Update src/video_core/renderer_vulkan/vk_buffer_cache.cpp

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.

src/common/settings.h

@@ -161,7 +161,7 @@ struct Values {
                                          Category::LibraryApplet};
     Setting<AppletMode> photo_viewer_applet_mode{
                                                  linkage, AppletMode::LLE, "photo_viewer_applet_mode", Category::LibraryApplet};
-    Setting<AppletMode> offline_web_applet_mode{linkage, AppletMode::LLE, "offline_web_applet_mode",
+    Setting<AppletMode> offline_web_applet_mode{linkage, AppletMode::HLE, "offline_web_applet_mode",
                                                 Category::LibraryApplet};
     Setting<AppletMode> login_share_applet_mode{linkage, AppletMode::HLE, "login_share_applet_mode",
                                                 Category::LibraryApplet};

src/core/device_memory_manager.h

@@ -1,6 +1,3 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
-// SPDX-License-Identifier: GPL-3.0-or-later
-
 // SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
@@ -112,9 +109,6 @@ public:
 
     void ReadBlock(DAddr address, void* dest_pointer, size_t size);
     void ReadBlockUnsafe(DAddr address, void* dest_pointer, size_t size);
-#ifdef YUZU_DEBUG
-    bool ReadBlockFastChecked(DAddr address, void* dest_pointer, size_t size);
-#endif
     void WriteBlock(DAddr address, const void* src_pointer, size_t size);
     void WriteBlockUnsafe(DAddr address, const void* src_pointer, size_t size);
 

src/core/device_memory_manager.inc

@@ -1,6 +1,3 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
-// SPDX-License-Identifier: GPL-3.0-or-later
-
 // SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
@@ -470,29 +467,6 @@ void DeviceMemoryManager<Traits>::ReadBlockUnsafe(DAddr address, void* dest_poin
         });
 }
 
-#ifdef YUZU_DEBUG
-template <typename Traits>
-bool DeviceMemoryManager<Traits>::ReadBlockFastChecked(DAddr address, void* dest_pointer,
-                                                       size_t size) {
-    bool success = true;
-    WalkBlock(
-        address, size,
-        [&](size_t copy_amount, DAddr current_vaddr) {
-            LOG_CRITICAL(Render, "DeviceMemory OOB/unmapped: addr=0x{:x} size={}", current_vaddr,
-                         size);
-            std::memset(dest_pointer, 0, copy_amount);
-            success = false;
-        },
-        [&](size_t copy_amount, const u8* const src_ptr) {
-            std::memcpy(dest_pointer, src_ptr, copy_amount);
-        },
-        [&](const std::size_t copy_amount) {
-            dest_pointer = static_cast<u8*>(dest_pointer) + copy_amount;
-        });
-    return success;
-}
-#endif
-
 template <typename Traits>
 void DeviceMemoryManager<Traits>::WriteBlockUnsafe(DAddr address, const void* src_pointer,
                                                    size_t size) {

