From 6544cdd4466797d575284178d2d4d839ecdce474 Mon Sep 17 00:00:00 2001
From: wildcard <wildcard@eden-emu.dev>
Date: Wed, 8 Oct 2025 18:58:20 +0200
Subject: [PATCH] [Vulkan] Introduce Ring buffers for UB

Create 3 ring buffers which rotates between buffers each frame to avoid GPU/CPU conflicts
BindMappedUniformBuffer first tries to allocate from the ring buffer and falls back to staging pool only if allocation is too large.
Note to testers:- please test the performance since it is primarily a performance optimization and also look for visual bugs.
---
 .../renderer_vulkan/vk_buffer_cache.cpp       | 42 +++++++++++++++++++
 .../renderer_vulkan/vk_buffer_cache.h         | 29 ++++++++++++-
 2 files changed, 69 insertions(+), 2 deletions(-)

diff --git a/src/video_core/renderer_vulkan/vk_buffer_cache.cpp b/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
index 55565e3d79..399b61f0a4 100644
--- a/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_buffer_cache.cpp
@@ -337,6 +337,11 @@ BufferCacheRuntime::BufferCacheRuntime(const Device& device_, MemoryAllocator& m
         uint8_pass = std::make_unique<Uint8Pass>(device, scheduler, descriptor_pool, staging_pool,
                                                  compute_pass_descriptor_queue);
     }
+    const u32 ubo_align = static_cast<u32>(
+            device.GetUniformBufferAlignment() /* if you have it */
+    );
+    // add the ability to change the size in settings in future
+    uniform_ring.Init(device, memory_allocator, 8 * 1024 * 1024 /* 8 MiB */, ubo_align ? ubo_align : 256);
     quad_array_index_buffer = std::make_shared<QuadArrayIndexBuffer>(device_, memory_allocator_,
                                                                      scheduler_, staging_pool_);
     quad_strip_index_buffer = std::make_shared<QuadStripIndexBuffer>(device_, memory_allocator_,
@@ -355,6 +360,42 @@ void BufferCacheRuntime::FreeDeferredStagingBuffer(StagingBufferRef& ref) {
     staging_pool.FreeDeferred(ref);
 }
 
+void BufferCacheRuntime::UniformRing::Init(const Device& device,
+                                           MemoryAllocator& alloc,
+                                           u64 bytes, u32 alignment) {
+    for (size_t i = 0; i < NUM_FRAMES; ++i) {
+        VkBufferCreateInfo ci{
+            .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+            .pNext = nullptr,
+            .flags = 0,
+            .size  = bytes,
+            .usage = VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT,
+            .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+            .queueFamilyIndexCount = 0,
+            .pQueueFamilyIndices = nullptr,
+        };
+        buffers[i] = alloc.CreateBuffer(ci, MemoryUsage::Upload);
+        mapped[i] = buffers[i].Mapped().data();
+    }
+    size   = bytes;
+    align  = alignment ? alignment : 256;
+    head   = 0;
+    current_frame = 0;
+}
+
+std::span<u8> BufferCacheRuntime::UniformRing::Alloc(u32 bytes, u32& out_offset) {
+    const u64 aligned = Common::AlignUp(head, static_cast<u64>(align));
+    u64 end = aligned + bytes;
+
+    if (end > size) {
+       return {}; // Fallback to staging pool
+    }
+
+    out_offset = static_cast<u32>(aligned);
+    head = end;
+    return {mapped[current_frame] + out_offset, bytes};
+}
+
 u64 BufferCacheRuntime::GetDeviceLocalMemory() const {
     return device.GetDeviceLocalMemory();
 }
@@ -375,6 +416,7 @@ void BufferCacheRuntime::TickFrame(Common::SlotVector<Buffer>& slot_buffers) noe
     for (auto it = slot_buffers.begin(); it != slot_buffers.end(); it++) {
         it->ResetUsageTracking();
     }
+    uniform_ring.BeginFrame();
 }
 
 void BufferCacheRuntime::Finish() {
diff --git a/src/video_core/renderer_vulkan/vk_buffer_cache.h b/src/video_core/renderer_vulkan/vk_buffer_cache.h
index efe960258c..185259a6db 100644
--- a/src/video_core/renderer_vulkan/vk_buffer_cache.h
+++ b/src/video_core/renderer_vulkan/vk_buffer_cache.h
@@ -124,8 +124,15 @@ public:
 
     void BindTransformFeedbackBuffers(VideoCommon::HostBindings<Buffer>& bindings);
 
-    std::span<u8> BindMappedUniformBuffer([[maybe_unused]] size_t stage,
-                                          [[maybe_unused]] u32 binding_index, u32 size) {
+    std::span<u8> BindMappedUniformBuffer([[maybe_unused]] size_t /*stage*/,
+                                          [[maybe_unused]] u32 /*binding_index*/,
+                                          u32 size) {
+        u32 offset = 0;
+        if (auto span = uniform_ring.Alloc(size, offset); !span.empty()) {
+            BindBuffer(*uniform_ring.buffers[uniform_ring.current_frame], offset, size);
+            return span;
+        }
+        // Fallback for giant requests
         const StagingBufferRef ref = staging_pool.Request(size, MemoryUsage::Upload);
         BindBuffer(ref.buffer, static_cast<u32>(ref.offset), size);
         return ref.mapped_span;
@@ -153,6 +160,24 @@ private:
     void ReserveNullBuffer();
     vk::Buffer CreateNullBuffer();
 
+    struct UniformRing {
+        static constexpr size_t NUM_FRAMES = 3;
+        std::array<vk::Buffer, NUM_FRAMES> buffers{};
+        std::array<u8*, NUM_FRAMES> mapped{};
+        u64 size = 0;
+        u64 head = 0;
+        u32 align = 256;
+        size_t current_frame = 0;
+
+        void Init(const Device& device, MemoryAllocator& alloc, u64 bytes, u32 alignment);
+        void BeginFrame() {
+            current_frame = (current_frame + 1) % NUM_FRAMES;
+            head = 0;
+        }
+        std::span<u8> Alloc(u32 bytes, u32& out_offset);
+    };
+    UniformRing uniform_ring;
+
     const Device& device;
     MemoryAllocator& memory_allocator;
     Scheduler& scheduler;