renderer: reduce synchronization and allocation overhead

Performance changes from #1495:

- Sync changed BDA ranges instead of scanning all mappings.
- Let VMA suballocate BDA buffers.
- Remove the renderer lock from presentation.
- Keep compatible pipeline caches across builds.
- Fold masked shifts that always produce zero.

src/graphics/host_gpu/memoryTracker.cpp

@@ -66,6 +66,18 @@ RegionManager* MemoryTracker::GetOrCreateRegion(uint64_t index) {
 	return ptr;
 }
 
+void MemoryTracker::QueueBdaSync(uint64_t vaddr, uint64_t size) {
+	const auto begin = Common::AlignDown(vaddr, TRACKER_PAGE_SIZE);
+	const auto end   = Common::AlignUp(vaddr + size, TRACKER_PAGE_SIZE);
+	std::lock_guard lock(m_bda_sync_mutex);
+	m_bda_sync.Add(begin, end - begin);
+}
+
+RangeSet MemoryTracker::TakeBdaSync() {
+	std::lock_guard lock(m_bda_sync_mutex);
+	return std::exchange(m_bda_sync, RangeSet {});
+}
+
 bool MemoryTracker::IsRegionCpuModified(uint64_t vaddr, uint64_t size) {
 	CheckNotInUploadCallback();
 	return Iterate<true>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
@@ -88,6 +100,7 @@ void MemoryTracker::MarkRegionAsCpuModified(uint64_t vaddr, uint64_t size) {
 		std::scoped_lock lock(manager->lock);
 		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
+	QueueBdaSync(vaddr, size);
 }
 
 void MemoryTracker::MarkRegionAsGpuModified(uint64_t vaddr, uint64_t size) {
@@ -128,6 +141,8 @@ void MemoryTracker::UntrackMemory(uint64_t vaddr, uint64_t size) {
 	Iterate<false>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
+	locks.clear();
+	QueueBdaSync(vaddr, size);
 }
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/memoryTracker.h

@@ -29,6 +29,10 @@ public:
 	void               MarkRegionAsGpuModified(uint64_t vaddr, uint64_t size);
 	void               UnmarkRegionAsGpuModified(uint64_t vaddr, uint64_t size);
 	void               UntrackMemory(uint64_t vaddr, uint64_t size);
+	// Queue after changing ownership; a concurrent drain must not consume an event
+	// before the dirty pages it describes become visible.
+	void                   QueueBdaSync(uint64_t vaddr, uint64_t size);
+	[[nodiscard]] RangeSet TakeBdaSync();
 	// Removes protection from a range and flushes GPU-owned data when required.
 	template <typename Flush>
 	void InvalidateRegion(uint64_t vaddr, uint64_t size, Flush&& on_flush) noexcept {
@@ -51,6 +55,7 @@ public:
 				on_flush();
 			}
 		});
+		QueueBdaSync(vaddr, size);
 	}
 #if KYTY_BUILD == KYTY_BUILD_DEBUG
 	void ValidateGpuDirtyPages(const RangeSet& dirty, uint64_t vaddr, uint64_t size,
@@ -149,6 +154,8 @@ private:
 	std::unique_ptr<std::atomic<RegionManager*>[]> m_regions;
 	std::vector<std::unique_ptr<RegionManager>>    m_region_storage;
 	std::mutex                                     m_region_mutex;
+	std::mutex                                     m_bda_sync_mutex;
+	RangeSet                                       m_bda_sync;
 	PageManager&                                   m_page_manager;
 };
 

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

@@ -126,6 +126,7 @@ void BufferCache::ChangeRegister(BufferId id) {
 		}
 		WriteDataBuffer(m_bda_pagetable_buffer, table_offset,
 		                addresses.data(), addresses.size() * sizeof(vk::DeviceAddress));
+		m_memory_tracker.QueueBdaSync(buffer.CpuAddress(), buffer.Size());
 	} else {
 		const auto found = m_buffers.find(buffer.CpuAddress());
 		EXIT_IF(found == m_buffers.end() || found->second != id);
@@ -685,6 +686,16 @@ void BufferCache::ProcessFaultBuffer() {
 	m_fault_manager.ProcessFaultBuffer();
 }
 
+void BufferCache::SynchronizeBda(const RangeSet& mapped) {
+	// Take before walking: writes queued during synchronization remain pending.
+	const auto pending = m_memory_tracker.TakeBdaSync();
+	pending.ForEach([&](uint64_t begin, uint64_t end) {
+		mapped.ForEachInRange(begin, end - begin, [this](uint64_t start, uint64_t finish) {
+			SynchronizeBuffersInRange(start, finish - start);
+		});
+	});
+}
+
 void BufferCache::SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size) {
 	const auto end = vaddr + size;
 	auto       it  = m_buffers.upper_bound(vaddr);

