[vk] Fast UBO: fix tracking, resize heuristics, add debug guard (#2695)

Co-authored-by: Ribbit <ribbit@placeholder.com>
Reviewed-on: #2695
Reviewed-by: CamilleLaVey <camillelavey99@gmail.com>
Co-authored-by: Ribbit <ribbit@eden-emu.dev>
Co-committed-by: Ribbit <ribbit@eden-emu.dev>

src/core/device_memory_manager.h

@@ -1,3 +1,6 @@
+// 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
 
@@ -109,6 +112,9 @@ 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

@@ -1,3 +1,6 @@
+// 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
 
@@ -467,6 +470,29 @@ 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

@@ -386,11 +386,10 @@ void BufferCache<P>::BindHostComputeBuffers() {
 template <class P>
 void BufferCache<P>::SetUniformBuffersState(const std::array<u32, NUM_STAGES>& mask,
                                             const UniformBufferSizes* sizes) {
-    if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
+    const bool mask_changed = channel_state->enabled_uniform_buffer_masks != mask;
-        if (channel_state->enabled_uniform_buffer_masks != mask) {
+    if (mask_changed) {
-            if constexpr (IS_OPENGL) {
+        channel_state->fast_bound_uniform_buffers.fill(0);
-                channel_state->fast_bound_uniform_buffers.fill(0);
+        if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
-            }
             channel_state->dirty_uniform_buffers.fill(~u32{0});
             channel_state->uniform_buffer_binding_sizes.fill({});
         }
@@ -806,7 +805,7 @@ 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);
                 }
@@ -815,13 +814,22 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
                 return;
             }
         }
-        if constexpr (IS_OPENGL) {
+        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;
-            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
@@ -830,7 +838,8 @@ 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
-    needs_bind |= HasFastUniformBufferBound(stage, binding_index);
+    const bool was_fast_bound = HasFastUniformBufferBound(stage, binding_index);
+    needs_bind |= was_fast_bound;
     if constexpr (HAS_PERSISTENT_UNIFORM_BUFFER_BINDINGS) {
         needs_bind |= channel_state->uniform_buffer_binding_sizes[stage][binding_index] != size;
     }
@@ -839,9 +848,6 @@ 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;
@@ -855,6 +861,7 @@ 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>
@@ -1789,12 +1796,7 @@ 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 {
-    if constexpr (IS_OPENGL) {
+    return ((channel_state->fast_bound_uniform_buffers[stage] >> binding_index) & 1u) != 0;
-        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

@@ -53,6 +53,7 @@ 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>;
@@ -137,8 +138,8 @@ public:
     u32 written_compute_texture_buffers = 0;
     u32 image_compute_texture_buffers = 0;
 
-    std::array<u32, 16> uniform_cache_hits{};
+    std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS> uniform_cache_hits{};
-    std::array<u32, 16> uniform_cache_shots{};
+    std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS> uniform_cache_shots{};
 
     u32 uniform_buffer_skip_cache_size = DEFAULT_SKIP_CACHE_SIZE;
 

src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp

@@ -25,12 +25,12 @@ namespace {
 
 using namespace Common::Literals;
 
-// Maximum potential alignment of a Vulkan buffer
+// Minimum alignment we want to enforce for the streaming ring
-constexpr VkDeviceSize MAX_ALIGNMENT = 256;
+constexpr VkDeviceSize MIN_STREAM_ALIGNMENT = 256;
 // Stream buffer size in bytes
 constexpr VkDeviceSize MAX_STREAM_BUFFER_SIZE = 128_MiB;
 
-size_t GetStreamBufferSize(const Device& device) {
+size_t GetStreamBufferSize(const Device& device, VkDeviceSize alignment) {
     VkDeviceSize size{0};
     if (device.HasDebuggingToolAttached()) {
         bool found_heap = false;
@@ -53,8 +53,9 @@ size_t GetStreamBufferSize(const Device& device) {
 
     // 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, MAX_ALIGNMENT), MAX_STREAM_BUFFER_SIZE);
+    const VkDeviceSize aligned =
-    const VkDeviceSize min_size = MAX_ALIGNMENT * StagingBufferPool::NUM_SYNCS;
+        (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));
 }
 } // Anonymous namespace
@@ -62,8 +63,10 @@ size_t GetStreamBufferSize(const Device& device) {
 StagingBufferPool::StagingBufferPool(const Device& device_, MemoryAllocator& memory_allocator_,
                                      Scheduler& scheduler_)
     : device{device_}, memory_allocator{memory_allocator_}, scheduler{scheduler_},
-      stream_buffer_size{GetStreamBufferSize(device)}, region_size{stream_buffer_size /
+      stream_alignment{std::max<VkDeviceSize>(device_.GetUniformBufferAlignment(),
-                                                                   StagingBufferPool::NUM_SYNCS} {
+                                              MIN_STREAM_ALIGNMENT)},
+      stream_buffer_size{GetStreamBufferSize(device_, stream_alignment)},
+      region_size{stream_buffer_size / StagingBufferPool::NUM_SYNCS} {
     VkBufferCreateInfo stream_ci = {
         .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
         .pNext = nullptr,
@@ -116,10 +119,11 @@ void StagingBufferPool::TickFrame() {
 }
 
 StagingBufferRef StagingBufferPool::GetStreamBuffer(size_t size) {
-    const size_t aligned_size = Common::AlignUp(size, MAX_ALIGNMENT);
+    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, MAX_ALIGNMENT);
+        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);
@@ -145,7 +149,7 @@ StagingBufferRef StagingBufferPool::GetStreamBuffer(size_t size) {
                   current_tick);
         used_iterator = 0;
         iterator = 0;
-        free_iterator = size;
+        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)) {
@@ -160,7 +164,7 @@ StagingBufferRef StagingBufferPool::GetStreamBuffer(size_t size) {
     iterator = new_iterator;
 
     if (!wraps) {
-        free_iterator = (std::max)(free_iterator, offset + size);
+        free_iterator = (std::max)(free_iterator, offset + aligned_size);
     }
 
     return StagingBufferRef{

src/video_core/renderer_vulkan/vk_staging_buffer_pool.h

@@ -1,3 +1,6 @@
+// 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
 
@@ -102,6 +105,7 @@ private:
     MemoryAllocator& memory_allocator;
     Scheduler& scheduler;
 
+    VkDeviceSize stream_alignment;
     vk::Buffer stream_buffer;
     std::span<u8> stream_pointer;
     VkDeviceSize stream_buffer_size;