vulkan: Implement rescaling shader patching

src/video_core/renderer_vulkan/pipeline_helper.h

@@ -20,6 +20,8 @@
 
 namespace Vulkan {
 
+constexpr size_t MAX_RESCALING_WORDS = 4;
+
 class DescriptorLayoutBuilder {
 public:
     DescriptorLayoutBuilder(const Device& device_) : device{&device_} {}
@@ -68,18 +70,26 @@ public:
     }
 
     vk::PipelineLayout CreatePipelineLayout(VkDescriptorSetLayout descriptor_set_layout) const {
+        const VkPushConstantRange range{
+            .stageFlags = static_cast<VkShaderStageFlags>(
+                is_compute ? VK_SHADER_STAGE_COMPUTE_BIT : VK_SHADER_STAGE_ALL_GRAPHICS),
+            .offset = 0,
+            .size = (is_compute ? 0 : sizeof(f32)) + sizeof(std::array<u32, MAX_RESCALING_WORDS>),
+        };
         return device->GetLogical().CreatePipelineLayout({
             .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
             .pNext = nullptr,
             .flags = 0,
             .setLayoutCount = descriptor_set_layout ? 1U : 0U,
             .pSetLayouts = bindings.empty() ? nullptr : &descriptor_set_layout,
-            .pushConstantRangeCount = 0,
-            .pPushConstantRanges = nullptr,
+            .pushConstantRangeCount = 1,
+            .pPushConstantRanges = &range,
         });
     }
 
     void Add(const Shader::Info& info, VkShaderStageFlags stage) {
+        is_compute |= (stage & VK_SHADER_STAGE_COMPUTE_BIT) != 0;
+
         Add(VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER, stage, info.constant_buffer_descriptors);
         Add(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, stage, info.storage_buffers_descriptors);
         Add(VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER, stage, info.texture_buffer_descriptors);
@@ -115,6 +125,7 @@ private:
     }
 
     const Device* device{};
+    bool is_compute{};
     boost::container::small_vector<VkDescriptorSetLayoutBinding, 32> bindings;
     boost::container::small_vector<VkDescriptorUpdateTemplateEntryKHR, 32> entries;
     u32 binding{};
@@ -122,21 +133,46 @@ private:
     size_t offset{};
 };
 
-inline void PushImageDescriptors(const Shader::Info& info, const VkSampler*& samplers,
-                                 const ImageId*& image_view_ids, TextureCache& texture_cache,
-                                 VKUpdateDescriptorQueue& update_descriptor_queue) {
-    for (const auto& desc : info.texture_buffer_descriptors) {
-        image_view_ids += desc.count;
+class RescalingPushConstant {
+public:
+    explicit RescalingPushConstant(u32 num_textures) noexcept {}
+
+    void PushTexture(bool is_rescaled) noexcept {
+        *texture_ptr |= is_rescaled ? texture_bit : 0;
+        texture_bit <<= 1;
+        if (texture_bit == 0) {
+            texture_bit = 1u;
+            ++texture_ptr;
         }
-    for (const auto& desc : info.image_buffer_descriptors) {
-        image_view_ids += desc.count;
     }
+
+    const std::array<u32, MAX_RESCALING_WORDS>& Data() const noexcept {
+        return words;
+    }
+
+private:
+    std::array<u32, MAX_RESCALING_WORDS> words{};
+    u32* texture_ptr{words.data()};
+    u32 texture_bit{1u};
+};
+
+inline void PushImageDescriptors(const Shader::Info& info, const VkSampler*& samplers,
+                                 const ImageId*& image_view_ids, TextureCache& texture_cache,
+                                 VKUpdateDescriptorQueue& update_descriptor_queue,
+                                 RescalingPushConstant& rescaling) {
+    static constexpr VideoCommon::ImageViewId NULL_IMAGE_VIEW_ID{0};
+    image_view_ids += Shader::NumDescriptors(info.texture_buffer_descriptors);
+    image_view_ids += Shader::NumDescriptors(info.image_buffer_descriptors);
     for (const auto& desc : info.texture_descriptors) {
         for (u32 index = 0; index < desc.count; ++index) {
+            const VideoCommon::ImageViewId image_view_id{*(image_view_ids++)};
             const VkSampler sampler{*(samplers++)};
-            ImageView& image_view{texture_cache.GetImageView(*(image_view_ids++))};
+            ImageView& image_view{texture_cache.GetImageView(image_view_id)};
+            const Image& image{texture_cache.GetImage(image_view.image_id)};
             const VkImageView vk_image_view{image_view.Handle(desc.type)};
             update_descriptor_queue.AddSampledImage(vk_image_view, sampler);
+            rescaling.PushTexture(image_view_id != NULL_IMAGE_VIEW_ID &&
+                                  True(image.flags & VideoCommon::ImageFlagBits::Rescaled));
         }
     }
     for (const auto& desc : info.image_descriptors) {