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

@@ -80,7 +80,10 @@ public:
 	[[nodiscard]] bool IsRegionCpuModified(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] bool IsRegionGpuModified(uint64_t vaddr, uint64_t size);
 	void               ProcessFaultBuffer();
-	void               SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size);
+	void               SynchronizeBda(const RangeSet& mapped);
+	void               QueueBdaSync(uint64_t vaddr, uint64_t size) {
+		m_memory_tracker.QueueBdaSync(vaddr, size);
+	}
 	void               RunGarbageCollector();
 
 private:
@@ -114,6 +117,7 @@ private:
 	void DeleteBuffer(BufferId id);
 	[[nodiscard]] bool SynchronizeBuffer(Buffer& buffer, uint64_t vaddr, uint64_t size,
 	                                     bool is_written, bool is_texel_buffer);
+	void SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] vk::Buffer UploadCopies(Buffer& buffer, std::span<vk::BufferCopy> copies,
 	                                      uint64_t total_size);
 	[[nodiscard]] bool SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uint64_t size);

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

@@ -67,11 +67,9 @@ Buffer::Buffer(GraphicContext& graphics, CommandScheduler& scheduler, MemoryUsag
 	buffer_info.usage       = flags;
 
 	const bool with_bda = bool(flags & vk::BufferUsageFlagBits::eShaderDeviceAddress);
-	const VmaAllocationCreateFlags bda_flag =
-	    with_bda ? VMA_ALLOCATION_CREATE_DEDICATED_MEMORY_BIT : 0;
 	VmaAllocationCreateInfo allocation_info {};
 	allocation_info.flags =
-	    VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT | bda_flag | AllocationFlags(usage);
+	    VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT | AllocationFlags(usage);
 	allocation_info.usage = AllocationUsage(usage);
 	allocation_info.preferredFlags = usage == MemoryUsage::DeviceLocal
 	                                     ? VkMemoryPropertyFlags {}

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

@@ -18,7 +18,6 @@
 #include "graphics/shader/recompiler/ShaderRecompiler.h"
 #include "graphics/shader/shaderCompiler.h"
 #include "kernel/memory.h"
-#include "kytyGitVersion.h"
 #include "loader/systemContent.h"
 
 #include <algorithm>
@@ -81,7 +80,8 @@ std::string DriverCacheSignature(const vk::PhysicalDeviceProperties& properties)
 		uuid[i * 2]     = hex[properties.pipelineCacheUUID[i] >> 4u];
 		uuid[i * 2 + 1] = hex[properties.pipelineCacheUUID[i] & 0xfu];
 	}
-	return fmt::format("KytyPC1:{}:{:08x}:{:08x}:{:08x}:{}\n", KYTY_GIT_REVISION,
+	// Driver cache entries already distinguish shader code and pipeline state across builds.
+	return fmt::format("KytyPC2:{:08x}:{:08x}:{:08x}:{}\n",
 	                   properties.vendorID, properties.deviceID, properties.driverVersion, uuid);
 }
 
@@ -516,17 +516,6 @@ void PipelineCache::InitializeDriverCache() {
 		PipelineCacheLog("Vulkan pipeline cache: disabled (non-Release build)");
 		return;
 	}
-	const std::string_view git_hash     = KYTY_GIT_HASH;
-	const std::string_view git_revision = KYTY_GIT_REVISION;
-	if (git_hash == "unknown" || git_revision == "unknown") {
-		PipelineCacheLog("Vulkan pipeline cache: disabled (unknown git revision)");
-		return;
-	}
-	if (git_hash.ends_with("-dirty")) {
-		PipelineCacheLog("Vulkan pipeline cache: disabled (dirty build)");
-		return;
-	}
-
 	m_driver_cache_path     = std::filesystem::path("_PipelineCache") / (title_id + ".bin");
 	const auto path         = Common::PathToString(m_driver_cache_path);
 	const bool cache_exists = Common::File::IsFileExisting(m_driver_cache_path);
@@ -559,7 +548,7 @@ void PipelineCache::InitializeDriverCache() {
 			    XXH3_64bits(initial_data.data(), initial_data.size()) != payload_hash) {
 				initial_data.clear();
 				PipelineCacheLog(
-				    "Vulkan pipeline cache: invalidating {} (driver, emulator, or data mismatch)",
+				    "Vulkan pipeline cache: invalidating {} (driver or data mismatch)",
 				    path);
 			}
 		} else {

src/graphics/host_gpu/renderer/renderContext.cpp

@@ -90,6 +90,7 @@ bool RenderContext::IsMapped(uint64_t vaddr, uint64_t size) const noexcept {
 void RenderContext::MapMemory(uint64_t vaddr, uint64_t size) {
 	std::lock_guard lock(m_mapped_ranges_mutex);
 	m_mapped_ranges.Add(vaddr, size);
+	m_buffer_cache.QueueBdaSync(vaddr, size);
 }
 
