const auto mapping = ConstantU32(state, mapping_offset);
+	const auto count = LoadMapping(mapping);
+	const auto entry_at = [&](uint32_t index) {
+		return Binary(state, spv::OpIAdd, TypeU32(state), mapping,
+		              Binary(state, spv::OpIAdd, TypeU32(state),
+		                     Binary(state, spv::OpShiftLeftLogical, TypeU32(state), index,
+		                            ConstantU32(state, 1u)), ConstantU32(state, 1u)));
+	};
+	auto low = ConstantU32(state, 0u);
+	auto high = count;
+	for (uint32_t iteration = 0; iteration < search_iterations; ++iteration) {
+		const auto active = Binary(state, spv::OpULessThan, TypeBool(state), low, high);
+		const auto mid = Binary(state, spv::OpShiftRightLogical, TypeU32(state),
+		                        Binary(state, spv::OpIAdd, TypeU32(state), low, high),
+		                        ConstantU32(state, 1u));
+		const auto probe = Select(state, TypeU32(state), active, mid, ConstantU32(state, 0u));
+		const auto less = Binary(state, spv::OpULessThan, TypeBool(state),
+		                         LoadMapping(entry_at(probe)), key);
+		low = Select(state, TypeU32(state),
+		             Binary(state, spv::OpLogicalAnd, TypeBool(state), active, less),
+		             Binary(state, spv::OpIAdd, TypeU32(state), mid, ConstantU32(state, 1u)), low);
+		high = Select(state, TypeU32(state),
+		              Binary(state, spv::OpLogicalAnd, TypeBool(state), active,
+		                     Unary(state, spv::OpLogicalNot, TypeBool(state), less)), mid, high);
+	}
+	const auto in_range = Binary(state, spv::OpULessThan, TypeBool(state), low, count);
+	const auto entry = entry_at(Select(state, TypeU32(state), in_range, low, ConstantU32(state, 0u)));
+	const auto match = Binary(state, spv::OpLogicalAnd, TypeBool(state), in_range,
+	                          Binary(state, spv::OpIEqual, TypeBool(state), LoadMapping(entry), key));
+	const auto resource = LoadMapping(Binary(state, spv::OpIAdd, TypeU32(state), entry,
+	                                         ConstantU32(state, 1u)));
+	return Select(state, TypeU32(state), match, resource, ConstantU32(state, default_resource));
+}
+
 uint32_t StorageBufferPackedStride(const EmitterState& state, const IR::MemoryInfo& mem) {
 	if (mem.resource >= state.program.info.buffers.size()) {
 		ExitDescriptorBindingFailure(state, IR::DescriptorBindingKind::Buffers, mem.resource,
@@ -47,6 +92,15 @@ Prospero::BufferFormat StorageBufferFormat(const EmitterState& state, const IR::
 	return state.program.info.buffers[mem.resource].descriptor_format;
 }
 
+uint32_t StorageBufferElementBits(const IR::Program& program, const IR::MemoryInfo& mem) {
+	if (!mem.formatted) return mem.data_bits;
+	const auto format = mem.typed ? Format::DecodeTBufferFormat(mem.data_format, mem.number_format)
+	                              : program.info.buffers[mem.resource].descriptor_format;
+	const auto info = Format::GetFormatInfo(format);
+	return info.type == Format::ComponentType::Unknown || info.packed_bitfield
+	           ? 32u : info.component_bits[0];
+}
+
 void EmitMemoryOffsets(EmitterState& state) {
 	for (uint32_t i = 0; i < state.program.bindings.memory_offset_count; i++) {
 		const auto word =
@@ -67,8 +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(
@@ -149,24 +203,18 @@ MemoryResourceAccess PrepareMemoryResourceAccess(EmitterState& state, const IR::
 			EXIT("physical address memory must use the BDA emitter\n");
 		case IR::ResourceKind::ScalarBuffer:
 		case IR::ResourceKind::Buffer: {
+			const auto bits = mem.kind == IR::ResourceKind::Buffer
+			                      ? StorageBufferElementBits(state.program, mem) : 32u;
+			const auto variable = bits == 8u ? state.storage_buffer_u8_variable
+			                      : bits == 16u ? state.storage_buffer_u16_variable
+			                                    : state.storage_buffer_variable;
 			access = PrepareStorageBufferResourceAccess(
-			    state, mem, state.storage_buffer_variable, TypeStorageBufferPointer(state));
-			access.index_offset = EmitBinaryU32(state, spv::OpShiftRightLogical, access.byte_offset,
-			                                    ConstantU32(state, 2u));
-			access.add_index_offset = true;
+			    state, mem, variable, TypeStorageBufferPointer(state, bits));
+			access.element_bits = bits;
 			return access;
 		}
 		default: EXIT("unsupported memory resource kind: %u\n", static_cast<unsigned>(mem.kind));
 	}
-	access.length = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpArrayLength, TypeU32(state), access.length,
-	                          access.object_pointer, 0);
-	return access;
-}
-
-uint32_t EmitMemoryElementIndex(EmitterState& state, const MemoryResourceAccess& access,
-                                uint32_t raw_index) {
-	return access.add_index_offset ? EmitAddU32(state, raw_index, access.index_offset) : raw_index;
 }
 
 uint32_t EmitMemoryElementInBounds(EmitterState& state, const MemoryResourceAccess& access,
@@ -179,12 +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);
@@ -195,7 +245,7 @@ uint32_t EmitMemoryElementPointer(EmitterState& state, const MemoryResourceAcces
 		return pointer;
 	}
 	return EmitStorageBufferElementPointer(state, access, index,
-	                                       TypeStorageBufferElementPointer(state));
+	                                       TypeStorageBufferElementPointer(state, access.element_bits));
 }
 
 uint32_t EmitStorageBufferElementPointer(EmitterState& state,
@@ -311,10 +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;
@@ -324,6 +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:
@@ -334,15 +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;
@@ -376,7 +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;
 		}
@@ -385,6 +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 {
@@ -394,15 +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);
@@ -417,6 +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/spirvEmitterMesh.cpp

@@ -25,43 +25,58 @@ uint32_t MeshLoad(EmitterState& state, uint32_t variable, spv::StorageClass stor
 }
 
 uint32_t MeshOutputType(EmitterState& state, IR::StageOutputKind kind) {
-	return kind == IR::StageOutputKind::Layer ? TypeU32(state) : TypeF32Vector(state, 4);
+	switch (kind) {
+		case IR::StageOutputKind::Layer: return TypeU32(state);
+		case IR::StageOutputKind::ClipDistance:
+		case IR::StageOutputKind::CullDistance: return TypeF32(state);
+		case IR::StageOutputKind::Position:
+		case IR::StageOutputKind::Parameter: return TypeF32Vector(state, 4);
+		default: EXIT("unsupported mesh output kind=%u\n", static_cast<uint32_t>(kind));
+	}
 }
 
 } // namespace
 
-void DefineMeshOutputs(EmitterState& state) {
+void DefineMeshOutputs(EmitterState& state, uint32_t clip_distance_count,
+                       uint32_t cull_distance_count) {
 	const auto& mesh = state.input_info.vertex->mesh;
 	for (auto& output: state.outputs) {
-		if (output.kind != IR::StageOutputKind::Position &&
-		    output.kind != IR::StageOutputKind::Parameter &&
-		    output.kind != IR::StageOutputKind::Layer) {
-			EXIT("unsupported mesh output kind=%u\n", static_cast<uint32_t>(output.kind));
+		const auto type = MeshOutputType(state, output.kind);
+		const bool clip = output.kind == IR::StageOutputKind::ClipDistance;
+		const bool cull = output.kind == IR::StageOutputKind::CullDistance;
+		auto& variable = clip ? state.clip_distance_variable
+		                 : cull ? state.cull_distance_variable : output.variable_id;
+		if (variable == 0) {
+			const auto element_type = clip || cull
+			                              ? state.builder.Type(spv::OpTypeArray, type,
+			                                                   ConstantU32(state, clip ? clip_distance_count
+			                                                                           : cull_distance_count))
+			                              : type;
+			variable = MeshArray(state, spv::StorageClassOutput, element_type,
+			                     output.kind == IR::StageOutputKind::Layer ? mesh.max_primitives
+			                                                             : mesh.max_vertices);
+			state.interface_variables.push_back(variable);
+			state.builder.AddName(variable, output.debug_name.c_str());
+			if (output.kind == IR::StageOutputKind::Parameter) {
+				state.builder.AddAnnotation(spv::OpDecorate, variable,
+				                            spv::DecorationLocation, output.location);
+			} else {
+				const auto builtin = clip ? spv::BuiltInClipDistance
+				                     : cull ? spv::BuiltInCullDistance
+				                     : output.kind == IR::StageOutputKind::Layer ? spv::BuiltInLayer
+				                                                                 : spv::BuiltInPosition;
+				state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationBuiltIn, builtin);
+			}
+			if (output.kind == IR::StageOutputKind::Layer) {
+				state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationPerPrimitiveEXT);
+			}
 		}
-		const auto type    = MeshOutputType(state, output.kind);
-		output.variable_id = MeshArray(
-		    state, spv::StorageClassOutput, type,
-		    output.kind == IR::StageOutputKind::Layer ? mesh.max_primitives : mesh.max_vertices);
+		output.variable_id = variable;
 		// Only Layer is read by another invocation, through the primitive's provoking vertex.
 		const bool shared = output.kind == IR::StageOutputKind::Layer;
 		output.mesh_data_variable =
 		    MeshArray(state, shared ? spv::StorageClassWorkgroup : spv::StorageClassPrivate, type,
 		              shared ? mesh.max_vertices : state.lane_count);
-		state.interface_variables.push_back(output.variable_id);
-		state.builder.AddName(output.variable_id, output.debug_name.c_str());
-		if (output.kind == IR::StageOutputKind::Parameter) {
-			state.builder.AddAnnotation(spv::OpDecorate, output.variable_id,
-			                            spv::DecorationLocation, output.location);
-		} else {
-			state.builder.AddAnnotation(spv::OpDecorate, output.variable_id, spv::DecorationBuiltIn,
-			                            output.kind == IR::StageOutputKind::Layer
-			                                ? spv::BuiltInLayer
-			                                : spv::BuiltInPosition);
-		}
-		if (output.kind == IR::StageOutputKind::Layer) {
-			state.builder.AddAnnotation(spv::OpDecorate, output.variable_id,
-			                            spv::DecorationPerPrimitiveEXT); // PerPrimitiveEXT
-		}
 	}
 	state.mesh_allocation = MeshArray(state, spv::StorageClassWorkgroup, TypeU32(state), 2);
 	state.mesh_primitive_data =
@@ -148,8 +163,16 @@ void EmitMeshEntryPoint(EmitterState& state) {
 				const auto value =
 				    MeshLoad(state, output.mesh_data_variable, spv::StorageClassPrivate, type,
 				             ConstantU32(state, half));
-				const auto pointer =
-				    MeshElement(state, output.variable_id, spv::StorageClassOutput, type, index);
+				uint32_t pointer;
+				if (output.kind == IR::StageOutputKind::ClipDistance ||
+				    output.kind == IR::StageOutputKind::CullDistance) {
+					pointer = state.builder.AllocateId();
+					state.builder.AddFunction(
+					    spv::OpAccessChain, TypePointer(state, spv::StorageClassOutput, type), pointer,
+					    output.variable_id, index, ConstantU32(state, output.index));
+				} else {
+					pointer = MeshElement(state, output.variable_id, spv::StorageClassOutput, type, index);
+				}
 				state.builder.AddFunction(spv::OpStore, pointer, value);
 			}
 		});

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

@@ -91,7 +91,8 @@ uint32_t TypeU32Vector(EmitterState& state, uint32_t components) {
 }
 
 uint32_t TypeU32Composite(EmitterState& state, uint32_t components) {
-	EXIT_IF(components < 2u || components > 4u);
+	EXIT_IF(components < 1u || components > 4u);
+	if (components == 1u) return TypeU32(state);
 	return components == 2u ? TypeU32Pair(state) : TypeU32Vector(state, components);
 }
 
@@ -122,46 +123,31 @@ uint32_t TypeFunction(EmitterState& state) {
 	return state.function_type;
 }
 
-uint32_t StorageRuntimeArrayType(EmitterState& state) {
-	return state.builder.DecoratedType(
-	    spv::OpTypeRuntimeArray,
-	    {{spv::OpDecorate, {spv::DecorationArrayStride, sizeof(uint32_t)}}}, TypeU32(state));
+uint32_t TypeStorageBufferElement(EmitterState& state, uint32_t bits) {
+	return bits == 32u ? TypeU32(state) : bits == 64u ? TypeScalarU64(state)
+	                                                   : state.builder.Type(spv::OpTypeInt, bits, 0);
 }
 
-uint32_t StorageBufferType(EmitterState& state) {
+uint32_t StorageRuntimeArrayType(EmitterState& state, uint32_t bits) {
+	return state.builder.DecoratedType(
+	    spv::OpTypeRuntimeArray,
+	    {{spv::OpDecorate, {spv::DecorationArrayStride, bits / 8u}}},
+	    TypeStorageBufferElement(state, bits));
+}
+
+uint32_t StorageBufferType(EmitterState& state, uint32_t bits = 32) {
 	return state.builder.DecoratedType(spv::OpTypeStruct,
 	                                   {{spv::OpMemberDecorate, {0, spv::DecorationOffset, 0}},
 	                                    {spv::OpDecorate, {spv::DecorationBlock}}},
-	                                   StorageRuntimeArrayType(state));
+	                                   StorageRuntimeArrayType(state, bits));
 }
 
-uint32_t TypeStorageBufferPointer(EmitterState& state) {
-	return TypePointer(state, spv::StorageClassStorageBuffer, StorageBufferType(state));
+uint32_t TypeStorageBufferPointer(EmitterState& state, uint32_t bits) {
+	return TypePointer(state, spv::StorageClassStorageBuffer, StorageBufferType(state, bits));
 }
 
-uint32_t TypeStorageBufferElementPointer(EmitterState& state) {
-	return TypePointer(state, spv::StorageClassStorageBuffer, TypeU32(state));
-}
-
-uint32_t StorageU64RuntimeArrayType(EmitterState& state) {
-	return state.builder.DecoratedType(
-	    spv::OpTypeRuntimeArray,
-	    {{spv::OpDecorate, {spv::DecorationArrayStride, sizeof(uint64_t)}}}, TypeScalarU64(state));
-}
-
-uint32_t StorageBufferU64Type(EmitterState& state) {
-	return state.builder.DecoratedType(spv::OpTypeStruct,
-	                                   {{spv::OpMemberDecorate, {0, spv::DecorationOffset, 0}},
-	                                    {spv::OpDecorate, {spv::DecorationBlock}}},
-	                                   StorageU64RuntimeArrayType(state));
-}
-
-uint32_t TypeStorageBufferU64Pointer(EmitterState& state) {
-	return TypePointer(state, spv::StorageClassStorageBuffer, StorageBufferU64Type(state));
-}
-
-uint32_t TypeStorageBufferU64ElementPointer(EmitterState& state) {
-	return TypePointer(state, spv::StorageClassStorageBuffer, TypeScalarU64(state));
+uint32_t TypeStorageBufferElementPointer(EmitterState& state, uint32_t bits) {
+	return TypePointer(state, spv::StorageClassStorageBuffer, TypeStorageBufferElement(state, bits));
 }
 
 uint32_t TypePhysicalU32Pointer(EmitterState& state) {
@@ -243,27 +229,34 @@ void DefineDescriptors(EmitterState& state) {
 			case IR::DescriptorBindingKind::Buffers:
 				state.storage_buffer_variable =
 				    Define(ArrayType(StorageBufferType(state)), "buffers");
+				if (state.requirements.buffer_u8) {
+					state.storage_buffer_u8_variable =
+					    Define(ArrayType(StorageBufferType(state, 8)), "buffers_u8");
+				}
+				if (state.requirements.buffer_u16) {
+					state.storage_buffer_u16_variable =
+					    Define(ArrayType(StorageBufferType(state, 16)), "buffers_u16");
+				}
 				if (state.requirements.buffer_int64_atomics) {
 					state.storage_buffer_u64_variable =
-					    Define(ArrayType(StorageBufferU64Type(state)), "buffers_u64");
-					state.builder.AddAnnotation(spv::OpDecorate, state.storage_buffer_variable,
-					                            spv::DecorationAliased);
-					state.builder.AddAnnotation(spv::OpDecorate, state.storage_buffer_u64_variable,
-					                            spv::DecorationAliased);
+					    Define(ArrayType(StorageBufferType(state, 64)), "buffers_u64");
 				}
-				if (state.requirements.coherent_buffers) {
+				for (const auto variable: {state.storage_buffer_variable, state.storage_buffer_u8_variable,
+				                           state.storage_buffer_u16_variable, state.storage_buffer_u64_variable}) {
+					if (variable == 0) continue;
+					if (state.storage_buffer_u8_variable != 0 || state.storage_buffer_u16_variable != 0 ||
+					    state.storage_buffer_u64_variable != 0) {
+						state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationAliased);
+					}
 					// RDNA2 stores publish to L2 even without GLC; every alias of the buffer
 					// must participate in visibility for cache-bypassing polling loads.
-					state.builder.AddAnnotation(spv::OpDecorate, state.storage_buffer_variable,
-					                            spv::DecorationCoherent);
-					if (state.storage_buffer_u64_variable != 0) {
-						state.builder.AddAnnotation(spv::OpDecorate, state.storage_buffer_u64_variable,
-						                            spv::DecorationCoherent);
+					if (state.requirements.coherent_buffers) {
+						state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationCoherent);
 					}
 				}
 				break;
 			case IR::DescriptorBindingKind::BdaPagetable:
-				state.bda_pagetable_variable = Define(StorageBufferU64Type(state), "bda_pagetable");
+				state.bda_pagetable_variable = Define(StorageBufferType(state, 64), "bda_pagetable");
 				break;
 			case IR::DescriptorBindingKind::FaultBuffer:
 				state.fault_buffer_variable = Define(StorageBufferType(state), "fault_buffer");
@@ -282,6 +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);
@@ -333,8 +337,9 @@ uint32_t ConstantU64(EmitterState& state, uint64_t value) {
 }
 
 uint32_t ConstantU32CompositeZero(EmitterState& state, uint32_t components) {
-	EXIT_IF(components < 2u || components > 4u);
+	EXIT_IF(components < 1u || components > 4u);
 	const auto            zero = ConstantU32(state, 0);
+	if (components == 1u) return zero;
 	std::vector<uint32_t> values(components, zero);
 	return state.builder.Constant(spv::OpConstantComposite, TypeU32Composite(state, components),
 	                              values);
@@ -414,6 +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;
@@ -426,15 +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;
@@ -465,6 +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;
@@ -541,7 +552,7 @@ void DefineOutputs(EmitterState& state) {
 		}
 	}
 	if (state.program.stage == ShaderType::Mesh) {
-		DefineMeshOutputs(state);
+		DefineMeshOutputs(state, clip_distance_count, cull_distance_count);
 		return;
 	}
 	if (state.program.stage == ShaderType::Vertex && clip_distance_count + cull_distance_count < 8u &&
@@ -667,6 +678,13 @@ void DefineModule(EmitterState& state) {
 