src/video_core/renderer_vulkan/vk_compute_pipeline.cpp

@@ -180,9 +180,11 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
     buffer_cache.UpdateComputeBuffers();
     buffer_cache.BindHostComputeBuffers();
 
+    RescalingPushConstant rescaling(num_textures);
     const VkSampler* samplers_it{samplers.data()};
     const ImageId* views_it{image_view_ids.data()};
-    PushImageDescriptors(info, samplers_it, views_it, texture_cache, update_descriptor_queue);
+    PushImageDescriptors(info, samplers_it, views_it, texture_cache, update_descriptor_queue,
+                         rescaling);
 
     if (!is_built.load(std::memory_order::relaxed)) {
         // Wait for the pipeline to be built
@@ -192,11 +194,15 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
         });
     }
     const void* const descriptor_data{update_descriptor_queue.UpdateData()};
-    scheduler.Record([this, descriptor_data](vk::CommandBuffer cmdbuf) {
+    scheduler.Record(
+        [this, descriptor_data, rescaling_data = rescaling.Data()](vk::CommandBuffer cmdbuf) {
             cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, *pipeline);
             if (!descriptor_set_layout) {
                 return;
             }
+            if (num_textures > 0) {
+                cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT, rescaling_data);
+            }
             const VkDescriptorSet descriptor_set{descriptor_allocator.Commit()};
             const vk::Device& dev{device.GetLogical()};
             dev.UpdateDescriptorSet(descriptor_set, *descriptor_update_template, descriptor_data);

src/video_core/renderer_vulkan/vk_compute_pipeline.h

@@ -59,6 +59,7 @@ private:
     vk::PipelineLayout pipeline_layout;
     vk::DescriptorUpdateTemplateKHR descriptor_update_template;
     vk::Pipeline pipeline;
+    u32 num_textures{};
 
     std::condition_variable build_condvar;
     std::mutex build_mutex;

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -235,6 +235,7 @@ GraphicsPipeline::GraphicsPipeline(
         stage_infos[stage] = *info;
         enabled_uniform_buffer_masks[stage] = info->constant_buffer_mask;
         std::ranges::copy(info->constant_buffer_used_sizes, uniform_buffer_sizes[stage].begin());
+        num_textures += Shader::NumDescriptors(info->texture_descriptors);
     }
     auto func{[this, shader_notify, &render_pass_cache, &descriptor_pool, pipeline_statistics] {
         DescriptorLayoutBuilder builder{MakeBuilder(device, stage_infos)};
@@ -428,12 +429,13 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
 
     update_descriptor_queue.Acquire();
 
+    RescalingPushConstant rescaling(num_textures);
     const VkSampler* samplers_it{samplers.data()};
     const ImageId* views_it{image_view_ids.data()};
     const auto prepare_stage{[&](size_t stage) LAMBDA_FORCEINLINE {
         buffer_cache.BindHostStageBuffers(stage);
         PushImageDescriptors(stage_infos[stage], samplers_it, views_it, texture_cache,
-                             update_descriptor_queue);
+                             update_descriptor_queue, rescaling);
     }};
     if constexpr (Spec::enabled_stages[0]) {
         prepare_stage(0);
@@ -450,10 +452,10 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
     if constexpr (Spec::enabled_stages[4]) {
         prepare_stage(4);
     }
-    ConfigureDraw();
+    ConfigureDraw(rescaling);
 }
 
-void GraphicsPipeline::ConfigureDraw() {
+void GraphicsPipeline::ConfigureDraw(const RescalingPushConstant& rescaling) {
     texture_cache.UpdateRenderTargets(false);
     scheduler.RequestRenderpass(texture_cache.GetFramebuffer());
 