 void RenderContext::UnmapMemory(uint64_t vaddr, uint64_t size) {
@@ -129,9 +130,7 @@ void RenderContext::PrepareBda() {
 		m_bda_logged = true;
 	}
 	std::shared_lock lock(m_mapped_ranges_mutex);
-	m_mapped_ranges.ForEach([this](uint64_t start, uint64_t end) {
-		m_buffer_cache.SynchronizeBuffersInRange(start, end - start);
-	});
+	m_buffer_cache.SynchronizeBda(m_mapped_ranges);
 	m_fault_process_pending = true;
 }
 

src/graphics/presentation/window/swapchain.cpp

@@ -945,7 +945,6 @@ Presenter::Frame& Presenter::PrepareBlankFrame(uint32_t width, uint32_t height,
 	auto*             frame  = m_impl->frames.Acquire({width, height}, format);
 	frame->hdr   = false;
 	frame->blank = true;
-	Common::LockGuard render_lock(m_impl->renderer.GetMutex());
 	frame->Configure(m_impl->window.graphic_ctx, {width, height}, format);
 	vk::ClearColorValue clear {};
 	clear.float32[3] = opaque ? 1.0f : 0.0f;
@@ -953,6 +952,7 @@ Presenter::Frame& Presenter::PrepareBlankFrame(uint32_t width, uint32_t height,
 		EXIT_IF(producer->IsInvalid());
 		frame->Clear(*producer, clear);
 	} else {
+		Common::LockGuard lock(m_impl->present_mutex);
 		auto& command = m_impl->present_scheduler.BeginCommand();
 		frame->Clear(command, clear);
 		frame->present_tick = m_impl->present_scheduler.Submit();
@@ -1036,18 +1036,17 @@ void Presenter::Impl::Present() {
 			RecoverSwapchain(status);
 			continue;
 		}
-		{
-			Common::LockGuard render_lock(renderer.GetMutex());
-			auto&             command = present_scheduler.BeginCommand();
-			const bool        draw_system_overlay =
-			    overlay_visual.active && swapchain.PrepareSystemOverlay();
-			swapchain.RecordPresentCommands(command, layers[0].frame, layers[1],
-			                                draw_system_overlay);
-			const auto tick = swapchain.Submit(present_scheduler);
-			for (const auto& layer: layers) {
-				if (layer.frame != nullptr) {
-					layer.frame->present_tick = tick;
-				}
+		// present_mutex owns the prepared frames and this scheduler. Only queue submission
+		// is shared with rendering, and it has its own queue_mutex.
+		auto&      command = present_scheduler.BeginCommand();
+		const bool draw_system_overlay =
+		    overlay_visual.active && swapchain.PrepareSystemOverlay();
+		swapchain.RecordPresentCommands(command, layers[0].frame, layers[1],
+		                                draw_system_overlay);
+		const auto tick = swapchain.Submit(present_scheduler);
+		for (const auto& layer: layers) {
+			if (layer.frame != nullptr) {
+				layer.frame->present_tick = tick;
 			}
 		}
 		status = swapchain.Present();

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

@@ -608,13 +608,23 @@ void FoldInstruction(Block& block, Block::iterator instruction,
 			    ReplaceBinaryIdentity(inst, Type::U32, UINT32_MAX)) return;
 			for (uint32_t arm = 0; arm < 2u; ++arm) {
 				const auto mask = Arg(inst, arm);
+				if (!IsImmediate(mask, Type::U32)) continue;
+				if (mask.U32() == 0u) {
+					Replace(inst, Value(0u));
+					return;
+				}
 				const auto* shifted = Arg(inst, arm ^ 1u).TryInstruction();
-				if (!IsImmediate(mask, Type::U32) || shifted == nullptr ||
-				    shifted->GetOpcode() != ValueOpcode::ShiftLeftLogical32 || !HasOneUse(*shifted)) continue;
+				if (shifted == nullptr || shifted->GetOpcode() != ValueOpcode::ShiftLeftLogical32) continue;
 				const auto shift = Arg(*shifted, 1);
-				if (!IsImmediate(shift, Type::U32) || shift.U32() >= 32u) continue;
+				if (!IsImmediate(shift, Type::U32)) continue;
+				const auto shifted_mask = mask.U32() >> (shift.U32() & 31u);
+				if (shifted_mask == 0u) {
+					Replace(inst, Value(0u));
+					return;
+				}
+				if (!HasOneUse(*shifted) || shift.U32() >= 32u) continue;
 				const auto masked = worklist.Prepend(block, instruction, ValueOpcode::BitwiseAnd32,
-				    {Arg(*shifted, 0), Value(mask.U32() >> shift.U32())});
+				    {Arg(*shifted, 0), Value(shifted_mask)});
 				Replace(inst, Value(&*worklist.Prepend(block, instruction, ValueOpcode::ShiftLeftLogical32,
 				    {Value(&*masked), shift})));
 				return;