feat: TAA 时域抗锯齿（velocity + 邻域 clamp + 子像素抖动）

- Velocity buffer：Forward 第二颜色附件，vertex 计算 clip 空间运动向量（UBO 尾部追加 projNoJitter/prevView/prevProj；上一帧实例矩阵 SSBO Set0 binding2，与当前实例数组平行对齐——转动的物体速度正确不鬼影） - 子像素抖动：SceneRenderer Halton(2,3)×8 序列，抖动只进光栅化投影 - TAA resolve pass：velocity 反推历史位置 + 3×3 邻域 clamp 去鬼影 + mix(cur,hist,0.9) 时域累积；TAAColor 自引用历史 - RG 历史读补 post-barrier：采样后恢复镜像 resting layout（TAA 的历史是 ColorAttachment layout，不恢复则下帧 entry oldLayout 脱节） - 管线：Forward(HDR+Velocity) → TAA(TAAColor) → Bloom/Tonemap 改源 TAAColor - 编译零错误、validation 零警告、8 秒冒烟通过（10 pass 全活跃）
2026-09-23 13:29:31 +08:00
parent a29aed2803
commit 27b7f8785c
18 changed files with 369 additions and 39 deletions
@@ -8,6 +8,7 @@
 #include <Engine/Renderer/Material.h>

 #include <array>
+#include <span>

 namespace Engine::Passes
 {
@@ -20,12 +21,21 @@ namespace Engine::Passes
            clearValue.color = vk::ClearColorValue(
                std::array<float, 4>{0.1f, 0.1f, 0.1f, 0.0f});

-            vk::RenderingAttachmentInfo colorAttachment;
-            colorAttachment.setImageView(ctx.View(m_HDRColor))
-                           .setImageLayout(vk::ImageLayout::eColorAttachmentOptimal)
-                           .setLoadOp(vk::AttachmentLoadOp::eClear)
-                           .setStoreOp(vk::AttachmentStoreOp::eStore)
-                           .setClearValue(clearValue);
+            vk::RenderingAttachmentInfo colorAttachments[2];
+            colorAttachments[0].setImageView(ctx.View(m_HDRColor))
+                               .setImageLayout(vk::ImageLayout::eColorAttachmentOptimal)
+                               .setLoadOp(vk::AttachmentLoadOp::eClear)
+                               .setStoreOp(vk::AttachmentStoreOp::eStore)
+                               .setClearValue(clearValue);
+
+            // velocity 附件：清 0 = 静止（背景无几何 → 历史原位采样）
+            vk::ClearValue velocityClear;
+            velocityClear.color = vk::ClearColorValue(std::array<float, 4>{0.f, 0.f, 0.f, 0.f});
+            colorAttachments[1].setImageView(ctx.View(m_Velocity))
+                               .setImageLayout(vk::ImageLayout::eColorAttachmentOptimal)
+                               .setLoadOp(vk::AttachmentLoadOp::eClear)
+                               .setStoreOp(vk::AttachmentStoreOp::eStore)
+                               .setClearValue(velocityClear);

            vk::RenderingAttachmentInfo depthAttachment;
            depthAttachment.setImageView(ctx.View(m_Depth))
@@ -35,10 +45,13 @@ namespace Engine::Passes
                           .setClearValue(vk::ClearValue(
                               vk::ClearDepthStencilValue(1.0f, 0)));

+            std::array<vk::RenderingAttachmentInfo, 2> colorArray = {
+                colorAttachments[0], colorAttachments[1]};
+
            vk::RenderingInfo renderingInfo;
            renderingInfo.setRenderArea({{0, 0}, ctx.extent})
                         .setLayerCount(1)
-                         .setColorAttachments(colorAttachment)
+                         .setColorAttachments(colorArray)
                         .setPDepthAttachment(&depthAttachment);

            cmd.beginRendering(renderingInfo);
@@ -95,8 +108,6 @@ namespace Engine::Passes
            cmd.setScissorWithCount(scissor);

