2nd step for the buffer address device implementation

src/video_core/CMakeLists.txt

@@ -158,6 +158,8 @@ add_library(video_core STATIC
     renderer_vulkan/vk_compute_pass.h
     renderer_vulkan/vk_compute_pipeline.cpp
     renderer_vulkan/vk_compute_pipeline.h
+    renderer_vulkan/vk_descriptor_buffer.cpp
+    renderer_vulkan/vk_descriptor_buffer.h
     renderer_vulkan/vk_descriptor_pool.cpp
     renderer_vulkan/vk_descriptor_pool.h
     renderer_vulkan/vk_fence_manager.cpp

src/video_core/renderer_vulkan/pipeline_helper.h

@@ -47,6 +47,93 @@ using Shader::Backend::SPIRV::NUM_TEXTURE_AND_IMAGE_SCALING_WORDS;
     return std::nullopt;
 }
 
+[[nodiscard]] inline VkDeviceSize DescriptorSizeForType(const Device& device,
+                                                        VkDescriptorType type) {
+    const auto& props = device.DescriptorBufferProperties();
+    const bool robust = device.IsRobustBufferAccessEnabled();
+    switch (type) {
+    case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
+        return robust ? props.robustUniformBufferDescriptorSize : props.uniformBufferDescriptorSize;
+    case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+        return robust ? props.robustStorageBufferDescriptorSize : props.storageBufferDescriptorSize;
+    case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
+        return robust ? props.robustUniformTexelBufferDescriptorSize
+                      : props.uniformTexelBufferDescriptorSize;
+    case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
+        return robust ? props.robustStorageTexelBufferDescriptorSize
+                      : props.storageTexelBufferDescriptorSize;
+    case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
+        return props.combinedImageSamplerDescriptorSize;
+    case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
+        return props.storageImageDescriptorSize;
+    default:
+        return 0;
+    }
+}
+
+struct DescriptorBufferBinding {
+    VkDescriptorType type;
+    u32 count;
+    VkDeviceSize offset;
+    VkDeviceSize stride;
+};
+
+struct DescriptorBufferLayout {
+    VkDeviceSize size{};
+    boost::container::small_vector<DescriptorBufferBinding, 32> bindings;
+
+    [[nodiscard]] bool Empty() const noexcept {
+        return bindings.empty();
+    }
+};
+
+inline void WriteDescriptorBuffer(const Device& device, const DescriptorBufferLayout& layout,
+                                  const DescriptorUpdateEntry* payload, u8* host) {
+    const vk::Device& dev = device.GetLogical();
+    for (const DescriptorBufferBinding& binding : layout.bindings) {
+        for (u32 index = 0; index < binding.count; ++index) {
+            const DescriptorUpdateEntry& entry = *(payload++);
+            const VkDescriptorAddressInfoEXT address_info{
+                .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_ADDRESS_INFO_EXT,
+                .pNext = nullptr,
+                .address = entry.address.address,
+                .range = entry.address.range,
+                .format = entry.address.format,
+            };
+            VkDescriptorGetInfoEXT get_info{
+                .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_GET_INFO_EXT,
+                .pNext = nullptr,
+                .type = binding.type,
+                .data{},
+            };
+            switch (binding.type) {
+            case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
+                get_info.data.pUniformBuffer = &address_info;
+                break;
+            case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+                get_info.data.pStorageBuffer = &address_info;
+                break;
+            case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
+                get_info.data.pUniformTexelBuffer = &address_info;
+                break;
+            case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
+                get_info.data.pStorageTexelBuffer = &address_info;
+                break;
+            case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
+                get_info.data.pCombinedImageSampler = &entry.image;
+                break;
+            case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
+                get_info.data.pStorageImage = &entry.image;
+                break;
+            default:
+                continue;
+            }
+            dev.GetDescriptorEXT(get_info, binding.stride,
+                                 host + binding.offset + index * binding.stride);
+        }
+    }
+}
+
 [[nodiscard]] inline u32 NumDescriptorEntries(const Shader::Info& info) {
     return Shader::NumDescriptors(info.constant_buffer_descriptors) +
            Shader::NumDescriptors(info.storage_buffers_descriptors) +
@@ -65,13 +152,44 @@ public:
                num_descriptors <= device->MaxPushDescriptors();
     }
 
