nvnflinger/gpu: implement applet capture

src/core/hle/service/am/system_buffer_manager.cpp

@@ -68,8 +68,12 @@ void SystemBufferManager::SetWindowVisibility(bool visible) {
 
 Result SystemBufferManager::WriteAppletCaptureBuffer(bool* out_was_written,
                                                      s32* out_fbshare_layer_index) {
-    // TODO
+    if (!m_buffer_sharing_enabled) {
-    R_SUCCEED();
+        return VI::ResultPermissionDenied;
+    }
+
+    return m_nvnflinger->GetSystemBufferManager().WriteAppletCaptureBuffer(out_was_written,
+                                                                           out_fbshare_layer_index);
 }
 
 } // namespace Service::AM

src/core/hle/service/nvnflinger/fb_share_buffer_manager.cpp

@@ -15,6 +15,7 @@
 #include "core/hle/service/vi/layer/vi_layer.h"
 #include "core/hle/service/vi/vi_results.h"
 #include "video_core/gpu.h"
+#include "video_core/host1x/host1x.h"
 
 namespace Service::Nvnflinger {
 
@@ -414,9 +415,30 @@ Result FbShareBufferManager::GetSharedFrameBufferAcquirableEvent(Kernel::KReadab
     R_SUCCEED();
 }
 
-Result FbShareBufferManager::WriteAppletCaptureBuffer(bool* out_was_written,
+Result FbShareBufferManager::WriteAppletCaptureBuffer(bool* out_was_written, s32* out_layer_index) {
-                                                      s32* out_layer_index) {
+    std::vector<u8> capture_buffer(m_system.GPU().GetAppletCaptureBuffer());
-    // TODO
+    Common::ScratchBuffer<u32> scratch;
+
+    // TODO: this could be optimized
+    s64 e = -1280 * 768 * 4;
+    for (auto& block : *m_buffer_page_group) {
+        u8* start = m_system.DeviceMemory().GetPointer<u8>(block.GetAddress());
+        u8* end = m_system.DeviceMemory().GetPointer<u8>(block.GetAddress() + block.GetSize());
+
+        for (; start < end; start++) {
+            *start = 0;
+
+            if (e >= 0 && e < static_cast<s64>(capture_buffer.size())) {
+                *start = capture_buffer[e];
+            }
+            e++;
+        }
+
+        m_system.GPU().Host1x().MemoryManager().ApplyOpOnPointer(start, scratch, [&](DAddr addr) {
+            m_system.GPU().InvalidateRegion(addr, end - start);
+        });
+    }
+
     *out_was_written = true;
     *out_layer_index = 1;
     R_SUCCEED();

src/video_core/CMakeLists.txt

@@ -18,6 +18,7 @@ add_library(video_core STATIC
     buffer_cache/usage_tracker.h
     buffer_cache/word_manager.h
     cache_types.h
+    capture.h
     cdma_pusher.cpp
     cdma_pusher.h
     compatible_formats.cpp

src/video_core/capture.h

@@ -0,0 +1,36 @@
+// SPDX-FileCopyrightText: Copyright 2024 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include "common/alignment.h"
+#include "common/bit_util.h"
+#include "common/common_types.h"
+#include "core/frontend/framebuffer_layout.h"
+#include "video_core/surface.h"
+
+namespace VideoCore::Capture {
+
+constexpr u32 BlockHeight = 4;
+constexpr u32 BlockDepth = 0;
+constexpr u32 BppLog2 = 2;
+
+constexpr auto PixelFormat = Surface::PixelFormat::B8G8R8A8_UNORM;
+
+constexpr auto LinearWidth = Layout::ScreenUndocked::Width;
+constexpr auto LinearHeight = Layout::ScreenUndocked::Height;
+constexpr auto LinearDepth = 1U;
+constexpr auto BytesPerPixel = 4U;
+
+constexpr auto TiledWidth = LinearWidth;
+constexpr auto TiledHeight = Common::AlignUpLog2(LinearHeight, BlockHeight + BlockDepth + BppLog2);
+constexpr auto TiledSize = TiledWidth * TiledHeight * (1 << BppLog2);
+
+constexpr Layout::FramebufferLayout Layout{
+    .width = LinearWidth,
+    .height = LinearHeight,
+    .screen = {0, 0, LinearWidth, LinearHeight},
+    .is_srgb = false,
+};
+
+} // namespace VideoCore::Capture