 	state.builder.RequireCapability(spv::CapabilityShader);
 	state.builder.RequireCapability(spv::CapabilitySignedZeroInfNanPreserve);
+	if (state.requirements.buffer_u8) {
+		state.builder.RequireExtension("SPV_KHR_8bit_storage");
+		state.builder.RequireCapability(spv::CapabilityStorageBuffer8BitAccess);
+	}
+	if (state.requirements.buffer_u16) {
+		state.builder.RequireCapability(spv::CapabilityStorageBuffer16BitAccess);
+	}
 	if (state.program.info.uses_dma) {
 		state.builder.RequireCapability(spv::CapabilityInt64);
 		state.builder.RequireCapability(spv::CapabilityPhysicalStorageBufferAddresses);
@@ -676,7 +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);
@@ -701,8 +720,9 @@ void DefineModule(EmitterState& state) {
 	if (state.requirements.image_gather_extended) {
 		state.builder.RequireCapability(spv::CapabilityImageGatherExtended);
 	}
-	if (state.lane_count == 2 || state.requirements.subgroup_ballot ||
-	    state.requirements.subgroup_shuffle || state.requirements.subgroup_local_invocation_id) {
+	if (state.lane_count == 2 || state.requirements.subgroup_barrier ||
+	    state.requirements.subgroup_ballot || state.requirements.subgroup_shuffle ||
+	    state.requirements.subgroup_local_invocation_id) {
 		state.builder.RequireCapability(spv::CapabilityGroupNonUniform);
 	}
 	if (state.lane_count == 2 || state.requirements.subgroup_ballot) {
@@ -737,15 +757,13 @@ void DefineModule(EmitterState& state) {
 	if (state.requirements.float64) {
 		EXIT_NOT_IMPLEMENTED(state.program.stage == ShaderType::Compute &&
 		                     state.input_info.compute->float_mode != 0xc0);
-		// MODE=0xc0 uses round-to-nearest-even and preserves FP64 input/output denormals.
+		// Use native rounding for MODE=0xc0, consistent with ordinary FP32 arithmetic.
 		state.builder.RequireCapability(spv::CapabilityFloat64);
-		state.builder.RequireCapability(spv::CapabilityRoundingModeRTE);
 		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeSignedZeroInfNanPreserve,
 		                               64u);
 		// FP64 denormal preservation is temporarily disabled.
 		// state.builder.RequireCapability(spv::CapabilityDenormPreserve);
 		// state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeDenormPreserve, 64u);
-		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeRoundingModeRTE, 32u);
 	}
 	if (const auto* cs = ShaderWorkgroupInput(state.program.stage, state.input_info)) {
 		uint32_t    local_x = state.requirements.compute_derivatives ? 2u : 1u;

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

@@ -315,6 +315,7 @@ void EmitBlock(ValueEmitContext& ctx, const IR::Block* block, EmitInstruction&&
 			auto& lane          = half == 0 ? ctx : *ctx.other_half;
 			ctx.state.lane_half = half;
 			if (half == 0 || (inst.GetOpcode() != IR::ValueOpcode::Barrier &&
+			                  inst.GetOpcode() != IR::ValueOpcode::StoreCompletion &&
 			                  inst.GetOpcode() != IR::ValueOpcode::MeshAllocate)) {
 				emit_instruction(lane, inst);
 			}
@@ -688,9 +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) {
@@ -751,6 +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;
 	}
@@ -228,7 +234,8 @@ Opcode DecodeMimgOpcode(uint32_t opcode, const MimgSampleInfo* sample, const Mim
 		case 0x09u: return Opcode::IMAGE_STORE_MIP;
 		case 0x0eu: return Opcode::IMAGE_GET_RESINFO;
 		case 0x60u: return Opcode::IMAGE_GET_LOD;
-		case 0xe6u: return Opcode::IMAGE_BVH_INTERSECT_RAY;
+		case 0xe6u:
+		case 0xe7u: return Opcode::IMAGE_BVH_INTERSECT_RAY;
 		default: return Opcode::UNSUPPORTED;
 	}
 }
@@ -245,7 +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);
 	}
@@ -258,6 +265,7 @@ uint32_t DecodeMimgAddressComponents(uint32_t opcode, ImageDimension dimension,
 
 	switch (opcode) {
 		case 0xe6u: return 11u;
+		case 0xe7u: return 12u;
 		case 0x0eu: return 1u;
 		case 0x01u:
 		case 0x09u: return ImageCoordComponents(dimension) + 1u;
@@ -368,7 +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");
 			}
@@ -379,10 +387,10 @@ void DecodeMimg(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	if (d16 && !supports_d16) {
 		SetUnsupported(inst, Family::MIMG, opcode, "MIMG opcode does not support D16 data");
 	}
-	if (opcode == 0xe6u &&
+	if (inst.opcode == Opcode::IMAGE_BVH_INTERSECT_RAY &&
 	    (a16 || !r128 || inst.dmask != 0xfu || (nsa_dwords != 0u && nsa_dwords != 3u))) {
 		SetUnsupported(inst, Family::MIMG, opcode,
-		               "BVH intersection requires eleven full-float ray DWORDs and R128/dmask:0xf");
+		               "BVH intersection requires full-float ray DWORDs and R128/dmask:0xf");
 	}
 
 	DecodeVectorGpr(vdata, inst.dst);

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

@@ -84,6 +84,7 @@ constexpr MemoryOpcodeInfo FLAT_OPCODE_LIST[] = {
     {0x18u, Opcode::FLAT_STORE_BYTE, 1, 8},     {0x1au, Opcode::FLAT_STORE_SHORT, 1, 16},
     {0x1cu, Opcode::FLAT_STORE_DWORD, 1, 32},   {0x1du, Opcode::FLAT_STORE_DWORDX2, 2, 32},
     {0x1eu, Opcode::FLAT_STORE_DWORDX4, 4, 32}, {0x1fu, Opcode::FLAT_STORE_DWORDX3, 3, 32},
+    {0x24u, Opcode::FLAT_LOAD_SHORT_D16, 1, 16},
 };
 
 constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
@@ -92,7 +93,8 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0x05u, Opcode::DS_MIN_I32, 1, 32},          {0x06u, Opcode::DS_MAX_I32, 1, 32},
     {0x07u, Opcode::DS_MIN_U32, 1, 32},          {0x08u, Opcode::DS_MAX_U32, 1, 32},
     {0x09u, Opcode::DS_AND_B32, 1, 32},          {0x0au, Opcode::DS_OR_B32, 1, 32},
-    {0x0bu, Opcode::DS_XOR_B32, 1, 32},          {0x0du, Opcode::DS_WRITE_B32, 1, 32},
+    {0x0bu, Opcode::DS_XOR_B32, 1, 32},          {0x0cu, Opcode::DS_MSKOR_B32, 1, 32},
+    {0x0du, Opcode::DS_WRITE_B32, 1, 32},
     {0x0eu, Opcode::DS_WRITE2_B32, 2, 32},       {0x0fu, Opcode::DS_WRITE2ST64_B32, 2, 32},
     {0x12u, Opcode::DS_MIN_F32, 1, 32},          {0x13u, Opcode::DS_MAX_F32, 1, 32},
     {0x1eu, Opcode::DS_WRITE_B8, 1, 8},          {0x1fu, Opcode::DS_WRITE_B16, 1, 16},
@@ -116,7 +118,7 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0xa6u, Opcode::DS_READ_U16_D16, 1, 16},
     {0xa7u, Opcode::DS_READ_U16_D16_HI, 1, 16},
     {0xb0u, Opcode::DS_WRITE_ADDTID_B32, 1, 32}, {0xb1u, Opcode::DS_READ_ADDTID_B32, 1, 32},
-    {0xb3u, Opcode::DS_BPERMUTE_B32, 1, 32},
+    {0xb2u, Opcode::DS_PERMUTE_B32, 1, 32},     {0xb3u, Opcode::DS_BPERMUTE_B32, 1, 32},
     {0xdeu, Opcode::DS_WRITE_B96, 3, 32},        {0xdfu, Opcode::DS_WRITE_B128, 4, 32},
     {0xfeu, Opcode::DS_READ_B96, 3, 32},         {0xffu, Opcode::DS_READ_B128, 4, 32},
 };
@@ -192,6 +194,7 @@ bool IsDsAtomicOpcode(Opcode opcode) {
 		case Opcode::DS_OR_RTN_B32:
 		case Opcode::DS_XOR_B32:
 		case Opcode::DS_XOR_RTN_B32:
+		case Opcode::DS_MSKOR_B32:
 		case Opcode::DS_WRXCHG_RTN_B32: return true;
 		default: return false;
 	}
@@ -203,8 +206,10 @@ 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: 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:
 		case Opcode::DS_CONSUME:
@@ -369,6 +374,10 @@ void DecodeFlat(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	}
 
 	DecodeVectorGpr(IsFlatStoreOpcode(inst.opcode) ? data : vdst, inst.dst);
+	if (inst.opcode == Opcode::FLAT_LOAD_SHORT_D16) {
+		// D16 loads reuse partial destinations to preserve the untouched high half.
+		inst.dst.sdwa_sel = 4u;
+	}
 	DecodeVectorGpr(addr, inst.src0);
 	inst.src_count = 1;
 	if (seg == 0u || saddr == 0x7du || saddr == 0x7fu) {
@@ -407,7 +416,8 @@ void DecodeDs(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
 		SetUnsupported(inst, Family::DS, opcode, "DS swizzle FFT mode is not implemented");
 	}
 	if (inst.gds &&
-	    (inst.opcode == Opcode::DS_SWIZZLE_B32 || inst.opcode == Opcode::DS_BPERMUTE_B32 ||
+	    (inst.opcode == Opcode::DS_SWIZZLE_B32 || inst.opcode == Opcode::DS_PERMUTE_B32 ||
+	     inst.opcode == Opcode::DS_BPERMUTE_B32 ||
 	     inst.opcode == Opcode::DS_WRITE_ADDTID_B32 ||
 	     inst.opcode == Opcode::DS_READ_ADDTID_B32)) {
 		SetUnsupported(inst, Family::DS, opcode, "DS lane operation is available only for LDS");

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

@@ -86,7 +86,7 @@ constexpr OpcodeMap SOPK_OPCODE_LIST[] = {
     {0x08u, Opcode::S_CMP_LE_I32}, {0x09u, Opcode::S_CMP_EQ_U32}, {0x0au, Opcode::S_CMP_LG_U32},
     {0x0bu, Opcode::S_CMP_GT_U32}, {0x0cu, Opcode::S_CMP_GE_U32}, {0x0du, Opcode::S_CMP_LT_U32},
     {0x0eu, Opcode::S_CMP_LE_U32}, {0x0fu, Opcode::S_ADD_I32},    {0x10u, Opcode::S_MULK_I32},
-    {0x13u, Opcode::S_SETREG_B32}, {0x17u, Opcode::S_WAITCNT},    {0x18u, Opcode::S_WAITCNT},
+    {0x13u, Opcode::S_SETREG_B32}, {0x17u, Opcode::S_WAITCNT_VSCNT}, {0x18u, Opcode::S_WAITCNT},
     {0x19u, Opcode::S_WAITCNT},    {0x1au, Opcode::S_WAITCNT},
     {0x1bu, Opcode::S_SUBVECTOR_LOOP_BEGIN}, {0x1cu, Opcode::S_SUBVECTOR_LOOP_END},
 };
@@ -234,6 +234,14 @@ void DecodeSopk(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 			inst.src_count         = 1;
 			return;
 		}
+		case Opcode::S_WAITCNT_VSCNT:
+			inst.dst.kind       = OperandKind::Null;
+			inst.src1           = inst.src0;
+			inst.src1.value     = word & 0xffffu;
+			inst.src1.signed_val = static_cast<int32_t>(inst.src1.value);
+			DecodeScalarSource(sdst, pc, inst.src0);
+			inst.src_count = 2;
+			return;
 		case Opcode::S_SETREG_B32:
 			inst.dst.kind        = OperandKind::Null;
 			inst.src1.kind       = OperandKind::LiteralConstant;

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

@@ -535,6 +535,11 @@ std::string InstructionToString(const Instruction& inst) {
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: s_setreg_b32 {}, {}", inst.pc,
 			                                               OperandToString(inst.src0).c_str(),
 			                                               OperandToString(inst.src1).c_str()));
+		case Opcode::S_WAITCNT_VSCNT:
+			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: {} {}, {}", inst.pc,
+			                                               magic_enum::enum_name(inst.opcode),
+			                                               OperandToString(inst.src0),
+			                                               OperandToString(inst.src1)));
 		case Opcode::S_NOP:
 		case Opcode::S_WAITCNT:
 		case Opcode::S_WAITCNT_DEPCTR:
@@ -658,6 +663,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::FLAT_LOAD_SBYTE:
 		case Opcode::FLAT_LOAD_USHORT:
 		case Opcode::FLAT_LOAD_SSHORT:
+		case Opcode::FLAT_LOAD_SHORT_D16:
 		case Opcode::FLAT_LOAD_DWORD:
 		case Opcode::FLAT_LOAD_DWORDX2:
 		case Opcode::FLAT_LOAD_DWORDX3:
@@ -692,10 +698,12 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::DS_OR_RTN_B32:
 		case Opcode::DS_XOR_B32:
 		case Opcode::DS_XOR_RTN_B32:
+		case Opcode::DS_MSKOR_B32:
 		case Opcode::DS_WRXCHG_RTN_B32:
 		case Opcode::DS_MIN_F32:
 		case Opcode::DS_MAX_F32:
 		case Opcode::DS_SWIZZLE_B32:
+		case Opcode::DS_PERMUTE_B32:
 		case Opcode::DS_BPERMUTE_B32:
 		case Opcode::DS_READ_I8:
 		case Opcode::DS_READ_U8:

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

@@ -186,10 +186,14 @@ enum class Opcode {
 	V_RCP_F32,
 	V_RCP_IFLAG_F32,
 	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,
@@ -220,6 +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,
@@ -247,6 +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,
@@ -445,6 +454,7 @@ enum class Opcode {
 	V_CMP_EQ_I64,
 	V_CMP_LT_I64,
 	V_CMP_LE_I64,
+	V_CMP_NE_I64,
 	V_CMP_LT_U64,
 	V_CMP_EQ_U64,
 	V_CMP_LE_U64,
@@ -522,6 +532,7 @@ enum class Opcode {
 	FLAT_LOAD_SBYTE,
 	FLAT_LOAD_USHORT,
 	FLAT_LOAD_SSHORT,
+	FLAT_LOAD_SHORT_D16,
 	FLAT_LOAD_DWORD,
 	FLAT_LOAD_DWORDX2,
 	FLAT_LOAD_DWORDX3,
@@ -556,10 +567,12 @@ enum class Opcode {
 	DS_OR_RTN_B32,
 	DS_XOR_B32,
 	DS_XOR_RTN_B32,
+	DS_MSKOR_B32,
 	DS_WRXCHG_RTN_B32,
 	DS_MIN_F32,
 	DS_MAX_F32,
 	DS_SWIZZLE_B32,
+	DS_PERMUTE_B32,
 	DS_BPERMUTE_B32,
 	DS_CONSUME,
 	DS_APPEND,
@@ -625,6 +638,7 @@ enum class Opcode {
 
 	S_NOP,
 	S_WAITCNT,
+	S_WAITCNT_VSCNT,
 	S_WAITCNT_DEPCTR,
 	S_BARRIER,
 	S_BRANCH,

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},
@@ -123,6 +126,7 @@ constexpr OpcodeMap VOP1_OPCODE_LIST[] = {
     {0x39u, Opcode::V_FFBH_U32},
     {0x3au, Opcode::V_FFBL_B32},
     {0x3bu, Opcode::V_FFBH_I32},
+    {0x3eu, Opcode::V_FRACT_F64},
     {0x3fu, Opcode::V_FREXP_EXP_I32_F32},
     {0x40u, Opcode::V_FREXP_MANT_F32},
     {0x42u, Opcode::V_MOVRELD_B32},
@@ -155,6 +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},
@@ -182,6 +189,7 @@ constexpr OpcodeMap VOP3_ENCODED_VOP1_OPCODE_LIST[] = {
     {0x39u, Opcode::V_FFBH_U32},
     {0x3au, Opcode::V_FFBL_B32},
     {0x3bu, Opcode::V_FFBH_I32},
+    {0x3eu, Opcode::V_FRACT_F64},
     {0x3fu, Opcode::V_FREXP_EXP_I32_F32},
     {0x40u, Opcode::V_FREXP_MANT_F32},
     {0x42u, Opcode::V_MOVRELD_B32},
@@ -254,6 +262,7 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0xc5u, Opcode::V_CMP_NE_U32},         {0xc6u, Opcode::V_CMP_GE_U32},
     {0xc7u, Opcode::V_CMP_T_U32},          {0xa2u, Opcode::V_CMP_EQ_I64, false},
     {0xa1u, Opcode::V_CMP_LT_I64, false},  {0xa3u, Opcode::V_CMP_LE_I64, false},
+    {0xa5u, Opcode::V_CMP_NE_I64, false},
     {0xb5u, Opcode::V_CMPX_NE_I64, false}, {0xd1u, Opcode::V_CMPX_LT_U32},
     {0xd2u, Opcode::V_CMPX_EQ_U32},        {0xd3u, Opcode::V_CMPX_LE_U32},
     {0xd4u, Opcode::V_CMPX_GT_U32},        {0xd5u, Opcode::V_CMPX_NE_U32},
@@ -289,6 +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},
@@ -472,11 +483,23 @@ bool IsNativeVop3B16BinaryOpcode(Opcode opcode) {
 bool IsVop1FloatResultOpcode(Opcode opcode);
 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_TRUNC_F64 || opcode == Opcode::V_CEIL_F64 ||
+	       opcode == Opcode::V_FLOOR_F64;
+}
+
 bool IsVop1FloatSourceOpcode(Opcode opcode) {
 	switch (opcode) {
 		case Opcode::V_CVT_F32_F64:
 		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:
@@ -551,6 +574,7 @@ constexpr Vop1SdwaRule VOP1_SDWA_RULES[] = {
     // partial destination may combine with any source selector.
     {Opcode::V_NOT_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), SdwaSelBytes() | SdwaSelWords(),
      SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), false},
+    {Opcode::V_BFREV_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
     {Opcode::V_FFBL_B32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
     {Opcode::V_FFBH_U32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), 0, 0, false},
     {Opcode::V_CVT_F32_F16, SdwaSelWords() | SdwaSelFull(), 0, 0, true},
@@ -647,6 +671,10 @@ bool SupportsVop1Clamp(Opcode opcode) {
 }
 
 bool ValidateVop1Sdwa(Instruction& inst, uint32_t opcode, uint32_t modifier) {
+	if (IsVop1Float64Opcode(inst.opcode)) {
+		SetUnsupported(inst, Family::VOP1, opcode, "FP64 instructions do not support SDWA");
+		return false;
+	}
 	const auto dst_sel   = (modifier >> 8u) & 0x7u;
 	const auto dst_u     = (modifier >> 11u) & 0x3u;
 	const auto clamp     = (modifier >> 13u) & 0x1u;
@@ -735,9 +763,7 @@ void ApplyDppModifier(Operand& operand, uint32_t modifier, uint32_t encoding) {
 
 void DecodeVop1Dpp(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
                    uint32_t opcode, uint32_t vdst, Instruction& inst) {
-	if (inst.opcode == Opcode::V_CVT_F64_I32 || inst.opcode == Opcode::V_CVT_F32_F64 ||
-	    inst.opcode == Opcode::V_CVT_F64_F32 || inst.opcode == Opcode::V_CVT_F64_U32 ||
-	    inst.opcode == Opcode::V_RCP_F64) {
+	if (IsVop1Float64Opcode(inst.opcode)) {
 		SetUnsupported(inst, Family::VOP1, opcode, "FP64 instructions do not support DPP");
 		return;
 	}
@@ -774,6 +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:
@@ -1031,6 +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      = {};
@@ -1038,6 +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;
@@ -1100,8 +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;
@@ -1152,7 +1185,8 @@ VopcSdwaFields DecodeVopcSdwaFields(uint32_t modifier) {
 }
 
 bool SupportsVopcSdwa(Opcode opcode) {
-	return opcode != Opcode::UNSUPPORTED && opcode != Opcode::V_CMP_EQ_F64 &&
+	return opcode != Opcode::UNSUPPORTED && opcode != Opcode::V_CMP_NE_I64 &&
+	       opcode != Opcode::V_CMP_EQ_F64 &&
 	       opcode != Opcode::V_CMP_LE_F64 && opcode != Opcode::V_CMPX_LE_F64 &&
 	       opcode != Opcode::V_CMPX_GE_F64;
 }
@@ -1221,6 +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:
@@ -1370,9 +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:
@@ -1403,6 +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:
@@ -1551,6 +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/Compare.cpp