src/video_core/buffer_cache/buffer_cache.h

@@ -386,10 +386,11 @@ void BufferCache<P>::BindHostComputeBuffers() {
 template <class P>
 void BufferCache<P>::SetUniformBuffersState(const std::array<u32, NUM_STAGES>& mask,
                                             const UniformBufferSizes* sizes) {
-    const bool mask_changed = channel_state->enabled_uniform_buffer_masks != mask;
-    if (mask_changed) {
-        channel_state->fast_bound_uniform_buffers.fill(0);
-        if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
+    if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
+        if (channel_state->enabled_uniform_buffer_masks != mask) {
+            if constexpr (IS_OPENGL) {
+                channel_state->fast_bound_uniform_buffers.fill(0);
+            }
             channel_state->dirty_uniform_buffers.fill(~u32{0});
             channel_state->uniform_buffer_binding_sizes.fill({});
         }
@@ -805,7 +806,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
                     channel_state->uniform_buffer_binding_sizes[stage][binding_index] != size;
                 if (should_fast_bind) {
                     // We only have to bind when the currently bound buffer is not the fast version
-                    channel_state->fast_bound_uniform_buffers[stage] |= 1u << binding_index;
+                    channel_state->fast_bound_uniform_buffers[stage] |= 1U << binding_index;
                     channel_state->uniform_buffer_binding_sizes[stage][binding_index] = size;
                     runtime.BindFastUniformBuffer(stage, binding_index, size);
                 }
@@ -814,22 +815,13 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
                 return;
             }
         }
-        channel_state->fast_bound_uniform_buffers[stage] |= 1u << binding_index;
-        channel_state->uniform_buffer_binding_sizes[stage][binding_index] = size;
+        if constexpr (IS_OPENGL) {
+            channel_state->fast_bound_uniform_buffers[stage] |= 1U << binding_index;
+            channel_state->uniform_buffer_binding_sizes[stage][binding_index] = size;
+        }
         // Stream buffer path to avoid stalling on non-Nvidia drivers or Vulkan
         const std::span<u8> span = runtime.BindMappedUniformBuffer(stage, binding_index, size);
-#ifdef YUZU_DEBUG
-        ASSERT(binding_index < NUM_GRAPHICS_UNIFORM_BUFFERS);
-        ASSERT(span.size() >= size && "UBO stream span too small");
-        if (!device_memory.ReadBlockFastChecked(device_addr, span.data(), size)) {
-            LOG_CRITICAL(Render, "DeviceMemory OOB/unmapped: addr=0x{:x} size={}", device_addr, size);
-            channel_state->fast_bound_uniform_buffers[stage] &= ~(1u << binding_index);
-            ASSERT(false);
-            return;
-        }
-#else
         device_memory.ReadBlockUnsafe(device_addr, span.data(), size);
-#endif
         return;
     }
     // Classic cached path
@@ -838,8 +830,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
     }
     // Skip binding if it's not needed and if the bound buffer is not the fast version
     // This exists to avoid instances where the fast buffer is bound and a GPU write happens
-    const bool was_fast_bound = HasFastUniformBufferBound(stage, binding_index);
-    needs_bind |= was_fast_bound;
+    needs_bind |= HasFastUniformBufferBound(stage, binding_index);
     if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
         needs_bind |= channel_state->uniform_buffer_binding_sizes[stage][binding_index] != size;
     }
@@ -848,6 +839,9 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
     }
     const u32 offset = buffer.Offset(device_addr);
     if constexpr (IS_OPENGL) {
+        // Fast buffer will be unbound
+        channel_state->fast_bound_uniform_buffers[stage] &= ~(1U << binding_index);
+
         // Mark the index as dirty if offset doesn't match
         const bool is_copy_bind = offset != 0 && !runtime.SupportsNonZeroUniformOffset();
         channel_state->dirty_uniform_buffers[stage] |= (is_copy_bind ? 1U : 0U) << index;
@@ -861,7 +855,6 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
     } else {
         runtime.BindUniformBuffer(buffer, offset, size);
     }
-    channel_state->fast_bound_uniform_buffers[stage] &= ~(1u << binding_index);
 }
 
 template <class P>
@@ -1796,7 +1789,12 @@ std::span<u8> BufferCache<P>::ImmediateBuffer(size_t wanted_capacity) {
 
 template <class P>
 bool BufferCache<P>::HasFastUniformBufferBound(size_t stage, u32 binding_index) const noexcept {
-    return ((channel_state->fast_bound_uniform_buffers[stage] >> binding_index) & 1u) != 0;
+    if constexpr (IS_OPENGL) {
+        return ((channel_state->fast_bound_uniform_buffers[stage] >> binding_index) & 1) != 0;
+    } else {
+        // Only OpenGL has fast uniform buffers
+        return false;
+    }
 }
 
 template <class P>

src/video_core/buffer_cache/buffer_cache_base.h

@@ -53,7 +53,6 @@ constexpr u32 NUM_COMPUTE_UNIFORM_BUFFERS = 8;
 constexpr u32 NUM_STORAGE_BUFFERS = 16;
 constexpr u32 NUM_TEXTURE_BUFFERS = 32;
 constexpr u32 NUM_STAGES = 5;
-static_assert(NUM_GRAPHICS_UNIFORM_BUFFERS <= 32, "fast bitmask must fit u32");
 
 using UniformBufferSizes = std::array<std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS>, NUM_STAGES>;
 using ComputeUniformBufferSizes = std::array<u32, NUM_COMPUTE_UNIFORM_BUFFERS>;
@@ -138,8 +137,8 @@ public:
     u32 written_compute_texture_buffers = 0;
     u32 image_compute_texture_buffers = 0;
 