src/video_core/gpu.cpp

@@ -347,6 +347,17 @@ struct GPU::Impl {
         WaitForSyncOperation(wait_fence);
     }
 
+    std::vector<u8> GetAppletCaptureBuffer() {
+        std::vector<u8> out;
+
+        const auto wait_fence =
+            RequestSyncOperation([&] { out = renderer->GetAppletCaptureBuffer(); });
+        gpu_thread.TickGPU();
+        WaitForSyncOperation(wait_fence);
+
+        return out;
+    }
+
     GPU& gpu;
     Core::System& system;
     Host1x::Host1x& host1x;
@@ -505,6 +516,10 @@ void GPU::RequestComposite(std::vector<Tegra::FramebufferConfig>&& layers,
     impl->RequestComposite(std::move(layers), std::move(fences));
 }
 
+std::vector<u8> GPU::GetAppletCaptureBuffer() {
+    return impl->GetAppletCaptureBuffer();
+}
+
 u64 GPU::GetTicks() const {
     return impl->GetTicks();
 }

src/video_core/gpu.h

@@ -215,6 +215,8 @@ public:
     void RequestComposite(std::vector<Tegra::FramebufferConfig>&& layers,
                           std::vector<Service::Nvidia::NvFence>&& fences);
 
+    std::vector<u8> GetAppletCaptureBuffer();
+
     /// Performs any additional setup necessary in order to begin GPU emulation.
     /// This can be used to launch any necessary threads and register any necessary
     /// core timing events.

src/video_core/renderer_base.h

@@ -40,6 +40,9 @@ public:
     /// Finalize rendering the guest frame and draw into the presentation texture
     virtual void Composite(std::span<const Tegra::FramebufferConfig> layers) = 0;
 
+    /// Get the tiled applet layer capture buffer
+    virtual std::vector<u8> GetAppletCaptureBuffer() = 0;
+
     [[nodiscard]] virtual RasterizerInterface* ReadRasterizer() = 0;
 
     [[nodiscard]] virtual std::string GetDeviceVendor() const = 0;

src/video_core/renderer_null/renderer_null.cpp

@@ -3,6 +3,7 @@
 
 #include "core/frontend/emu_window.h"
 #include "core/frontend/graphics_context.h"
+#include "video_core/capture.h"
 #include "video_core/renderer_null/renderer_null.h"
 
 namespace Null {
@@ -22,4 +23,8 @@ void RendererNull::Composite(std::span<const Tegra::FramebufferConfig> framebuff
     render_window.OnFrameDisplayed();
 }
 
+std::vector<u8> RendererNull::GetAppletCaptureBuffer() {
+    return std::vector<u8>(VideoCore::Capture::TiledSize);
+}
+
 } // namespace Null

src/video_core/renderer_null/renderer_null.h

@@ -19,6 +19,8 @@ public:
 
     void Composite(std::span<const Tegra::FramebufferConfig> framebuffer) override;
 
+    std::vector<u8> GetAppletCaptureBuffer() override;
+
     VideoCore::RasterizerInterface* ReadRasterizer() override {
         return &m_rasterizer;
     }

src/video_core/renderer_opengl/gl_blit_screen.cpp

@@ -20,7 +20,7 @@ BlitScreen::BlitScreen(RasterizerOpenGL& rasterizer_,
 BlitScreen::~BlitScreen() = default;
 
 void BlitScreen::DrawScreen(std::span<const Tegra::FramebufferConfig> framebuffers,
-                            const Layout::FramebufferLayout& layout) {
+                            const Layout::FramebufferLayout& layout, bool invert_y) {
     // TODO: Signal state tracker about these changes
     state_tracker.NotifyScreenDrawVertexArray();
     state_tracker.NotifyPolygonModes();
@@ -60,7 +60,7 @@ void BlitScreen::DrawScreen(std::span<const Tegra::FramebufferConfig> framebuffe
     }
 