@@ -31,6 +31,7 @@ void Translator::EmitIntegerCompare(const Decoder::Instruction& inst, IR::ValueO
 	const bool signed_64 = opcode == IR::ValueOpcode::SLessThan64 ||
 	                       opcode == IR::ValueOpcode::SLessThanEqual64 ||
 	                       inst.opcode == Decoder::Opcode::V_CMP_EQ_I64 ||
+	                       inst.opcode == Decoder::Opcode::V_CMP_NE_I64 ||
 	                       inst.opcode == Decoder::Opcode::V_CMPX_NE_I64;
 	const auto read = [&](const Decoder::Operand& operand) {
 		// RDNA2 expands signed 64-bit integer literals by sign extension.

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

@@ -181,8 +181,12 @@ void Translator::EmitControlNop() {
 	ir.Emit(IR::ValueOpcode::ControlNop);
 }
 
-void Translator::EmitWaitcnt() {
-	ir.Emit(IR::ValueOpcode::Waitcnt);
+void Translator::S_WAITCNT_VSCNT(const Decoder::Instruction& inst) {
+	const auto count = inst.src1.value & 63u;
+	if (inst.src0.kind != Decoder::OperandKind::Null || (count != 0u && count != 63u)) {
+		EXIT("unsupported partial or register-based S_WAITCNT_VSCNT at 0x%08x", inst.pc);
+	}
+	if (count == 0u) ir.Emit(IR::ValueOpcode::StoreCompletion);
 }
 
 void Translator::S_BARRIER() {
@@ -252,13 +256,14 @@ void Translator::ScalarSelect64(const Decoder::Instruction& inst,
 	const auto selected_mask_valid =
 	    IR::U1(ir.Emit(IR::ValueOpcode::SelectU1,
 	                   {condition, ReadMaskValid(inst.src0), ReadMaskValid(false_source)}));
-	if (IsExecOrVcc(inst.dst)) {
-		WriteMask(inst.dst, selected_mask, true);
-		return;
-	}
 	WriteU32Pair(inst.dst,
 	             {ir.Select(condition, lhs[0], rhs[0]), ir.Select(condition, lhs[1], rhs[1])});
-	if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
+	// Scalar selects preserve the raw pair even when EXEC/VCC holds an address.
+	if (inst.dst.kind == Decoder::OperandKind::ExecLo) {
+		ir.SetExec(selected_mask);
+	} else if (inst.dst.kind == Decoder::OperandKind::VccLo) {
+		ir.SetVcc(selected_mask);
+	} else if (inst.dst.kind == Decoder::OperandKind::Sgpr) {
 		const auto dst = static_cast<IR::ScalarReg>(inst.dst.reg);
 		ir.SetThreadBitScalarReg(dst, selected_mask);
 		ir.SetScalarMaskTag(dst, selected_mask_valid);

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

@@ -483,10 +483,11 @@ void Translator::S_BITCMP_B64(const Decoder::Instruction& inst, bool expected) {
 	WriteCompareResult(inst.dst, ir.IEqual(bit, IR::U32(IR::Value(expected ? 1u : 0u))));
 }
 
-void Translator::V_ALIGNBIT_B32(const Decoder::Instruction& inst) {
-	const auto hi      = ReadU32(inst.src0);
-	const auto lo      = ReadU32(inst.src1);
-	const auto shift   = ir.BitwiseAnd(ReadU32(inst.src2), IR::U32(IR::Value(31u)));
+void Translator::V_ALIGN_B32(const Decoder::Instruction& inst, bool byte_offset) {
+	const auto hi = ReadU32(inst.src0);
+	const auto lo = ReadU32(inst.src1);
+	auto shift    = ir.BitwiseAnd(ReadU32(inst.src2), IR::U32(IR::Value(byte_offset ? 3u : 31u)));
+	if (byte_offset) shift = ir.ShiftLeftLogical(shift, IR::U32(IR::Value(3u)));
 	const auto lo_part = ir.ShiftRightLogical(lo, shift);
 	const auto inverse =
 	    ir.BitwiseAnd(ir.ISub(IR::U32(IR::Value(32u)), shift), IR::U32(IR::Value(31u)));
@@ -496,21 +497,6 @@ void Translator::V_ALIGNBIT_B32(const Decoder::Instruction& inst) {
 	WriteOperand(DestinationOperand(inst), ir.BitwiseOr(lo_part, hi_part));
 }
 
-void Translator::V_ALIGNBYTE_B32(const Decoder::Instruction& inst) {
-	const auto hi           = ReadU32(inst.src0);
-	const auto lo           = ReadU32(inst.src1);
-	const auto byte_offset  = ir.BitwiseAnd(ReadU32(inst.src2), IR::U32(IR::Value(31u)));
-	const auto bit_offset   = ir.ShiftLeftLogical(byte_offset, IR::U32(IR::Value(3u)));
-	const auto concatenated = ir.ConstructU64(lo, hi);
-	const auto shifted =
-	    IR::U64(ir.Emit(IR::ValueOpcode::ShiftRightLogical64,
-	                    {concatenated, ir.BitwiseAnd(bit_offset, IR::U32(IR::Value(63u)))}));
-	const auto in_range =
-	    IR::U1(ir.Emit(IR::ValueOpcode::ULessThan32, {byte_offset, IR::Value(8u)}));
-	WriteOperand(DestinationOperand(inst),
-	             ir.Select(in_range, ExtractU64(shifted)[0], IR::U32(IR::Value(0u))));
-}
-
 void Translator::V_LSHL_ADD_U32(const Decoder::Instruction& inst) {
 	const auto shift  = ir.BitwiseAnd(ReadU32(inst.src1), IR::U32(IR::Value(31u)));
 	const auto result = ir.IAdd(ir.ShiftLeftLogical(ReadU32(inst.src0), shift), ReadU32(inst.src2));

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

@@ -239,9 +239,6 @@ Decoder::Operand MemorySourceAt(const Decoder::Instruction& decoded, uint32_t in
 			case Decoder::Opcode::DS_READ_ADDTID_B32: return MakeM0Operand();
 			case Decoder::Opcode::DS_WRITE_ADDTID_B32:
 				return index == 0u ? decoded.src1 : MakeM0Operand();
-			case Decoder::Opcode::DS_MIN_F32:
-			case Decoder::Opcode::DS_MAX_F32:
-				return index == 0u ? decoded.src1 : index == 1u ? decoded.src0 : decoded.src2;
 			case Decoder::Opcode::DS_WRITE_B8:
 			case Decoder::Opcode::DS_WRITE_B16:
 			case Decoder::Opcode::DS_WRITE_B8_D16_HI:
@@ -255,12 +252,7 @@ Decoder::Operand MemorySourceAt(const Decoder::Instruction& decoded, uint32_t in
 			case Decoder::Opcode::DS_WRITE2_B64:
 			case Decoder::Opcode::DS_WRITE2ST64_B64:
 				return index == 0u ? decoded.src1 : index == 1u ? decoded.src0 : decoded.src2;
-			default:
-				if (decoded.opcode >= Decoder::Opcode::DS_ADD_U32 &&
-				    decoded.opcode <= Decoder::Opcode::DS_WRXCHG_RTN_B32) {
-					return index == 0u ? decoded.src1 : decoded.src0;
-				}
-				return decoded.src0;
+			default: return decoded.src0;
 		}
 	}
 	return DecodedSourceAt(decoded, index);
@@ -554,7 +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));
@@ -580,12 +575,13 @@ void Translator::IMAGE_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode
 void Translator::DS_ATOMIC(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
                            bool returns_value) {
 	const auto memory  = MemoryInfoFromDecoded(inst);
-	const auto address = ReadU32(MemorySourceAt(inst, 1));
-	const auto data_src = MemorySourceAt(inst, 0);
+	const auto address = ReadU32(inst.src0);
 	const IR::Value value =
-	    memory.data_dwords == 2u ? IR::Value(ReadU64(data_src)) : IR::Value(ReadU32(data_src));
-	const auto result  = ir.Emit(opcode, {address, value, ir.GetExec()},
-	                             AddMemoryInfo(memory, inst.pc));
+	    memory.data_dwords == 2u ? IR::Value(ReadU64(inst.src1)) : IR::Value(ReadU32(inst.src1));
+	const auto flags = AddMemoryInfo(memory, inst.pc);
+	const auto result = inst.src_count == 3u
+	                        ? ir.Emit(opcode, {address, value, ReadU32(inst.src2), ir.GetExec()}, flags)
+	                        : ir.Emit(opcode, {address, value, ir.GetExec()}, flags);
 	if (returns_value) {
 		WriteOperand(inst.dst, result);
 	}
@@ -857,14 +853,6 @@ void Translator::DS_WRITE2(const Decoder::Instruction& inst) {
 	}
 }
 
-void Translator::DS_MINMAX_F32(const Decoder::Instruction& inst, IR::ValueOpcode opcode) {
-	const auto memory = MemoryInfoFromDecoded(inst);
-	ir.Emit(opcode,
-	        {ReadU32(MemorySourceAt(inst, 1)), ReadU32(MemorySourceAt(inst, 0)),
-	         ReadU32(MemorySourceAt(inst, 2)), ir.GetExec()},
-	        AddMemoryInfo(memory, inst.pc));
-}
-
 void Translator::DS_APPEND_CONSUME(const Decoder::Instruction& inst, IR::ValueOpcode opcode) {
 	const auto memory = MemoryInfoFromDecoded(inst);
 	WriteOperand(inst.dst, ir.Emit(opcode,
@@ -894,9 +882,10 @@ void Translator::DS_SWIZZLE_B32(const Decoder::Instruction& inst) {
 	                                ReadU32(MemorySourceAt(inst, 1)), ir.GetExec()}));
 }
 
-void Translator::DS_BPERMUTE_B32(const Decoder::Instruction& inst) {
+void Translator::DS_PERMUTE(const Decoder::Instruction& inst, bool backward) {
 	const auto address = ir.IAdd(ReadU32(inst.src0), IR::U32(IR::Value(inst.offset)));
-	WriteOperand(inst.dst, ir.Emit(IR::ValueOpcode::BpermuteU32,
+	WriteOperand(inst.dst, ir.Emit(backward ? IR::ValueOpcode::BpermuteU32
+	                                       : IR::ValueOpcode::PermuteU32,
 	                               {ReadU32(inst.src1), address, ir.GetExec()}));
 }
 
@@ -904,7 +893,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
 		case Decoder::Opcode::IMAGE_BVH_INTERSECT_RAY: {
 			const auto result = ir.Emit(IR::ValueOpcode::BvhIntersect,
-			    {ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()});
+			    {ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()},
+			    inst.image_address_components - 10u);
 			for (uint32_t component = 0; component < 4; ++component) {
 				WriteOperand(OffsetOperand(inst.dst, component),
 				    ir.Emit(IR::ValueOpcode::CompositeExtractU32x4, {result, IR::Value(component)}));
@@ -1046,30 +1036,38 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicXor32, false);
 		case Decoder::Opcode::DS_XOR_RTN_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicXor32, true);
+		case Decoder::Opcode::DS_MSKOR_B32:
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicMaskedOr32, false);
 		case Decoder::Opcode::DS_WRXCHG_RTN_B32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicSwap32, true);
 
 		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:
@@ -1079,6 +1077,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::FLAT_LOAD_SBYTE:
 		case Decoder::Opcode::FLAT_LOAD_USHORT:
 		case Decoder::Opcode::FLAT_LOAD_SSHORT:
+		case Decoder::Opcode::FLAT_LOAD_SHORT_D16:
 		case Decoder::Opcode::FLAT_LOAD_DWORD:
 		case Decoder::Opcode::FLAT_LOAD_DWORDX2:
 		case Decoder::Opcode::FLAT_LOAD_DWORDX3:
@@ -1108,11 +1107,12 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::IMAGE_GATHER4H: return IMAGE_GATHER(inst);
 
 		case Decoder::Opcode::DS_MIN_F32:
-			return DS_MINMAX_F32(inst, IR::ValueOpcode::SharedAtomicFMin32);
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicFMin32, false);
 		case Decoder::Opcode::DS_MAX_F32:
-			return DS_MINMAX_F32(inst, IR::ValueOpcode::SharedAtomicFMax32);
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicFMax32, false);
 		case Decoder::Opcode::DS_SWIZZLE_B32: return DS_SWIZZLE_B32(inst);
-		case Decoder::Opcode::DS_BPERMUTE_B32: return DS_BPERMUTE_B32(inst);
+		case Decoder::Opcode::DS_PERMUTE_B32: return DS_PERMUTE(inst, false);
+		case Decoder::Opcode::DS_BPERMUTE_B32: return DS_PERMUTE(inst, true);
 		case Decoder::Opcode::DS_CONSUME:
 			return DS_APPEND_CONSUME(inst, IR::ValueOpcode::DataConsume);
 		case Decoder::Opcode::DS_APPEND:

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

@@ -18,7 +18,8 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_CSELECT_B64: ScalarSelect64(inst, inst.src1); return;
 		case O::S_CMOV_B64: ScalarSelect64(inst, inst.dst); return;
 		case O::S_SETREG_B32: EmitControlNop(); return;
-		case O::S_WAITCNT: EmitWaitcnt(); return;
+		case O::S_WAITCNT_VSCNT: S_WAITCNT_VSCNT(inst); return;
+		case O::S_WAITCNT: return;
 
 		case O::S_AND_SAVEEXEC_B32:
 			S_SAVEEXEC(inst, IR::ValueOpcode::LogicalAnd, false, false, false);
@@ -224,7 +225,7 @@ void Translator::EmitScalar(const Decoder::Instruction& inst) {
 		case O::S_SLEEP:
 		case O::S_SETPRIO:
 		case O::S_TRAP: EmitControlNop(); return;
-		case O::S_WAITCNT_DEPCTR: EmitWaitcnt(); return;
+		case O::S_WAITCNT_DEPCTR: return;
 		case O::S_BARRIER: S_BARRIER(); return;
 		case O::S_SENDMSG: S_SENDMSG(inst); return;
 		case O::S_TTRACEDATA: S_TTRACEDATA(); return;

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

@@ -895,8 +895,14 @@ 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: include_vector(inst.dst, 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:
 		case Decoder::Opcode::V_CMP_LE_F64:
@@ -905,6 +911,7 @@ void IncludeInstructionVectorRegisters(const Decoder::Instruction& inst, uint32_
 		case Decoder::Opcode::V_CMP_EQ_I64:
 		case Decoder::Opcode::V_CMP_LT_I64:
 		case Decoder::Opcode::V_CMP_LE_I64:
+		case Decoder::Opcode::V_CMP_NE_I64:
 		case Decoder::Opcode::V_CMP_LT_U64:
 		case Decoder::Opcode::V_CMP_EQ_U64:
 		case Decoder::Opcode::V_CMP_LE_U64:
@@ -1104,6 +1111,16 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				                                  IR::U32(IR::Value(total_threads)));
 			}
 		}
+		if (options.stage == ShaderType::Compute &&
+		    options.input_info.compute->dispatch_thread_dimensions) {
+			for (uint32_t axis = 0; axis < 3u; axis++) {
+				const auto extent = IR::U32(entry_ir.Emit(
+				    IR::ValueOpcode::GetDispatchThreadExtent, {IR::Value(axis)}));
+				initial_exec = entry_ir.LogicalAnd(
+				    initial_exec,
+				    entry_ir.ULessThan(builtin(IR::StageInputKind::GlobalInvocationId, axis), extent));
+			}
+		}
 		entry_ir.SetExec(initial_exec);
 		const auto initial_mask = entry_ir.Emit(IR::ValueOpcode::Ballot, {initial_exec});
 		entry_ir.SetExecLo(entry_ir.CompositeExtract(initial_mask, 0));

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);
@@ -105,11 +106,10 @@ private:
 	void DS_READ2(const Decoder::Instruction& inst);
 	void DS_WRITE(const Decoder::Instruction& inst);
 	void DS_WRITE2(const Decoder::Instruction& inst);
-	void DS_MINMAX_F32(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
 	void DS_APPEND_CONSUME(const Decoder::Instruction& inst, IR::ValueOpcode opcode);
 	void DS_ADDTID(const Decoder::Instruction& inst, bool write);
 	void DS_SWIZZLE_B32(const Decoder::Instruction& inst);
-	void DS_BPERMUTE_B32(const Decoder::Instruction& inst);
+	void DS_PERMUTE(const Decoder::Instruction& inst, bool backward);
 