+    bool CanUseDescriptorBuffer() const noexcept {
+        return device->IsExtDescriptorBufferSupported() && !bindings.empty() &&
+               !CanUsePushDescriptor() &&
+               device->DescriptorBufferProperties().combinedImageSamplerDescriptorSingleArray;
+    }
+
+    DescriptorBufferLayout MakeDescriptorBufferLayout(VkDescriptorSetLayout layout) const {
+        DescriptorBufferLayout result;
+        if (!layout) {
+            return result;
+        }
+        const vk::Device& dev = device->GetLogical();
+        result.size = dev.GetDescriptorSetLayoutSizeEXT(layout);
+        result.bindings.reserve(bindings.size());
+        for (const VkDescriptorSetLayoutBinding& binding : bindings) {
+            result.bindings.push_back(DescriptorBufferBinding{
+                .type = binding.descriptorType,
+                .count = binding.descriptorCount,
+                .offset = dev.GetDescriptorSetLayoutBindingOffsetEXT(layout, binding.binding),
+                .stride = DescriptorSizeForType(*device, binding.descriptorType),
+            });
+        }
+        return result;
+    }
+
     // TODO(crueter): utilize layout binding flags
-    vk::DescriptorSetLayout CreateDescriptorSetLayout(bool use_push_descriptor) const {
+    vk::DescriptorSetLayout CreateDescriptorSetLayout(bool use_push_descriptor,
+                                                      bool use_descriptor_buffer = false) const {
         if (bindings.empty()) {
             return nullptr;
         }
-        const VkDescriptorSetLayoutCreateFlags flags =
-            use_push_descriptor ? VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR : 0;
+        VkDescriptorSetLayoutCreateFlags flags = 0;
+        if (use_push_descriptor) {
+            flags |= VK_DESCRIPTOR_SET_LAYOUT_CREATE_PUSH_DESCRIPTOR_BIT_KHR;
+        }
+        if (use_descriptor_buffer) {
+            flags |= VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT;
+        }
         return device->GetLogical().CreateDescriptorSetLayout({
             .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
             .pNext = nullptr,

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -94,6 +94,9 @@ Buffer::Buffer(BufferCacheRuntime& runtime, VideoCommon::NullBufferParams null_p
     device = &runtime.device;
     buffer = runtime.CreateNullBuffer();
     is_null = true;
+    if (device->IsBufferDeviceAddressSupported()) {
+        device_address = device->GetLogical().GetBufferDeviceAddress(*buffer);
+    }
 }
 
 Buffer::Buffer(BufferCacheRuntime& runtime, DAddr cpu_addr_, u64 size_bytes_)
@@ -369,6 +372,10 @@ StagingBufferRef BufferCacheRuntime::DownloadStagingBuffer(size_t size, bool def
     return staging_pool.Request(size, MemoryUsage::Download, deferred);
 }
 
+VkFormat BufferCacheRuntime::TexelBufferFormat(VideoCore::Surface::PixelFormat format) const {
+    return MaxwellToVK::SurfaceFormat(device, FormatType::Buffer, false, format).format;
+}
+
 void BufferCacheRuntime::FreeDeferredStagingBuffer(StagingBufferRef& ref) {
     staging_pool.FreeDeferred(ref);
 }
@@ -717,6 +724,9 @@ vk::Buffer BufferCacheRuntime::CreateNullBuffer() {
     if (device.IsExtTransformFeedbackSupported()) {
         create_info.usage |= VK_BUFFER_USAGE_TRANSFORM_FEEDBACK_BUFFER_BIT_EXT;
     }
+    if (device.IsBufferDeviceAddressSupported()) {
+        create_info.usage |= VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT;
+    }
     vk::Buffer ret = memory_allocator.CreateBuffer(create_info, MemoryUsage::DeviceLocal);
     if (device.HasDebuggingToolAttached()) {
         ret.SetObjectNameEXT("Null buffer");

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -156,22 +156,25 @@ public:
                                           [[maybe_unused]] u32 binding_index,
                                           u32 size) {
         const StagingBufferRef ref = staging_pool.Request(size, MemoryUsage::Upload);
-        BindBuffer(ref.buffer, static_cast<u32>(ref.offset), size);
+        guest_descriptor_queue.AddBuffer(ref.buffer, ref.device_address,
+                                         static_cast<u32>(ref.offset), size);
         return ref.mapped_span;
     }
 
-    void BindUniformBuffer(VkBuffer buffer, u32 offset, u32 size) {
+    void BindUniformBuffer(const Buffer& buffer, u32 offset, u32 size) {
         BindBuffer(buffer, offset, size);
     }
 
-    void BindStorageBuffer(VkBuffer buffer, u32 offset, u32 size,
+    void BindStorageBuffer(const Buffer& buffer, u32 offset, u32 size,
                            [[maybe_unused]] bool is_written) {
         BindBuffer(buffer, offset, size);
     }
 
     void BindTextureBuffer(Buffer& buffer, u32 offset, u32 size,
                            VideoCore::Surface::PixelFormat format) {
-        guest_descriptor_queue.AddTexelBuffer(buffer.View(offset, size, format));
+        guest_descriptor_queue.AddTexelBuffer(buffer.View(offset, size, format),
+                                              buffer.DeviceAddress(), offset, size,
+                                              TexelBufferFormat(format));
     }
 
     bool ShouldLimitDynamicStorageBuffers() const {
@@ -183,14 +186,17 @@ public:
     }
 
 private:
-    void BindBuffer(VkBuffer buffer, u32 offset, u32 size) {
-        if (buffer == VK_NULL_HANDLE) {
-            guest_descriptor_queue.AddBuffer(buffer, 0, VK_WHOLE_SIZE);
+    void BindBuffer(const Buffer& buffer, u32 offset, u32 size) {
+        const VkBuffer handle = buffer.Handle();
+        if (handle == VK_NULL_HANDLE) {
+            guest_descriptor_queue.AddBuffer(handle, 0, 0, VK_WHOLE_SIZE);
         } else {
-            guest_descriptor_queue.AddBuffer(buffer, offset, size);
+            guest_descriptor_queue.AddBuffer(handle, buffer.DeviceAddress(), offset, size);
         }
     }
 
