[vk] Remove UniformRing and vkCmdResetQueryPool

Fixes perfomance regression on Xenoblade Chronicles DE and Pokemon Scarlet (among other games)
It should be investigated the reason why such perfomance loss (more than ~10% in some case)

At core it partially reverted the following commits:

Reverts "[vk] Introduce Ring Buffers for Uniform Buffer (#2698)"
  This reverts commit 776958c79d.

Revert "[vk] Bring Vulkan closer to Spec (#180)"
  This reverts commit c8d6f23129.

Revert "[VK] PR 180 extension (#257)"
  This reverts commit 444b9f361e.

Revert "[vk] Fixes regression of PR #180 vk_scheduler.cpp for AMD GPU and Windows OS (#3071)"
  This reverts commit be218cc020.

Signed-off-by: Caio Oliveira <caiooliveirafarias0@gmail.com>

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -338,11 +338,6 @@ 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() //check if the device has it
-    );
-    // add the ability to change the size in settings in future
-    uniform_ring.Init(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_,
@@ -361,41 +356,6 @@ void BufferCacheRuntime::FreeDeferredStagingBuffer(StagingBufferRef& ref) {
     staging_pool.FreeDeferred(ref);
 }
 
-void BufferCacheRuntime::UniformRing::Init(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();
 }
@@ -416,7 +376,6 @@ 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() {

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -127,15 +127,9 @@ public:
 
     void BindTransformFeedbackBuffers(VideoCommon::HostBindings<Buffer>& bindings);
 
-    std::span<u8> BindMappedUniformBuffer([[maybe_unused]] size_t /*stage*/,
-                                          [[maybe_unused]] u32 /*binding_index*/,
+    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;
@@ -163,24 +157,6 @@ 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(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;

src/video_core/renderer_vulkan/vk_query_cache.cpp

@@ -156,19 +156,6 @@ public:
 
         ReserveHostQuery();
 
-        // Ensure outside render pass
-        scheduler.RequestOutsideRenderPassOperationContext();
-
-        // Reset query pool outside render pass
-        scheduler.Record([query_pool = current_query_pool,
-                                 query_index = current_bank_slot](vk::CommandBuffer cmdbuf) {
-            cmdbuf.ResetQueryPool(query_pool, static_cast<u32>(query_index), 1);
-        });
-
-        // Manually restart the render pass (required for vkCmdClearAttachments, etc.)
-        scheduler.RequestRenderpass(texture_cache.GetFramebuffer());
-
-        // Begin query inside the newly started render pass
         scheduler.Record([query_pool = current_query_pool,
                                  query_index = current_bank_slot](vk::CommandBuffer cmdbuf) {
             const bool use_precise = Settings::IsGPULevelHigh();

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -283,8 +283,6 @@ void Scheduler::EndRenderPass()
                        images = renderpass_images,
                        ranges = renderpass_image_ranges](vk::CommandBuffer cmdbuf) {
             std::array<VkImageMemoryBarrier, 9> barriers;
-            VkPipelineStageFlags src_stages = 0;
-
             for (size_t i = 0; i < num_images; ++i) {
                 const VkImageSubresourceRange& range = ranges[i];
                 const bool is_color = (range.aspectMask & VK_IMAGE_ASPECT_COLOR_BIT) != 0;
@@ -293,20 +291,14 @@ void Scheduler::EndRenderPass()
                                                  | VK_IMAGE_ASPECT_STENCIL_BIT)) !=0;
 
                 VkAccessFlags src_access = 0;
-                VkPipelineStageFlags this_stage = 0;
 
-                if (is_color) {
+                if (is_color)
                     src_access |= VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
-                    this_stage |= VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
-                }
-
-                if (is_depth_stencil) {
+                else if (is_depth_stencil)
                     src_access |= VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
-                    this_stage |= VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT
-                                  | VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT;
-                }
-
-                src_stages |= this_stage;
+                else
+                    src_access |= VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT
+                                  | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT;
 
                 barriers[i] = VkImageMemoryBarrier{
                         .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
@@ -325,15 +317,10 @@ void Scheduler::EndRenderPass()
                         .subresourceRange = range,
                 };
             }
-
-            // Graft: ensure explicit fragment tests + color output stages are always synchronized (AMD/Windows fix)
-            src_stages |= VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
-                    VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
-                    VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT;
-
             cmdbuf.EndRenderPass();
-
-            cmdbuf.PipelineBarrier(src_stages,
+            cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
+                                   VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
+                                   VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
                                    VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
                                    0,
                                    nullptr,

src/video_core/vulkan_common/vulkan_wrapper.cpp

@@ -119,7 +119,6 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkCmdEndConditionalRenderingEXT);
     X(vkCmdEndQuery);
     X(vkCmdEndRenderPass);
-    X(vkCmdResetQueryPool);
     X(vkCmdEndTransformFeedbackEXT);
     X(vkCmdEndDebugUtilsLabelEXT);
     X(vkCmdFillBuffer);

src/video_core/vulkan_common/vulkan_wrapper.h

@@ -219,7 +219,6 @@ struct DeviceDispatch : InstanceDispatch {
     PFN_vkCmdEndConditionalRenderingEXT vkCmdEndConditionalRenderingEXT{};
     PFN_vkCmdEndDebugUtilsLabelEXT vkCmdEndDebugUtilsLabelEXT{};
     PFN_vkCmdEndQuery vkCmdEndQuery{};
-    PFN_vkCmdResetQueryPool vkCmdResetQueryPool{};
     PFN_vkCmdEndRenderPass vkCmdEndRenderPass{};
     PFN_vkCmdEndTransformFeedbackEXT vkCmdEndTransformFeedbackEXT{};
     PFN_vkCmdFillBuffer vkCmdFillBuffer{};
@@ -1140,9 +1139,6 @@ public:
     VkCommandBuffer operator*() const noexcept {
         return handle;
     }
-    void ResetQueryPool(VkQueryPool query_pool, uint32_t first, uint32_t count) const noexcept {
-        dld->vkCmdResetQueryPool(handle, query_pool, first, count);
-    }
     void Begin(const VkCommandBufferBeginInfo& begin_info) const {
         Check(dld->vkBeginCommandBuffer(handle, &begin_info));
     }