 	IR::F32 SelectF32(IR::U1 condition, IR::F32 true_value, IR::F32 false_value);
 	IR::U32 ConvertF32ToU32Saturated(IR::F32 value, float upper_bound, float safe_upper,
@@ -208,8 +208,7 @@ private:
 	void    V_BFI_B32(const Decoder::Instruction& inst);
 	void    S_BITCMP_B32(const Decoder::Instruction& inst, bool expected);
 	void    S_BITCMP_B64(const Decoder::Instruction& inst, bool expected);
-	void    V_ALIGNBIT_B32(const Decoder::Instruction& inst);
-	void    V_ALIGNBYTE_B32(const Decoder::Instruction& inst);
+	void    V_ALIGN_B32(const Decoder::Instruction& inst, bool byte_offset);
 	void    V_LSHL_ADD_U32(const Decoder::Instruction& inst);
 	void    V_ADD_LSHL_U32(const Decoder::Instruction& inst);
 	void    V_XAD_U32(const Decoder::Instruction& inst);
@@ -229,7 +228,7 @@ private:
 	void ScalarMinMax32(const Decoder::Instruction& inst, IR::ValueOpcode value_opcode,
 	                    IR::ValueOpcode compare_opcode);
 	void EmitControlNop();
-	void EmitWaitcnt();
+	void S_WAITCNT_VSCNT(const Decoder::Instruction& inst);
 	void S_BARRIER();
 	void S_SENDMSG(const Decoder::Instruction& inst);
 	void S_TTRACEDATA();

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

@@ -106,6 +106,7 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CMP_GT_U64:
 			EmitIntegerCompare(inst, IR::ValueOpcode::UGreaterThan64, IR::Type::U64, false, false);
 			return;
+		case O::V_CMP_NE_I64:
 		case O::V_CMP_NE_U64:
 			EmitIntegerCompare(inst, IR::ValueOpcode::INotEqual64, IR::Type::U64, false, false);
 			return;
@@ -453,10 +454,14 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_RCP_F64: return FloatUnary(inst, IR::ValueOpcode::FPRecip64);
 		case O::V_RCP_IFLAG_F32: return FloatUnary(inst, IR::ValueOpcode::FPRecipIFlag32);
 		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);
@@ -469,13 +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);
@@ -591,8 +601,8 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_BFE_U32: return V_BFE_U32(inst, false);
 		case O::V_BFE_I32: return V_BFE_U32(inst, true);
 		case O::V_BFI_B32: return V_BFI_B32(inst);
-		case O::V_ALIGNBIT_B32: return V_ALIGNBIT_B32(inst);
-		case O::V_ALIGNBYTE_B32: return V_ALIGNBYTE_B32(inst);
+		case O::V_ALIGNBIT_B32: return V_ALIGN_B32(inst, false);
+		case O::V_ALIGNBYTE_B32: return V_ALIGN_B32(inst, true);
 		case O::V_LSHL_ADD_U32: return V_LSHL_ADD_U32(inst);
 		case O::V_ADD_LSHL_U32: return V_ADD_LSHL_U32(inst);
 		case O::V_XAD_U32: return V_XAD_U32(inst);

src/graphics/shader/recompiler/ir/Program.cpp

@@ -190,6 +190,18 @@ Value ResolveInvariantPhi(const ResourcePlan& program, Value value) {
 	return invariant;
 }
 
+Value ResolveActiveU32(Value value, Value active) {
+	for (uint32_t depth = 0; depth <= 32; ++depth) {
+		value = value.Resolve();
+		const auto* inst = value.TryInstruction();
+		if (inst == nullptr || inst->GetOpcode() != ValueOpcode::SelectU32 ||
+		    inst->Arg(0).Resolve() != active)
+			return value;
+		value = inst->Arg(1);
+	}
+	return {};
+}
+
 bool HasShaderMemoryWrites(const Program& program) {
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
@@ -399,7 +411,8 @@ void ValidateProgram(const Program& program, bool require_ssa) {
 					                        ValueOpcodeName(inst.GetOpcode())));
 				}
 				const auto& memory = program.memory_info[memory_index];
-				if (memory.kind != ResourceKind::ScalarBuffer) {
+				if (memory.kind != ResourceKind::ScalarBuffer &&
+				    memory.kind != ResourceKind::IndirectBuffer) {
 					return Fail(fmt::format("{} has an invalid scalar-memory resource kind",
 					                        ValueOpcodeName(inst.GetOpcode())));
 				}
@@ -457,7 +470,7 @@ void ValidateProgram(const Program& program, bool require_ssa) {
 				}
 				if (memory.kind == ResourceKind::IndirectBuffer &&
 				    !memory.SupportsIndirectBufferLoad(inst.GetOpcode())) {
-					return Fail("indirect buffer requires a raw DWORD x2/x3/x4 load");
+					return Fail("indirect buffer requires a scalar, raw DWORD x1/x2/x3/x4, or formatted X load");
 				}
 				if (buffer_components > 1u &&
 				    (!vector_buffer || memory.data_bits != 32u ||

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

@@ -73,9 +73,13 @@ struct MemoryInfo {
 	bool                    planning_only                                         = false;
 
 	[[nodiscard]] bool SupportsIndirectBufferLoad(ValueOpcode opcode) const {
-		return !formatted && !typed && data_bits == 32u &&
-		       (opcode == ValueOpcode::LoadBufferU32x2 || opcode == ValueOpcode::LoadBufferU32x3 ||
-		        opcode == ValueOpcode::LoadBufferU32x4);
+		return !typed && data_bits == 32u &&
+		       (formatted ? opcode == ValueOpcode::LoadBufferU32
+		                  : opcode == ValueOpcode::ReadConstBuffer ||
+		                        opcode == ValueOpcode::LoadBufferU32 ||
+		                        opcode == ValueOpcode::LoadBufferU32x2 ||
+		                        opcode == ValueOpcode::LoadBufferU32x3 ||
+		                        opcode == ValueOpcode::LoadBufferU32x4);
 	}
 
 	bool operator==(const MemoryInfo& other) const = default;
@@ -97,6 +101,7 @@ struct ExportInfo {
 
 struct BufferResource {
 	static constexpr uint32_t NoImageAlias = UINT32_MAX;
+	static constexpr uint32_t NoIndirectBuffer = UINT32_MAX;
 
 	uint32_t               source             = 0;
 	uint32_t               first_use_pc       = 0;
@@ -110,6 +115,10 @@ struct BufferResource {
 	bool                   atomic             = false;
 	bool                   formatted          = false;
 	bool                   scalar             = false;
+	uint32_t               indirect_root              = NoIndirectBuffer;
+	uint32_t               indirect_mapping_offset    = 0;
+	uint32_t               indirect_search_iterations = 0;
+	std::vector<uint32_t>  indirect_resources;
 
 	bool operator==(const BufferResource& other) const = default;
 };
@@ -190,6 +199,7 @@ enum class StageInputKind {
 	BaryCoordSmoothCentroid,
 	BaryCoordNoPerspective,
 	WorkgroupId,
+	NumWorkgroups,
 	LocalInvocationId,
 	LocalInvocationIndex,
 	GlobalInvocationId,
@@ -292,11 +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;
@@ -421,6 +432,7 @@ struct DescriptorBinding {
 
 struct BindingLayout {
 	uint32_t                       push_data_start_dword = PushData::NoStart;
+	uint32_t                       dispatch_thread_dword = PushData::NoStart;
 	uint32_t                       memory_offset_dword = 0;
 	uint32_t                       memory_offset_count = 0;
 	std::vector<uint32_t>          user_data_registers;
@@ -472,22 +484,26 @@ struct BlockInfo {
 };
 
 struct DescriptorSource {
-	struct IndirectImage {
+	struct IndirectDescriptor {
 		uint32_t material_source = UINT32_MAX;
 		uint32_t table_source    = 0;
 		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;
 		Value    selector_mask;
 		std::vector<uint32_t> sources;
 
-		bool operator==(const IndirectImage& other) const = default;
+		bool operator==(const IndirectDescriptor& other) const = default;
 	};
 
 	std::array<Value, 8>         dwords {};
 	uint32_t                     dword_count = 0;
-	std::optional<IndirectImage> indirect_image;
+	std::optional<IndirectDescriptor> indirect_descriptor;
 
 	bool operator==(const DescriptorSource& other) const = default;
 };
@@ -610,6 +626,7 @@ void  ValidateProgram(const Program& program, bool require_ssa);
 void  ResolveControlFlowIdentities(Program& program);
 bool  EquivalentValue(const ResourcePlan& program, Value left, Value right);
 Value ResolveInvariantPhi(const ResourcePlan& program, Value value);
+Value ResolveActiveU32(Value value, Value active);
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR
 

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

@@ -93,6 +93,7 @@ bool HasSideEffects(ValueOpcode opcode) {
 		case ValueOpcode::SetAttribute:
 		case ValueOpcode::SetTessellationAttribute:
 		case ValueOpcode::MeshAllocate:
+		case ValueOpcode::StoreCompletion:
 		case ValueOpcode::Barrier: return true;
 		default: return false;
 	}
@@ -177,7 +178,8 @@ SharedAccess SharedAccessOf(ValueOpcode opcode) {
 		case ValueOpcode::SharedAtomicAnd32:
 		case ValueOpcode::SharedAtomicOr32:
 		case ValueOpcode::SharedAtomicOr64:
-		case ValueOpcode::SharedAtomicXor32: return SharedAccess::Atomic;
+		case ValueOpcode::SharedAtomicXor32:
+		case ValueOpcode::SharedAtomicMaskedOr32: return SharedAccess::Atomic;
 		case ValueOpcode::DataAppend: return SharedAccess::Append;
 		case ValueOpcode::DataConsume: return SharedAccess::Consume;
 		default: return SharedAccess::None;
@@ -222,15 +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

@@ -8,6 +8,7 @@ VALUE_OPCODE(ConditionRef, U1, U1)
 
 // Translation and register-state pseudo operations.
 VALUE_OPCODE(GetUserData, U32, ScalarReg)
+VALUE_OPCODE(GetDispatchThreadExtent, U32, U32)
 VALUE_OPCODE(GetShaderBase, U64)
 VALUE_OPCODE(MeshDrawParameter, U32, U32)
 VALUE_OPCODE(MeshAllocate, Void, U32)
@@ -178,6 +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)
@@ -185,6 +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)
@@ -201,7 +205,11 @@ 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)
 
 VALUE_OPCODE(LaneId, U32)
 VALUE_OPCODE(Ballot, U32x4, U1)
@@ -209,6 +217,7 @@ VALUE_OPCODE(ReadFirstLane, U32, U32, U1)
 VALUE_OPCODE(ReadLane, U32, U32, U32)
 VALUE_OPCODE(WriteLane, U32, U32, U32, U32)
 VALUE_OPCODE(Permlane16U32, U32, U32, U32, U32, U1)
+VALUE_OPCODE(PermuteU32, U32, U32, U32, U1)
 VALUE_OPCODE(BpermuteU32, U32, U32, U32, U1)
 
 // Typed shader resources and memory operations. Descriptor state is represented by SSA values;
@@ -271,8 +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)
@@ -287,6 +296,7 @@ VALUE_OPCODE(SharedAtomicAnd32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicOr32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicOr64, Void, U32, U64, U1)
 VALUE_OPCODE(SharedAtomicXor32, U32, U32, U32, U1)
+VALUE_OPCODE(SharedAtomicMaskedOr32, Void, U32, U32, U32, U1)
 VALUE_OPCODE(DataAppend, U32, U32, U1)
 VALUE_OPCODE(DataConsume, U32, U32, U1)
 VALUE_OPCODE(SwizzleU32, U32, U32, U32, U1)
@@ -300,15 +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)
@@ -316,7 +332,7 @@ VALUE_OPCODE(GetInterpolationParameter, U32, U32, U32, U32)
 VALUE_OPCODE(SetAttribute, Void, U32x4, U1)
 
 VALUE_OPCODE(ControlNop, Void)
-VALUE_OPCODE(Waitcnt, Void)
+VALUE_OPCODE(StoreCompletion, Void)
 VALUE_OPCODE(Barrier, Void)
 VALUE_OPCODE(Sendmsg, Void)
 VALUE_OPCODE(TtraceData, Void)

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

@@ -15,11 +15,16 @@ namespace {
 	std::abort();
 }
 
-std::vector<uint32_t> CollectUserData(const Program& program) {
+void CollectShaderData(const Program& program, BindingLayout& layout) {
 	std::array<bool, NumScalarRegs> registers {};
+	bool uses_dispatch_threads = false;
 	for (const auto* block: program.blocks) {
 		for (const auto& inst: *block) {
-			if (inst.GetOpcode() != ValueOpcode::GetUserData || !inst.HasUses()) {
+			if (!inst.HasUses()) {
+				continue;
+			}
+			uses_dispatch_threads |= inst.GetOpcode() == ValueOpcode::GetDispatchThreadExtent;
+			if (inst.GetOpcode() != ValueOpcode::GetUserData) {
 				continue;
 			}
 			if (inst.Arg(0).GetType() != Type::ScalarReg) {
@@ -32,13 +37,16 @@ std::vector<uint32_t> CollectUserData(const Program& program) {
 			registers[index] = true;
 		}
 	}
-	std::vector<uint32_t> result;
 	for (uint32_t index = 0; index < registers.size(); index++) {
 		if (registers[index]) {
-			result.push_back(index);
+			layout.user_data_registers.push_back(index);
 		}
 	}
-	return result;
+	layout.memory_offset_dword = static_cast<uint32_t>(layout.user_data_registers.size());
+	if (uses_dispatch_threads) {
+		layout.dispatch_thread_dword = layout.memory_offset_dword;
+		layout.memory_offset_dword += 3u;
+	}
 }
 
 void AddBinding(BindingLayout& layout, DescriptorBindingKind kind,
@@ -48,14 +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;
@@ -66,14 +75,19 @@ 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;
+				for (const auto child: program.info.buffers[memory.resource].indirect_resources) {
+					live_buffers.at(child) = true;
+				}
 			}
 		}
 	}
@@ -82,20 +96,22 @@ bool CollectMemoryResources(const Program& program, std::vector<uint32_t>& buffe
 			buffers.push_back(i);
 		}
 	}
-	return uses_gds;
+	return shared;
 }
 
 bool UsesFlattenedSrt(const Program& program) {
+	const auto uses_mapping = [](const auto& resource) {
+		return resource.indirect_search_iterations != 0u;
+	};
 	return std::ranges::any_of(program.blocks, [](const Block* block) {
 		return std::ranges::any_of(*block, [](const Inst& inst) {
 			return inst.GetOpcode() == ValueOpcode::ReadConst;
 		});
-	}) || std::ranges::any_of(program.info.images, [](const ImageResource& image) {
-		return image.indirect_search_iterations != 0u;
-	});
+	}) || std::ranges::any_of(program.info.buffers, uses_mapping) ||
+	       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"
@@ -103,9 +119,8 @@ void AllocateBindings(Program& program, uint32_t push_data_start_dword) {
 	}
 	BindingLayout next;
 	std::vector<uint32_t> buffers;
-	const bool            uses_gds = CollectMemoryResources(program, buffers);
-	next.user_data_registers = CollectUserData(program);
-	next.memory_offset_dword = static_cast<uint32_t>(next.user_data_registers.size());
+	const auto shared = CollectMemoryResources(program, buffers);
+	CollectShaderData(program, next);
 	next.memory_offset_count       = static_cast<uint32_t>(buffers.size());
 	next.push_data_start_dword =
 	    PushData::StartFor(push_data_start_dword, next.ShaderDataDwords());
@@ -148,9 +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/ConstantPropagation.cpp

@@ -199,6 +199,11 @@ void FoldInstruction(Block& block, Block::iterator instruction,
                       std::unordered_set<Inst*>& lowered_ancillary) {
 	auto& inst = *instruction;
 	switch (inst.GetOpcode()) {
+		case ValueOpcode::StoreBufferU32: {
+			const auto data = ResolveActiveU32(Arg(inst, 4), Arg(inst, 5));
+			if (!data.IsEmpty() && data != inst.Arg(4)) inst.SetArg(4, data);
+			return;
+		}
 		case ValueOpcode::Phi: FoldPhi(inst); return;
 		case ValueOpcode::SelectU1:
 			if (!FoldSelect(inst) && IsImmediate(Arg(inst, 2), Type::U1) &&
@@ -503,6 +508,17 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 			}
 			return;
 		}
+		case ValueOpcode::BitReverse32: {
+			const auto value = Arg(inst, 0);
+			if (IsImmediate(value, Type::U32)) {
+				uint32_t source = value.U32(), reversed = 0;
+				for (uint32_t bit = 0; bit < 32u; ++bit, source >>= 1u) {
+					reversed = (reversed << 1u) | (source & 1u);
+				}
+				Replace(inst, Value(reversed));
+			}
+			return;
+		}
 		case ValueOpcode::BitCount32: {
 			const auto value = Arg(inst, 0);
 			if (IsImmediate(value, Type::U32)) {

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

@@ -20,7 +20,7 @@ namespace Libs::Graphics::ShaderRecompiler::IR {
 namespace {
 
 constexpr uint64_t AddressMask            = 0x0000ffffffffffffull;
-constexpr uint64_t MaxIndirectImageProbes = 65536u;
+constexpr uint64_t MaxIndirectDescriptorProbes = 65536u;
 
 bool SpecializationFail(std::string_view message) {
 	std::fprintf(stderr, "shader resource specialization failed: %.*s\n",
@@ -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
@@ -224,14 +225,44 @@ bool ReadScalarTable(uint64_t base, uint64_t size, uint64_t dynamic_offset,
 	       runtime.read_specialization_memory(runtime.userdata, address, prefix);
 }
 
