service/gsp: Implement saving of framebuffers in SaveVramSysArea. (#6821)

* service/gsp: Implement saving of framebuffers in SaveVramSysArea.

* Address review comments.

* service/apt: Separate capture info and capture buffer info.

The former is used with the RequestForSysApplet message and GetCaptureInfo.
The latter is used with SendCaptureBufferInfo and ReceiveCaptureBufferInfo.

src/core/hle/service/apt/applet_manager.cpp

@@ -232,7 +232,7 @@ void AppletManager::CancelAndSendParameter(const MessageParameter& parameter) {
                       parameter.sender_id);
 
             if (parameter.buffer.size() >= sizeof(CaptureBufferInfo)) {
-                SendCaptureBufferInfo(parameter.buffer);
+                SetCaptureInfo(parameter.buffer);
                 CaptureFrameBuffers();
             }
 
@@ -511,6 +511,8 @@ ResultCode AppletManager::PrepareToStartLibraryApplet(AppletId applet_id) {
     last_library_launcher_slot = active_slot;
     last_prepared_library_applet = applet_id;
 
+    capture_buffer_info.reset();
+
     auto cfg = Service::CFG::GetModule(system);
     auto process =
         NS::LaunchTitle(FS::MediaType::NAND, GetTitleIdForApplet(applet_id, cfg->GetRegionValue()));
@@ -849,6 +851,9 @@ ResultCode AppletManager::PrepareToJumpToHomeMenu() {
     }
 
     last_jump_to_home_slot = active_slot;
+
+    capture_buffer_info.reset();
+
     if (last_jump_to_home_slot == AppletSlot::Application) {
         EnsureHomeMenuLoaded();
     }
@@ -1250,6 +1255,8 @@ ResultCode AppletManager::PrepareToStartApplication(u64 title_id, FS::MediaType
     app_start_parameters->next_title_id = title_id;
     app_start_parameters->next_media_type = media_type;
 
+    capture_buffer_info.reset();
+
     return RESULT_SUCCESS;
 }
 
@@ -1359,48 +1366,74 @@ void AppletManager::EnsureHomeMenuLoaded() {
     }
 }
 
-static void CaptureFrameBuffer(Core::System& system, u32 capture_offset, VAddr src, u32 height,
+static u32 GetDisplayBufferModePixelSize(DisplayBufferMode mode) {
-                               u32 format) {
+    switch (mode) {
-    static constexpr auto screen_capture_base_vaddr = static_cast<VAddr>(0x1F500000);
+    // NOTE: APT does in fact use pixel size 3 for R8G8B8A8 captures.
-    static constexpr auto screen_width = 240;
+    case DisplayBufferMode::R8G8B8A8:
-    static constexpr auto screen_width_pow2 = 256;
+    case DisplayBufferMode::R8G8B8:
-    const auto bpp = format < 2 ? 3 : 2;
+        return 3;
+    case DisplayBufferMode::R5G6B5:
+    case DisplayBufferMode::R5G5B5A1:
+    case DisplayBufferMode::R4G4B4A4:
+        return 2;
+    case DisplayBufferMode::Unimportable:
+        return 0;
+    default:
+        UNREACHABLE_MSG("Unknown display buffer mode {}", mode);
+        return 0;
+    }
+}
 
-    Memory::RasterizerFlushVirtualRegion(src, screen_width * height * bpp,
+static void CaptureFrameBuffer(Core::System& system, u32 capture_offset, VAddr src, u32 height,
+                               DisplayBufferMode mode) {
+    const auto bpp = GetDisplayBufferModePixelSize(mode);
+    if (bpp == 0) {
+        return;
+    }
+
+    Memory::RasterizerFlushVirtualRegion(src, GSP::FRAMEBUFFER_WIDTH * height * bpp,
                                          Memory::FlushMode::Flush);
 