     CreateWindowAdapt();
-    window_adapt->DrawToFramebuffer(program_manager, layers, framebuffers, layout);
+    window_adapt->DrawToFramebuffer(program_manager, layers, framebuffers, layout, invert_y);
 
     // TODO
     // program_manager.RestoreGuestPipeline();

src/video_core/renderer_opengl/gl_blit_screen.h

@@ -51,7 +51,7 @@ public:
 
     /// Draws the emulated screens to the emulator window.
     void DrawScreen(std::span<const Tegra::FramebufferConfig> framebuffers,
-                    const Layout::FramebufferLayout& layout);
+                    const Layout::FramebufferLayout& layout, bool invert_y);
 
 private:
     void CreateWindowAdapt();

src/video_core/renderer_opengl/present/layer.cpp

@@ -34,7 +34,7 @@ GLuint Layer::ConfigureDraw(std::array<GLfloat, 3 * 2>& out_matrix,
                             std::array<ScreenRectVertex, 4>& out_vertices,
                             ProgramManager& program_manager,
                             const Tegra::FramebufferConfig& framebuffer,
-                            const Layout::FramebufferLayout& layout) {
+                            const Layout::FramebufferLayout& layout, bool invert_y) {
     FramebufferTextureInfo info = PrepareRenderTarget(framebuffer);
     auto crop = Tegra::NormalizeCrop(framebuffer, info.width, info.height);
     GLuint texture = info.display_texture;
@@ -83,10 +83,15 @@ GLuint Layer::ConfigureDraw(std::array<GLfloat, 3 * 2>& out_matrix,
     const auto w = screen.GetWidth();
     const auto h = screen.GetHeight();
 
-    out_vertices[0] = ScreenRectVertex(x, y, crop.left, crop.top);
+    const auto left = crop.left;
-    out_vertices[1] = ScreenRectVertex(x + w, y, crop.right, crop.top);
+    const auto right = crop.right;
-    out_vertices[2] = ScreenRectVertex(x, y + h, crop.left, crop.bottom);
+    const auto top = invert_y ? crop.bottom : crop.top;
-    out_vertices[3] = ScreenRectVertex(x + w, y + h, crop.right, crop.bottom);
+    const auto bottom = invert_y ? crop.top : crop.bottom;
+
+    out_vertices[0] = ScreenRectVertex(x, y, left, top);
+    out_vertices[1] = ScreenRectVertex(x + w, y, right, top);
+    out_vertices[2] = ScreenRectVertex(x, y + h, left, bottom);
+    out_vertices[3] = ScreenRectVertex(x + w, y + h, right, bottom);
 
     return texture;
 }

src/video_core/renderer_opengl/present/layer.h

@@ -51,7 +51,7 @@ public:
                          std::array<ScreenRectVertex, 4>& out_vertices,
                          ProgramManager& program_manager,
                          const Tegra::FramebufferConfig& framebuffer,
-                         const Layout::FramebufferLayout& layout);
+                         const Layout::FramebufferLayout& layout, bool invert_y);
 
 private:
     /// Loads framebuffer from emulated memory into the active OpenGL texture.