-bool MaterializeIndirectImage(const ResourcePlan& program,
-                              const DescriptorSource::IndirectImage& indirect,
-                              uint32_t image_index,
-                              const SrtRuntime& runtime, SrtWalker& clean,
-                              ResourceSnapshot& snapshot,
-                              ResourceSpecialization& specialization) {
+bool NormalizeIndirectStoreBuffer(DescriptorValue& value) {
+	ShaderBufferResource descriptor;
+	if (!DecodeBufferDescriptor(value, descriptor)) return false;
+	if (descriptor.Type() != 0u || descriptor.Format() == Prospero::BufferFormat::kInvalid ||
+	    descriptor.Base48() == 0u || descriptor.GetSize() == 0u) {
+		value.dwords.fill(0u);
+	} else if (descriptor.Stride() != 1u || descriptor.SwizzleEnabled() || descriptor.AddTid() ||
+	           descriptor.OutOfBounds() != 3u) {
+		return SpecializationFail("indirect buffer stores require linear byte bounds");
+	}
+	return true;
+}
+
+template <typename Specialization, typename Normalize>
+bool MaterializeIndirectDescriptor(const ResourcePlan&                         program,
+                                   const DescriptorSource::IndirectDescriptor& indirect,
+                                   uint32_t resource_index, uint32_t dword_count,
+                                   const SrtRuntime& runtime, SrtWalker& clean,
+                                   ResourceSnapshot&             snapshot,
+                                   std::vector<DescriptorValue>& descriptors,
+                                   std::vector<Specialization>&  specializations,
+                                   uint32_t maximum_resources, Normalize&& normalize) {
+	const auto  descriptor_bytes = dword_count * sizeof(uint32_t);
 	const auto& sources = indirect.sources;
 	auto& keys = program.material_keys;
+	const auto read_keys = [&](const ShaderBufferResource& material, uint64_t first,
+	                           uint64_t step, uint64_t count) {
+		keys.resize(count);
+		if (step == 4u) {
+			return ReadScalarTable(material.Base48(), material.GetSize(), first, runtime, keys);
+		}
+		for (auto& key: keys) {
+			if (!ReadScalarTable(material.Base48(), material.GetSize(), first, runtime, {&key, 1}))
+				return false;
+			first += step;
+		}
+		return true;
+	};
 	uint64_t table_base = 0;
 	uint64_t table_size = UINT64_MAX; // Scalar addresses have no buffer descriptor bounds.
 	if (sources.empty()) {
@@ -254,15 +285,41 @@ bool MaterializeIndirectImage(const ResourcePlan& program,
 		}
 		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 > MaxIndirectImageProbes ||
-			    uint64_t {indirect.table_offset} + uint64_t {key_count} * 32u > 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);
 			std::iota(keys.begin(), keys.end(), 0u);
+		} else if (!indirect.selector_first.IsEmpty()) {
+			ShaderBufferResource material;
+			uint32_t             first = 0, count = 0;
+			if (!DecodeBufferDescriptor(material_value, material) || material.Type() != 0u ||
+			    (indirect.selector_shift != 0u && (material.Base48() & 3u) != 0u) ||
+			    material.SwizzleEnabled() ||
+			    material.AddTid() || material.OutOfBounds() != 0u ||
+			    uint64_t {indirect.selector_offset} + 4u > material.Stride() ||
+			    !clean.Evaluate(indirect.selector_first, first) ||
+			    !clean.Evaluate(indirect.key_count, count) || count > MaxIndirectDescriptorProbes ||
+			    uint64_t {first} + count > material.NumRecords())
+				return false;
+			if (count != 0u && (uint64_t {first} + count - 1u) * material.Stride() +
+			                           indirect.selector_offset + 4u >
+			                       uint64_t {UINT32_MAX} + 1u)
+				return false;
+			if (!read_keys(material, uint64_t {first} * material.Stride() + indirect.selector_offset,
+			               material.Stride(), count)) return false;
 		} else if (!indirect.selector_mask.IsEmpty()) {
 			uint32_t mask = 0;
 			uint32_t count = 0;
@@ -286,12 +343,9 @@ bool MaterializeIndirectImage(const ResourcePlan& program,
 				keys.push_back(key);
 				mask &= mask - 1u;
 			}
-			std::ranges::sort(keys);
-			keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
 		} else {
 			ShaderBufferResource material;
-			if (!DecodeBufferDescriptor(material_value, material) || table_value.dword_count != 4u ||
-			    material.Stride() != indirect.selector_stride) {
+			if (!DecodeBufferDescriptor(material_value, material) || table_value.dword_count != 4u) {
 				return false;
 			}
 			// The first aligned offset includes the immediate added after shader U32 arithmetic.
@@ -301,70 +355,68 @@ bool MaterializeIndirectImage(const ResourcePlan& program,
 			const auto size = material.GetSize();
 			const auto limit = std::min(first + (uint64_t {1} << 32u) - step, size >= 4u ? size - 4u : 0u);
 			const auto probe_count = size >= 4u && first <= limit ? (limit - first) / step + 1u : 0u;
-			if (probe_count > MaxIndirectImageProbes) {
+			if (probe_count > MaxIndirectDescriptorProbes) {
 				return false;
 			}
-			keys.reserve(static_cast<size_t>(probe_count) + 1u);
-			keys.push_back(0u);
-			for (uint64_t probe = 0, offset = first; probe < probe_count; ++probe, offset += step) {
-				uint32_t key = 0;
-				if (!ReadScalarTable(material.Base48(), size, offset, runtime, {&key, 1})) {
-					return false;
-				}
-				keys.push_back(key);
+			if (!read_keys(material, first, step, probe_count)) return false;
+			if (indirect.selector_shift == 0u) keys.push_back(0u);
+		}
+		if (indirect.material_source != UINT32_MAX) {
+			if (indirect.selector_shift != 0u) {
+				for (auto& key: keys) key >>= indirect.selector_shift;
+				keys.push_back(0u); // An out-of-range material load returns zero.
 			}
 			std::ranges::sort(keys);
 			keys.erase(std::unique(keys.begin(), keys.end()), keys.end());
 		}
 	}
 
-	const auto children_begin = snapshot.images.size();
+	const auto children_begin = descriptors.size();
 	const auto mapping_offset = snapshot.flattened_srt.size();
-	const auto root_image = specialization.images[image_index];
+	const auto root_resource  = specializations[resource_index];
 	const auto key_count = sources.empty() ? keys.size() : sources.size();
 	snapshot.flattened_srt.resize(mapping_offset + 1u + key_count * 2u);
 	snapshot.flattened_srt[mapping_offset] = static_cast<uint32_t>(key_count);
 	for (uint32_t entry = 0; entry < key_count; ++entry) {
 		const auto key = sources.empty() ? keys[entry] : entry;
 		DescriptorValue candidate;
-		candidate.dword_count = 8u;
+		candidate.dword_count = dword_count;
 		if (sources.empty()) {
-			const auto table_offset = (key << 5u) + indirect.table_offset;
-			if (!ReadScalarTable(table_base, table_size, table_offset, runtime, candidate.dwords)) {
+			const auto 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;
 			}
 		} else if (!clean.EvaluateDescriptor(sources[entry], candidate)) {
 			return false;
 		}
-		if (NullImageDescriptor(candidate) ||
-		    !ValidImageDescriptor(candidate, program.info.images[image_index].r128)) {
-			candidate.dwords.fill(0);
-		}
+		if (!normalize(candidate)) return false;
 		uint32_t ordinal = 0;
 		if (entry == 0) {
-			snapshot.images[image_index] = candidate;
-		} else if (snapshot.images[image_index] != candidate) {
-			const auto found = std::find(snapshot.images.begin() + children_begin,
-			                             snapshot.images.end(), candidate);
-			ordinal = static_cast<uint32_t>(found - snapshot.images.begin() - children_begin + 1u);
-			if (found == snapshot.images.end()) {
-				if (snapshot.images.size() >= ShaderInfo::MaxImages) {
+			descriptors[resource_index] = candidate;
+		} else if (descriptors[resource_index] != candidate) {
+			const auto found =
+			    std::find(descriptors.begin() + children_begin, descriptors.end(), candidate);
+			ordinal = static_cast<uint32_t>(found - descriptors.begin() - children_begin + 1u);
+			if (found == descriptors.end()) {
+				if (descriptors.size() >= maximum_resources) {
 					return false;
 				}
-				snapshot.images.push_back(candidate);
-				auto child = root_image;
-				child.indirect_root = image_index;
-				specialization.images.push_back(child);
+				descriptors.push_back(candidate);
+				auto child          = root_resource;
+				child.indirect_root = resource_index;
+				specializations.push_back(child);
 			}
 		}
 		snapshot.flattened_srt[mapping_offset + 1u + entry * 2u] = key;
 		snapshot.flattened_srt[mapping_offset + 2u + entry * 2u] = ordinal;
 	}
-	if (snapshot.images.size() == children_begin) {
+	if (descriptors.size() == children_begin) {
 		snapshot.flattened_srt.resize(mapping_offset);
 	} else {
-		auto& root = specialization.images[image_index];
-		root.indirect_root = image_index;
+		auto& root                      = specializations[resource_index];
+		root.indirect_root              = resource_index;
 		root.indirect_mapping_offset = static_cast<uint32_t>(mapping_offset);
 		root.indirect_search_iterations = std::bit_width(key_count);
 	}
@@ -383,10 +435,21 @@ struct SamplerPlan {
 	uint32_t                                                  sampler_count = 0;
 };
 
+template <typename T, typename Keep>
+void CompactImages(std::vector<T>& images, Keep&& keep) {
+	size_t count = 0;
+	for (size_t index = 0; index < images.size(); ++index) {
+		if (!keep(index)) continue;
+		if (count != index) images[count] = std::move(images[index]);
+		++count;
+	}
+	images.resize(count);
+}
+
 struct ImageRemap {
 	explicit ImageRemap(const ResourceSpecialization& specialization)
-	    : source_count(static_cast<uint32_t>(specialization.images.size())) {
-		EXIT_IF(specialization.images.size() > indices.size());
+	    : indices(specialization.images.size()),
+	      source_count(static_cast<uint32_t>(specialization.images.size())) {
 		for (uint32_t index = 0; index < source_count; index++) {
 			indices[index] = specialization.images[index].fmask ? UINT32_MAX : count++;
 		}
@@ -403,16 +466,11 @@ struct ImageRemap {
 		if (count == source_count) {
 			return;
 		}
-		for (uint32_t index = 0; index < source_count; index++) {
-			if (indices[index] != UINT32_MAX && indices[index] != index) {
-				images[indices[index]] = std::move(images[index]);
-			}
-		}
-		images.resize(count);
+		CompactImages(images, [&](size_t index) { return indices[index] != UINT32_MAX; });
 	}
 
 private:
-	std::array<uint32_t, ShaderInfo::MaxImages> indices;
+	std::vector<uint32_t> indices;
 	uint32_t                                    source_count;
 	uint32_t                                    count = 0;
 };
@@ -559,7 +617,7 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 			}
 		}
 	}
-	ImageRemap(specialization).Apply(snapshot.images);
+	CompactImages(snapshot.images, [&](size_t index) { return !specialization.images[index].fmask; });
 	return true;
 }
 
@@ -605,6 +663,7 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 		}
 	}
 	std::vector<ResourceBlock> blocks(program.blocks.size());
+	std::vector<std::vector<uint32_t>> predecessors(blocks.size());
 	for (uint32_t i = 0; i < blocks.size(); i++) {
 		auto&                 block      = blocks[i];
 		const auto&           info       = program.block_info[i];
@@ -628,6 +687,7 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 				return {};
 			}
 			block.successors.push_back(found->second);
+			predecessors[found->second].push_back(i);
 		}
 		for (const auto& inst: *program.blocks[i]) {
 			const auto op     = inst.GetOpcode();
@@ -657,7 +717,30 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 		block.sources.erase(std::unique(block.sources.begin(), block.sources.end()),
 		                    block.sources.end());
 	}
-	for (auto& block: blocks) block.condition = predicate(block.condition);
+	// Host branches only matter when they guard a later resource access. Leave
+	// data-only branches (for example a mutable flag checked before a trap) on the GPU.
+	std::vector<uint8_t> resource_reachable(blocks.size());
+	std::vector<uint32_t> pending;
+	for (uint32_t i = 0; i < blocks.size(); ++i) {
+		if (!blocks[i].sources.empty() || !blocks[i].srt_reads.empty()) {
+			resource_reachable[i] = 1;
+			pending.push_back(i);
+		}
+	}
+	while (!pending.empty()) {
+		const auto index = pending.back();
+		pending.pop_back();
+		for (const auto predecessor: predecessors[index]) {
+			if (resource_reachable[predecessor]) continue;
+			resource_reachable[predecessor] = 1;
+			pending.push_back(predecessor);
+		}
+	}
+	for (auto& block: blocks) {
+		block.condition = std::ranges::any_of(block.successors, [&](uint32_t successor) {
+			return resource_reachable[successor] != 0;
+		}) ? predicate(block.condition) : Value {};
+	}
 	if (std::ranges::none_of(
 	        blocks, [](const ResourceBlock& block) { return !block.condition.IsEmpty(); })) {
 		return {};
@@ -665,11 +748,33 @@ static std::vector<ResourceBlock> ResourceControlFlow(const Program& program, co
 	return blocks;
 }
 
+// ResolveComputeBufferFill proves complete workgroups and matching requested extents, so every
+// dispatch-bound predicate accepted here is true for every invocation covered by the fill.
+static bool IsFullDispatchPredicate(Value value, uint32_t depth = 0) {
+	value = value.Resolve();
+	if (value == Value(true)) return true;
+	const auto* inst = value.TryInstruction();
+	if (inst == nullptr || depth > 32) return false;
+	if (inst->GetOpcode() == ValueOpcode::LogicalAnd) {
+		return IsFullDispatchPredicate(inst->Arg(0), depth + 1) &&
+		       IsFullDispatchPredicate(inst->Arg(1), depth + 1);
+	}
+	if (inst->GetOpcode() != ValueOpcode::ULessThan32) return false;
+	const auto* global = inst->Arg(0).Resolve().TryInstruction();
+	const auto* extent = inst->Arg(1).Resolve().TryInstruction();
+	if (global == nullptr || global->GetOpcode() != ValueOpcode::GetBuiltin ||
+	    global->Arg(0).Resolve() != Value(static_cast<uint32_t>(StageInputKind::GlobalInvocationId)) ||
+	    extent == nullptr || extent->GetOpcode() != ValueOpcode::GetDispatchThreadExtent)
+		return false;
+	const auto axis = global->Arg(1).Resolve();
+	return axis.IsImmediate() && axis.U32() < 3 && extent->Arg(0).Resolve() == axis;
+}
+
 // Nonnegative affine coefficients for constant, local and workgroup coordinates. Reject modular
 // arithmetic that could wrap; runtime coverage also bounds the largest invocation index.
-static std::optional<std::array<uint64_t, 3>> FillIndex(Value value, uint32_t axis,
+static std::optional<std::array<uint64_t, 3>> FillIndex(Value value, uint32_t axis, Value guard,
                                                       uint32_t depth = 0) {
-	value = value.Resolve();
+	value = ResolveActiveU32(value, guard);
 	if (depth > 32 || value.GetType() != Type::U32) {
 		return {};
 	}
@@ -693,8 +798,8 @@ static std::optional<std::array<uint64_t, 3>> FillIndex(Value value, uint32_t ax
 	    op != ValueOpcode::ShiftLeftLogical32) {
 		return {};
 	}
-	auto left  = FillIndex(inst->Arg(0), axis, depth + 1);
-	auto right = FillIndex(inst->Arg(1), axis, depth + 1);
+	auto left  = FillIndex(inst->Arg(0), axis, guard, depth + 1);
+	auto right = FillIndex(inst->Arg(1), axis, guard, depth + 1);
 	if (!left || !right) {
 		return {};
 	}
@@ -750,6 +855,7 @@ static UniformFillPlan AnalyzeUniformFill(const Program& program) {
 	UniformFillPlan result;
 	result.fill.resource = memory.resource;
 	Value data;
+	Value guard(true);
 	if (store->GetOpcode() == ValueOpcode::ImageWrite) {
 		if (program.info.images.size() != 1 || memory.dmask != 1 || memory.data_bits != 32 ||
 		    memory.image_has_mip || memory.image_sample_flags != 0 || memory.image_r128 ||
@@ -760,7 +866,7 @@ static UniformFillPlan AnalyzeUniformFill(const Program& program) {
 		const auto* address = store->Arg(1).ResolveInstruction();
 		if (address == nullptr || address->GetOpcode() != ValueOpcode::MakeImageAddress) return {};
 		for (uint32_t axis = 0; axis < 3; ++axis) {
-			const auto index = FillIndex(address->Arg(axis), axis);
+			const auto index = FillIndex(address->Arg(axis), axis, guard);
 			if (!index || (*index)[0] != 0) return {};
 			if (axis < 2) {
 				if ((*index)[1] != 1 || (*index)[2] == 0) return {};
@@ -782,20 +888,22 @@ static UniformFillPlan AnalyzeUniformFill(const Program& program) {
 		                             ValueOpcode::StoreBufferU32x3, ValueOpcode::StoreBufferU32x4};
 		const auto           store_op = std::ranges::find(stores, op);
 		if (store_op == stores.end() || store->Arg(2).Resolve() != Value(0u) ||
-		    store->Arg(3).Resolve() != Value(0u) || store->Arg(5).Resolve() != Value(true))
+		    store->Arg(3).Resolve() != Value(0u))
 			return {};
+		guard = store->Arg(5).Resolve();
+		if (!IsFullDispatchPredicate(guard)) return {};
 		if (!memory.formatted || memory.typed || !memory.idxen || memory.offen || memory.offset != 0 ||
 		    memory.data_bits != 32 ||
 		    memory.data_dwords != static_cast<uint32_t>(store_op - stores.begin() + 1))
 			return {};
-		const auto address = FillIndex(store->Arg(1), 0);
+		const auto address = FillIndex(store->Arg(1), 0, guard);
 		if (!address || (*address)[0] != 0 || (*address)[1] != 1 || (*address)[2] == 0) return {};
 		result.fill.kind = UniformFillKind::Buffer;
 		result.fill.group_stride[0] = static_cast<uint32_t>((*address)[2]);
 		result.fill.words = memory.data_dwords;
 		data = store->Arg(4);
 	}
-	data = data.Resolve();
+	data = ResolveActiveU32(data, guard);
 	const auto*          vector = data.TryInstruction();
 	constexpr std::array composites {ValueOpcode::CompositeConstructU32x2,
 	                                 ValueOpcode::CompositeConstructU32x3,
@@ -804,7 +912,7 @@ static UniformFillPlan AnalyzeUniformFill(const Program& program) {
 	    (vector == nullptr || vector->GetOpcode() != composites[result.fill.words - 2]))
 		return {};
 	for (uint32_t i = 0; i < result.fill.words; ++i) {
-		const auto word = result.fill.words == 1 ? data : vector->Arg(i);
+		const auto word = ResolveActiveU32(result.fill.words == 1 ? data : vector->Arg(i), guard);
 		if (word.GetType() != Type::U32 ||
 		    !ValidateRuntimeValue(program, word, RuntimeValueType::Integer))
 			return {};
@@ -831,6 +939,11 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		if (source == nullptr) {
 			return value;
 		}
+		if (source->GetOpcode() == ValueOpcode::ReadFirstLane &&
+		    ValidateRuntimeValue(program, source->Arg(0))) {
+			// Uniform values need no new EXEC context; preserve their shared evaluation memo.
+			return Clone(source->Arg(0));
+		}
 		if (source->GetOpcode() == ValueOpcode::Phi) {
 			const auto invariant = ResolveInvariantPhi(program, value);
 			if (!invariant.IsEmpty() && invariant != value) {
@@ -840,6 +953,10 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		if (const auto found = cloned.find(source); found != cloned.end()) {
 			return Value(found->second);
 		}
+		if (source->GetOpcode() == ValueOpcode::LoadBufferU32) {
+			plan.requires_specialization_memory = true;
+			plan.capture_specialization_reads   = true;
+		}
 		auto& target =
 		    plan.value_storage.emplace_back(source->GetOpcode(), source->Flags<uint64_t>());
 		cloned.emplace(source, &target);
@@ -859,10 +976,13 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 	for (const auto& source: program.descriptor_sources) {
 		auto& target          = plan.descriptor_sources.emplace_back();
 		target.dword_count    = source.dword_count;
-		target.indirect_image = source.indirect_image;
-		if (target.indirect_image.has_value()) {
-			target.indirect_image->key_count = Clone(target.indirect_image->key_count);
-			target.indirect_image->selector_mask = Clone(target.indirect_image->selector_mask);
+		target.indirect_descriptor = source.indirect_descriptor;
+		if (target.indirect_descriptor.has_value()) {
+			target.indirect_descriptor->key_count = Clone(target.indirect_descriptor->key_count);
+			target.indirect_descriptor->selector_first =
+			    Clone(target.indirect_descriptor->selector_first);
+			target.indirect_descriptor->selector_mask =
+			    Clone(target.indirect_descriptor->selector_mask);
 		}
 		for (uint32_t dword = 0; dword < source.dword_count; dword++) {
 			target.dwords[dword] = Clone(source.dwords[dword]);
@@ -909,35 +1029,35 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		plan.uniform_fill.values[i] = Clone(plan.uniform_fill.values[i]);
 	}
 	plan.clean_flat_slots.resize(plan.srt_reads.size());
-	bool capture_image_reads = false;
-	for (const auto& image: plan.info.images) {
-		const auto* source = Source(plan, image.source);
-		if (source == nullptr || !source->indirect_image.has_value()) {
-			continue;
-		}
+	bool capture_indirect_reads = false;
+	for (const auto& source: plan.descriptor_sources) {
+		if (!source.indirect_descriptor.has_value()) continue;
+		const auto& indirect                = *source.indirect_descriptor;
 		plan.requires_specialization_memory = true;
-		capture_image_reads |= !source->indirect_image->selector_mask.IsEmpty() ||
-		                       !source->indirect_image->sources.empty();
-		MarkCleanFlatSlots(plan, Source(plan, source->indirect_image->material_source),
-		                   plan.clean_flat_slots, source->indirect_image->selector_mask);
-		if (source->indirect_image->sources.empty()) {
-			MarkCleanFlatSlots(plan, Source(plan, source->indirect_image->table_source),
-			                   plan.clean_flat_slots);
+		capture_indirect_reads |= indirect.material_source != UINT32_MAX || !indirect.sources.empty();
+		MarkCleanFlatSlots(plan, Source(plan, indirect.material_source), plan.clean_flat_slots,
+		                   indirect.selector_mask);
+		MarkCleanFlatSlots(plan, nullptr, plan.clean_flat_slots, indirect.selector_first);
+		MarkCleanFlatSlots(plan, nullptr, plan.clean_flat_slots, indirect.key_count);
+		if (indirect.sources.empty()) {
+			MarkCleanFlatSlots(plan, Source(plan, indirect.table_source), plan.clean_flat_slots);
 		} else {
-			for (const auto candidate: source->indirect_image->sources) {
+			for (const auto candidate: indirect.sources) {
 				MarkCleanFlatSlots(plan, Source(plan, candidate), plan.clean_flat_slots);
 			}
 		}
 	}
-	plan.capture_specialization_reads = capture_image_reads || !plan.control_flow.empty();
-	// Writable descriptor addresses must come from clean backing for the alias proof.
-	for (const auto& buffer: plan.info.buffers) {
-		if (buffer.written) MarkCleanFlatSlots(plan, Source(plan, buffer.source), plan.clean_flat_slots);
+	plan.capture_specialization_reads |= capture_indirect_reads || !plan.control_flow.empty();
+	if (plan.capture_specialization_reads) {
+		// Clean writable addresses prove that shader writes cannot overlap captured reads.
+		for (const auto& buffer: plan.info.buffers) {
+			if (buffer.written) MarkCleanFlatSlots(plan, Source(plan, buffer.source), plan.clean_flat_slots);
+		}
+		for (const auto& image: plan.info.images) {
+			if (image.written) MarkCleanFlatSlots(plan, Source(plan, image.source), plan.clean_flat_slots);
+		}
 	}
-	for (const auto& image: plan.info.images) {
-		if (image.written) MarkCleanFlatSlots(plan, Source(plan, image.source), plan.clean_flat_slots);
-	}
-	if (capture_image_reads) plan.resource_tracking_complete &= !program.has_address_writes;
+	if (capture_indirect_reads) plan.resource_tracking_complete &= !program.has_address_writes;
 	return plan;
 }
 