+    VkFormat TexelBufferFormat(VideoCore::Surface::PixelFormat format) const;
+
     void ReserveNullBuffer();
     vk::Buffer CreateNullBuffer();
 

src/video_core/renderer_vulkan/vk_compute_pipeline.cpp

@@ -34,12 +34,14 @@ using Tegra::Texture::TexturePair;
 ComputePipeline::ComputePipeline(const Device& device_, Scheduler& scheduler, vk::PipelineCache& pipeline_cache_,
                                  DescriptorPool& descriptor_pool,
                                  GuestDescriptorQueue& guest_descriptor_queue_,
+                                 DescriptorBufferRing& descriptor_buffer_ring_,
                                  Common::ThreadWorker* thread_worker,
                                  PipelineStatistics* pipeline_statistics,
                                  VideoCore::ShaderNotify* shader_notify, const Shader::Info& info_,
                                  vk::ShaderModule spv_module_, u64 shader_hash_)
     : device{device_},
-      pipeline_cache(pipeline_cache_), guest_descriptor_queue{guest_descriptor_queue_}, info{info_},
+      pipeline_cache(pipeline_cache_), guest_descriptor_queue{guest_descriptor_queue_},
+      descriptor_buffer_ring{descriptor_buffer_ring_}, info{info_},
       shader_hash{shader_hash_}, spv_module(std::move(spv_module_)) {
     if (shader_notify) {
         shader_notify->MarkShaderBuilding();
@@ -48,18 +50,35 @@ ComputePipeline::ComputePipeline(const Device& device_, Scheduler& scheduler, vk
                 uniform_buffer_sizes.begin());
     num_descriptor_entries = NumDescriptorEntries(info);
 
-    auto func{[this, &scheduler, &descriptor_pool, shader_notify, pipeline_statistics] {
-        DescriptorLayoutBuilder builder{device};
-        builder.Add(info, VK_SHADER_STAGE_COMPUTE_BIT);
+    DescriptorLayoutBuilder builder{device};
+    builder.Add(info, VK_SHADER_STAGE_COMPUTE_BIT);
 
-        uses_push_descriptor = builder.CanUsePushDescriptor();
-        descriptor_set_layout = builder.CreateDescriptorSetLayout(uses_push_descriptor);
-        pipeline_layout = builder.CreatePipelineLayout(*descriptor_set_layout);
+    uses_push_descriptor = builder.CanUsePushDescriptor();
+    uses_descriptor_buffer = builder.CanUseDescriptorBuffer() && descriptor_buffer_ring.IsValid();
+    descriptor_set_layout =
+        builder.CreateDescriptorSetLayout(uses_push_descriptor, uses_descriptor_buffer);
+    if (uses_descriptor_buffer) {
+        descriptor_buffer_layout = builder.MakeDescriptorBufferLayout(*descriptor_set_layout);
+        if (descriptor_buffer_layout.size > DescriptorBufferRing::MaxAllocationSize()) {
+            LOG_WARNING(Render_Vulkan,
+                        "Compute shader {:016X} needs {} descriptor bytes, falling back to sets",
+                        shader_hash, descriptor_buffer_layout.size);
+            uses_descriptor_buffer = false;
+            descriptor_buffer_layout = {};
+            descriptor_set_layout = builder.CreateDescriptorSetLayout(false);
+        }
+    }
+    pipeline_layout = builder.CreatePipelineLayout(*descriptor_set_layout);
+    if (!uses_descriptor_buffer) {
         descriptor_update_template =
             builder.CreateTemplate(*descriptor_set_layout, *pipeline_layout, uses_push_descriptor);
         if (!uses_push_descriptor) {
-            descriptor_allocator = descriptor_pool.Allocator(device, scheduler, *descriptor_set_layout, info);
+            descriptor_allocator =
+                descriptor_pool.Allocator(device, scheduler, *descriptor_set_layout, info);
         }
+    }
+
+    auto func{[this, shader_notify, pipeline_statistics] {
         const VkPipelineShaderStageRequiredSubgroupSizeCreateInfoEXT subgroup_size_ci{
             .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_REQUIRED_SUBGROUP_SIZE_CREATE_INFO_EXT,
             .pNext = nullptr,
@@ -69,6 +88,9 @@ ComputePipeline::ComputePipeline(const Device& device_, Scheduler& scheduler, vk
         if (device.IsKhrPipelineExecutablePropertiesEnabled() && Settings::values.renderer_debug.GetValue()) {
             flags |= VK_PIPELINE_CREATE_CAPTURE_STATISTICS_BIT_KHR;
         }
+        if (uses_descriptor_buffer) {
+            flags |= VK_PIPELINE_CREATE_DESCRIPTOR_BUFFER_BIT_EXT;
+        }
         VkPipelineCreationFeedback creation_feedback{};
         const VkPipelineCreationFeedbackCreateInfo feedback_ci{
             .sType = VK_STRUCTURE_TYPE_PIPELINE_CREATION_FEEDBACK_CREATE_INFO,
@@ -144,7 +166,7 @@ ComputePipeline::ComputePipeline(const Device& device_, Scheduler& scheduler, vk
 void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
                                 Tegra::MemoryManager& gpu_memory, Scheduler& scheduler,
                                 BufferCache& buffer_cache, TextureCache& texture_cache) {
-    guest_descriptor_queue.Acquire(scheduler, num_descriptor_entries);
+    guest_descriptor_queue.Acquire(scheduler, num_descriptor_entries, uses_descriptor_buffer);
 
     buffer_cache.SetComputeUniformBufferState(info.constant_buffer_mask, &uniform_buffer_sizes);
     buffer_cache.UnbindComputeStorageBuffers();
@@ -265,9 +287,22 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
         GPU::Logging::GPULogger::GetInstance().LogPipelineBind(true, "compute pipeline");
     }
 
