renderer_vulkan: Rewrite descriptor management

* Switch to batched vkUpdateDescriptorSets from cached descriptor sets with templates

Author: raphaelthegreat

externals/vulkan-headers

@@ -160,8 +160,8 @@ if (ENABLE_VULKAN)
         renderer_vulkan/vk_blit_helper.h
         renderer_vulkan/vk_common.cpp
         renderer_vulkan/vk_common.h
-        renderer_vulkan/vk_descriptor_pool.cpp
-        renderer_vulkan/vk_descriptor_pool.h
+        renderer_vulkan/vk_descriptor_update_queue.cpp
+        renderer_vulkan/vk_descriptor_update_queue.h
         renderer_vulkan/vk_graphics_pipeline.cpp
         renderer_vulkan/vk_graphics_pipeline.h
         renderer_vulkan/vk_master_semaphore.cpp

src/video_core/CMakeLists.txt

@@ -385,7 +385,7 @@ std::vector<FileUtil::FSTEntry> CustomTexManager::GetTextures(u64 title_id) {
 }
 
 void CustomTexManager::CreateWorkers() {
-    const std::size_t num_workers = std::max(std::thread::hardware_concurrency(), 2U) - 1;
+    const std::size_t num_workers = std::max(std::thread::hardware_concurrency(), 2U) >> 1;
     workers = std::make_unique<Common::ThreadWorker>(num_workers, "Custom textures");
 }
 

src/video_core/custom_textures/custom_tex_manager.cpp

@@ -176,15 +176,15 @@ struct TexturingRegs {
     INSERT_PADDING_WORDS(0x9);
 
     struct FullTextureConfig {
-        const bool enabled;
+        const u32 enabled;
         const TextureConfig config;
         const TextureFormat format;
     };
     const std::array<FullTextureConfig, 3> GetTextures() const {
         return {{
-            {static_cast<bool>(main_config.texture0_enable), texture0, texture0_format},
-            {static_cast<bool>(main_config.texture1_enable), texture1, texture1_format},
-            {static_cast<bool>(main_config.texture2_enable), texture2, texture2_format},
+            {main_config.texture0_enable, texture0, texture0_format},
+            {main_config.texture1_enable, texture1, texture1_format},
+            {main_config.texture2_enable, texture2, texture2_format},
         }};
     }
 

src/video_core/pica/regs_texturing.h

@@ -600,14 +600,43 @@ typename T::Surface& RasterizerCache<T>::GetTextureCube(const TextureCubeConfig&
     auto [it, new_surface] = texture_cube_cache.try_emplace(config);
     TextureCube& cube = it->second;
 
+    const std::array addresses = {config.px, config.nx, config.py, config.ny, config.pz, config.nz};
+
     if (new_surface) {
+        Pica::Texture::TextureInfo info = {
+            .width = config.width,
+            .height = config.width,
+            .format = config.format,
+        };
+        info.SetDefaultStride();
+
+        u32 res_scale = 1;
+        for (u32 i = 0; i < addresses.size(); i++) {
+            if (!addresses[i]) {
+                continue;
+            }
+
+            SurfaceId& face_id = cube.face_ids[i];
+            if (!face_id) {
+                info.physical_address = addresses[i];
+                face_id = GetTextureSurface(info, config.levels - 1);
+                Surface& surface = slot_surfaces[face_id];
+                ASSERT_MSG(
+                    surface.levels >= config.levels,
+                    "Texture cube face levels are not enough to validate the levels requested");
+                surface.flags |= SurfaceFlagBits::Tracked;
+            }
+            Surface& surface = slot_surfaces[face_id];
+            res_scale = std::max(surface.res_scale, res_scale);
+        }
+
         SurfaceParams cube_params = {
             .addr = config.px,
             .width = config.width,
             .height = config.width,
             .stride = config.width,
             .levels = config.levels,
-            .res_scale = filter != Settings::TextureFilter::None ? resolution_scale_factor : 1,
+            .res_scale = res_scale,
             .texture_type = TextureType::CubeMap,
             .pixel_format = PixelFormatFromTextureFormat(config.format),
             .type = SurfaceType::Texture,
@@ -616,48 +645,30 @@ typename T::Surface& RasterizerCache<T>::GetTextureCube(const TextureCubeConfig&
         cube.surface_id = CreateSurface(cube_params);
     }
 