            vk::Bool32 blendEnable = VK_FALSE;
-            cmd.setColorBlendEnableEXT(0, blendEnable);
-
            vk::ColorBlendEquationEXT blendEquation;
            blendEquation.setSrcColorBlendFactor(vk::BlendFactor::eOne)
                         .setDstColorBlendFactor(vk::BlendFactor::eZero)
@@ -104,14 +115,20 @@ namespace Engine::Passes
                         .setSrcAlphaBlendFactor(vk::BlendFactor::eOne)
                         .setDstAlphaBlendFactor(vk::BlendFactor::eZero)
                         .setAlphaBlendOp(vk::BlendOp::eAdd);
-            cmd.setColorBlendEquationEXT(0, blendEquation);

            vk::ColorComponentFlags colorWriteMask =
                vk::ColorComponentFlagBits::eR |
                vk::ColorComponentFlagBits::eG |
                vk::ColorComponentFlagBits::eB |
                vk::ColorComponentFlagBits::eA;
-            cmd.setColorWriteMaskEXT(0, colorWriteMask);
+
+            // 两个颜色附件（HDR + velocity）都是无混合直写
+            cmd.setColorBlendEnableEXT(0,
+                std::array<vk::Bool32, 2>{blendEnable, blendEnable});
+            cmd.setColorBlendEquationEXT(0,
+                std::array<vk::ColorBlendEquationEXT, 2>{blendEquation, blendEquation});
+            cmd.setColorWriteMaskEXT(0,
+                std::array<vk::ColorComponentFlags, 2>{colorWriteMask, colorWriteMask});

            // ---- 绘制：GPUDriven 按 mesh 分 batch（相机视锥剔除后）----
            // 实例矩阵已由 Renderer 写入 ObjectBuffer（per-slot SSBO）；
@@ -127,6 +144,7 @@ namespace Engine::Passes
            cmd.endRendering();
        })
        .Write(m_HDRColor, RG::ResourceUse::ColorAttachment)
+        .Write(m_Velocity, RG::ResourceUse::ColorAttachment)
        .Write(m_Depth, RG::ResourceUse::DepthAttachment)
        .Read(m_ShadowMap, RG::ResourceUse::ShaderRead,
              vk::ShaderStageFlagBits::eFragment);
@@ -68,14 +68,21 @@ namespace Engine::Passes
            vk::ShaderStageFlagBits::eFragment,
            std::vector<vk::DescriptorSetLayout>{setLayout},
            static_cast<u32>(sizeof(FullscreenPush)));
+        auto taaF = shaders.CreateFromFile(
+            shaderDir + "taa.frag.spv",
+            vk::ShaderStageFlagBits::eFragment,
+            std::vector<vk::DescriptorSetLayout>{setLayout},
+            static_cast<u32>(sizeof(FullscreenPush)));

        vert = vertShader.Handle;
        tonemapFrag = tonemapF.Handle;
        extractFrag = extractF.Handle;
        blurFrag = blurF.Handle;
        diffFrag = diffF.Handle;
+        taaFrag = taaF.Handle;

-        if (!vert || !tonemapFrag || !extractFrag || !blurFrag || !diffFrag) {
+        if (!vert || !tonemapFrag || !extractFrag || !blurFrag ||
+            !diffFrag || !taaFrag) {
            ENGINE_FATAL("FullscreenGpu: failed to create shader objects");
        }
    }
@@ -89,6 +96,7 @@ namespace Engine::Passes
        if (extractFrag) vkDevice.destroyShaderEXT(extractFrag);
        if (blurFrag) vkDevice.destroyShaderEXT(blurFrag);
        if (diffFrag) vkDevice.destroyShaderEXT(diffFrag);
+        if (taaFrag) vkDevice.destroyShaderEXT(taaFrag);