src/video_core/renderer_opengl/present/window_adapt_pass.cpp

@@ -37,7 +37,7 @@ WindowAdaptPass::~WindowAdaptPass() = default;
 
 void WindowAdaptPass::DrawToFramebuffer(ProgramManager& program_manager, std::list<Layer>& layers,
                                         std::span<const Tegra::FramebufferConfig> framebuffers,
-                                        const Layout::FramebufferLayout& layout) {
+                                        const Layout::FramebufferLayout& layout, bool invert_y) {
     GLint old_read_fb;
     GLint old_draw_fb;
     glGetIntegerv(GL_READ_FRAMEBUFFER_BINDING, &old_read_fb);
@@ -51,7 +51,7 @@ void WindowAdaptPass::DrawToFramebuffer(ProgramManager& program_manager, std::li
     auto layer_it = layers.begin();
     for (size_t i = 0; i < layer_count; i++) {
         textures[i] = layer_it->ConfigureDraw(matrices[i], vertices[i], program_manager,
-                                              framebuffers[i], layout);
+                                              framebuffers[i], layout, invert_y);
         layer_it++;
     }
 

src/video_core/renderer_opengl/present/window_adapt_pass.h

@@ -31,7 +31,7 @@ public:
 
     void DrawToFramebuffer(ProgramManager& program_manager, std::list<Layer>& layers,
                            std::span<const Tegra::FramebufferConfig> framebuffers,
-                           const Layout::FramebufferLayout& layout);
+                           const Layout::FramebufferLayout& layout, bool invert_y);
 
 private:
     const Device& device;

src/video_core/renderer_opengl/renderer_opengl.cpp

@@ -16,6 +16,7 @@
 #include "core/core_timing.h"
 #include "core/frontend/emu_window.h"
 #include "core/telemetry_session.h"
+#include "video_core/capture.h"
 #include "video_core/renderer_opengl/gl_blit_screen.h"
 #include "video_core/renderer_opengl/gl_rasterizer.h"
 #include "video_core/renderer_opengl/gl_shader_manager.h"
@@ -121,6 +122,13 @@ RendererOpenGL::RendererOpenGL(Core::TelemetrySession& telemetry_session_,
     }
     blit_screen = std::make_unique<BlitScreen>(rasterizer, device_memory, state_tracker,
                                                program_manager, device);
+    blit_applet = std::make_unique<BlitScreen>(rasterizer, device_memory, state_tracker,
+                                               program_manager, device);
+    capture_framebuffer.Create();
+    capture_renderbuffer.Create();
+    glBindRenderbuffer(GL_RENDERBUFFER, capture_renderbuffer.handle);
+    glRenderbufferStorage(GL_RENDERBUFFER, GL_SRGB8, VideoCore::Capture::LinearWidth,
+                          VideoCore::Capture::LinearHeight);
 }
 
 RendererOpenGL::~RendererOpenGL() = default;
@@ -130,10 +138,11 @@ void RendererOpenGL::Composite(std::span<const Tegra::FramebufferConfig> framebu
         return;
     }
 
+    RenderAppletCaptureLayer(framebuffers);
     RenderScreenshot(framebuffers);
 
     state_tracker.BindFramebuffer(0);
-    blit_screen->DrawScreen(framebuffers, emu_window.GetFramebufferLayout());
+    blit_screen->DrawScreen(framebuffers, emu_window.GetFramebufferLayout(), false);
 
     ++m_current_frame;
 
@@ -159,11 +168,8 @@ void RendererOpenGL::AddTelemetryFields() {
     telemetry_session.AddField(user_system, "GPU_OpenGL_Version", std::string(gl_version));
 }
 
-void RendererOpenGL::RenderScreenshot(std::span<const Tegra::FramebufferConfig> framebuffers) {
+void RendererOpenGL::RenderToBuffer(std::span<const Tegra::FramebufferConfig> framebuffers,
-    if (!renderer_settings.screenshot_requested) {
+                                    const Layout::FramebufferLayout& layout, void* dst) {
-        return;
-    }
-
     GLint old_read_fb;
     GLint old_draw_fb;
     glGetIntegerv(GL_READ_FRAMEBUFFER_BINDING, &old_read_fb);
@@ -173,29 +179,86 @@ void RendererOpenGL::RenderScreenshot(std::span<const Tegra::FramebufferConfig>
     screenshot_framebuffer.Create();
     glBindFramebuffer(GL_FRAMEBUFFER, screenshot_framebuffer.handle);
 
