[memory, vulkan] Initial implementation for Unified Memory

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -374,6 +374,93 @@ BufferCacheRuntime::BufferCacheRuntime(const Device& device_, MemoryAllocator& m
                                                                      scheduler_, staging_pool_);
 }
 
+void BufferCacheRuntime::TryEnableUnifiedMemory(void* base, size_t size,
+                                                std::span<AHardwareBuffer* const> hardware_buffers,
+                                                size_t hardware_buffer_window,
+                                                size_t hardware_buffer_base) {
+    unified_memory = std::make_unique<HostMemoryImport>(
+        device, base, size, hardware_buffers, hardware_buffer_window, hardware_buffer_base);
+    if (!unified_memory->IsValid()) {
+        unified_memory.reset();
+    }
+}
+
+void BufferCacheRuntime::CopyToUnifiedMemory(
+    size_t window_index, VkBuffer src_buffer,
+    std::span<const VideoCommon::BufferCopy> copies) {
+    if (!unified_memory || src_buffer == VK_NULL_HANDLE || copies.empty() ||
+        window_index >= unified_memory->GetWindowCount()) {
+        return;
+    }
+    const VkBuffer dst_buffer = unified_memory->GetWindowBuffer(window_index);
+    if (dst_buffer == VK_NULL_HANDLE) {
+        return;
+    }
+    VkDeviceSize covered_begin = std::numeric_limits<VkDeviceSize>::max();
+    VkDeviceSize covered_end = 0;
+    for (const VideoCommon::BufferCopy& copy : copies) {
+        covered_begin = (std::min)(covered_begin, static_cast<VkDeviceSize>(copy.dst_offset));
+        covered_end = (std::max)(covered_end,
+                                 static_cast<VkDeviceSize>(copy.dst_offset + copy.size));
+    }
+
+    boost::container::small_vector<VkBufferCopy, 8> vk_copies(copies.size());
+    std::ranges::transform(copies, vk_copies.begin(), MakeBufferCopy);
+
+    const bool foreign = unified_memory->NeedsForeignOwnershipTransfer();
+    const u32 queue_family = device.GetGraphicsFamily();
+
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([src_buffer, dst_buffer, vk_copies, foreign, queue_family, covered_begin,
+                      covered_end](vk::CommandBuffer cmdbuf) {
+        if (foreign) {
+            const VkBufferMemoryBarrier acquire{
+                .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
+                .pNext = nullptr,
+                .srcAccessMask = 0,
+                .dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
+                .dstQueueFamilyIndex = queue_family,
+                .buffer = dst_buffer,
+                .offset = covered_begin,
+                .size = covered_end - covered_begin,
+            };
+            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT,
+                                   VK_PIPELINE_STAGE_TRANSFER_BIT, 0, acquire);
+        }
+        cmdbuf.CopyBuffer(src_buffer, dst_buffer, VideoCommon::FixSmallVectorADL(vk_copies));
+        if (foreign) {
+            const VkBufferMemoryBarrier release{
+                .sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER,
+                .pNext = nullptr,
+                .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+                .dstAccessMask = 0,
+                .srcQueueFamilyIndex = queue_family,
+                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_FOREIGN_EXT,
+                .buffer = dst_buffer,
+                .offset = covered_begin,
+                .size = covered_end - covered_begin,
+            };
+            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT,
+                                   VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, 0, release);
+        }
+    });
+}
+
+void BufferCacheRuntime::UnifiedMemoryHostBarrier() {
+    static constexpr VkMemoryBarrier HOST_BARRIER{
+        .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+        .pNext = nullptr,
+        .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+        .dstAccessMask = VK_ACCESS_HOST_READ_BIT,
+    };
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([](vk::CommandBuffer cmdbuf) {
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_HOST_BIT, 0,
+                               HOST_BARRIER);
+    });
+}
+
 StagingBufferRef BufferCacheRuntime::UploadStagingBuffer(size_t size) {
     return staging_pool.Request(size, MemoryUsage::Upload);
 }

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -7,6 +7,8 @@
 #pragma once
 
 #include <limits>
+#include <memory>
+#include <span>
 
 #include <boost/container/small_vector.hpp>
 
@@ -110,6 +112,31 @@ public:
 
     void TickFrame(Common::SlotVector<Buffer>& slot_buffers) noexcept;
 
+    void TryEnableUnifiedMemory(void* base, size_t size,
+                                std::span<AHardwareBuffer* const> hardware_buffers,
+                                size_t hardware_buffer_window, size_t hardware_buffer_base);
+
+    [[nodiscard]] bool HasUnifiedMemory() const noexcept {
+        return unified_memory != nullptr && unified_memory->IsValid();
+    }
+
+    [[nodiscard]] u64 UnifiedMemorySize() const noexcept {
+        return unified_memory ? unified_memory->GetSize() : 0;
+    }
+
+    [[nodiscard]] u64 UnifiedMemoryBase() const noexcept {
+        return unified_memory ? unified_memory->GetBaseOffset() : 0;
+    }
+
+    [[nodiscard]] u64 UnifiedMemoryWindowSize() const noexcept {
+        return unified_memory ? unified_memory->GetWindowSize() : 0;
+    }
+
+    void CopyToUnifiedMemory(size_t window_index, VkBuffer src_buffer,
+                             std::span<const VideoCommon::BufferCopy> copies);
+
+    void UnifiedMemoryHostBarrier();
+
     u64 CurrentTick();
 
     u64 KnownGpuTick();
@@ -257,6 +284,7 @@ private:
     std::shared_ptr<QuadStripIndexBuffer> quad_strip_index_buffer;
 
     vk::Buffer null_buffer;
+    std::unique_ptr<HostMemoryImport> unified_memory;
 
     std::unique_ptr<Uint8Pass> uint8_pass;
     QuadIndexedPass quad_index_pass;
@@ -283,6 +311,7 @@ struct BufferCacheParams {
     static constexpr bool USE_MEMORY_MAPS = true;
     static constexpr bool SEPARATE_IMAGE_BUFFER_BINDINGS = false;
     static constexpr bool USE_MEMORY_MAPS_FOR_UPLOADS = true;
+    static constexpr bool USE_UNIFIED_MEMORY = true;
 };
 
 using BufferCache = VideoCommon::BufferCache<BufferCacheParams>;

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -225,6 +225,13 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache, device, scheduler),
       wfi_event(device.GetLogical().CreateEvent()) {
     scheduler.SetQueryCache(query_cache);
+    if (Settings::values.use_unified_memory.GetValue() && device_memory.IsBackingShared()) {
+        buffer_cache_runtime.TryEnableUnifiedMemory(
+            device_memory.GetPhysicalBase(), device_memory.GetPhysicalSize(),
+            device_memory.GetBackingHardwareBuffers(),
+            device_memory.GetBackingHardwareBufferWindowSize(),
+            device_memory.GetBackingHardwareBufferBase());
+    }
 }
 
 RasterizerVulkan::~RasterizerVulkan() {