nvnflinger/gpu: implement blending

6 files changed, 102 insertions, 41 deletions

diff --git a/src/video_core/renderer_vulkan/present/util.cpp b/src/video_core/renderer_vulkan/present/util.cpp
index 6ee16595d..7f27c7c1b 100644
--- a/src/video_core/renderer_vulkan/present/util.cpp
+++ b/src/video_core/renderer_vulkan/present/util.cpp
@@ -362,10 +362,10 @@ vk::PipelineLayout CreateWrappedPipelineLayout(const Device& device,
     });
 }
 
-vk::Pipeline CreateWrappedPipeline(const Device& device, vk::RenderPass& renderpass,
-                                   vk::PipelineLayout& layout,
-                                   std::tuple<vk::ShaderModule&, vk::ShaderModule&> shaders,
-                                   bool enable_blending) {
+static vk::Pipeline CreateWrappedPipelineImpl(
+    const Device& device, vk::RenderPass& renderpass, vk::PipelineLayout& layout,
+    std::tuple<vk::ShaderModule&, vk::ShaderModule&> shaders,
+    VkPipelineColorBlendAttachmentState blending) {
     const std::array<VkPipelineShaderStageCreateInfo, 2> shader_stages{{
         {
             .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
@@ -443,30 +443,6 @@ vk::Pipeline CreateWrappedPipeline(const Device& device, vk::RenderPass& renderp
         .alphaToOneEnable = VK_FALSE,
     };
 
-    constexpr VkPipelineColorBlendAttachmentState color_blend_attachment_disabled{
-        .blendEnable = VK_FALSE,
-        .srcColorBlendFactor = VK_BLEND_FACTOR_ZERO,
-        .dstColorBlendFactor = VK_BLEND_FACTOR_ZERO,
-        .colorBlendOp = VK_BLEND_OP_ADD,
-        .srcAlphaBlendFactor = VK_BLEND_FACTOR_ZERO,
-        .dstAlphaBlendFactor = VK_BLEND_FACTOR_ZERO,
-        .alphaBlendOp = VK_BLEND_OP_ADD,
-        .colorWriteMask = VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT |
-                          VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT,
-    };
-
-    constexpr VkPipelineColorBlendAttachmentState color_blend_attachment_enabled{
-        .blendEnable = VK_TRUE,
-        .srcColorBlendFactor = VK_BLEND_FACTOR_SRC_ALPHA,
-        .dstColorBlendFactor = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA,
-        .colorBlendOp = VK_BLEND_OP_ADD,
-        .srcAlphaBlendFactor = VK_BLEND_FACTOR_ONE,
-        .dstAlphaBlendFactor = VK_BLEND_FACTOR_ZERO,
-        .alphaBlendOp = VK_BLEND_OP_ADD,
-        .colorWriteMask = VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT |
-                          VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT,
-    };
-
     const VkPipelineColorBlendStateCreateInfo color_blend_ci{
         .sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
         .pNext = nullptr,
@@ -474,8 +450,7 @@ vk::Pipeline CreateWrappedPipeline(const Device& device, vk::RenderPass& renderp
         .logicOpEnable = VK_FALSE,
         .logicOp = VK_LOGIC_OP_COPY,
         .attachmentCount = 1,
-        .pAttachments =
-            enable_blending ? &color_blend_attachment_enabled : &color_blend_attachment_disabled,
+        .pAttachments = &blending,
         .blendConstants = {0.0f, 0.0f, 0.0f, 0.0f},
     };
 
@@ -515,6 +490,63 @@ vk::Pipeline CreateWrappedPipeline(const Device& device, vk::RenderPass& renderp
     });
 }
 