-    auto dst_vaddr = screen_capture_base_vaddr + capture_offset;
+    // Address in VRAM that APT copies framebuffer captures to.
+    constexpr VAddr screen_capture_base_vaddr = Memory::VRAM_VADDR + 0x500000;
+    const auto dst_vaddr = screen_capture_base_vaddr + capture_offset;
     auto dst_ptr = system.Memory().GetPointer(dst_vaddr);
+    if (!dst_ptr) {
+        LOG_ERROR(Service_APT,
+                  "Could not retrieve framebuffer capture destination buffer, skipping screen.");
+        return;
+    }
+
     const auto src_ptr = system.Memory().GetPointer(src);
+    if (!src_ptr) {
+        LOG_ERROR(Service_APT,
+                  "Could not retrieve framebuffer capture source buffer, skipping screen.");
+        return;
+    }
+
     for (u32 y = 0; y < height; y++) {
-        for (u32 x = 0; x < screen_width; x++) {
+        for (u32 x = 0; x < GSP::FRAMEBUFFER_WIDTH; x++) {
-            auto dst_offset =
+            const auto dst_offset = VideoCore::GetMortonOffset(x, y, bpp) +
-                VideoCore::GetMortonOffset(x, y, bpp) + (y & ~7) * screen_width_pow2 * bpp;
+                                    (y & ~7) * GSP::FRAMEBUFFER_WIDTH_POW2 * bpp;
-            auto src_offset = bpp * (screen_width * y + x);
+            const auto src_offset = bpp * (GSP::FRAMEBUFFER_WIDTH * y + x);
             std::memcpy(dst_ptr + dst_offset, src_ptr + src_offset, bpp);
         }
     }
 
-    Memory::RasterizerFlushVirtualRegion(dst_vaddr, screen_width_pow2 * height * bpp,
+    Memory::RasterizerFlushVirtualRegion(dst_vaddr, GSP::FRAMEBUFFER_WIDTH_POW2 * height * bpp,
                                          Memory::FlushMode::Invalidate);
 }
 
 void AppletManager::CaptureFrameBuffers() {
-    auto gsp =
+    CaptureFrameBuffer(system, capture_info->bottom_screen_left_offset,
-        Core::System::GetInstance().ServiceManager().GetService<Service::GSP::GSP_GPU>("gsp::Gpu");
+                       GSP::FRAMEBUFFER_SAVE_AREA_BOTTOM, GSP::BOTTOM_FRAMEBUFFER_HEIGHT,
-    auto active_thread_id = gsp->GetActiveThreadId();
-    auto top_screen = gsp->GetFrameBufferInfo(active_thread_id, 0);
-    auto bottom_screen = gsp->GetFrameBufferInfo(active_thread_id, 1);
-
-    auto top_fb = top_screen->framebuffer_info[top_screen->index];
-    auto bottom_fb = bottom_screen->framebuffer_info[bottom_screen->index];
-
-    CaptureFrameBuffer(system, capture_info->bottom_screen_left_offset, bottom_fb.address_left, 320,
                        capture_info->bottom_screen_format);
-    CaptureFrameBuffer(system, capture_info->top_screen_left_offset, top_fb.address_left, 400,
+    CaptureFrameBuffer(system, capture_info->top_screen_left_offset,
+                       GSP::FRAMEBUFFER_SAVE_AREA_TOP_LEFT, GSP::TOP_FRAMEBUFFER_HEIGHT,
                        capture_info->top_screen_format);
     if (capture_info->is_3d) {
-        CaptureFrameBuffer(system, capture_info->top_screen_right_offset, top_fb.address_right, 400,
+        CaptureFrameBuffer(system, capture_info->top_screen_right_offset,
+                           GSP::FRAMEBUFFER_SAVE_AREA_TOP_RIGHT, GSP::TOP_FRAMEBUFFER_HEIGHT,
                            capture_info->top_screen_format);
     }
 }

src/core/hle/service/apt/applet_manager.h

@@ -210,6 +210,15 @@ private:
     friend class boost::serialization::access;
 };
 
+enum class DisplayBufferMode : u32_le {
+    R8G8B8A8 = 0,
+    R8G8B8 = 1,
+    R5G6B5 = 2,
+    R5G5B5A1 = 3,
+    R4G4B4A4 = 4,
+    Unimportable = 0xFFFFFFFF,
+};
+
 /// Used by the application to pass information about the current framebuffer to applets.
 struct CaptureBufferInfo {
     u32_le size;
@@ -217,10 +226,10 @@ struct CaptureBufferInfo {
     INSERT_PADDING_BYTES(0x3); // Padding for alignment
     u32_le top_screen_left_offset;
     u32_le top_screen_right_offset;
-    u32_le top_screen_format;
+    DisplayBufferMode top_screen_format;
     u32_le bottom_screen_left_offset;
     u32_le bottom_screen_right_offset;
-    u32_le bottom_screen_format;
+    DisplayBufferMode bottom_screen_format;
 