-    std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS> uniform_cache_hits{};
-    std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS> uniform_cache_shots{};
+    std::array<u32, 16> uniform_cache_hits{};
+    std::array<u32, 16> uniform_cache_shots{};
 
     u32 uniform_buffer_skip_cache_size = DEFAULT_SKIP_CACHE_SIZE;
 

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() //check if the device has 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() {

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;

src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp

@@ -25,48 +25,35 @@ namespace {
 
 using namespace Common::Literals;
 
-// Minimum alignment we want to enforce for the streaming ring
-constexpr VkDeviceSize MIN_STREAM_ALIGNMENT = 256;
+// Maximum potential alignment of a Vulkan buffer
+constexpr VkDeviceSize MAX_ALIGNMENT = 256;
 // Stream buffer size in bytes
 constexpr VkDeviceSize MAX_STREAM_BUFFER_SIZE = 128_MiB;
 
-size_t GetStreamBufferSize(const Device& device, VkDeviceSize alignment) {
+size_t GetStreamBufferSize(const Device& device) {
     VkDeviceSize size{0};
     if (device.HasDebuggingToolAttached()) {
-        bool found_heap = false;
-        ForEachDeviceLocalHostVisibleHeap(device, [&size, &found_heap](size_t /*index*/, VkMemoryHeap& heap) {
+        ForEachDeviceLocalHostVisibleHeap(device, [&size](size_t index, VkMemoryHeap& heap) {
             size = (std::max)(size, heap.size);
-            found_heap = true;
         });
-        // If no suitable heap was found fall back to the default cap to avoid creating a zero-sized stream buffer.
-        if (!found_heap) {
-            size = MAX_STREAM_BUFFER_SIZE;
-        } else if (size <= 256_MiB) {
-            // If rebar is not supported, cut the max heap size to 40%. This will allow 2 captures to be
-            // loaded at the same time in RenderDoc. If rebar is supported, this shouldn't be an issue
-            // as the heap will be much larger.
+        // If rebar is not supported, cut the max heap size to 40%. This will allow 2 captures to be
+        // loaded at the same time in RenderDoc. If rebar is supported, this shouldn't be an issue
+        // as the heap will be much larger.
+        if (size <= 256_MiB) {
             size = size * 40 / 100;
         }
     } else {
         size = MAX_STREAM_BUFFER_SIZE;
     }
-
-    // Clamp to the configured maximum, align up for safety, and ensure a sane minimum so
-    // region_size (stream_buffer_size / NUM_SYNCS) never becomes zero.
-    const VkDeviceSize aligned =
-        (std::min)(Common::AlignUp(size, alignment), MAX_STREAM_BUFFER_SIZE);
-    const VkDeviceSize min_size = alignment * StagingBufferPool::NUM_SYNCS;
-    return static_cast<size_t>((std::max)(aligned, min_size));
+    return (std::min)(Common::AlignUp(size, MAX_ALIGNMENT), MAX_STREAM_BUFFER_SIZE);
 }
 } // Anonymous namespace
 
 StagingBufferPool::StagingBufferPool(const Device& device_, MemoryAllocator& memory_allocator_,
                                      Scheduler& scheduler_)
     : device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_},
-      stream_alignment{std::max<VkDeviceSize>(device_.GetUniformBufferAlignment(),
-                                              MIN_STREAM_ALIGNMENT)},
-      stream_buffer_size{GetStreamBufferSize(device_, stream_alignment)},
-      region_size{stream_buffer_size / StagingBufferPool::NUM_SYNCS} {
+      stream_buffer_size{GetStreamBufferSize(device)}, region_size{stream_buffer_size /
+                                                                   StagingBufferPool::NUM_SYNCS} {
     VkBufferCreateInfo stream_ci = {
         .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
         .pNext = nullptr,
@@ -119,54 +106,31 @@ void StagingBufferPool::TickFrame() {
 }
 
