2nd step on the bindless implementation

src/video_core/renderer_vulkan/pipeline_helper.h

@@ -177,7 +177,6 @@ public:
         return result;
     }
 
-    // TODO(crueter): utilize layout binding flags
     vk::DescriptorSetLayout CreateDescriptorSetLayout(bool use_push_descriptor,
                                                       bool use_descriptor_buffer = false) const {
         if (bindings.empty()) {
@@ -190,9 +189,22 @@ public:
         if (use_descriptor_buffer) {
             flags |= VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT;
         }
+        boost::container::small_vector<VkDescriptorBindingFlags, 32> binding_flags;
+        VkDescriptorSetLayoutBindingFlagsCreateInfo binding_flags_ci{};
+        const void* pnext = nullptr;
+        if (!use_push_descriptor && device->IsDescriptorBindingPartiallyBoundSupported()) {
+            binding_flags.assign(bindings.size(), VK_DESCRIPTOR_BINDING_PARTIALLY_BOUND_BIT);
+            binding_flags_ci = {
+                .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO,
+                .pNext = nullptr,
+                .bindingCount = static_cast<u32>(binding_flags.size()),
+                .pBindingFlags = binding_flags.data(),
+            };
+            pnext = &binding_flags_ci;
+        }
         return device->GetLogical().CreateDescriptorSetLayout({
             .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
-            .pNext = nullptr,
+            .pNext = pnext,
             .flags = flags,
             .bindingCount = static_cast<u32>(bindings.size()),
             .pBindings = bindings.data(),

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -251,13 +251,15 @@ GraphicsPipeline::GraphicsPipeline(
     Scheduler& scheduler_, BufferCache& buffer_cache_, TextureCache& texture_cache_,
     vk::PipelineCache& pipeline_cache_, VideoCore::ShaderNotify* shader_notify,
     const Device& device_, DescriptorPool& descriptor_pool,
-    GuestDescriptorQueue& guest_descriptor_queue_, Common::ThreadWorker* worker_thread,
+    GuestDescriptorQueue& guest_descriptor_queue_, DescriptorBufferRing& descriptor_buffer_ring_,
+    Common::ThreadWorker* worker_thread,
     PipelineStatistics* pipeline_statistics, RenderPassCache& render_pass_cache,
     const GraphicsPipelineCacheKey& key_, std::array<vk::ShaderModule, NUM_STAGES> stages,
     const std::array<const Shader::Info*, NUM_STAGES>& infos)
     : key{key_}, device{device_}, texture_cache{texture_cache_}, buffer_cache{buffer_cache_},
       pipeline_cache(pipeline_cache_), scheduler{scheduler_},
-      guest_descriptor_queue{guest_descriptor_queue_}, spv_modules{std::move(stages)} {
+      guest_descriptor_queue{guest_descriptor_queue_},
+      descriptor_buffer_ring{descriptor_buffer_ring_}, spv_modules{std::move(stages)} {
     if (shader_notify) {
         shader_notify->MarkShaderBuilding();
     }
@@ -277,20 +279,36 @@ GraphicsPipeline::GraphicsPipeline(
         num_descriptor_entries += NumDescriptorEntries(*info);
     }
     fragment_has_color0_output = stage_infos[NUM_STAGES - 1].stores_frag_color[0];
-    auto func{[this, shader_notify, &render_pass_cache, &descriptor_pool, pipeline_statistics] {
-        DescriptorLayoutBuilder builder{MakeBuilder(device, stage_infos)};
-        uses_push_descriptor = builder.CanUsePushDescriptor();
-        descriptor_set_layout = builder.CreateDescriptorSetLayout(uses_push_descriptor);
 
-        if (!uses_push_descriptor) {
-            descriptor_allocator = descriptor_pool.Allocator(device, scheduler, *descriptor_set_layout, stage_infos);
+    DescriptorLayoutBuilder builder{MakeBuilder(device, stage_infos)};
+    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,
+                        "Graphics pipeline {:016X} needs {} descriptor bytes, falling back to sets",
+                        key.Hash(), descriptor_buffer_layout.size);
+            uses_descriptor_buffer = false;
+            descriptor_buffer_layout = {};
+            descriptor_set_layout = builder.CreateDescriptorSetLayout(false);
         }
+    }
 
-        const VkDescriptorSetLayout set_layout{*descriptor_set_layout};
-        pipeline_layout = builder.CreatePipelineLayout(set_layout);
+    const VkDescriptorSetLayout set_layout{*descriptor_set_layout};
+    pipeline_layout = builder.CreatePipelineLayout(set_layout);
+    if (!uses_descriptor_buffer) {
         descriptor_update_template =
             builder.CreateTemplate(set_layout, *pipeline_layout, uses_push_descriptor);
+        if (!uses_push_descriptor) {
+            descriptor_allocator =
+                descriptor_pool.Allocator(device, scheduler, set_layout, stage_infos);
+        }
+    }
 
+    auto func{[this, shader_notify, &render_pass_cache, pipeline_statistics] {
         VkRenderPass render_pass{};
         if (!device.IsKhrDynamicRenderingSupported()) {
             render_pass = render_pass_cache.Get(MakeRenderPassKey(key.state, device));
@@ -500,7 +518,7 @@ bool GraphicsPipeline::ConfigureImpl(bool is_indexed) {
     buffer_cache.UpdateGraphicsBuffers(is_indexed);
     buffer_cache.BindHostGeometryBuffers(is_indexed);
 
-    guest_descriptor_queue.Acquire(scheduler, num_descriptor_entries);
+    guest_descriptor_queue.Acquire(scheduler, num_descriptor_entries, uses_descriptor_buffer);
 