-    const void* const descriptor_data{guest_descriptor_queue.UpdateData()};
+    const DescriptorUpdateEntry* const descriptor_data{guest_descriptor_queue.UpdateData()};
+    VkDeviceSize descriptor_buffer_offset{};
+    bool descriptor_buffer_ready{false};
+    if (uses_descriptor_buffer) {
+        const DescriptorBufferRing::Allocation alloc{
+            descriptor_buffer_ring.Allocate(scheduler, descriptor_buffer_layout.size)};
+        if (alloc.host) {
+            WriteDescriptorBuffer(device, descriptor_buffer_layout, descriptor_data, alloc.host);
+            descriptor_buffer_offset = alloc.offset;
+            descriptor_buffer_ready = true;
+        }
+    }
+
     const bool is_rescaling = !info.texture_descriptors.empty() || !info.image_descriptors.empty();
-    scheduler.Record([this, descriptor_data, is_rescaling,
+    scheduler.Record([this, descriptor_data, is_rescaling, descriptor_buffer_offset,
+                      descriptor_buffer_ready,
                       rescaling_data = rescaling.Data()](vk::CommandBuffer cmdbuf) {
         if (!pipeline) {
             return;
@@ -281,7 +316,17 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
                                  RESCALING_LAYOUT_WORDS_OFFSET, sizeof(rescaling_data),
                                  rescaling_data.data());
         }
-        if (uses_push_descriptor) {
+        if (uses_descriptor_buffer) {
+            if (!descriptor_buffer_ready) {
+                return;
+            }
+            const VkDescriptorBufferBindingInfoEXT binding_info{
+                descriptor_buffer_ring.BindingInfo()};
+            cmdbuf.BindDescriptorBuffersEXT(binding_info);
+            const u32 buffer_index{};
+            cmdbuf.SetDescriptorBufferOffsetsEXT(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline_layout,
+                                                 0, buffer_index, descriptor_buffer_offset);
+        } else if (uses_push_descriptor) {
             cmdbuf.PushDescriptorSetWithTemplateKHR(*descriptor_update_template, *pipeline_layout,
                                                     0, descriptor_data);
         } else {

src/video_core/renderer_vulkan/vk_compute_pipeline.h

@@ -13,7 +13,9 @@
 #include "common/common_types.h"
 #include "common/thread_worker.h"
 #include "shader_recompiler/shader_info.h"
+#include "video_core/renderer_vulkan/pipeline_helper.h"
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
+#include "video_core/renderer_vulkan/vk_descriptor_buffer.h"
 #include "video_core/renderer_vulkan/vk_descriptor_pool.h"
 #include "video_core/renderer_vulkan/vk_texture_cache.h"
 #include "video_core/renderer_vulkan/vk_update_descriptor.h"
@@ -34,6 +36,7 @@ public:
     explicit ComputePipeline(const Device& device, Scheduler& scheduler, vk::PipelineCache& pipeline_cache,
                              DescriptorPool& descriptor_pool,
                              GuestDescriptorQueue& guest_descriptor_queue,
+                             DescriptorBufferRing& descriptor_buffer_ring,
                              Common::ThreadWorker* thread_worker,
                              PipelineStatistics* pipeline_statistics,
                              VideoCore::ShaderNotify* shader_notify, const Shader::Info& info,
@@ -56,6 +59,7 @@ private:
     const Device& device;
     vk::PipelineCache& pipeline_cache;
     GuestDescriptorQueue& guest_descriptor_queue;
+    DescriptorBufferRing& descriptor_buffer_ring;
     Shader::Info info;
     u64 shader_hash{};
     u32 num_descriptor_entries{};
@@ -65,6 +69,8 @@ private:
     vk::ShaderModule spv_module;
     vk::DescriptorSetLayout descriptor_set_layout;
     bool uses_push_descriptor{false};
+    bool uses_descriptor_buffer{false};
+    DescriptorBufferLayout descriptor_buffer_layout;
     DescriptorAllocator descriptor_allocator;
     vk::PipelineLayout pipeline_layout;
     vk::DescriptorUpdateTemplate descriptor_update_template;

