[Vulkan] Phase 1: Descriptor Set caching

wip

src/video_core/renderer_vulkan/vk_descriptor_pool.cpp

@@ -96,9 +96,30 @@ VkDescriptorSet DescriptorAllocator::Commit() {
     const size_t index = CommitResource();
     return sets[index / SETS_GROW_RATE][index % SETS_GROW_RATE];
 }
-
+VkDescriptorSet DescriptorAllocator::CommitWithTracking(u64 current_frame, const void* descriptor_data) {
+     const size_t index = CommitResource();
+     const size_t group = index / SETS_GROW_RATE;
+     const size_t slot  = index % SETS_GROW_RATE;
+     set_states[group][slot].set               = sets[group][slot];
+     set_states[group][slot].last_update_frame = current_frame;
+     set_states[group][slot].last_data_ptr     = descriptor_data;
+     return set_states[group][slot].set;
+     }
+bool DescriptorAllocator::NeedsUpdate(VkDescriptorSet set, u64 current_frame, const void* descriptor_data) const {
+        for (const auto& group : set_states) {
+            for (const auto& st : group) {
+                if (st.set == set) {
+                    // Update if pointer changed or the set hasn't been updated this frame
+                    return st.last_data_ptr != descriptor_data || st.last_update_frame != current_frame;
+                }
+            }
+        }
+        return true;
+ }
 void DescriptorAllocator::Allocate(size_t begin, size_t end) {
-    sets.push_back(AllocateDescriptors(end - begin));
+    const size_t count = end - begin;
+    sets.push_back(AllocateDescriptors(count));
+    set_states.emplace_back(count); // create parallel state storage
 }
 
 vk::DescriptorSets DescriptorAllocator::AllocateDescriptors(size_t count) {

src/video_core/renderer_vulkan/vk_descriptor_pool.h

@@ -6,7 +6,7 @@
 #include <shared_mutex>
 #include <span>
 #include <vector>
-
+#include "common/common_types.h"
 #include "shader_recompiler/shader_info.h"
 #include "video_core/renderer_vulkan/vk_resource_pool.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
@@ -44,7 +44,15 @@ public:
     DescriptorAllocator(const DescriptorAllocator&) = delete;
 
     VkDescriptorSet Commit();
-
+    // commit + remember when/with what we last updated this set
+    struct DescriptorSetState {
+           VkDescriptorSet set = VK_NULL_HANDLE;
+           u64 last_update_frame = 0;
+           const void* last_data_ptr = nullptr; // fast pointer compare
+    };
+
+    VkDescriptorSet CommitWithTracking(u64 current_frame, const void* descriptor_data);
+    bool NeedsUpdate(VkDescriptorSet set, u64 current_frame, const void* descriptor_data) const;
 private:
     explicit DescriptorAllocator(const Device& device_, MasterSemaphore& master_semaphore_,
                                  DescriptorBank& bank_, VkDescriptorSetLayout layout_);
@@ -58,6 +66,7 @@ private:
     VkDescriptorSetLayout layout{};
 
     std::vector<vk::DescriptorSets> sets;
+    std::vector<std::vector<DescriptorSetState>> set_states;
 };
 
 class DescriptorPool {

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -515,7 +515,25 @@ void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling,
     const bool update_rescaling{scheduler.UpdateRescaling(is_rescaling)};
     const bool bind_pipeline{scheduler.UpdateGraphicsPipeline(this)};
     const void* const descriptor_data{guest_descriptor_queue.UpdateData()};
-    scheduler.Record([this, descriptor_data, bind_pipeline, rescaling_data = rescaling.Data(),
+
+    // allocate/bind descriptor set and only update if needed
+    VkDescriptorSet descriptor_set = VK_NULL_HANDLE;
+    bool needs_update = false;
+    if (descriptor_set_layout) {
+        if (uses_push_descriptor) {
+            // handled below via push descriptor
+        } else {
+            descriptor_set = descriptor_allocator.CommitWithTracking(current_frame_number, descriptor_data);
+            needs_update = descriptor_allocator.NeedsUpdate(descriptor_set, current_frame_number, descriptor_data);
+            if (needs_update) {
+                const vk::Device& dev{device.GetLogical()};
+                dev.UpdateDescriptorSet(descriptor_set, *descriptor_update_template, descriptor_data);
+                } else {
+                }
+            }
+        }
+
+    scheduler.Record([this, descriptor_data, descriptor_set, bind_pipeline, 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) {
@@ -541,12 +559,8 @@ void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling,
             return;
         }
         if (uses_push_descriptor) {
-            cmdbuf.PushDescriptorSetWithTemplateKHR(*descriptor_update_template, *pipeline_layout,
-                                                    0, descriptor_data);
+          cmdbuf.PushDescriptorSetWithTemplateKHR(*descriptor_update_template, *pipeline_layout, 0, descriptor_data);
         } else {
-            const VkDescriptorSet descriptor_set{descriptor_allocator.Commit()};
-            const vk::Device& dev{device.GetLogical()};
-            dev.UpdateDescriptorSet(descriptor_set, *descriptor_update_template, descriptor_data);
             cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_GRAPHICS, *pipeline_layout, 0,
                                       descriptor_set, nullptr);
         }

src/video_core/renderer_vulkan/vk_graphics_pipeline.h

@@ -116,7 +116,7 @@ public:
         maxwell3d = maxwell3d_;
         gpu_memory = gpu_memory_;
     }
-
+    void SetFrameNumber(u64 frame) { current_frame_number = frame; }
 private:
     template <typename Spec>
     bool ConfigureImpl(bool is_indexed);
@@ -160,6 +160,7 @@ private:
     std::mutex build_mutex;
     std::atomic_bool is_built{false};
     bool uses_push_descriptor{false};
+    u64 current_frame_number{0};
 };
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -234,6 +234,7 @@ void RasterizerVulkan::PrepareDraw(bool is_indexed, Func&& draw_func) {
     std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
     // update engine as channel may be different.
     pipeline->SetEngine(maxwell3d, gpu_memory);
+    pipeline->SetFrameNumber(current_frame_number);
     if (!pipeline->Configure(is_indexed))
         return;
 
@@ -771,6 +772,7 @@ void RasterizerVulkan::FlushCommands() {
 }
 
 void RasterizerVulkan::TickFrame() {
+    current_frame_number++;
     draw_counter = 0;
     guest_descriptor_queue.TickFrame();
     compute_pass_descriptor_queue.TickFrame();

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -160,7 +160,7 @@ private:
     void UpdateDynamicStates();
 
     void HandleTransformFeedback();
-
+    u64 current_frame_number{0};
     void UpdateViewportsState(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateScissorsState(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateDepthBias(Tegra::Engines::Maxwell3D::Regs& regs);