 private:
     template <class Archive>
@@ -333,16 +342,27 @@ public:
         }
         return buffer;
     }
+    void SetCaptureInfo(std::vector<u8> buffer) {
+        ASSERT_MSG(buffer.size() >= sizeof(CaptureBufferInfo), "CaptureBufferInfo is too small.");
+
+        capture_info.emplace();
+        std::memcpy(&capture_info.get(), buffer.data(), sizeof(CaptureBufferInfo));
+    }
+
     std::vector<u8> ReceiveCaptureBufferInfo() {
-        std::vector<u8> buffer = GetCaptureInfo();
+        std::vector<u8> buffer;
-        capture_info.reset();
+        if (capture_buffer_info) {
+            buffer.resize(sizeof(CaptureBufferInfo));
+            std::memcpy(buffer.data(), &capture_buffer_info.get(), sizeof(CaptureBufferInfo));
+            capture_buffer_info.reset();
+        }
         return buffer;
     }
     void SendCaptureBufferInfo(std::vector<u8> buffer) {
         ASSERT_MSG(buffer.size() >= sizeof(CaptureBufferInfo), "CaptureBufferInfo is too small.");
 
-        capture_info.emplace();
+        capture_buffer_info.emplace();
-        std::memcpy(&capture_info.get(), buffer.data(), sizeof(CaptureBufferInfo));
+        std::memcpy(&capture_buffer_info.get(), buffer.data(), sizeof(CaptureBufferInfo));
     }
 
     ResultCode PrepareToStartApplication(u64 title_id, FS::MediaType media_type);
@@ -395,6 +415,7 @@ private:
     boost::optional<DeliverArg> deliver_arg{};
 
     boost::optional<CaptureBufferInfo> capture_info;
+    boost::optional<CaptureBufferInfo> capture_buffer_info;
 
     static constexpr std::size_t NumAppletSlot = 4;
 
@@ -500,6 +521,8 @@ private:
             ar& delayed_parameter;
             ar& app_start_parameters;
             ar& deliver_arg;
+            ar& capture_info;
+            ar& capture_buffer_info;
             ar& active_slot;
             ar& last_library_launcher_slot;
             ar& last_prepared_library_applet;

src/core/hle/service/gsp/gsp_gpu.cpp

@@ -704,18 +704,73 @@ void GSP_GPU::ImportDisplayCaptureInfo(Kernel::HLERequestContext& ctx) {
     LOG_WARNING(Service_GSP, "called");
 }
 
+static void CopyFrameBuffer(Core::System& system, VAddr dst, VAddr src, u32 stride, u32 lines) {
+    auto dst_ptr = system.Memory().GetPointer(dst);
+    const auto src_ptr = system.Memory().GetPointer(src);
+    if (!dst_ptr || !src_ptr) {
+        LOG_WARNING(Service_GSP,
+                    "Could not resolve pointers for framebuffer capture, skipping screen.");
+        return;
+    }
+
+    Memory::RasterizerFlushVirtualRegion(src, stride * lines, Memory::FlushMode::Flush);
+    std::memcpy(dst_ptr, src_ptr, stride * lines);
+    Memory::RasterizerFlushVirtualRegion(dst, stride * lines, Memory::FlushMode::Invalidate);
+}
+
 void GSP_GPU::SaveVramSysArea(Kernel::HLERequestContext& ctx) {
     IPC::RequestParser rp(ctx);
 
+    if (active_thread_id == std::numeric_limits<u32>::max()) {
+        LOG_WARNING(Service_GSP, "Called without an active thread.");
+
+        // TODO: Find the right error code.
+        IPC::RequestBuilder rb = rp.MakeBuilder(1, 0);
+        rb.Push(-1);
+        return;
+    }
+
     LOG_INFO(Service_GSP, "called");
 