src/video_core/renderer_vulkan/vk_descriptor_buffer.cpp

@@ -0,0 +1,98 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include <algorithm>
+
+#include "common/alignment.h"
+#include "common/assert.h"
+#include "common/logging.h"
+#include "video_core/renderer_vulkan/vk_descriptor_buffer.h"
+#include "video_core/renderer_vulkan/vk_scheduler.h"
+#include "video_core/vulkan_common/vulkan_device.h"
+
+namespace Vulkan {
+
+DescriptorBufferRing::DescriptorBufferRing(const Device& device_,
+                                           MemoryAllocator& memory_allocator)
+    : device{device_} {
+    if (!device.IsExtDescriptorBufferSupported() || !device.IsBufferDeviceAddressSupported()) {
+        return;
+    }
+    alignment = std::max<VkDeviceSize>(
+        device.DescriptorBufferProperties().descriptorBufferOffsetAlignment, 1);
+    const VkDeviceSize total = FRAME_SIZE * FRAMES_IN_FLIGHT + alignment;
+    const VkBufferCreateInfo buffer_ci{
+        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
+        .pNext = nullptr,
+        .flags = 0,
+        .size = total,
+        .usage = VK_BUFFER_USAGE_RESOURCE_DESCRIPTOR_BUFFER_BIT_EXT |
+                 VK_BUFFER_USAGE_SAMPLER_DESCRIPTOR_BUFFER_BIT_EXT |
+                 VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT,
+        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
+        .queueFamilyIndexCount = 0,
+        .pQueueFamilyIndices = nullptr,
+    };
+    buffer = memory_allocator.CreateBuffer(buffer_ci, MemoryUsage::Upload);
+    if (!buffer.IsHostVisible()) {
+        LOG_WARNING(Render_Vulkan, "Descriptor buffer is not host visible, disabling");
+        buffer.reset();
+        return;
+    }
+    if (device.HasDebuggingToolAttached()) {
+        buffer.SetObjectNameEXT("Descriptor buffer");
+    }
+    const VkDeviceAddress raw_address = device.GetLogical().GetBufferDeviceAddress(*buffer);
+    base_address = Common::AlignUp(raw_address, alignment);
+    base_host = buffer.Mapped().data() + (base_address - raw_address);
+}
+
+DescriptorBufferRing::~DescriptorBufferRing() = default;
+
+void DescriptorBufferRing::TickFrame() {
+    if (++frame_index >= FRAMES_IN_FLIGHT) {
+        frame_index = 0;
+    }
+    frame_start = static_cast<VkDeviceSize>(frame_index) * FRAME_SIZE;
+    cursor = 0;
+    frame_reused = true;
+}
+
+DescriptorBufferRing::Allocation DescriptorBufferRing::Allocate(Scheduler& scheduler,
+                                                                VkDeviceSize size) {
+    ASSERT(buffer);
+    const VkDeviceSize needed = Common::AlignUp(size, alignment);
+    if (needed > FRAME_SIZE) {
+        LOG_ERROR(Render_Vulkan, "Descriptor set of {} bytes exceeds frame capacity {}", needed,
+                  FRAME_SIZE);
+        return Allocation{};
+    }
+    if (frame_reused) {
+        frame_reused = false;
+        scheduler.Wait(frame_ticks[frame_index]);
+    }
+    if (cursor + needed > FRAME_SIZE) {
+        LOG_WARNING(Render_Vulkan, "Descriptor buffer frame exhausted, stalling on the GPU");
+        scheduler.Finish();
+        cursor = 0;
+    }
+    const VkDeviceSize offset = frame_start + cursor;
+    cursor += needed;
+    frame_ticks[frame_index] = scheduler.CurrentTick();
+    return Allocation{
+        .host = base_host + offset,
+        .offset = offset,
+    };
+}
+
+VkDescriptorBufferBindingInfoEXT DescriptorBufferRing::BindingInfo() const noexcept {
+    return VkDescriptorBufferBindingInfoEXT{
+        .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_BUFFER_BINDING_INFO_EXT,
+        .pNext = nullptr,
+        .address = base_address,
+        .usage = VK_BUFFER_USAGE_RESOURCE_DESCRIPTOR_BUFFER_BIT_EXT |
+                 VK_BUFFER_USAGE_SAMPLER_DESCRIPTOR_BUFFER_BIT_EXT,
+    };
+}
+
+} // namespace Vulkan

src/video_core/renderer_vulkan/vk_descriptor_buffer.h

@@ -0,0 +1,57 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#pragma once
+
+#include <array>
+
+#include "common/common_types.h"
+#include "video_core/vulkan_common/vulkan_memory_allocator.h"
+#include "video_core/vulkan_common/vulkan_wrapper.h"
+
+namespace Vulkan {
+
+class Device;
+class Scheduler;
+
+class DescriptorBufferRing final {
+    static constexpr size_t FRAMES_IN_FLIGHT = 8;
+    static constexpr VkDeviceSize FRAME_SIZE = 512 * 1024;
+
+public:
+    explicit DescriptorBufferRing(const Device& device_, MemoryAllocator& memory_allocator);
+    ~DescriptorBufferRing();
+
+    struct Allocation {
+        u8* host{};
+        VkDeviceSize offset{};
+    };
+
+    [[nodiscard]] static constexpr VkDeviceSize MaxAllocationSize() noexcept {
+        return FRAME_SIZE;
+    }
+
+    void TickFrame();
+
+    [[nodiscard]] Allocation Allocate(Scheduler& scheduler, VkDeviceSize size);
+
+    [[nodiscard]] VkDescriptorBufferBindingInfoEXT BindingInfo() const noexcept;
+
+    [[nodiscard]] bool IsValid() const noexcept {
+        return static_cast<bool>(buffer);
+    }
+
+private:
+    const Device& device;
+    vk::Buffer buffer;
+    VkDeviceAddress base_address{};
+    u8* base_host{};
+    VkDeviceSize alignment{1};
+    size_t frame_index{};
+    VkDeviceSize frame_start{};
+    VkDeviceSize cursor{};
+    std::array<u64, FRAMES_IN_FLIGHT> frame_ticks{};
+    bool frame_reused{};
+};
+
+} // namespace Vulkan