@@ -464,12 +466,23 @@ void GraphicsPipeline::ConfigureDraw() {
             build_condvar.wait(lock, [this] { return is_built.load(std::memory_order::relaxed); });
         });
     }
+    const bool is_rescaling{texture_cache.IsRescaling()};
+    const bool update_rescaling{scheduler.UpdateRescaling(is_rescaling)};
     const bool bind_pipeline{scheduler.UpdateGraphicsPipeline(this)};
     const void* const descriptor_data{update_descriptor_queue.UpdateData()};
-    scheduler.Record([this, descriptor_data, bind_pipeline](vk::CommandBuffer cmdbuf) {
+    scheduler.Record([this, descriptor_data, bind_pipeline, rescaling_data = rescaling.Data(),
+                      is_rescaling, update_rescaling](vk::CommandBuffer cmdbuf) {
         if (bind_pipeline) {
             cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_GRAPHICS, *pipeline);
         }
+        if (update_rescaling) {
+            const f32 config_down_factor{Settings::values.resolution_info.down_factor};
+            const float scale_down_factor{is_rescaling ? config_down_factor : 1.0f};
+            cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_ALL_GRAPHICS, 0,
+                                 sizeof(scale_down_factor), &scale_down_factor);
+        }
+        cmdbuf.PushConstants(*pipeline_layout, VK_SHADER_STAGE_ALL_GRAPHICS, sizeof(f32),
+                             sizeof(rescaling_data), rescaling_data.data());
         if (!descriptor_set_layout) {
             return;
         }

src/video_core/renderer_vulkan/vk_graphics_pipeline.h

@@ -62,6 +62,7 @@ namespace Vulkan {
 class Device;
 class PipelineStatistics;
 class RenderPassCache;
+class RescalingPushConstant;
 class VKScheduler;
 class VKUpdateDescriptorQueue;
 
@@ -113,7 +114,7 @@ private:
     template <typename Spec>
     void ConfigureImpl(bool is_indexed);
 
-    void ConfigureDraw();
+    void ConfigureDraw(const RescalingPushConstant& rescaling);
 
     void MakePipeline(VkRenderPass render_pass);
 
@@ -138,6 +139,7 @@ private:
     std::array<Shader::Info, NUM_STAGES> stage_infos;
     std::array<u32, 5> enabled_uniform_buffer_masks{};
     VideoCommon::UniformBufferSizes uniform_buffer_sizes{};
+    u32 num_textures{};
 
     vk::DescriptorSetLayout descriptor_set_layout;
     DescriptorAllocator descriptor_allocator;

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -139,6 +139,9 @@ Shader::RuntimeInfo MakeRuntimeInfo(std::span<const Shader::IR::Program> program
     } else {
         info.previous_stage_stores.mask.set();
     }
+    for (const auto& stage : programs) {
+        info.num_textures += Shader::NumDescriptors(stage.info.texture_descriptors);
+    }
     const Shader::Stage stage{program.stage};
     const bool has_geometry{key.unique_hashes[4] != 0 && !programs[4].is_geometry_passthrough};
     const bool gl_ndc{key.state.ndc_minus_one_to_one != 0};

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -128,6 +128,15 @@ bool VKScheduler::UpdateGraphicsPipeline(GraphicsPipeline* pipeline) {
     return true;
 }
 
+bool VKScheduler::UpdateRescaling(bool is_rescaling) {
+    if (state.rescaling_defined && is_rescaling == state.is_rescaling) {
+        return false;
+    }
+    state.rescaling_defined = true;
+    state.is_rescaling = is_rescaling;
+    return true;
+}
+
 void VKScheduler::WorkerThread(std::stop_token stop_token) {
     Common::SetCurrentThreadName("yuzu:VulkanWorker");
     do {
@@ -227,6 +236,7 @@ void VKScheduler::AllocateNewContext() {
 
 void VKScheduler::InvalidateState() {
     state.graphics_pipeline = nullptr;
+    state.rescaling_defined = false;
     state_tracker.InvalidateCommandBufferState();
 }
 

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -56,6 +56,9 @@ public:
     /// Update the pipeline to the current execution context.
     bool UpdateGraphicsPipeline(GraphicsPipeline* pipeline);
 
+    /// Update the rescaling state. Returns true if the state has to be updated.
+    bool UpdateRescaling(bool is_rescaling);
+
     /// Invalidates current command buffer state except for render passes
     void InvalidateState();
 
@@ -185,6 +188,8 @@ private:
         VkFramebuffer framebuffer = nullptr;
         VkExtent2D render_area = {0, 0};
         GraphicsPipeline* graphics_pipeline = nullptr;
+        bool is_rescaling = false;
+        bool rescaling_defined = false;
     };
 
     void WorkerThread(std::stop_token stop_token);