-    // TODO: This should also DMA framebuffers into VRAM and save LCD register state.
+    // TODO: This should also save LCD register state.
     Memory::RasterizerFlushVirtualRegion(Memory::VRAM_VADDR, Memory::VRAM_SIZE,
                                          Memory::FlushMode::Flush);
-    auto vram = system.Memory().GetPointer(Memory::VRAM_VADDR);
+    const auto vram = system.Memory().GetPointer(Memory::VRAM_VADDR);
     saved_vram.emplace(std::vector<u8>(Memory::VRAM_SIZE));
     std::memcpy(saved_vram.get().data(), vram, Memory::VRAM_SIZE);
 
+    const auto top_screen = GetFrameBufferInfo(active_thread_id, 0);
+    if (top_screen) {
+        const auto top_fb = top_screen->framebuffer_info[top_screen->index];
+        if (top_fb.address_left) {
+            CopyFrameBuffer(system, FRAMEBUFFER_SAVE_AREA_TOP_LEFT, top_fb.address_left,
+                            top_fb.stride, TOP_FRAMEBUFFER_HEIGHT);
+        } else {
+            LOG_WARNING(Service_GSP, "No framebuffer bound to top left screen, skipping capture.");
+        }
+        if (top_fb.address_right) {
+            CopyFrameBuffer(system, FRAMEBUFFER_SAVE_AREA_TOP_RIGHT, top_fb.address_right,
+                            top_fb.stride, TOP_FRAMEBUFFER_HEIGHT);
+        } else {
+            LOG_WARNING(Service_GSP, "No framebuffer bound to top right screen, skipping capture.");
+        }
+    } else {
+        LOG_WARNING(Service_GSP, "No top screen bound, skipping capture.");
+    }
+
+    const auto bottom_screen = GetFrameBufferInfo(active_thread_id, 1);
+    if (bottom_screen) {
+        const auto bottom_fb = bottom_screen->framebuffer_info[bottom_screen->index];
+        if (bottom_fb.address_left) {
+            CopyFrameBuffer(system, FRAMEBUFFER_SAVE_AREA_BOTTOM, bottom_fb.address_left,
+                            bottom_fb.stride, BOTTOM_FRAMEBUFFER_HEIGHT);
+        } else {
+            LOG_WARNING(Service_GSP, "No framebuffer bound to bottom screen, skipping capture.");
+        }
+    } else {
+        LOG_WARNING(Service_GSP, "No bottom screen bound, skipping capture.");
+    }
+
     IPC::RequestBuilder rb = rp.MakeBuilder(1, 0);
     rb.Push(RESULT_SUCCESS);
 }
@@ -819,7 +874,7 @@ SessionData* GSP_GPU::FindRegisteredThreadData(u32 thread_id) {
     return nullptr;
 }
 
-GSP_GPU::GSP_GPU(Core::System& system) : ServiceFramework("gsp::Gpu", 2), system(system) {
+GSP_GPU::GSP_GPU(Core::System& system) : ServiceFramework("gsp::Gpu", 4), system(system) {
     static const FunctionInfo functions[] = {
         // clang-format off
         {0x0001, &GSP_GPU::WriteHWRegs, "WriteHWRegs"},

src/core/hle/service/gsp/gsp_gpu.h

@@ -186,6 +186,17 @@ struct CommandBuffer {
 };
 static_assert(sizeof(CommandBuffer) == 0x200, "CommandBuffer struct has incorrect size");
 
+constexpr u32 FRAMEBUFFER_WIDTH = 240;
+constexpr u32 FRAMEBUFFER_WIDTH_POW2 = 256;
+constexpr u32 TOP_FRAMEBUFFER_HEIGHT = 400;
+constexpr u32 BOTTOM_FRAMEBUFFER_HEIGHT = 320;
+constexpr u32 FRAMEBUFFER_HEIGHT_POW2 = 512;
+
+// These are the VRAM addresses that GSP copies framebuffers to in SaveVramSysArea.
+constexpr VAddr FRAMEBUFFER_SAVE_AREA_TOP_LEFT = Memory::VRAM_VADDR + 0x273000;
+constexpr VAddr FRAMEBUFFER_SAVE_AREA_TOP_RIGHT = Memory::VRAM_VADDR + 0x2B9800;
+constexpr VAddr FRAMEBUFFER_SAVE_AREA_BOTTOM = Memory::VRAM_VADDR + 0x4C7800;
+
 class GSP_GPU;
 
 class SessionData : public Kernel::SessionRequestHandler::SessionDataBase {