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -340,10 +340,12 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
                              const Device& device_, Scheduler& scheduler_,
                              DescriptorPool& descriptor_pool_,
                              GuestDescriptorQueue& guest_descriptor_queue_,
+                             DescriptorBufferRing& descriptor_buffer_ring_,
                              RenderPassCache& render_pass_cache_, BufferCache& buffer_cache_,
                              TextureCache& texture_cache_, VideoCore::ShaderNotify& shader_notify_)
     : VideoCommon::ShaderCache{device_memory_}, device{device_}, scheduler{scheduler_},
       descriptor_pool{descriptor_pool_}, guest_descriptor_queue{guest_descriptor_queue_},
+      descriptor_buffer_ring{descriptor_buffer_ring_},
       render_pass_cache{render_pass_cache_}, buffer_cache{buffer_cache_},
       texture_cache{texture_cache_}, shader_notify{shader_notify_},
       use_asynchronous_shaders{Settings::values.use_asynchronous_shaders.GetValue()},
@@ -964,7 +966,8 @@ std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline(
     }
     Common::ThreadWorker* const thread_worker{build_in_parallel ? &workers : nullptr};
     return std::make_unique<ComputePipeline>(device, scheduler, vulkan_pipeline_cache, descriptor_pool,
-                                             guest_descriptor_queue, thread_worker, statistics,
+                                             guest_descriptor_queue, descriptor_buffer_ring,
+                                             thread_worker, statistics,
                                              &shader_notify, program.info, std::move(spv_module),
                                              key.unique_hash);
 

src/video_core/renderer_vulkan/vk_pipeline_cache.h

@@ -105,6 +105,7 @@ public:
     explicit PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_, const Device& device,
                            Scheduler& scheduler, DescriptorPool& descriptor_pool,
                            GuestDescriptorQueue& guest_descriptor_queue,
+                           DescriptorBufferRing& descriptor_buffer_ring,
                            RenderPassCache& render_pass_cache, BufferCache& buffer_cache,
                            TextureCache& texture_cache, VideoCore::ShaderNotify& shader_notify_);
     ~PipelineCache();
@@ -147,6 +148,7 @@ private:
     Scheduler& scheduler;
     DescriptorPool& descriptor_pool;
     GuestDescriptorQueue& guest_descriptor_queue;
+    DescriptorBufferRing& descriptor_buffer_ring;
     RenderPassCache& render_pass_cache;
     BufferCache& buffer_cache;
     TextureCache& texture_cache;

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -204,8 +204,10 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
     : gpu{gpu_}, device_memory{device_memory_}, device{device_},
       memory_allocator{memory_allocator_}, state_tracker{state_tracker_}, scheduler{scheduler_},
       staging_pool(device, memory_allocator, scheduler), descriptor_pool(device, scheduler),
-      guest_descriptor_queue(device, UpdateDescriptorQueue::GUEST_FRAME_PAYLOAD_SIZE),
+      guest_descriptor_queue(device, UpdateDescriptorQueue::GUEST_FRAME_PAYLOAD_SIZE,
+                             device.IsExtDescriptorBufferSupported()),
       compute_pass_descriptor_queue(device, UpdateDescriptorQueue::COMPUTE_FRAME_PAYLOAD_SIZE),
+      descriptor_buffer_ring(device, memory_allocator),
       blit_image(device, scheduler, state_tracker, descriptor_pool), render_pass_cache(device),
       texture_cache_runtime{
           device,     scheduler,         memory_allocator, staging_pool,
@@ -218,7 +220,8 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
                           staging_pool, compute_pass_descriptor_queue, descriptor_pool, texture_cache),
       query_cache(gpu, *this, device_memory, query_cache_runtime),
       pipeline_cache(device_memory, device, scheduler, descriptor_pool, guest_descriptor_queue,
-                     render_pass_cache, buffer_cache, texture_cache, gpu.ShaderNotify()),
+                     descriptor_buffer_ring, render_pass_cache, buffer_cache, texture_cache,
+                     gpu.ShaderNotify()),
       accelerate_dma(buffer_cache, texture_cache, scheduler),
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache, device, scheduler),
       wfi_event(device.GetLogical().CreateEvent()) {
@@ -910,6 +913,7 @@ void RasterizerVulkan::TickFrame() {
     vk::TickDeletionQueue(master_semaphore.KnownGpuTick());
     guest_descriptor_queue.TickFrame();
     compute_pass_descriptor_queue.TickFrame();
+    descriptor_buffer_ring.TickFrame();
     fence_manager.TickFrame();
     staging_pool.TickFrame();
     {