@@ -959,7 +1079,8 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		observed.read_memory = CaptureOrdinaryRead;
 	}
 	SrtWalker clean(program, CleanRuntime(observed));
-	SrtWalker walker(program, observed, program.clean_flat_slots, &clean);
+	SrtWalker walker(program, observed, program.clean_flat_slots,
+	                 capture_reads || program.requires_specialization_memory ? &clean : nullptr);
 	if (!walker.RefreshFlatBuffer(snapshot.flattened_srt)) {
 		return false;
 	}
@@ -981,19 +1102,34 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			return false;
 		}
 		if (active.empty() || active[source]) {
-			return (written ? clean : walker).EvaluateDescriptor(source, value);
+			return (capture_reads && written ? clean : walker).EvaluateDescriptor(source, value);
 		}
 		value = {};
 		value.dword_count = program.descriptor_sources[source].dword_count;
 		return true;
 	};
 	snapshot.buffers.resize(program.info.buffers.size());
-	specialization.buffers.clear();
-	specialization.buffers.reserve(program.info.buffers.size());
-	for (uint32_t i = 0; i < program.info.buffers.size(); ++i) {
-		if (!evaluate(program.info.buffers[i].source, snapshot.buffers[i],
-		              program.info.buffers[i].written)) {
-			return false;
+	specialization.buffers.assign(program.info.buffers.size(), {});
+	for (uint32_t i = 0; i < snapshot.buffers.size(); ++i) {
+		const auto base_index =
+		    i < program.info.buffers.size() ? i : specialization.buffers[i].indirect_root;
+		if (base_index >= program.info.buffers.size()) return false;
+		const auto& base = program.info.buffers[base_index];
+		if (i < program.info.buffers.size()) {
+			const auto* source = Source(program, base.source);
+			if (source == nullptr) return false;
+			if (source->indirect_descriptor.has_value()) {
+				snapshot.buffers[i] = {.dword_count = 4u};
+				if (active.empty() || active[base.source]) {
+					if (!MaterializeIndirectDescriptor(
+					        program, *source->indirect_descriptor, i, 4u, observed, clean, snapshot,
+					        snapshot.buffers, specialization.buffers, ShaderInfo::MaxBuffers,
+					        NormalizeIndirectStoreBuffer))
+						return false;
+				}
+			} else if (!evaluate(base.source, snapshot.buffers[i], base.written)) {
+				return false;
+			}
 		}
 		auto&                descriptor_value = snapshot.buffers[i];
 		ShaderBufferResource descriptor;
@@ -1012,15 +1148,12 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		} else if (!swizzle) {
 			packed_stride &= ~(3u << 16u);
 		}
-		specialization.buffers.push_back({
-		    .packed_stride     = packed_stride,
-		    .descriptor_format = program.info.buffers[i].formatted
-		                             ? descriptor.Format()
-		                             : Prospero::BufferFormat::kInvalid,
-		    .descriptor_swizzle =
-		        program.info.buffers[i].formatted ? descriptor.DstSelXYZW() : DstSel(4, 5, 6, 7),
-		    .zero_stride_oob = descriptor.OutOfBounds() == 0u && stride == 0u,
-		});
+		auto& buffer         = specialization.buffers[i];
+		buffer.packed_stride = packed_stride;
+		buffer.descriptor_format =
+		    base.formatted ? descriptor.Format() : Prospero::BufferFormat::kInvalid;
+		buffer.descriptor_swizzle = base.formatted ? descriptor.DstSelXYZW() : DstSel(4, 5, 6, 7);
+		buffer.zero_stride_oob    = descriptor.OutOfBounds() == 0u && stride == 0u;
 	}
 	snapshot.images.resize(program.info.images.size());
 	specialization.images.resize(program.info.images.size());
@@ -1041,13 +1174,19 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		if (source == nullptr) {
 			return false;
 		}
-		if (source->indirect_image.has_value()) {
+		if (source->indirect_descriptor.has_value()) {
 			snapshot.images[i] = {.dword_count = 8u};
 			if (!active.empty() && !active[image.source]) {
 				continue;
 			}
-			const auto& indirect = *source->indirect_image;
-			if (!MaterializeIndirectImage(program, indirect, i, observed, clean, snapshot, specialization)) {
+			const auto& indirect = *source->indirect_descriptor;
+			if (!MaterializeIndirectDescriptor(
+			        program, indirect, i, 8u, observed, clean, snapshot, snapshot.images,
+			        specialization.images, UINT32_MAX, [&](DescriptorValue& value) {
+				        if (NullImageDescriptor(value) || !ValidImageDescriptor(value, image.r128))
+					        value.dwords.fill(0u);
+				        return true;
+			        })) {
 				return false;
 			}
 		} else {
@@ -1082,14 +1221,33 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 void ApplyResourceSpecialization(Program& program, const ResourceSpecialization& specialization) {
 	EXIT_IF(!program.resource_tracking_complete || program.shader_info_complete ||
 	        program.binding_layout_complete);
-	EXIT_IF(program.info.buffers.size() != specialization.buffers.size() ||
+	EXIT_IF(program.info.buffers.size() > specialization.buffers.size() ||
 	        program.info.images.size() > specialization.images.size());
 
 	auto& buffers = program.info.buffers;
-	for (size_t index = 0; index < buffers.size(); index++) {
-		buffers[index].packed_stride      = specialization.buffers[index].packed_stride;
-		buffers[index].descriptor_format  = specialization.buffers[index].descriptor_format;
-		buffers[index].descriptor_swizzle = specialization.buffers[index].descriptor_swizzle;
+	const auto original_buffer_count = buffers.size();
+	buffers.reserve(specialization.buffers.size());
+	for (uint32_t index = 0; index < specialization.buffers.size(); ++index) {
+		const auto& source = specialization.buffers[index];
+		if (index >= buffers.size()) {
+			EXIT_IF(source.indirect_root >= original_buffer_count);
+			buffers.push_back(buffers[source.indirect_root]);
+		}
+		auto& buffer                      = buffers[index];
+		buffer.packed_stride              = source.packed_stride;
+		buffer.descriptor_format          = source.descriptor_format;
+		buffer.descriptor_swizzle         = source.descriptor_swizzle;
+		buffer.indirect_root              = source.indirect_root;
+		buffer.indirect_mapping_offset    = source.indirect_mapping_offset;
+		buffer.indirect_search_iterations = source.indirect_search_iterations;
+		buffer.indirect_resources.clear();
+	}
+	for (uint32_t index = 0; index < buffers.size(); ++index) {
+		const auto root = buffers[index].indirect_root;
+		if (root != BufferResource::NoIndirectBuffer) {
+			EXIT_IF(root >= buffers.size());
+			buffers[root].indirect_resources.push_back(index);
+		}
 	}
 	auto& images = program.info.images;
 	const auto original_image_count = images.size();

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

@@ -13,6 +13,9 @@ struct ResourceSpecialization {
 		Prospero::BufferFormat descriptor_format               = Prospero::BufferFormat::kInvalid;
 		uint32_t               descriptor_swizzle              = DstSel(4, 5, 6, 7);
 		bool                   zero_stride_oob                 = false;
+		uint32_t               indirect_root                   = BufferResource::NoIndirectBuffer;
+		uint32_t               indirect_mapping_offset         = 0;
+		uint32_t               indirect_search_iterations      = 0;
 		bool                   operator==(const Buffer&) const = default;
 	};
 

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.cpp

@@ -23,9 +23,6 @@ namespace {
 constexpr uint64_t AddressMask = 0x0000ffffffffffffull;
 
 bool AddSignedAddress(uint64_t base, int64_t offset, uint64_t& result) {
-	if (base > AddressMask) {
-		return false;
-	}
 	if (offset < 0) {
 		const auto magnitude = uint64_t {0} - static_cast<uint64_t>(offset);
 		if (magnitude > base) {
@@ -35,7 +32,7 @@ bool AddSignedAddress(uint64_t base, int64_t offset, uint64_t& result) {
 		return true;
 	}
 	const auto magnitude = static_cast<uint64_t>(offset);
-	if (magnitude > AddressMask - base) {
+	if (magnitude > UINT64_MAX - base) {
 		return false;
 	}
 	result = base + magnitude;
@@ -44,16 +41,29 @@ bool AddSignedAddress(uint64_t base, int64_t offset, uint64_t& result) {
 
 bool IsRawRead(const ResourcePlan& values, const Inst& inst) {
 	const auto op = inst.GetOpcode();
-	if (op != ValueOpcode::LoadAddressU32 && op != ValueOpcode::ReadConstBuffer) {
+	if (op != ValueOpcode::LoadAddressU32 && op != ValueOpcode::ReadConstBuffer &&
+	    op != ValueOpcode::LoadBufferU32) {
 		return false;
 	}
 	const auto index = inst.Flags<MemoryFlags>().index;
 	if (index >= values.memory_info.size()) {
 		return false;
 	}
-	const auto kind = values.memory_info[index].kind;
-	return (op == ValueOpcode::LoadAddressU32 && kind == ResourceKind::ScalarAddress) ||
-	       (op == ValueOpcode::ReadConstBuffer && kind == ResourceKind::ScalarBuffer);
+	const auto& memory = values.memory_info[index];
+	if (op != ValueOpcode::LoadBufferU32) {
+		return (op == ValueOpcode::LoadAddressU32 && memory.kind == ResourceKind::ScalarAddress) ||
+		       (op == ValueOpcode::ReadConstBuffer && memory.kind == ResourceKind::ScalarBuffer);
+	}
+	if (inst.NumArgs() != 5u || memory.kind != ResourceKind::Buffer || memory.typed ||
+	    memory.formatted || memory.coherent || memory.data_bits != 32u ||
+	    memory.data_dwords != 1u || memory.idxen || memory.offen || memory.offset != 0u ||
+	    inst.Arg(4).GetType() != Type::U1) {
+		return false;
+	}
+	for (uint32_t arg = 1; arg <= 3; ++arg) {
+		if (inst.Arg(arg).Resolve() != Value(0u)) return false;
+	}
+	return true;
 }
 