-    const Layout::FramebufferLayout layout{renderer_settings.screenshot_framebuffer_layout};
-
     GLuint renderbuffer;
     glGenRenderbuffers(1, &renderbuffer);
     glBindRenderbuffer(GL_RENDERBUFFER, renderbuffer);
     glRenderbufferStorage(GL_RENDERBUFFER, GL_SRGB8, layout.width, layout.height);
     glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, renderbuffer);
 
-    blit_screen->DrawScreen(framebuffers, layout);
+    blit_screen->DrawScreen(framebuffers, layout, false);
 
     glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
     glPixelStorei(GL_PACK_ROW_LENGTH, 0);
-    glReadPixels(0, 0, layout.width, layout.height, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV,
+    glReadPixels(0, 0, layout.width, layout.height, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, dst);
-                 renderer_settings.screenshot_bits);
 
     screenshot_framebuffer.Release();
     glDeleteRenderbuffers(1, &renderbuffer);
 
     glBindFramebuffer(GL_READ_FRAMEBUFFER, old_read_fb);
     glBindFramebuffer(GL_DRAW_FRAMEBUFFER, old_draw_fb);
+}
+
+void RendererOpenGL::RenderScreenshot(std::span<const Tegra::FramebufferConfig> framebuffers) {
+    if (!renderer_settings.screenshot_requested) {
+        return;
+    }
+
+    RenderToBuffer(framebuffers, renderer_settings.screenshot_framebuffer_layout,
+                   renderer_settings.screenshot_bits);
 
     renderer_settings.screenshot_complete_callback(true);
     renderer_settings.screenshot_requested = false;
 }
 
+void RendererOpenGL::RenderAppletCaptureLayer(
+    std::span<const Tegra::FramebufferConfig> framebuffers) {
+    GLint old_read_fb;
+    GLint old_draw_fb;
+    glGetIntegerv(GL_READ_FRAMEBUFFER_BINDING, &old_read_fb);
+    glGetIntegerv(GL_DRAW_FRAMEBUFFER_BINDING, &old_draw_fb);
+
+    glBindFramebuffer(GL_FRAMEBUFFER, capture_framebuffer.handle);
+    glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER,
+                              capture_renderbuffer.handle);
+
+    blit_applet->DrawScreen(framebuffers, VideoCore::Capture::Layout, true);
+
+    glBindFramebuffer(GL_READ_FRAMEBUFFER, old_read_fb);
+    glBindFramebuffer(GL_DRAW_FRAMEBUFFER, old_draw_fb);
+}
+
+std::vector<u8> RendererOpenGL::GetAppletCaptureBuffer() {
+    using namespace VideoCore::Capture;
+
+    std::vector<u8> linear(TiledSize);
+    std::vector<u8> out(TiledSize);
+
+    GLint old_read_fb;
+    GLint old_draw_fb;
+    GLint old_pixel_pack_buffer;
+    GLint old_pack_row_length;
+    glGetIntegerv(GL_READ_FRAMEBUFFER_BINDING, &old_read_fb);
+    glGetIntegerv(GL_DRAW_FRAMEBUFFER_BINDING, &old_draw_fb);
+    glGetIntegerv(GL_PIXEL_PACK_BUFFER_BINDING, &old_pixel_pack_buffer);
+    glGetIntegerv(GL_PACK_ROW_LENGTH, &old_pack_row_length);
+
+    glBindFramebuffer(GL_FRAMEBUFFER, capture_framebuffer.handle);
+    glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER,
+                              capture_renderbuffer.handle);
+    glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+    glPixelStorei(GL_PACK_ROW_LENGTH, 0);
+    glReadPixels(0, 0, LinearWidth, LinearHeight, GL_RGBA, GL_UNSIGNED_INT_8_8_8_8_REV,
+                 linear.data());
+
+    glBindFramebuffer(GL_READ_FRAMEBUFFER, old_read_fb);
+    glBindFramebuffer(GL_DRAW_FRAMEBUFFER, old_draw_fb);
+    glBindBuffer(GL_PIXEL_PACK_BUFFER, old_pixel_pack_buffer);
+    glPixelStorei(GL_PACK_ROW_LENGTH, old_pack_row_length);
+
+    Tegra::Texture::SwizzleTexture(out, linear, BytesPerPixel, LinearWidth, LinearHeight,
+                                   LinearDepth, BlockHeight, BlockDepth);
+
+    return out;
+}
+
 } // namespace OpenGL