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -17,6 +17,7 @@
 #include "video_core/rasterizer_interface.h"
 #include "video_core/renderer_vulkan/blit_image.h"
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
+#include "video_core/renderer_vulkan/vk_descriptor_buffer.h"
 #include "video_core/renderer_vulkan/vk_descriptor_pool.h"
 #include "video_core/renderer_vulkan/vk_fence_manager.h"
 #include "video_core/renderer_vulkan/vk_pipeline_cache.h"
@@ -206,6 +207,7 @@ private:
     DescriptorPool descriptor_pool;
     GuestDescriptorQueue guest_descriptor_queue;
     ComputePassDescriptorQueue compute_pass_descriptor_queue;
+    DescriptorBufferRing descriptor_buffer_ring;
     BlitImageHelper blit_image;
     RenderPassCache render_pass_cache;
 

src/video_core/renderer_vulkan/vk_staging_buffer_pool.cpp

@@ -84,10 +84,16 @@ StagingBufferPool::StagingBufferPool(const Device& device_, MemoryAllocator& mem
     if (device.IsExtTransformFeedbackSupported()) {
         stream_ci.usage |= VK_BUFFER_USAGE_TRANSFORM_FEEDBACK_BUFFER_BIT_EXT;
     }
+    if (device.IsBufferDeviceAddressSupported()) {
+        stream_ci.usage |= VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT;
+    }
     stream_buffer = memory_allocator.CreateBuffer(stream_ci, MemoryUsage::Stream);
     if (device.HasDebuggingToolAttached()) {
         stream_buffer.SetObjectNameEXT("Stream Buffer");
     }
+    if (device.IsBufferDeviceAddressSupported()) {
+        stream_buffer_address = device.GetLogical().GetBufferDeviceAddress(*stream_buffer);
+    }
     stream_pointer = stream_buffer.Mapped();
     ASSERT_MSG(!stream_pointer.empty(), "Stream buffer must be host visible!");
 }
@@ -149,6 +155,7 @@ StagingBufferRef StagingBufferPool::GetStreamBuffer(size_t size) {
     iterator = Common::AlignUp(iterator + size, MAX_ALIGNMENT);
     return StagingBufferRef{
         .buffer = *stream_buffer,
+        .device_address = stream_buffer_address,
         .offset = static_cast<VkDeviceSize>(offset),
         .mapped_span = stream_pointer.subspan(offset, size),
         .usage{},
@@ -212,14 +219,22 @@ StagingBufferRef StagingBufferPool::CreateStagingBuffer(size_t size, MemoryUsage
     if (device.IsExtTransformFeedbackSupported()) {
         buffer_ci.usage |= VK_BUFFER_USAGE_TRANSFORM_FEEDBACK_BUFFER_BIT_EXT;
     }
+    if (device.IsBufferDeviceAddressSupported()) {
+        buffer_ci.usage |= VK_BUFFER_USAGE_SHADER_DEVICE_ADDRESS_BIT;
+    }
     vk::Buffer buffer = memory_allocator.CreateBuffer(buffer_ci, usage);
     if (device.HasDebuggingToolAttached()) {
         ++buffer_index;
         buffer.SetObjectNameEXT(fmt::format("Staging Buffer {}", buffer_index).c_str());
     }
     const std::span<u8> mapped_span = buffer.Mapped();
+    const VkDeviceAddress buffer_address =
+        device.IsBufferDeviceAddressSupported()
+            ? device.GetLogical().GetBufferDeviceAddress(*buffer)
+            : VkDeviceAddress{};
     StagingBuffer& entry = GetCache(usage)[log2_size].entries.emplace_back(StagingBuffer{
         .buffer = std::move(buffer),
+        .device_address = buffer_address,
         .mapped_span = mapped_span,
         .usage = usage,
         .log2_level = log2_size,

src/video_core/renderer_vulkan/vk_staging_buffer_pool.h

@@ -21,6 +21,7 @@ class Scheduler;
 
 struct StagingBufferRef {
     VkBuffer buffer;
+    VkDeviceAddress device_address;
     VkDeviceSize offset;
     std::span<u8> mapped_span;
     MemoryUsage usage;
@@ -55,6 +56,7 @@ private:
 
     struct StagingBuffer {
         vk::Buffer buffer;
+        VkDeviceAddress device_address;
         std::span<u8> mapped_span;
         MemoryUsage usage;
         u32 log2_level;
@@ -65,6 +67,7 @@ private:
         StagingBufferRef Ref() const noexcept {
             return {
                 .buffer = *buffer,
+                .device_address = device_address,
                 .offset = 0,
                 .mapped_span = mapped_span,
                 .usage = usage,
@@ -108,6 +111,7 @@ private:
     Scheduler& scheduler;
 
     vk::Buffer stream_buffer;
+    VkDeviceAddress stream_buffer_address{};
     std::span<u8> stream_pointer;
     VkDeviceSize stream_buffer_size;
     VkDeviceSize region_size;

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -55,7 +55,7 @@ namespace {
 constexpr bool ENABLE_MSAA_RESOLVE_CONSUME = true;
 constexpr bool ENABLE_MSAA_COLOR_DISCARD = true;
 constexpr bool ENABLE_MSAA_DEPTH_DISCARD = true;
-constexpr bool ENABLE_MSAA_DEPTH_RESOLVE = true;
+constexpr bool ENABLE_MSAA_DEPTH_RESOLVE = false;
 
 constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
     if (color == std::array<float, 4>{0, 0, 0, 0}) {

src/video_core/renderer_vulkan/vk_update_descriptor.cpp

@@ -16,8 +16,10 @@
 
 namespace Vulkan {
 
-UpdateDescriptorQueue::UpdateDescriptorQueue(const Device& device_, size_t frame_payload_size_)
+UpdateDescriptorQueue::UpdateDescriptorQueue(const Device& device_, size_t frame_payload_size_,
+                                             bool supports_descriptor_buffer_)
     : device{device_}, frame_payload_size{frame_payload_size_},
+      supports_descriptor_buffer{supports_descriptor_buffer_},
       payload(frame_payload_size_ * FRAMES_IN_FLIGHT)
 {
     payload_start = payload.data();
@@ -34,7 +36,9 @@ void UpdateDescriptorQueue::TickFrame() {
     payload_cursor = payload_start;
 }
 