-    const u32 scaled_size = slot_surfaces[cube.surface_id].GetScaledWidth();
-    const std::array addresses = {config.px, config.nx, config.py, config.ny, config.pz, config.nz};
-
-    Pica::Texture::TextureInfo info = {
-        .width = config.width,
-        .height = config.width,
-        .format = config.format,
-    };
-    info.SetDefaultStride();
-
+    Surface& cube_surface = slot_surfaces[cube.surface_id];
     for (u32 i = 0; i < addresses.size(); i++) {
         if (!addresses[i]) {
             continue;
         }
-
-        SurfaceId& face_id = cube.face_ids[i];
-        if (!face_id) {
-            info.physical_address = addresses[i];
-            face_id = GetTextureSurface(info, config.levels - 1);
-            ASSERT_MSG(slot_surfaces[face_id].levels >= config.levels,
-                       "Texture cube face levels are not enough to validate the levels requested");
-        }
-        Surface& surface = slot_surfaces[face_id];
-        surface.flags |= SurfaceFlagBits::Tracked;
+        Surface& surface = slot_surfaces[cube.face_ids[i]];
         if (cube.ticks[i] == surface.modification_tick) {
             continue;
         }
         cube.ticks[i] = surface.modification_tick;
-        Surface& cube_surface = slot_surfaces[cube.surface_id];
+        boost::container::small_vector<TextureCopy, 8> upload_copies;
         for (u32 level = 0; level < config.levels; level++) {
-            const u32 width_lod = scaled_size >> level;
-            const TextureCopy texture_copy = {
+            const u32 width_lod = surface.GetScaledWidth() >> level;
+            upload_copies.push_back({
                 .src_level = level,
                 .dst_level = level,
                 .src_layer = 0,
                 .dst_layer = i,
                 .src_offset = {0, 0},
                 .dst_offset = {0, 0},
                 .extent = {width_lod, width_lod},
-            };
-            runtime.CopyTextures(surface, cube_surface, texture_copy);
+            });
         }
+        runtime.CopyTextures(surface, cube_surface, upload_copies);
     }
 
     return slot_surfaces[cube.surface_id];

src/video_core/rasterizer_cache/rasterizer_cache.h

@@ -260,16 +260,19 @@ void TextureRuntime::ClearTexture(Surface& surface, const VideoCore::TextureClea
 }
 
 bool TextureRuntime::CopyTextures(Surface& source, Surface& dest,
-                                  const VideoCore::TextureCopy& copy) {
+                                  std::span<const VideoCore::TextureCopy> copies) {
     const GLenum src_textarget = source.texture_type == VideoCore::TextureType::CubeMap
                                      ? GL_TEXTURE_CUBE_MAP
                                      : GL_TEXTURE_2D;
     const GLenum dest_textarget =
         dest.texture_type == VideoCore::TextureType::CubeMap ? GL_TEXTURE_CUBE_MAP : GL_TEXTURE_2D;
-    glCopyImageSubData(source.Handle(), src_textarget, copy.src_level, copy.src_offset.x,
-                       copy.src_offset.y, copy.src_layer, dest.Handle(), dest_textarget,
-                       copy.dst_level, copy.dst_offset.x, copy.dst_offset.y, copy.dst_layer,
-                       copy.extent.width, copy.extent.height, 1);
+
+    for (const auto& copy : copies) {
+        glCopyImageSubData(source.Handle(), src_textarget, copy.src_level, copy.src_offset.x,
+                           copy.src_offset.y, copy.src_layer, dest.Handle(), dest_textarget,
+                           copy.dst_level, copy.dst_offset.x, copy.dst_offset.y, copy.dst_layer,
+                           copy.extent.width, copy.extent.height, 1);
+    }
     return true;
 }
 

src/video_core/renderer_opengl/gl_texture_runtime.cpp

@@ -65,7 +65,12 @@ class TextureRuntime {
     void ClearTexture(Surface& surface, const VideoCore::TextureClear& clear);
 
     /// Copies a rectangle of source to another rectange of dest
-    bool CopyTextures(Surface& source, Surface& dest, const VideoCore::TextureCopy& copy);
+    bool CopyTextures(Surface& source, Surface& dest,
+                      std::span<const VideoCore::TextureCopy> copies);
+
+    bool CopyTextures(Surface& source, Surface& dest, const VideoCore::TextureCopy& copy) {
+        return CopyTextures(source, dest, std::array{copy});
+    }
 
     /// Blits a rectangle of source to another rectange of dest
     bool BlitTextures(Surface& source, Surface& dest, const VideoCore::TextureBlit& blit);