src/video_core/renderer_opengl/renderer_opengl.h

@@ -42,6 +42,8 @@ public:
 
     void Composite(std::span<const Tegra::FramebufferConfig> framebuffers) override;
 
+    std::vector<u8> GetAppletCaptureBuffer() override;
+
     VideoCore::RasterizerInterface* ReadRasterizer() override {
         return &rasterizer;
     }
@@ -52,7 +54,11 @@ public:
 
 private:
     void AddTelemetryFields();
+
+    void RenderToBuffer(std::span<const Tegra::FramebufferConfig> framebuffers,
+                        const Layout::FramebufferLayout& layout, void* dst);
     void RenderScreenshot(std::span<const Tegra::FramebufferConfig> framebuffers);
+    void RenderAppletCaptureLayer(std::span<const Tegra::FramebufferConfig> framebuffers);
 
     Core::TelemetrySession& telemetry_session;
     Core::Frontend::EmuWindow& emu_window;
@@ -64,8 +70,11 @@ private:
     ProgramManager program_manager;
     RasterizerOpenGL rasterizer;
     OGLFramebuffer screenshot_framebuffer;
+    OGLFramebuffer capture_framebuffer;
+    OGLRenderbuffer capture_renderbuffer;
 
     std::unique_ptr<BlitScreen> blit_screen;
+    std::unique_ptr<BlitScreen> blit_applet;
 };
 
 } // namespace OpenGL

src/video_core/renderer_vulkan/renderer_vulkan.cpp

@@ -19,6 +19,7 @@
 #include "core/core_timing.h"
 #include "core/frontend/graphics_context.h"
 #include "core/telemetry_session.h"
+#include "video_core/capture.h"
 #include "video_core/gpu.h"
 #include "video_core/renderer_vulkan/present/util.h"
 #include "video_core/renderer_vulkan/renderer_vulkan.h"
@@ -38,6 +39,20 @@
 
 namespace Vulkan {
 namespace {
+
+constexpr VkExtent2D CaptureImageSize{
+    .width = VideoCore::Capture::LinearWidth,
+    .height = VideoCore::Capture::LinearHeight,
+};
+
+constexpr VkExtent3D CaptureImageExtent{
+    .width = VideoCore::Capture::LinearWidth,
+    .height = VideoCore::Capture::LinearHeight,
+    .depth = VideoCore::Capture::LinearDepth,
+};
+
+constexpr VkFormat CaptureFormat = VK_FORMAT_A8B8G8R8_UNORM_PACK32;
+
 std::string GetReadableVersion(u32 version) {
     return fmt::format("{}.{}.{}", VK_VERSION_MAJOR(version), VK_VERSION_MINOR(version),
                        VK_VERSION_PATCH(version));
@@ -100,11 +115,11 @@ RendererVulkan::RendererVulkan(Core::TelemetrySession& telemetry_session_,
       present_manager(instance, render_window, device, memory_allocator, scheduler, swapchain,
                       surface),
       blit_swapchain(device_memory, device, memory_allocator, present_manager, scheduler),
-      blit_screenshot(device_memory, device, memory_allocator, present_manager, scheduler),
+      blit_capture(device_memory, device, memory_allocator, present_manager, scheduler),
-      blit_application_layer(device_memory, device, memory_allocator, present_manager, scheduler),
+      blit_applet(device_memory, device, memory_allocator, present_manager, scheduler),
       rasterizer(render_window, gpu, device_memory, device, memory_allocator, state_tracker,
                  scheduler),
-      application_frame() {
+      applet_frame() {
     if (Settings::values.renderer_force_max_clock.GetValue() && device.ShouldBoostClocks()) {
         turbo_mode.emplace(instance, dld);
         scheduler.RegisterOnSubmit([this] { turbo_mode->QueueSubmitted(); });
@@ -127,6 +142,8 @@ void RendererVulkan::Composite(std::span<const Tegra::FramebufferConfig> framebu
 