     RescalingPushConstant rescaling;
     RenderAreaPushConstant render_area;
@@ -570,7 +588,7 @@ void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling,
 
     const void* const descriptor_data{guest_descriptor_queue.UpdateData()};
     bool update_descriptors = true;
-    if (descriptor_set_layout && !uses_push_descriptor) {
+    if (descriptor_set_layout && !uses_push_descriptor && !uses_descriptor_buffer) {
         const auto* const entries = static_cast<const DescriptorUpdateEntry*>(descriptor_data);
         update_descriptors =
             bind_pipeline || last_descriptor_payload.size() != num_descriptor_entries ||
@@ -580,7 +598,21 @@ void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling,
             last_descriptor_payload.assign(entries, entries + num_descriptor_entries);
         }
     }
+    VkDeviceSize descriptor_buffer_offset{};
+    bool descriptor_buffer_ready{false};
+    if (descriptor_set_layout && uses_descriptor_buffer) {
+        const DescriptorBufferRing::Allocation alloc{
+            descriptor_buffer_ring.Allocate(scheduler, descriptor_buffer_layout.size)};
+        if (alloc.host) {
+            WriteDescriptorBuffer(device, descriptor_buffer_layout,
+                                  static_cast<const DescriptorUpdateEntry*>(descriptor_data),
+                                  alloc.host);
+            descriptor_buffer_offset = alloc.offset;
+            descriptor_buffer_ready = true;
+        }
+    }
     scheduler.Record([this, descriptor_data, bind_pipeline, update_descriptors,
+                      descriptor_buffer_offset, descriptor_buffer_ready,
                       rescaling_data = rescaling.Data(), is_rescaling, update_rescaling,
                       uses_render_area = render_area.uses_render_area,
                       render_area_data = render_area.words](vk::CommandBuffer cmdbuf) {
@@ -608,7 +640,17 @@ void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling,
         if (!descriptor_set_layout) {
             return;
         }
-        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_GRAPHICS, *pipeline_layout,
+                                                 0, buffer_index, descriptor_buffer_offset);
+        } else if (uses_push_descriptor) {
             cmdbuf.PushDescriptorSetWithTemplateKHR(*descriptor_update_template, *pipeline_layout,
                                                     0, descriptor_data);
         } else if (update_descriptors) {
@@ -1017,6 +1059,9 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) {
     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;
+    }
 
     const RenderPassKey renderpass_key{MakeRenderPassKey(key.state, device)};
     std::array<VkFormat, Maxwell::NumRenderTargets> color_attachment_formats{};

src/video_core/renderer_vulkan/vk_graphics_pipeline.h

@@ -18,7 +18,9 @@
 #include "shader_recompiler/shader_info.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/renderer_vulkan/fixed_pipeline_state.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"
@@ -78,7 +80,8 @@ public:
         Scheduler& scheduler, BufferCache& buffer_cache, TextureCache& texture_cache,
         vk::PipelineCache& pipeline_cache, VideoCore::ShaderNotify* shader_notify,
         const Device& device, DescriptorPool& descriptor_pool,
-        GuestDescriptorQueue& guest_descriptor_queue, Common::ThreadWorker* worker_thread,
+        GuestDescriptorQueue& guest_descriptor_queue,
+        DescriptorBufferRing& descriptor_buffer_ring, Common::ThreadWorker* worker_thread,
         PipelineStatistics* pipeline_statistics, RenderPassCache& render_pass_cache,
         const GraphicsPipelineCacheKey& key, std::array<vk::ShaderModule, NUM_STAGES> stages,
         const std::array<const Shader::Info*, NUM_STAGES>& infos);
@@ -150,6 +153,7 @@ private:
     vk::PipelineCache& pipeline_cache;
     Scheduler& scheduler;
     GuestDescriptorQueue& guest_descriptor_queue;
+    DescriptorBufferRing& descriptor_buffer_ring;
 
     bool (*configure_func)(GraphicsPipeline*, bool){};
 
@@ -172,12 +176,14 @@ private:
     vk::DescriptorUpdateTemplate descriptor_update_template;
     vk::Pipeline pipeline;
 
+    DescriptorBufferLayout descriptor_buffer_layout;
     std::vector<DescriptorUpdateEntry> last_descriptor_payload;
 
     std::condition_variable build_condvar;
     std::mutex build_mutex;
     std::atomic_bool is_built{false};
     bool uses_push_descriptor{false};
+    bool uses_descriptor_buffer{false};
 };
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -873,8 +873,8 @@ std::unique_ptr<GraphicsPipeline> PipelineCache::CreateGraphicsPipeline(
     Common::ThreadWorker* const thread_worker{build_in_parallel ? &workers : nullptr};
     return std::make_unique<GraphicsPipeline>(
         scheduler, buffer_cache, texture_cache, vulkan_pipeline_cache, &shader_notify, device,
-        descriptor_pool, guest_descriptor_queue, thread_worker, statistics, render_pass_cache, key,
-        std::move(modules), infos);
+        descriptor_pool, guest_descriptor_queue, descriptor_buffer_ring, thread_worker, statistics,
+        render_pass_cache, key, std::move(modules), infos);
 
 } catch (const Shader::Exception& exception) {
     auto hash = key.Hash();

src/video_core/vulkan_common/vulkan_device.h

@@ -905,6 +905,10 @@ FN_MAX_LIMIT_LIST
         return extensions.external_memory_host;
     }
 
+    bool IsDescriptorBindingPartiallyBoundSupported() const {
+        return features.descriptor_indexing.descriptorBindingPartiallyBound;
+    }
+
     u64 GetMinImportedHostPointerAlignment() const {
         return properties.external_memory_host.minImportedHostPointerAlignment;
     }