 StagingBufferRef StagingBufferPool::GetStreamBuffer(size_t size) {
-    const size_t alignment = static_cast<size_t>(stream_alignment);
-    const size_t aligned_size = Common::AlignUp(size, alignment);
-    const bool wraps = iterator + size >= stream_buffer_size;
-    const size_t new_iterator =
-        wraps ? aligned_size : Common::AlignUp(iterator + size, alignment);
-    const size_t begin_region = wraps ? 0 : Region(iterator);
-    const size_t last_byte = new_iterator == 0 ? 0 : new_iterator - 1;
-    const size_t end_region = (std::min)(Region(last_byte) + 1, NUM_SYNCS);
-    const size_t guard_begin = (std::min)(Region(free_iterator) + 1, NUM_SYNCS);
-
-    if (!wraps) {
-        if (guard_begin < end_region && AreRegionsActive(guard_begin, end_region)) {
-            // Avoid waiting for the previous usages to be free
-            return GetStagingBuffer(size, MemoryUsage::Upload);
-        }
-    } else if (guard_begin < NUM_SYNCS && AreRegionsActive(guard_begin, NUM_SYNCS)) {
+    if (AreRegionsActive(Region(free_iterator) + 1,
+                         (std::min)(Region(iterator + size) + 1, NUM_SYNCS))) {
         // Avoid waiting for the previous usages to be free
         return GetStagingBuffer(size, MemoryUsage::Upload);
     }
-
     const u64 current_tick = scheduler.CurrentTick();
     std::fill(sync_ticks.begin() + Region(used_iterator), sync_ticks.begin() + Region(iterator),
               current_tick);
     used_iterator = iterator;
+    free_iterator = (std::max)(free_iterator, iterator + size);
 
-    if (wraps) {
+    if (iterator + size >= stream_buffer_size) {
         std::fill(sync_ticks.begin() + Region(used_iterator), sync_ticks.begin() + NUM_SYNCS,
                   current_tick);
         used_iterator = 0;
         iterator = 0;
-        free_iterator = aligned_size;
-        const size_t head_last_byte = aligned_size == 0 ? 0 : aligned_size - 1;
-        const size_t head_end_region = (std::min)(Region(head_last_byte) + 1, NUM_SYNCS);
-        if (AreRegionsActive(0, head_end_region)) {
+        free_iterator = size;
+
+        if (AreRegionsActive(0, Region(size) + 1)) {
             // Avoid waiting for the previous usages to be free
             return GetStagingBuffer(size, MemoryUsage::Upload);
         }
     }
-
-    std::fill(sync_ticks.begin() + begin_region, sync_ticks.begin() + end_region, current_tick);
-
-    const size_t offset = wraps ? 0 : iterator;
-    iterator = new_iterator;
-
-    if (!wraps) {
-        free_iterator = (std::max)(free_iterator, offset + aligned_size);
-    }
-
+    const size_t offset = iterator;
+    iterator = Common::AlignUp(iterator + size, MAX_ALIGNMENT);
     return StagingBufferRef{
         .buffer = *stream_buffer,
         .offset = static_cast<VkDeviceSize>(offset),

src/video_core/renderer_vulkan/vk_staging_buffer_pool.h

@@ -1,6 +1,3 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
-// SPDX-License-Identifier: GPL-3.0-or-later
-
 // SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
 
@@ -105,7 +102,6 @@ private:
     MemoryAllocator& memory_allocator;
     Scheduler& scheduler;
 
-    VkDeviceSize stream_alignment;
     vk::Buffer stream_buffer;
     std::span<u8> stream_pointer;
     VkDeviceSize stream_buffer_size;

src/yuzu/configuration/configure_debug.cpp

@@ -83,8 +83,7 @@ void ConfigureDebug::SetConfiguration() {
 #ifdef YUZU_USE_QT_WEB_ENGINE
     ui->disable_web_applet->setChecked(UISettings::values.disable_web_applet.GetValue());
 #else
-    ui->disable_web_applet->setEnabled(false);
-    ui->disable_web_applet->setText(tr("Web applet not compiled"));
+    ui->disable_web_applet->setVisible(false);
 #endif
 }