     SCOPE_EXIT({ render_window.OnFrameDisplayed(); });
 
+    RenderAppletCaptureLayer(framebuffers);
+
     if (!render_window.IsShown()) {
         return;
     }
@@ -169,30 +186,20 @@ void RendererVulkan::Report() const {
     telemetry_session.AddField(field, "GPU_Vulkan_Extensions", extensions);
 }
 
-void Vulkan::RendererVulkan::RenderScreenshot(
+vk::Buffer RendererVulkan::RenderToBuffer(std::span<const Tegra::FramebufferConfig> framebuffers,
-    std::span<const Tegra::FramebufferConfig> framebuffers) {
+                                          const Layout::FramebufferLayout& layout, VkFormat format,
-    if (!renderer_settings.screenshot_requested) {
+                                          VkDeviceSize buffer_size) {
-        return;
-    }
-
-    constexpr VkFormat ScreenshotFormat{VK_FORMAT_B8G8R8A8_UNORM};
-    const Layout::FramebufferLayout layout{renderer_settings.screenshot_framebuffer_layout};
-
     auto frame = [&]() {
         Frame f{};
-        f.image = CreateWrappedImage(memory_allocator, VkExtent2D{layout.width, layout.height},
+        f.image =
-                                     ScreenshotFormat);
+            CreateWrappedImage(memory_allocator, VkExtent2D{layout.width, layout.height}, format);
-        f.image_view = CreateWrappedImageView(device, f.image, ScreenshotFormat);
+        f.image_view = CreateWrappedImageView(device, f.image, format);
-        f.framebuffer = blit_screenshot.CreateFramebuffer(layout, *f.image_view, ScreenshotFormat);
+        f.framebuffer = blit_capture.CreateFramebuffer(layout, *f.image_view, format);
         return f;
     }();
 
-    blit_screenshot.DrawToFrame(rasterizer, &frame, framebuffers, layout, 1,
+    auto dst_buffer = CreateWrappedBuffer(memory_allocator, buffer_size, MemoryUsage::Download);
-                                VK_FORMAT_B8G8R8A8_UNORM);
+    blit_capture.DrawToFrame(rasterizer, &frame, framebuffers, layout, 1, format);
-
-    const auto dst_buffer = CreateWrappedBuffer(
-        memory_allocator, static_cast<VkDeviceSize>(layout.width * layout.height * 4),
-        MemoryUsage::Download);
 
     scheduler.RequestOutsideRenderPassOperationContext();
     scheduler.Record([&](vk::CommandBuffer cmdbuf) {
@@ -200,15 +207,68 @@ void Vulkan::RendererVulkan::RenderScreenshot(
                            VkExtent3D{layout.width, layout.height, 1});
     });
 
-    // Ensure the copy is fully completed before saving the screenshot
+    // Ensure the copy is fully completed before saving the capture
     scheduler.Finish();
 