+vk::Pipeline CreateWrappedPipeline(const Device& device, vk::RenderPass& renderpass,
+                                   vk::PipelineLayout& layout,
+                                   std::tuple<vk::ShaderModule&, vk::ShaderModule&> shaders) {
+    constexpr VkPipelineColorBlendAttachmentState color_blend_attachment_disabled{
+        .blendEnable = VK_FALSE,
+        .srcColorBlendFactor = VK_BLEND_FACTOR_ZERO,
+        .dstColorBlendFactor = VK_BLEND_FACTOR_ZERO,
+        .colorBlendOp = VK_BLEND_OP_ADD,
+        .srcAlphaBlendFactor = VK_BLEND_FACTOR_ZERO,
+        .dstAlphaBlendFactor = VK_BLEND_FACTOR_ZERO,
+        .alphaBlendOp = VK_BLEND_OP_ADD,
+        .colorWriteMask = VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT |
+                          VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT,
+    };
+
+    return CreateWrappedPipelineImpl(device, renderpass, layout, shaders,
+                                     color_blend_attachment_disabled);
+}
+
+vk::Pipeline CreateWrappedPremultipliedBlendingPipeline(
+    const Device& device, vk::RenderPass& renderpass, vk::PipelineLayout& layout,
+    std::tuple<vk::ShaderModule&, vk::ShaderModule&> shaders) {
+    constexpr VkPipelineColorBlendAttachmentState color_blend_attachment_premultiplied{
+        .blendEnable = VK_TRUE,
+        .srcColorBlendFactor = VK_BLEND_FACTOR_ONE,
+        .dstColorBlendFactor = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA,
+        .colorBlendOp = VK_BLEND_OP_ADD,
+        .srcAlphaBlendFactor = VK_BLEND_FACTOR_ONE,
+        .dstAlphaBlendFactor = VK_BLEND_FACTOR_ZERO,
+        .alphaBlendOp = VK_BLEND_OP_ADD,
+        .colorWriteMask = VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT |
+                          VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT,
+    };
+
+    return CreateWrappedPipelineImpl(device, renderpass, layout, shaders,
+                                     color_blend_attachment_premultiplied);
+}
+
+vk::Pipeline CreateWrappedCoverageBlendingPipeline(
+    const Device& device, vk::RenderPass& renderpass, vk::PipelineLayout& layout,
+    std::tuple<vk::ShaderModule&, vk::ShaderModule&> shaders) {
+    constexpr VkPipelineColorBlendAttachmentState color_blend_attachment_coverage{
+        .blendEnable = VK_TRUE,
+        .srcColorBlendFactor = VK_BLEND_FACTOR_SRC_ALPHA,
+        .dstColorBlendFactor = VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA,
+        .colorBlendOp = VK_BLEND_OP_ADD,
+        .srcAlphaBlendFactor = VK_BLEND_FACTOR_ONE,
+        .dstAlphaBlendFactor = VK_BLEND_FACTOR_ZERO,
+        .alphaBlendOp = VK_BLEND_OP_ADD,
+        .colorWriteMask = VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT |
+                          VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT,
+    };
+
+    return CreateWrappedPipelineImpl(device, renderpass, layout, shaders,
+                                     color_blend_attachment_coverage);
+}
+
 VkWriteDescriptorSet CreateWriteDescriptorSet(std::vector<VkDescriptorImageInfo>& images,
                                               VkSampler sampler, VkImageView view,
                                               VkDescriptorSet set, u32 binding) {
diff --git a/src/video_core/renderer_vulkan/present/util.h b/src/video_core/renderer_vulkan/present/util.h
index 1104aaa15..5b22f0fa8 100644
--- a/src/video_core/renderer_vulkan/present/util.h
+++ b/src/video_core/renderer_vulkan/present/util.h
@@ -42,8 +42,13 @@ vk::PipelineLayout CreateWrappedPipelineLayout(const Device& device,
                                                vk::DescriptorSetLayout& layout);
 vk::Pipeline CreateWrappedPipeline(const Device& device, vk::RenderPass& renderpass,
                                    vk::PipelineLayout& layout,
-                                   std::tuple<vk::ShaderModule&, vk::ShaderModule&> shaders,
-                                   bool enable_blending = false);
+                                   std::tuple<vk::ShaderModule&, vk::ShaderModule&> shaders);
+vk::Pipeline CreateWrappedPremultipliedBlendingPipeline(
+    const Device& device, vk::RenderPass& renderpass, vk::PipelineLayout& layout,
+    std::tuple<vk::ShaderModule&, vk::ShaderModule&> shaders);
+vk::Pipeline CreateWrappedCoverageBlendingPipeline(
+    const Device& device, vk::RenderPass& renderpass, vk::PipelineLayout& layout,
+    std::tuple<vk::ShaderModule&, vk::ShaderModule&> shaders);
 VkWriteDescriptorSet CreateWriteDescriptorSet(std::vector<VkDescriptorImageInfo>& images,
                                               VkSampler sampler, VkImageView view,
                                               VkDescriptorSet set, u32 binding);
diff --git a/src/video_core/renderer_vulkan/present/window_adapt_pass.cpp b/src/video_core/renderer_vulkan/present/window_adapt_pass.cpp
index c5db0230d..22ffacf11 100644
--- a/src/video_core/renderer_vulkan/present/window_adapt_pass.cpp
+++ b/src/video_core/renderer_vulkan/present/window_adapt_pass.cpp
@@ -22,7 +22,7 @@ WindowAdaptPass::WindowAdaptPass(const Device& device_, VkFormat frame_format,
     CreatePipelineLayout();
     CreateVertexShader();
     CreateRenderPass(frame_format);
-    CreatePipeline();
+    CreatePipelines();
 }
 