-void UpdateDescriptorQueue::Acquire(Scheduler& scheduler, size_t required_entries) {
+void UpdateDescriptorQueue::Acquire(Scheduler& scheduler, size_t required_entries,
+                                    bool use_descriptor_buffer_) {
+    use_descriptor_buffer = supports_descriptor_buffer && use_descriptor_buffer_;
     static constexpr size_t DEFAULT_REQUIRED_ENTRIES = 0x400;
     const size_t reserve = required_entries > 0 ? required_entries : DEFAULT_REQUIRED_ENTRIES;
     ASSERT_MSG(reserve < frame_payload_size, "Descriptor reservation {} >= frame capacity {}",

src/video_core/renderer_vulkan/vk_update_descriptor.h

@@ -15,15 +15,23 @@ namespace Vulkan {
 class Device;
 class Scheduler;
 
+struct DescriptorAddress {
+    VkDeviceAddress address;
+    VkDeviceSize range;
+    VkFormat format;
+};
+
 union DescriptorUpdateEntry {
     DescriptorUpdateEntry() = default;
     DescriptorUpdateEntry(VkDescriptorImageInfo image_) : image{image_} {}
     DescriptorUpdateEntry(VkDescriptorBufferInfo buffer_) : buffer{buffer_} {}
     DescriptorUpdateEntry(VkBufferView texel_buffer_) : texel_buffer{texel_buffer_} {}
+    DescriptorUpdateEntry(DescriptorAddress address_) : address{address_} {}
     std::monostate empty{};
     VkDescriptorImageInfo image;
     VkDescriptorBufferInfo buffer;
     VkBufferView texel_buffer;
+    DescriptorAddress address;
 };
 
 class UpdateDescriptorQueue final {
@@ -35,11 +43,17 @@ public:
     static constexpr size_t GUEST_FRAME_PAYLOAD_SIZE = 0x80000;
     static constexpr size_t COMPUTE_FRAME_PAYLOAD_SIZE = 0x20000;
 
-    explicit UpdateDescriptorQueue(const Device& device_, size_t frame_payload_size_);
+    explicit UpdateDescriptorQueue(const Device& device_, size_t frame_payload_size_,
+                                   bool supports_descriptor_buffer_ = false);
     ~UpdateDescriptorQueue();
 
+    [[nodiscard]] bool UsesDescriptorBuffer() const noexcept {
+        return use_descriptor_buffer;
+    }
+
     void TickFrame();
-    void Acquire(Scheduler& scheduler, size_t required_entries = 0);
+    void Acquire(Scheduler& scheduler, size_t required_entries = 0,
+                 bool use_descriptor_buffer_ = false);
 
     const DescriptorUpdateEntry* UpdateData() const noexcept {
         return upload_start;
@@ -69,13 +83,41 @@ public:
         };
     }
 
+    void AddBuffer(VkBuffer buffer, VkDeviceAddress base_address, VkDeviceSize offset,
+                   VkDeviceSize size) {
+        if (!use_descriptor_buffer) {
+            AddBuffer(buffer, offset, size);
+            return;
+        }
+        *(payload_cursor++) = DescriptorAddress{
+            .address = base_address == 0 ? 0 : base_address + offset,
+            .range = base_address == 0 ? VK_WHOLE_SIZE : size,
+            .format = VK_FORMAT_UNDEFINED,
+        };
+    }
+
     void AddTexelBuffer(VkBufferView texel_buffer) {
         *(payload_cursor++) = texel_buffer;
     }
 
+    void AddTexelBuffer(VkBufferView texel_buffer, VkDeviceAddress base_address,
+                        VkDeviceSize offset, VkDeviceSize size, VkFormat format) {
+        if (!use_descriptor_buffer) {
+            AddTexelBuffer(texel_buffer);
+            return;
+        }
+        *(payload_cursor++) = DescriptorAddress{
+            .address = base_address == 0 ? 0 : base_address + offset,
+            .range = base_address == 0 ? VK_WHOLE_SIZE : size,
+            .format = format,
+        };
+    }
+
 private:
     const Device& device;
     const size_t frame_payload_size;
+    const bool supports_descriptor_buffer;
+    bool use_descriptor_buffer{false};
     size_t frame_index{0};
     DescriptorUpdateEntry* payload_cursor = nullptr;
     DescriptorUpdateEntry* payload_start = nullptr;

src/video_core/vulkan_common/vulkan_device.h

@@ -489,6 +489,10 @@ FN_MAX_LIMIT_LIST
         return extensions.pipeline_creation_feedback;
     }
 
+    bool IsRobustBufferAccessEnabled() const {
+        return features.features.robustBufferAccess == VK_TRUE;
+    }
+
     bool IsExtDescriptorBufferSupported() const {
         return extensions.descriptor_buffer;
     }