-    // Copy backing image data to the QImage screenshot buffer
+    // Copy backing image data to the capture buffer
     dst_buffer.Invalidate();
+    return dst_buffer;
+}
+
+void RendererVulkan::RenderScreenshot(std::span<const Tegra::FramebufferConfig> framebuffers) {
+    if (!renderer_settings.screenshot_requested) {
+        return;
+    }
+
+    const auto& layout{renderer_settings.screenshot_framebuffer_layout};
+    const auto dst_buffer = RenderToBuffer(framebuffers, layout, VK_FORMAT_B8G8R8A8_UNORM,
+                                           layout.width * layout.height * 4);
+
     std::memcpy(renderer_settings.screenshot_bits, dst_buffer.Mapped().data(),
                 dst_buffer.Mapped().size());
     renderer_settings.screenshot_complete_callback(false);
     renderer_settings.screenshot_requested = false;
 }
 
+std::vector<u8> RendererVulkan::GetAppletCaptureBuffer() {
+    using namespace VideoCore::Capture;
+
+    std::vector<u8> out(VideoCore::Capture::TiledSize);
+
+    if (!applet_frame.image) {
+        return out;
+    }
+
+    const auto dst_buffer =
+        CreateWrappedBuffer(memory_allocator, VideoCore::Capture::TiledSize, MemoryUsage::Download);
+
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([&](vk::CommandBuffer cmdbuf) {
+        DownloadColorImage(cmdbuf, *applet_frame.image, *dst_buffer, CaptureImageExtent);
+    });
+
+    // Ensure the copy is fully completed before writing the capture
+    scheduler.Finish();
+
+    // Swizzle image data to the capture buffer
+    dst_buffer.Invalidate();
+    Tegra::Texture::SwizzleTexture(out, dst_buffer.Mapped(), BytesPerPixel, LinearWidth,
+                                   LinearHeight, LinearDepth, BlockHeight, BlockDepth);
+
+    return out;
+}
+
+void RendererVulkan::RenderAppletCaptureLayer(
+    std::span<const Tegra::FramebufferConfig> framebuffers) {
+    if (!applet_frame.image) {
+        applet_frame.image = CreateWrappedImage(memory_allocator, CaptureImageSize, CaptureFormat);
+        applet_frame.image_view = CreateWrappedImageView(device, applet_frame.image, CaptureFormat);
+        applet_frame.framebuffer = blit_applet.CreateFramebuffer(
+            VideoCore::Capture::Layout, *applet_frame.image_view, CaptureFormat);
+    }
+
+    blit_applet.DrawToFrame(rasterizer, &applet_frame, framebuffers, VideoCore::Capture::Layout, 1,
+                            CaptureFormat);
+}
+
 } // namespace Vulkan

src/video_core/renderer_vulkan/renderer_vulkan.h

@@ -48,6 +48,8 @@ public:
 
     void Composite(std::span<const Tegra::FramebufferConfig> framebuffers) override;
 
+    std::vector<u8> GetAppletCaptureBuffer() override;
+
     VideoCore::RasterizerInterface* ReadRasterizer() override {
         return &rasterizer;
     }
@@ -59,7 +61,11 @@ public:
 private:
     void Report() const;
 
+    vk::Buffer RenderToBuffer(std::span<const Tegra::FramebufferConfig> framebuffers,
+                              const Layout::FramebufferLayout& layout, VkFormat format,
+                              VkDeviceSize buffer_size);
     void RenderScreenshot(std::span<const Tegra::FramebufferConfig> framebuffers);
+    void RenderAppletCaptureLayer(std::span<const Tegra::FramebufferConfig> framebuffers);
 
     Core::TelemetrySession& telemetry_session;
     Tegra::MaxwellDeviceMemoryManager& device_memory;
@@ -79,12 +85,12 @@ private:
     Swapchain swapchain;
     PresentManager present_manager;
     BlitScreen blit_swapchain;
-    BlitScreen blit_screenshot;
+    BlitScreen blit_capture;
-    BlitScreen blit_application_layer;
+    BlitScreen blit_applet;
     RasterizerVulkan rasterizer;
     std::optional<TurboMode> turbo_mode;
 
-    Frame application_frame;
+    Frame applet_frame;
 };
 
 } // namespace Vulkan