 WindowAdaptPass::~WindowAdaptPass() = default;
@@ -34,7 +34,6 @@ void WindowAdaptPass::Draw(RasterizerVulkan& rasterizer, Scheduler& scheduler, s
 
     const VkFramebuffer host_framebuffer{*dst->framebuffer};
     const VkRenderPass renderpass{*render_pass};
-    const VkPipeline graphics_pipeline{*pipeline};
     const VkPipelineLayout graphics_pipeline_layout{*pipeline_layout};
     const VkExtent2D render_area{
         .width = dst->width,
@@ -44,9 +43,23 @@ void WindowAdaptPass::Draw(RasterizerVulkan& rasterizer, Scheduler& scheduler, s
     const size_t layer_count = configs.size();
     std::vector<PresentPushConstants> push_constants(layer_count);
     std::vector<VkDescriptorSet> descriptor_sets(layer_count);
+    std::vector<VkPipeline> graphics_pipelines(layer_count);
 
     auto layer_it = layers.begin();
     for (size_t i = 0; i < layer_count; i++) {
+        switch (configs[i].blending) {
+        case Tegra::BlendMode::Opaque:
+        default:
+            graphics_pipelines[i] = *opaque_pipeline;
+            break;
+        case Tegra::BlendMode::Premultiplied:
+            graphics_pipelines[i] = *premultiplied_pipeline;
+            break;
+        case Tegra::BlendMode::Coverage:
+            graphics_pipelines[i] = *coverage_pipeline;
+            break;
+        }
+
         layer_it->ConfigureDraw(&push_constants[i], &descriptor_sets[i], rasterizer, *sampler,
                                 image_index, configs[i], layout);
         layer_it++;
@@ -77,8 +90,8 @@ void WindowAdaptPass::Draw(RasterizerVulkan& rasterizer, Scheduler& scheduler, s
         BeginRenderPass(cmdbuf, renderpass, host_framebuffer, render_area);
         cmdbuf.ClearAttachments({clear_attachment}, {clear_rect});
 
-        cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_GRAPHICS, graphics_pipeline);
         for (size_t i = 0; i < layer_count; i++) {
+            cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_GRAPHICS, graphics_pipelines[i]);
             cmdbuf.PushConstants(graphics_pipeline_layout, VK_SHADER_STAGE_VERTEX_BIT,
                                  push_constants[i]);
             cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_GRAPHICS, graphics_pipeline_layout, 0,
@@ -129,9 +142,13 @@ void WindowAdaptPass::CreateRenderPass(VkFormat frame_format) {
     render_pass = CreateWrappedRenderPass(device, frame_format, VK_IMAGE_LAYOUT_UNDEFINED);
 }
 