 bool IsDescriptorHandle(ValueOpcode opcode) {
@@ -92,6 +102,7 @@ bool IsRuntimeUniformOp(ValueOpcode op) {
 		case ValueOpcode::ISub64:
 		case ValueOpcode::IMul32:
 		case ValueOpcode::IMul64:
+		case ValueOpcode::UMulHi:
 		case ValueOpcode::UMin32:
 		case ValueOpcode::ShiftLeftLogical32:
 		case ValueOpcode::ShiftLeftLogical64:
@@ -108,6 +119,7 @@ bool IsRuntimeUniformOp(ValueOpcode op) {
 		case ValueOpcode::SelectU32:
 		case ValueOpcode::SelectF32:
 		case ValueOpcode::ULessThan32:
+		case ValueOpcode::ULessThanEqual32:
 		case ValueOpcode::IEqual32:
 		case ValueOpcode::UGreaterThan32:
 		case ValueOpcode::SGreaterThanEqual32:
@@ -120,6 +132,7 @@ bool IsRuntimeUniformOp(ValueOpcode op) {
 		case ValueOpcode::FPOrdGreaterThanEqual32:
 		case ValueOpcode::FPIsNan32:
 		case ValueOpcode::FPMul32:
+		case ValueOpcode::FPRecipIFlag32:
 		case ValueOpcode::FPTrunc32: return true;
 		default: return false;
 	}
@@ -160,6 +173,7 @@ private:
 				default: return false;
 			}
 		}
+		if (require_uniform && !m_active_mask.IsEmpty() && value == m_active_mask) return true;
 		// Integer-only dependency checks do not depend on the active EXEC mask.
 		if (!require_uniform && m_validated_dependencies.contains(inst)) return true;
 		if (!m_visiting.insert(inst).second) {
@@ -249,7 +263,8 @@ private:
 			}
 			return finish(true);
 		}
-		if (op == ValueOpcode::LoadAddressU32 || op == ValueOpcode::ReadConstBuffer) {
+		if (op == ValueOpcode::LoadAddressU32 || op == ValueOpcode::ReadConstBuffer ||
+		    op == ValueOpcode::LoadBufferU32) {
 			const auto  expected = op == ValueOpcode::LoadAddressU32
 			                           ? ValueOpcode::GetAddressResource
 			                           : ValueOpcode::GetBufferResource;
@@ -258,6 +273,11 @@ private:
 			    handle->GetOpcode() != expected) {
 				return finish(false);
 			}
+			if (op == ValueOpcode::LoadBufferU32) {
+				const auto guard = inst->Arg(4).Resolve();
+				return finish(Validate(inst->Arg(0)) &&
+				              ((!m_active_mask.IsEmpty() && guard == m_active_mask) || Validate(guard)));
+			}
 		} else if (op == ValueOpcode::CompositeExtractU64) {
 			const auto index = inst->NumArgs() == 2 ? inst->Arg(1).Resolve() : Value {};
 			if (!index.IsImmediate() || index.GetType() != Type::U32 || index.U32() >= 2u) {
@@ -344,6 +364,10 @@ bool SrtWalker::EvaluateWide(Value value, uint64_t& result) {
 			default: return false;
 		}
 	}
+	if (!m_active_mask.IsEmpty() && value == m_active_mask) {
+		result = 1u;
+		return true;
+	}
 	auto* inst = value.TryInstruction();
 	if (inst == nullptr) {
 		return false;
@@ -433,6 +457,16 @@ bool SrtWalker::EvaluateRawRead(const Inst& inst, uint64_t& result) {
 		return false;
 	}
 	const auto& mem    = m_program.memory_info[flags.index];
+	const bool vector = inst.GetOpcode() == ValueOpcode::LoadBufferU32;
+	const auto guard = vector ? inst.Arg(4).Resolve() : Value {};
+	if (vector && (guard.IsImmediate() || guard != m_active_mask)) {
+		uint64_t enabled = 0;
+		if (!Arg(inst, 4, enabled)) return false;
+		if (enabled == 0u) {
+			result = 0u;
+			return true;
+		}
+	}
 	const auto* handle = inst.Arg(0).ResolveInstruction();
 	if (handle == nullptr) {
 		return false;
@@ -443,10 +477,10 @@ bool SrtWalker::EvaluateRawRead(const Inst& inst, uint64_t& result) {
 	if (!Arg(*handle, 0, low) || !Arg(*handle, 1, high) || !Arg(inst, 1, offset)) {
 		return false;
 	}
-	const auto base      = ((high << 32u) | static_cast<uint32_t>(low)) & AddressMask;
+	const auto base      = (high << 32u) | static_cast<uint32_t>(low);
 	const auto immediate = static_cast<int64_t>(static_cast<int32_t>(mem.offset));
 	uint64_t   address   = 0;
-	if (inst.GetOpcode() == ValueOpcode::ReadConstBuffer) {
+	if (inst.GetOpcode() == ValueOpcode::ReadConstBuffer || vector) {
 		uint64_t records = 0;
 		uint64_t word3   = 0;
 		if (handle->NumArgs() != 4u || !Arg(*handle, 2, records) || !Arg(*handle, 3, word3)) {
@@ -458,13 +492,22 @@ bool SrtWalker::EvaluateRawRead(const Inst& inst, uint64_t& result) {
 		const auto byte_offset =
 		    (static_cast<uint64_t>(immediate) & ~uint64_t {3}) + (static_cast<uint32_t>(offset) & ~3u);
 		const auto stride  = (static_cast<uint32_t>(high) >> 16u) & 0x3fffu;
+		if (vector) {
+			// Only the uniform, unswizzled structured DWORD address is evaluated on the host.
+			if ((high & (1u << 31u)) != 0u || (word3 & ((1u << 23u) | 0xf0000000u)) != 0u)
+				return false;
+			if (stride == 0u || records == 0u || ((word3 >> 12u) & 0x7fu) == 0u) {
+				result = 0u;
+				return true;
+			}
+		}
 		const auto size = stride == 0u
 		                      ? static_cast<uint64_t>(static_cast<uint32_t>(records))
 		                      : static_cast<uint64_t>(stride) * static_cast<uint32_t>(records);
 		if (byte_offset > size || size - byte_offset < sizeof(uint32_t)) {
 			return false;
 		}
-		address = (base & ~uint64_t {3}) + byte_offset;
+		address = (base & AddressMask & ~uint64_t {3}) + byte_offset;
 	} else {
 		const auto relative = (immediate & ~int64_t {3}) +
 		                      static_cast<int64_t>(static_cast<uint32_t>(offset) & ~3u);
@@ -473,11 +516,13 @@ bool SrtWalker::EvaluateRawRead(const Inst& inst, uint64_t& result) {
 		}
 	}
 	uint32_t word = 0;
-	if (m_runtime.read_memory != nullptr) {
-		if (!m_runtime.read_memory(m_runtime.userdata, address, {&word, 1})) {
+	const auto reader = vector ? m_runtime.read_specialization_memory : m_runtime.read_memory;
+	if (reader != nullptr) {
+		if (!reader(m_runtime.userdata, address, {&word, 1})) {
 			return false;
 		}
 	} else {
+		if (vector) return false;
 		std::memcpy(&word, reinterpret_cast<const void*>(address), sizeof(word));
 	}
 	result = word;
@@ -505,6 +550,9 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 		case ValueOpcode::GetShaderBase: result = m_runtime.shader_base; return true;
 		case ValueOpcode::Phi: return EvaluatePhi(inst, result);
 		case ValueOpcode::ReadFirstLane: {
+			if (!m_active_mask.IsEmpty() && inst.Arg(1).Resolve() == m_active_mask) {
+				return EvaluateWide(inst.Arg(0), result);
+			}
 			const auto clean_runtime = CleanRuntime(m_runtime);
 			SrtWalker  clean_active(m_program, clean_runtime, {}, nullptr, inst.Arg(1));
 			SrtWalker  active(m_program, m_runtime, m_clean_flat_slots, &clean_active,
@@ -537,7 +585,12 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 		}
 		case ValueOpcode::LoadAddressU32:
 		case ValueOpcode::ReadConstBuffer:
+		case ValueOpcode::LoadBufferU32:
 			if (IsRawRead(m_program, inst)) {
+				if (inst.GetOpcode() == ValueOpcode::LoadBufferU32 && m_clean_evaluator != nullptr &&
+				    m_clean_evaluator->m_active_mask == m_active_mask) {
+					return m_clean_evaluator->EvaluateWide(Value(const_cast<Inst*>(&inst)), result);
+				}
 				return EvaluateRawRead(inst, result);
 			}
 			break;
@@ -577,6 +630,12 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 				return true;
 			}
 			return false;
+		case ValueOpcode::UMulHi:
+			if (binary()) {
+				result = (static_cast<uint64_t>(static_cast<uint32_t>(a)) * static_cast<uint32_t>(b)) >> 32u;
+				return true;
+			}
+			return false;
 		case ValueOpcode::UMin32:
 			if (binary()) {
 				result = std::min(static_cast<uint32_t>(a), static_cast<uint32_t>(b));
@@ -612,6 +671,15 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 				return true;
 			}
 			return false;
+		case ValueOpcode::FPRecipIFlag32:
+			if (Arg(inst, 0, a)) {
+				const auto exponent = (a >> 23u) & 0xffu;
+				// Normal positive powers of two have exact normal reciprocals in every FP mode.
+				if ((a & 0x807fffffu) != 0u || exponent == 0u || exponent >= 254u) return false;
+				result = (254u - exponent) << 23u;
+				return true;
+			}
+			return false;
 		case ValueOpcode::FPIsNan32:
 			if (Arg(inst, 0, a)) {
 				result = std::isnan(Float32(a));
@@ -782,6 +850,12 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 				return true;
 			}
 			return false;
+		case ValueOpcode::ULessThanEqual32:
+			if (binary()) {
+				result = static_cast<uint32_t>(a) <= static_cast<uint32_t>(b);
+				return true;
+			}
+			return false;
 		case ValueOpcode::UGreaterThan32:
 			if (binary()) {
 				result = static_cast<uint32_t>(a) > static_cast<uint32_t>(b);

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/kernel/memory.cpp

@@ -166,7 +166,8 @@ static bool VirtualRangesOverlap(uint64_t left_start, uint64_t left_size, uint64
 }
 
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
-static uint32_t g_test_backing_store_unmaps_before_failure = UINT32_MAX;
+static uint32_t        g_test_backing_store_unmaps_before_failure = UINT32_MAX;
+static callback_func_t g_test_before_backing_map                   = nullptr;
 #endif
 
 #include "memoryAddressSpace.inc"
@@ -880,13 +881,15 @@ bool TryReadBacking(uint64_t vaddr, void* data, uint64_t size) {
 	       g_guest_address_space->TryReadBacking(vaddr, data, size);
 }
 
-bool TryReadGpuCleanBacking(uint64_t vaddr, void* data, uint64_t size) {
+bool TryReadBufferBacking(uint64_t vaddr, void* data, uint64_t size) {
 	if (g_gpu_resources != nullptr && IsGpuAddressRange(vaddr, size)) {
-		if (!Graphics::GuestGpu::IsGpuThread() ||
-		    GetGpuResources().GetBufferCache().HasGpuDirtyBytes(vaddr, size) ||
-		    GetGpuResources().GetTextureCache().IsRegionGpuModified(vaddr, size)) {
+		if (!Graphics::GuestGpu::IsGpuThread()) {
 			return false;
 		}
+		auto& buffers = GetGpuResources().GetBufferCache();
+		if (buffers.HasGpuDirtyBytes(vaddr, size)) {
+			buffers.ReadMemory(vaddr, size);
+		}
 	}
 	return TryReadBacking(vaddr, data, size);
 }
@@ -3024,6 +3027,22 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 		return g_guest_address_space->MapBacking(target_addr, len, direct_memory_start, mode,
 		                                         &shared_failure);
 	};
+	std::vector<PhysicalMemory::AllocatedBlock> replaced_mappings;
+	auto restore_mappings = [&](size_t count, bool restore_views) {
+		for (size_t index = 0; index < count; ++index) {
+			const auto& old = replaced_mappings[index];
+			if (restore_views) {
+				EXIT_IF(!g_guest_address_space->MapBacking(old.map_vaddr, old.map_size,
+				                                           old.start_addr, old.mode));
+			}
+			EXIT_IF(!g_physical_memory->Map(old.map_vaddr, old.start_addr, old.map_size,
+			                               old.prot, old.mode, old.gpu_mode));
+			g_physical_memory->SetVirtualRangeName(old.map_vaddr, old.map_size, old.name);
+			g_physical_memory->SetVirtualRangeMemoryType(old.map_vaddr, old.map_size,
+			                                          old.memory_type);
+		}
+		MapGpuRange(in_addr, len);
+	};
 	if (fixed) {
 		if (in_addr == 0 || (in_addr & (PAGE_SIZE - 1u)) != 0 ||
 		    (alignment != 0 && in_addr % alignment != 0)) {
@@ -3034,22 +3053,68 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 		}
 
 		std::vector<VirtualRanges::Range> reserved_ranges;
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX
 		if (g_virtual_ranges->QuerySpan(in_addr, len, &reserved_ranges) &&
 		    std::all_of(reserved_ranges.begin(), reserved_ranges.end(), [](const auto& range) {
-			    return range.type == VirtualRangeType::Reserved;
+			    return range.type == VirtualRangeType::Direct;
 		    })) {
+			uint64_t covered = 0;
+			{
+				Common::LockGuard lock(g_physical_memory->GetMutex());
+				for (const auto& mapping: g_physical_memory->GetMappings()) {
+					const auto start = std::max(in_addr, mapping.map_vaddr);
+					const auto end = std::min(in_addr + len, mapping.map_vaddr + mapping.map_size);
+					if (start >= end) {
+						continue;
+					}
+					auto old = mapping;
+					old.start_addr += start - old.map_vaddr;
+					old.map_vaddr = start;
+					old.size = old.map_size = end - start;
+					covered += old.map_size;
+					replaced_mappings.push_back(old);
+				}
+			}
+			if (covered != len) {
+				return KERNEL_ERROR_EBUSY;
+			}
 			UnmapGpuRange(in_addr, len);
-			reserved_target = true;
-		}
-		if (!reserved_target && ReplaceFixedRangeWithReserved(in_addr, len)) {
-			reserved_target = true;
-		}
-		if (!reserved_target) {
-			return KERNEL_ERROR_ENOMEM;
-		}
-		if (map_shared_fixed(in_addr)) {
+			for (size_t index = 0; index < replaced_mappings.size(); ++index) {
+				const auto& old = replaced_mappings[index];
+				GpuAccessMode old_gpu_mode = GpuAccessMode::NoAccess;
+				if (!g_physical_memory->Unmap(old.map_vaddr, old.map_size, &old_gpu_mode)) {
+					restore_mappings(index, false);
+					return KERNEL_ERROR_EBUSY;
+				}
+			}
+			// Keep the old view accessible until mmap atomically replaces it. In particular,
+			// do not publish a temporary reserved/PROT_NONE range between two direct views.
+			if (!map_shared_fixed(in_addr)) {
+				restore_mappings(replaced_mappings.size(), true);
+				return KERNEL_ERROR_ENOMEM;
+			}
 			out_addr       = in_addr;
 			shared_backing = true;
+		} else
+#endif
+		{
+			if (g_virtual_ranges->QuerySpan(in_addr, len, &reserved_ranges) &&
+			    std::all_of(reserved_ranges.begin(), reserved_ranges.end(), [](const auto& range) {
+				    return range.type == VirtualRangeType::Reserved;
+			    })) {
+				UnmapGpuRange(in_addr, len);
+				reserved_target = true;
+			}
+			if (!reserved_target && ReplaceFixedRangeWithReserved(in_addr, len)) {
+				reserved_target = true;
+			}
+			if (!reserved_target) {
+				return KERNEL_ERROR_ENOMEM;
+			}
+			if (map_shared_fixed(in_addr)) {
+				out_addr       = in_addr;
+				shared_backing = true;
+			}
 		}
 	} else {
 		constexpr size_t DEFAULT_ALIGNMENT = 0x4000;
@@ -3105,15 +3170,22 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 
 	if (!g_physical_memory->Map(out_addr, direct_memory_start, len, prot, mode, gpu_mode)) {
 		LOGF_COLOR(Log::Color::Red, "\t [Fail]\n");
-		EXIT_IF(!g_guest_address_space->UnmapBacking(out_addr, len));
+		if (replaced_mappings.empty()) {
+			EXIT_IF(!g_guest_address_space->UnmapBacking(out_addr, len));
+		} else {
+			restore_mappings(replaced_mappings.size(), true);
+		}
 		return KERNEL_ERROR_EBUSY;
 	}
 
 	PhysicalMemory::AllocatedBlock mapped_block {};
 	g_physical_memory->Find(direct_memory_start, false, &mapped_block);
-	const bool published = reserved_target
+	const bool published = !replaced_mappings.empty() || reserved_target
 	                           ? g_virtual_ranges->ReplaceSpan(
-	                                 out_addr, len, VirtualRangeType::Reserved, direct_memory_start,
+	                                 out_addr, len, replaced_mappings.empty()
+	                                                    ? VirtualRangeType::Reserved
+	                                                    : VirtualRangeType::Direct,
+	                                 direct_memory_start,
 	                                 prot, mapped_block.memory_type, VirtualRangeType::Direct, "")
 	                           : g_virtual_ranges->Add(out_addr, len, direct_memory_start, prot,
 	                                                   mapped_block.memory_type,
@@ -3121,7 +3193,11 @@ int KYTY_SYSV_ABI KernelMapDirectMemory(void** addr, size_t len, int prot, int f
 	if (!published) {
 		GpuAccessMode rollback_gpu_mode = GpuAccessMode::NoAccess;
 		EXIT_IF(!g_physical_memory->Unmap(out_addr, len, &rollback_gpu_mode));
-		EXIT_IF(!g_guest_address_space->UnmapBacking(out_addr, len));
+		if (replaced_mappings.empty()) {
+			EXIT_IF(!g_guest_address_space->UnmapBacking(out_addr, len));
+		} else {
+			restore_mappings(replaced_mappings.size(), true);
+		}
 		return KERNEL_ERROR_EBUSY;
 	}
 
@@ -3504,6 +3580,10 @@ int KYTY_SYSV_ABI KernelReserveVirtualRange(void** addr, size_t len, int flags,
 }
 
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+void TestBeforeNextBackingMap(callback_func_t callback) {
+	g_test_before_backing_map = callback;
+}
+
 void TestFailNextPhysicalMemoryUnmap() {
 	TestFailPhysicalMemoryUnmapAfter(0);
 }

src/kernel/memory.h

@@ -114,7 +114,7 @@ int AllocateDirectMemory(int64_t search_start, int64_t search_end, size_t size,
 int MapAutomaticMemory(uint64_t vaddr, size_t size, int type, int prot);
 bool                   TryWriteBacking(uint64_t vaddr, const void* data, uint64_t size);
 bool                   TryReadBacking(uint64_t vaddr, void* data, uint64_t size);
-bool                   TryReadGpuCleanBacking(uint64_t vaddr, void* data, uint64_t size);
+bool                   TryReadBufferBacking(uint64_t vaddr, void* data, uint64_t size);
 bool                   TryReadSparseBacking(uint64_t vaddr, void* data, uint64_t size);
 [[nodiscard]] uint64_t ClampRangeSize(uint64_t vaddr, uint64_t size);
 void                   WriteBacking(uint64_t vaddr, const void* data, uint64_t size) noexcept;
@@ -191,6 +191,7 @@ bool ProtectGuestHostMemory(uint64_t vaddr, uint64_t size, Common::VirtualMemory
 bool FreeGuestMemory(uint64_t vaddr, uint64_t size);
 
 #if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+void     TestBeforeNextBackingMap(callback_func_t callback);
 void     TestFailNextPhysicalMemoryUnmap();
 void     TestFailPhysicalMemoryUnmapAfter(uint32_t successful_unmaps);
 void     TestFailGuestBackingStoreUnmapAfter(uint32_t successful_unmaps);

src/kernel/memoryAddressSpace.inc

@@ -231,9 +231,8 @@ public:
 		return true;
 	}
 
-	bool MapExistingPlaceholderFixed(uint64_t vaddr, uint64_t size, uint64_t backing_offset,
-	                                 VirtualMemory::Mode mode,
-	                                 FailureReason*      failure_reason = nullptr) {
+	bool MapFixed(uint64_t vaddr, uint64_t size, uint64_t backing_offset,
+	              VirtualMemory::Mode mode, FailureReason* failure_reason = nullptr) {
 		if (!IsAvailable() || !RangeOk(backing_offset, size)) {
 			SetFailure(failure_reason, !IsAvailable() ? FailureReason::BackingUnavailable
 			                                          : FailureReason::PhysicalRangeOutOfBounds);
@@ -257,14 +256,13 @@ public:
 		SetFailure(failure_reason, FailureReason::None);
 		return true;
 #else
+		std::lock_guard lock(m_mutex);
 		void* mapped = MapView(vaddr, size, backing_offset, mode, failure_reason);
 		if (mapped == nullptr || reinterpret_cast<uint64_t>(mapped) != vaddr) {
 			return false;
 		}
-		{
-			std::lock_guard lock(m_mutex);
-			EXIT_IF(!m_maps.emplace(vaddr, MapEntry {vaddr, size, backing_offset, mode}).second);
-		}
+		RemoveMappingsUnlocked(vaddr, size);
+		EXIT_IF(!m_maps.emplace(vaddr, MapEntry {vaddr, size, backing_offset, mode}).second);
 		SetFailure(failure_reason, FailureReason::None);
 		return true;
 #endif
@@ -302,8 +300,7 @@ public:
 			bool part_placeholder_preserved = false;
 			if (!UnmapOne(entry.vaddr, entry.size, &part_placeholder_preserved)) {
 				for (auto it = removed.rbegin(); it != removed.rend(); ++it) {
-					EXIT_IF(!MapExistingPlaceholderFixed(it->vaddr, it->size, it->backing_offset,
-					                                     it->mode));
+					EXIT_IF(!MapFixed(it->vaddr, it->size, it->backing_offset, it->mode));
 				}
 				if (placeholder_preserved != nullptr) {
 					*placeholder_preserved = false;
@@ -320,8 +317,7 @@ public:
 		// views into one placeholder before publishing it there.
 		if (removed.size() > 1 && !CoalescePlaceholders(vaddr, size)) {
 			for (auto it = removed.rbegin(); it != removed.rend(); ++it) {
-				EXIT_IF(!MapExistingPlaceholderFixed(it->vaddr, it->size, it->backing_offset,
-				                                     it->mode));
+				EXIT_IF(!MapFixed(it->vaddr, it->size, it->backing_offset, it->mode));
 			}
 			return false;
 		}
@@ -363,12 +359,10 @@ public:
 			return false;
 		}
 