src/video_core/renderer_opengl/gl_texture_runtime.h

@@ -54,21 +54,21 @@ RendererVulkan::RendererVulkan(Core::System& system, Pica::PicaCore& pica_,
                                Frontend::EmuWindow& window, Frontend::EmuWindow* secondary_window)
     : RendererBase{system, window, secondary_window}, memory{system.Memory()}, pica{pica_},
       instance{system.TelemetrySession(), window, Settings::values.physical_device.GetValue()},
-      scheduler{instance}, render_manager{instance, scheduler}, pool{instance},
-      main_window{window, instance, scheduler},
+      scheduler{instance}, render_manager{instance, scheduler}, main_window{window, instance,
+                                                                            scheduler},
       vertex_buffer{instance, scheduler, vk::BufferUsageFlagBits::eVertexBuffer,
                     VERTEX_BUFFER_SIZE},
-      rasterizer{memory,
-                 pica,
-                 system.CustomTexManager(),
-                 *this,
-                 render_window,
-                 instance,
-                 scheduler,
-                 pool,
-                 render_manager,
-                 main_window.ImageCount()},
-      present_set_provider{instance, pool, PRESENT_BINDINGS} {
+      update_queue{instance}, rasterizer{memory,
+                                         pica,
+                                         system.CustomTexManager(),
+                                         *this,
+                                         render_window,
+                                         instance,
+                                         scheduler,
+                                         render_manager,
+                                         update_queue,
+                                         main_window.ImageCount()},
+      present_heap{instance, scheduler.GetMasterSemaphore(), PRESENT_BINDINGS, 32} {
     CompileShaders();
     BuildLayouts();
     BuildPipelines();
@@ -127,16 +127,14 @@ void RendererVulkan::PrepareRendertarget() {
 
 void RendererVulkan::PrepareDraw(Frame* frame, const Layout::FramebufferLayout& layout) {
     const auto sampler = present_samplers[!Settings::values.filter_mode.GetValue()];
-    std::transform(screen_infos.begin(), screen_infos.end(), present_textures.begin(),
-                   [&](auto& info) {
-                       return DescriptorData{vk::DescriptorImageInfo{sampler, info.image_view,
-                                                                     vk::ImageLayout::eGeneral}};
-                   });
-
-    const auto descriptor_set = present_set_provider.Acquire(present_textures);
+    const auto present_set = present_heap.Commit();
+    for (u32 index = 0; index < screen_infos.size(); index++) {
+        update_queue.AddImageSampler(present_set, 0, index, screen_infos[index].image_view,
+                                     sampler);
+    }
 
     render_manager.EndRendering();
-    scheduler.Record([this, layout, frame, descriptor_set, renderpass = main_window.Renderpass(),
+    scheduler.Record([this, layout, frame, present_set, renderpass = main_window.Renderpass(),
                       index = current_pipeline](vk::CommandBuffer cmdbuf) {
         const vk::Viewport viewport = {
             .x = 0.0f,
@@ -171,7 +169,7 @@ void RendererVulkan::PrepareDraw(Frame* frame, const Layout::FramebufferLayout&
 
         cmdbuf.beginRenderPass(renderpass_begin_info, vk::SubpassContents::eInline);
         cmdbuf.bindPipeline(vk::PipelineBindPoint::eGraphics, present_pipelines[index]);
-        cmdbuf.bindDescriptorSets(vk::PipelineBindPoint::eGraphics, layout, 0, descriptor_set, {});
+        cmdbuf.bindDescriptorSets(vk::PipelineBindPoint::eGraphics, layout, 0, present_set, {});
     });
 }
 
@@ -258,7 +256,7 @@ void RendererVulkan::BuildLayouts() {
         .size = sizeof(PresentUniformData),
     };
 
-    const auto descriptor_set_layout = present_set_provider.Layout();
+    const auto descriptor_set_layout = present_heap.Layout();
     const vk::PipelineLayoutCreateInfo layout_info = {
         .setLayoutCount = 1,
         .pSetLayouts = &descriptor_set_layout,

src/video_core/renderer_vulkan/renderer_vulkan.cpp

@@ -7,7 +7,6 @@
 #include "common/common_types.h"
 #include "common/math_util.h"
 #include "video_core/renderer_base.h"
-#include "video_core/renderer_vulkan/vk_descriptor_pool.h"
 #include "video_core/renderer_vulkan/vk_instance.h"
 #include "video_core/renderer_vulkan/vk_present_window.h"
 #include "video_core/renderer_vulkan/vk_rasterizer.h"
@@ -119,22 +118,21 @@ class RendererVulkan : public VideoCore::RendererBase {
     Instance instance;
     Scheduler scheduler;
     RenderManager render_manager;
-    DescriptorPool pool;
     PresentWindow main_window;
     StreamBuffer vertex_buffer;
+    DescriptorUpdateQueue update_queue;
     RasterizerVulkan rasterizer;
     std::unique_ptr<PresentWindow> second_window;
 
+    DescriptorHeap present_heap;
     vk::UniquePipelineLayout present_pipeline_layout;
-    DescriptorSetProvider present_set_provider;
     std::array<vk::Pipeline, PRESENT_PIPELINES> present_pipelines;
     std::array<vk::ShaderModule, PRESENT_PIPELINES> present_shaders;
     std::array<vk::Sampler, 2> present_samplers;
     vk::ShaderModule present_vertex_shader;
     u32 current_pipeline = 0;
 
     std::array<ScreenInfo, 3> screen_infos{};
-    std::array<DescriptorData, 3> present_textures{};
     PresentUniformData draw_info{};
     vk::ClearColorValue clear_color{};
 };