-void WindowAdaptPass::CreatePipeline() {
-    pipeline = CreateWrappedPipeline(device, render_pass, pipeline_layout,
-                                     std::tie(vertex_shader, fragment_shader), false);
+void WindowAdaptPass::CreatePipelines() {
+    opaque_pipeline = CreateWrappedPipeline(device, render_pass, pipeline_layout,
+                                            std::tie(vertex_shader, fragment_shader));
+    premultiplied_pipeline = CreateWrappedPremultipliedBlendingPipeline(
+        device, render_pass, pipeline_layout, std::tie(vertex_shader, fragment_shader));
+    coverage_pipeline = CreateWrappedCoverageBlendingPipeline(
+        device, render_pass, pipeline_layout, std::tie(vertex_shader, fragment_shader));
 }
 
 } // namespace Vulkan
diff --git a/src/video_core/renderer_vulkan/present/window_adapt_pass.h b/src/video_core/renderer_vulkan/present/window_adapt_pass.h
index 0e2edfc31..cf667a4fc 100644
--- a/src/video_core/renderer_vulkan/present/window_adapt_pass.h
+++ b/src/video_core/renderer_vulkan/present/window_adapt_pass.h
@@ -42,7 +42,7 @@ private:
     void CreatePipelineLayout();
     void CreateVertexShader();
     void CreateRenderPass(VkFormat frame_format);
-    void CreatePipeline();
+    void CreatePipelines();
 
 private:
     const Device& device;
@@ -52,7 +52,9 @@ private:
     vk::ShaderModule vertex_shader;
     vk::ShaderModule fragment_shader;
     vk::RenderPass render_pass;
-    vk::Pipeline pipeline;
+    vk::Pipeline opaque_pipeline;
+    vk::Pipeline premultiplied_pipeline;
+    vk::Pipeline coverage_pipeline;
 };
 
 } // namespace Vulkan
diff --git a/src/video_core/renderer_vulkan/renderer_vulkan.cpp b/src/video_core/renderer_vulkan/renderer_vulkan.cpp
index 48a105327..c7c234fd8 100644
--- a/src/video_core/renderer_vulkan/renderer_vulkan.cpp
+++ b/src/video_core/renderer_vulkan/renderer_vulkan.cpp
@@ -101,8 +101,10 @@ RendererVulkan::RendererVulkan(Core::TelemetrySession& telemetry_session_,
                       surface),
       blit_swapchain(device_memory, device, memory_allocator, present_manager, scheduler),
       blit_screenshot(device_memory, device, memory_allocator, present_manager, scheduler),
+      blit_application_layer(device_memory, device, memory_allocator, present_manager, scheduler),
       rasterizer(render_window, gpu, device_memory, device, memory_allocator, state_tracker,
-                 scheduler) {
+                 scheduler),
+      application_frame() {
     if (Settings::values.renderer_force_max_clock.GetValue() && device.ShouldBoostClocks()) {
         turbo_mode.emplace(instance, dld);
         scheduler.RegisterOnSubmit([this] { turbo_mode->QueueSubmitted(); });
diff --git a/src/video_core/renderer_vulkan/renderer_vulkan.h b/src/video_core/renderer_vulkan/renderer_vulkan.h
index c6d8a0f21..ed9c7af7f 100644
--- a/src/video_core/renderer_vulkan/renderer_vulkan.h
+++ b/src/video_core/renderer_vulkan/renderer_vulkan.h
@@ -80,8 +80,11 @@ private:
     PresentManager present_manager;
     BlitScreen blit_swapchain;
     BlitScreen blit_screenshot;
+    BlitScreen blit_application_layer;
     RasterizerVulkan rasterizer;
     std::optional<TurboMode> turbo_mode;
+
+    Frame application_frame;
 };
 
 } // namespace Vulkan