        vkDevice.destroySampler(sampler);
        vkDevice.destroyPipelineLayout(layout);
@@ -15,15 +15,19 @@ namespace Engine::Passes
 {
    SceneGpu::SceneGpu(Device& device, DescriptorManager& descriptors,
                       ShaderManager& shaders,
-                       UniformBuffer& ubo, UniformBuffer& objectBuffer)
+                       UniformBuffer& ubo, UniformBuffer& objectBuffer,
+                       UniformBuffer& prevObjectBuffer)
        : m_Device(device), m_Descriptors(descriptors)
    {
        // Set 0: UBO + GPUDriven 实例数据（SSBO，vertex 读取）
+        // + 上一帧实例矩阵（TAA velocity，vertex 读取）
        std::vector<DescriptorBinding> uboBindings = {
            {0, vk::DescriptorType::eUniformBuffer,
             vk::ShaderStageFlagBits::eVertex | vk::ShaderStageFlagBits::eFragment},
            {1, vk::DescriptorType::eStorageBuffer,
             vk::ShaderStageFlagBits::eVertex},
+            {2, vk::DescriptorType::eStorageBuffer,
+             vk::ShaderStageFlagBits::eVertex},
        };
        uboSetLayout = descriptors.CreateLayout(uboBindings);

@@ -103,6 +107,9 @@ namespace Engine::Passes
                uboSets[slot], 0, ubo.Get(slot), ubo.GetSize());
            descriptors.UpdateStorageBuffer(
                uboSets[slot], 1, objectBuffer.Get(slot), objectBuffer.GetSize());
+            descriptors.UpdateStorageBuffer(
+                uboSets[slot], 2, prevObjectBuffer.Get(slot),
+                prevObjectBuffer.GetSize());

            shadowSets[slot] = descriptors.AllocateDescriptorSet(shadowSetLayout);
        }
@@ -0,0 +1,51 @@
+//
+// Created by atdunbg on 2026/9/23.
+//
+
+#include <Engine/Renderer/Passes/TAAPass.h>
+#include <Engine/Core/FrameContext.h>
+
+namespace Engine::Passes
+{
+    TAAPass::TAAPass(FullscreenGpu& gpu, f32 historyAlpha,
+                     RG::TextureHandle hdrColor, RG::TextureHandle velocity,
+                     RG::TextureHandle taaColor)
+        : m_Gpu(gpu), m_HistoryAlpha(historyAlpha),
+          m_HDR(hdrColor), m_Velocity(velocity), m_TAAColor(taaColor)
+    {
+        m_Sets.resize(kMaxFramesInFlight);
+        for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
+            m_Sets[slot] = m_Gpu.AllocateSet();
+        }
+    }
+
+    void TAAPass::Register(RG::RenderGraph& rg)
+    {
+        rg.AddPass("TAA",
+            [this](RG::RenderGraphContext& ctx) {
+                m_Gpu.RecordFullscreen(ctx, ctx.View(m_TAAColor), m_Gpu.taaFrag,
+                                       m_Sets[ctx.frame.frameSlot],
+                                       FullscreenPush{0.0f, m_HistoryAlpha, 0.0f, 0.0f});
+            })
+            .Read(m_HDR, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            .Read(m_Velocity, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            // 自引用历史：TAAColor 的上一帧槽位就是累积历史
+            .ReadHistory(m_TAAColor, RG::ResourceUse::ShaderRead,
+                         vk::ShaderStageFlagBits::eFragment)
+            .Write(m_TAAColor, RG::ResourceUse::ColorAttachment);
+    }
+
+    void TAAPass::OnTargetsRebuilt(RG::RenderGraph& rg)
+    {
+        for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
+            m_Gpu.UpdateSetInput(m_Sets[slot], 0, rg, m_HDR, slot);
+            m_Gpu.UpdateSetInput(m_Sets[slot], 1, rg, m_Velocity, slot);
+            // binding2 = 历史（上一帧槽位镜像）
+            m_Gpu.UpdateSetInput(m_Sets[slot], 2, rg, m_TAAColor,
+                                 rg.HistorySlot(slot));
+        }
+    }
+
+} // namespace Engine::Passes