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 		const uint64_t left_size  = vaddr - old.vaddr;
 		const uint64_t right_addr = vaddr + size;
-		const uint64_t old_end    = old.vaddr + old.size;
-		const uint64_t right_size = old_end - right_addr;
-
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+		const uint64_t right_size = old.vaddr + old.size - right_addr;
 		const bool exact_unmap = (vaddr == old.vaddr && size == old.size);
 		if (exact_unmap) {
 			const bool unmapped =
@@ -393,14 +387,11 @@ public:
 				ok = SplitPlaceholder(vaddr, size) && ok;
 			}
 			if (left_size != 0) {
-				ok = MapExistingPlaceholderFixed(old.vaddr, left_size, old.backing_offset,
-				                                 old.mode) &&
-				     ok;
+				ok = MapFixed(old.vaddr, left_size, old.backing_offset, old.mode) && ok;
 			}
 			if (right_size != 0) {
-				ok = MapExistingPlaceholderFixed(right_addr, right_size,
-				                                 old.backing_offset + (right_addr - old.vaddr),
-				                                 old.mode) &&
+				ok = MapFixed(right_addr, right_size,
+				              old.backing_offset + (right_addr - old.vaddr), old.mode) &&
 				     ok;
 			}
 			if (ok) {
@@ -418,8 +409,7 @@ public:
 			}
 			EXIT_IF(VirtualFree(reinterpret_cast<void*>(old.vaddr), old.size,
 			                    MEM_RELEASE | MEM_COALESCE_PLACEHOLDERS) == 0);
-			EXIT_IF(
-			    !MapExistingPlaceholderFixed(old.vaddr, old.size, old.backing_offset, old.mode));
+			EXIT_IF(!MapFixed(old.vaddr, old.size, old.backing_offset, old.mode));
 			return false;
 		}
 		std::lock_guard lock(m_mutex);
@@ -433,16 +423,7 @@ public:
 		}
 
 		std::lock_guard lock(m_mutex);
-		m_maps.erase(old.vaddr);
-		if (left_size != 0) {
-			m_maps.emplace(old.vaddr,
-			               MapEntry {old.vaddr, left_size, old.backing_offset, old.mode});
-		}
-		if (right_size != 0) {
-			m_maps.emplace(right_addr,
-			               MapEntry {right_addr, right_size,
-			                         old.backing_offset + right_addr - old.vaddr, old.mode});
-		}
+		RemoveMappingsUnlocked(vaddr, size);
 
 		if (placeholder_preserved != nullptr) {
 			*placeholder_preserved = true;
@@ -516,6 +497,33 @@ private:
 		VirtualMemory::Mode mode           = VirtualMemory::Mode::NoAccess;
 	};
 
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX
+	void RemoveMappingsUnlocked(uint64_t vaddr, uint64_t size) {
+		const auto end = vaddr + size;
+		auto       it = m_maps.upper_bound(vaddr);
+		if (it != m_maps.begin()) {
+			--it;
+		}
+		while (it != m_maps.end() && it->first < end) {
+			const auto old     = it->second;
+			const auto old_end = old.vaddr + old.size;
+			if (old_end <= vaddr) {
+				++it;
+				continue;
+			}
+			it = m_maps.erase(it);
+			if (old.vaddr < vaddr) {
+				m_maps.emplace(old.vaddr,
+				               MapEntry {old.vaddr, vaddr - old.vaddr, old.backing_offset, old.mode});
+			}
+			if (end < old_end) {
+				m_maps.emplace(end, MapEntry {end, old_end - end,
+				                             old.backing_offset + end - old.vaddr, old.mode});
+			}
+		}
+	}
+#endif
+
 	bool FindContainingUnlocked(uint64_t vaddr, uint64_t size, MapEntry* out) const {
 		EXIT_IF(out == nullptr);
 
@@ -621,6 +629,12 @@ private:
 
 	void* MapView(uint64_t vaddr, uint64_t size, uint64_t backing_offset, VirtualMemory::Mode mode,
 	              FailureReason* failure_reason) {
+#if defined(KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS)
+		if (const auto callback = g_test_before_backing_map) {
+			g_test_before_backing_map = nullptr;
+			callback(vaddr, size);
+		}
+#endif
 		void* ptr = mmap(reinterpret_cast<void*>(vaddr), size, GetProtect(mode),
 		                 MAP_SHARED | MAP_FIXED, m_fd, static_cast<off_t>(backing_offset));
 		if (ptr == MAP_FAILED) {
@@ -778,7 +792,12 @@ public:
 	                VirtualMemory::Mode               mode,
 	                GuestBackingStore::FailureReason* failure_reason = nullptr) {
 		std::lock_guard lock(m_mutex);
-		if (!ConsumeUnlocked(vaddr, size)) {
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX
+		const bool replacing = ContainsMappedUnlocked(vaddr, size, MappingType::Backing);
+#else
+		constexpr bool replacing = false;
+#endif
+		if (!replacing && !ConsumeUnlocked(vaddr, size)) {
 			if (failure_reason != nullptr) {
 				*failure_reason = GuestBackingStore::FailureReason::AddressRangeUnavailable;
 			}
@@ -787,12 +806,15 @@ public:
 		const auto host_mode = VirtualMemory::IsExecute(mode)
 		                           ? VirtualMemory::Mode::ExecuteReadWrite
 		                           : VirtualMemory::Mode::ReadWrite;
-		if (m_backing->MapExistingPlaceholderFixed(vaddr, size, backing_offset, host_mode,
-		                                           failure_reason)) {
-			EXIT_IF(!AddMappedUnlocked(vaddr, size, MappingType::Backing));
+		if (m_backing->MapFixed(vaddr, size, backing_offset, host_mode, failure_reason)) {
+			if (!replacing) {
+				EXIT_IF(!AddMappedUnlocked(vaddr, size, MappingType::Backing));
+			}
 			return true;
 		}
-		AddFreeUnlocked(vaddr, size);
+		if (!replacing) {
+			AddFreeUnlocked(vaddr, size);
+		}
 		return false;
 	}
 

src/launcher/CMakeLists.txt

@@ -21,6 +21,10 @@ add_executable(launcher WIN32
 
 
 target_link_libraries(launcher Qt6::Concurrent Qt6::Network Qt6::Widgets Vulkan::Headers SDL3::SDL3-static kyty_git_version common)
+if(LINUX)
+	find_package(Qt6 COMPONENTS DBus REQUIRED)
+	target_link_libraries(launcher Qt6::DBus)
+endif()
 if(WIN32)
 target_link_libraries(launcher setupapi)
 endif()

src/launcher/forms/configuration_edit_dialog.ui

@@ -359,16 +359,16 @@
        <layout class="QHBoxLayout" name="controllerVolumeLayout">
         <item>
          <widget class="QSlider" name="slider_controller_volume">
-          <property name="toolTip"><string>Scale the DualSense controller speaker over USB and Bluetooth</string></property>
+          <property name="toolTip"><string>Scale the DualSense speaker over USB and Bluetooth. 50% is the previous default; 100% adds 6 dB of gain.</string></property>
           <property name="maximum"><number>100</number></property>
-          <property name="value"><number>100</number></property>
+          <property name="value"><number>50</number></property>
           <property name="orientation"><enum>Qt::Horizontal</enum></property>
          </widget>
         </item>
         <item>
          <widget class="QLabel" name="label_controller_volume_value">
           <property name="minimumSize"><size><width>40</width><height>0</height></size></property>
-          <property name="text"><string>100%</string></property>
+          <property name="text"><string>50%</string></property>
          </widget>
         </item>
        </layout>

src/launcher/include/configuration.h

@@ -48,7 +48,7 @@ QString EnumToText(T value) {
 
 struct ControllerSettings {
 	QString color;
-	int     speaker_volume      = 100;
+	int     speaker_volume      = 50;
 	int     vibration_intensity = 100;
 
 	void WriteSettings(QSettings* s) const {
@@ -60,13 +60,13 @@ struct ControllerSettings {
 	void ReadSettings(QSettings* s) {
 		const QColor saved_color(s->value("controller_color").toString());
 		color = saved_color.isValid() ? saved_color.name(QColor::HexRgb) : QString {};
-		const auto read_percent = [s](const char* key) {
+		const auto read_percent = [s](const char* key, int fallback) {
 			bool      ok    = false;
-			const int value = s->value(key, 100).toInt(&ok);
-			return ok ? qBound(0, value, 100) : 100;
+			const int value = s->value(key, fallback).toInt(&ok);
+			return ok ? qBound(0, value, 100) : fallback;
 		};
-		speaker_volume      = read_percent("controller_speaker_volume");
-		vibration_intensity = read_percent("controller_vibration_intensity");
+		speaker_volume      = read_percent("controller_speaker_volume", 50);
+		vibration_intensity = read_percent("controller_vibration_intensity", 100);
 	}
 };
 

src/launcher/src/configurationListWidget.cpp

@@ -36,6 +36,7 @@
 #include <QPainter>
 #include <QPalette>
 #include <QPointer>
+#include <QProcess>
 #include <QPushButton>
 #include <QRegularExpression>
 #include <QSet>
@@ -47,6 +48,12 @@
 #include <QUrl>
 #include <QtCore>
 
+#ifdef __linux__
+#include <QDBusConnection>
+#include <QDBusMessage>
+#include <QDBusPendingCallWatcher>
+#endif
+
 #include <memory>
 
 #include "ui_configuration_list_widget.h"
@@ -837,16 +844,52 @@ void ConfigurationListWidget::open_game_folder() {
 		return;
 	}
 
-	const auto base = item->GetInfo().basedir;
-	const QDir game_dir(GameContent::IsArchive(base) ? QFileInfo(base).absolutePath() : base);
+	const auto base       = item->GetInfo().basedir;
+	const bool is_archive = GameContent::IsArchive(base);
+	const QDir game_dir(is_archive ? QFileInfo(base).absolutePath() : base);
 	if (!game_dir.exists()) {
 		QMessageBox::warning(this, tr("Open game folder"), tr("Game folder does not exist."));
 		return;
 	}
 
-	if (!QDesktopServices::openUrl(QUrl::fromLocalFile(game_dir.absolutePath()))) {
-		QMessageBox::warning(this, tr("Open game folder"), tr("Could not open game folder."));
+	const auto open_directory = [this, game_dir] {
+		if (!QDesktopServices::openUrl(QUrl::fromLocalFile(game_dir.absolutePath()))) {
+			QMessageBox::warning(this, tr("Open game folder"), tr("Could not open game folder."));
+		}
+	};
+	if (is_archive) {
+		const auto path = QFileInfo(base).absoluteFilePath();
+#if defined(_WIN32)
+		QProcess explorer;
+		explorer.setProgram("explorer.exe");
+		explorer.setNativeArguments(
+		    QStringLiteral("/select,\"%1\"").arg(QDir::toNativeSeparators(path)));
+		if (explorer.startDetached()) {
+			return;
+		}
+#elif defined(__APPLE__)
+		if (QProcess::startDetached("/usr/bin/open", {"-R", path})) {
+			return;
+		}
+#elif defined(__linux__)
+		auto request = QDBusMessage::createMethodCall("org.freedesktop.FileManager1",
+		                                              "/org/freedesktop/FileManager1",
+		                                              "org.freedesktop.FileManager1", "ShowItems");
+		request << QStringList {QUrl::fromLocalFile(path).toString(QUrl::FullyEncoded)}
+		        << QString();
+		auto* watcher = new QDBusPendingCallWatcher(
+		    QDBusConnection::sessionBus().asyncCall(request, 5000), this);
+		connect(watcher, &QDBusPendingCallWatcher::finished, this,
+		        [open_directory](QDBusPendingCallWatcher* call) {
+			        call->deleteLater();
+			        if (call->isError()) {
+				        open_directory();
+			        }
+		        });
+		return;
+#endif
 	}
+	open_directory();
 }
 
 void ConfigurationListWidget::remove_save_data() {

src/libs/controller.cpp

@@ -224,7 +224,9 @@ void GameController::CycleSetting(Setting setting) {
 float GameController::GetSettingScale(Setting setting) const {
 	const auto step = m_setting_steps[static_cast<size_t>(setting)].load(std::memory_order_relaxed);
 	if (setting == Setting::SpeakerVolume) {
-		return SPEAKER_VOLUME[step] * (Config::GetControllerSpeakerVolume() / 100.0f);
+		// The old speaker level matches the PS5 slider's midpoint. Give the upper
+		// half of the global slider up to 6 dB of extra software gain.
+		return SPEAKER_VOLUME[step] * (Config::GetControllerSpeakerVolume() / 50.0f);
 	}
 	if (setting == Setting::VibrationIntensity) {
 		return INTENSITY[step] * (Config::GetControllerVibrationIntensity() / 100.0f);

src/libs/dualSenseHaptics.cpp

@@ -178,7 +178,11 @@ SDL_AudioDeviceID FindDevice() {
 	for (int i = 0; i < count; i++) {
 		const char*   name = SDL_GetAudioDeviceName(devices[i]);
 		SDL_AudioSpec spec {};
-		if (name != nullptr && SDL_strcasestr(name, "DualSense") != nullptr &&
+		const bool is_dualsense_name =
+		    name != nullptr &&
+		    (SDL_strcasestr(name, "DualSense") != nullptr ||
+		     SDL_strcasestr(name, "Wireless Controller") != nullptr);
+		if (is_dualsense_name &&
 		    SDL_GetAudioDeviceFormat(devices[i], &spec, nullptr) && spec.channels == 4) {
 			if (device != 0) {
 				device = 0; // Multiple matching endpoints cannot be associated reliably either.

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/guestInstructionPatcher.cpp

@@ -590,7 +590,10 @@ DecodedFunction DecodeFunction(uintptr_t function_start, uintptr_t function_end,
 	return function;
 }
 
-RedZoneMask TranslateRedZoneMask(const RedZoneMask& mask, s64 stack_pointer_delta) {
+std::optional<RedZoneMask> TranslateRedZoneMask(const RedZoneMask& mask, s64 stack_pointer_delta) {
+	if (stack_pointer_delta == 0) {
+		return mask;
+	}
 	RedZoneMask translated;
 	for (size_t bit = 0; bit < GuestRedZoneSize; ++bit) {
 		if (!mask.test(bit)) {
@@ -598,9 +601,12 @@ RedZoneMask TranslateRedZoneMask(const RedZoneMask& mask, s64 stack_pointer_delt
 		}
 		const s64 after_offset  = static_cast<s64>(bit) - static_cast<s64>(GuestRedZoneSize);
 		const s64 before_offset = after_offset + stack_pointer_delta;
-		if (before_offset >= -static_cast<s64>(GuestRedZoneSize) && before_offset < 0) {
-			translated.set(static_cast<size_t>(before_offset + static_cast<s64>(GuestRedZoneSize)));
+		if (before_offset < -static_cast<s64>(GuestRedZoneSize) || before_offset >= 0) {
+			// A later inverse RSP adjustment can bring this byte back into the red zone.
+			// Losing its stack-slot identity makes the bounded analysis inconclusive.
+			return std::nullopt;
 		}
+		translated.set(static_cast<size_t>(before_offset + static_cast<s64>(GuestRedZoneSize)));
 	}
 	return translated;
 }
@@ -610,13 +616,16 @@ void AnalyzeRedZoneLiveness(DecodedFunction& function) {
 		return;
 	}
 
+	const auto protect_function = [&] {
+		for (auto& [_, decoded]: function.instructions) {
+			decoded.red_zone_live.set();
+		}
+	};
 	if (function.has_indirect_branch || function.requires_conservative_red_zone_tracking ||
 	    std::ranges::any_of(function.branch_targets, [&function](uintptr_t target) {
 		    return !function.instructions.contains(target);
 	    })) {
-		for (auto& [_, decoded]: function.instructions) {
-			decoded.red_zone_live.set();
-		}
+		protect_function();
 		return;
 	}
 
@@ -653,12 +662,14 @@ void AnalyzeRedZoneLiveness(DecodedFunction& function) {
 				add_successor(next_address);
 			}
 
-			const RedZoneMask translated_live_out =
-			    decoded.stack_pointer_delta.has_value()
-			        ? TranslateRedZoneMask(live_out, *decoded.stack_pointer_delta)
-			        : live_out;
+			const auto translated_live_out =
+			    TranslateRedZoneMask(live_out, decoded.stack_pointer_delta.value_or(0));
+			if (!translated_live_out) {
+				protect_function();
+				return;
+			}
 			const RedZoneMask new_live_in =
-			    decoded.red_zone_use | (translated_live_out & ~decoded.red_zone_def);
+			    decoded.red_zone_use | (*translated_live_out & ~decoded.red_zone_def);
 			if (new_live_in != live_in[reverse_index]) {
 				live_in[reverse_index] = new_live_in;
 				changed                = true;

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);

src/main.cpp

@@ -54,7 +54,7 @@ static void PrintUsage() {
 	         Config::DEFAULT_USER_ID);
 	::printf("  --mic <name>                        Capture from this microphone; omit for silence.\n");
 	::printf("  --controller-color <#RRGGBB>        Override the controller lightbar color.\n");
-	::printf("  --controller-volume <0-100>         DualSense speaker volume. Default: 100.\n");
+	::printf("  --controller-volume <0-100>         DualSense speaker volume. Default: 50.\n");
 	::printf("  --controller-vibration <0-100>      DualSense vibration intensity. Default: 100.\n");
 	::printf(
 	    "  --present-mode <value>               Fifo, Mailbox, or Immediate. Default: Mailbox.\n");

tests/ControllerSettingsTests.cpp

@@ -112,7 +112,7 @@ struct Controller {
 		Initialize();
 		Connect(1);
 		Check(GetSettingScale(Setting::SpeakerVolume) ==
-		              Config::GetControllerSpeakerVolume() / 100.0f &&
+		              Config::GetControllerSpeakerVolume() / 50.0f &&
 		          GetSettingScale(Setting::VibrationIntensity) ==
 		              Config::GetControllerVibrationIntensity() / 100.0f &&
 		          GetSettingScale(Setting::TriggerEffectIntensity) == 1.0f,
@@ -165,7 +165,7 @@ void TestSettingCycles() {
 
 void TestGlobalControllerLevels() {
 	Config::ConfigOptions options;
-	options.controller_speaker_volume      = 50;
+	options.controller_speaker_volume      = 25;
 	options.controller_vibration_intensity = 25;
 	Config::Load(options);
 	{
@@ -190,6 +190,23 @@ void TestGlobalControllerLevels() {
 		Check(rumble.back().large == 17 * 257 && rumble.back().small == 8 * 257,
 		      "cycling vibration did not apply the combined intensity to cached rumble");
 	}
+	options.controller_speaker_volume = 50;
+	Config::Load(options);
+	{
+		Controller controller;
+		Check(GetSettingScale(Setting::SpeakerVolume) == 1.0f,
+		      "speaker midpoint did not preserve the previous default level");
+	}
+	options.controller_speaker_volume = 100;
+	Config::Load(options);
+	{
+		Controller controller;
+		Check(GetSettingScale(Setting::SpeakerVolume) == 2.0f,
+		      "speaker maximum did not add 6 dB of gain");
+		CycleSetting(Setting::SpeakerVolume);
+		Check(GetSettingScale(Setting::SpeakerVolume) == 0.0f,
+		      "speaker maximum prevented the hotkey from muting");
+	}
 	options.controller_speaker_volume      = 0;
 	options.controller_vibration_intensity = 0;
 	Config::Load(options);