// t5de-renderer.exe: the renderer process. Started by the adapter (t5deRenderEngine.dll) for each of the client's
// 3D views: receives the client's drawing as commands through a shared memory channel, and draws it with Diligent
// into a child window it places in the client's window (over the right half, for now, to compare with the client's
// own renderer on the left).

#include <windows.h>

#include <d3d11.h>
#include <dxgi.h>
#include <shellapi.h>

#include <cstdint>
#include <cstring>
#include <filesystem>
#include <format>
#include <string>
#include <string_view>
#include <vector>

#include "EngineFactoryD3D11.h"
#include "RefCntAutoPtr.hpp"
#include "RenderDeviceD3D11.h"
#include "log.hpp"
#include "scene.hpp"
#include "t5de/protocol/channel.hpp"

namespace {

using namespace Diligent;
using t5de::protocol::Channel;
using t5de::protocol::View;

using t5de::log::info;

// Diligent's messages (shader compilation errors, invalid states, ...), into the log.
void DILIGENT_CALL_TYPE diligent_message(DEBUG_MESSAGE_SEVERITY severity, const Char* message, const Char* function,
                                         const Char*, int)
{
    static constexpr const char* severities[] = {"info", "warning", "error", "fatal error"};
    info("Diligent {}: {}{}{}", severities[severity < 4 ? severity : 3], function != nullptr ? function : "",
        function != nullptr ? ": " : "", message != nullptr ? message : "");
}

struct Arguments {
    std::wstring channel;
    DWORD parent = 0;
};

Arguments parse_arguments()
{
    Arguments arguments;
    int count = 0;
    wchar_t** argv = CommandLineToArgvW(GetCommandLineW(), &count);
    for (int i = 1; i + 1 < count; i += 2) {
        const std::wstring_view name = argv[i];
        if (name == L"--channel") {
            arguments.channel = argv[i + 1];
        } else if (name == L"--parent") {
            arguments.parent = static_cast<DWORD>(std::wcstoul(argv[i + 1], nullptr, 10));
        }
    }
    LocalFree(argv);
    return arguments;
}

LRESULT CALLBACK window_procedure(HWND window, UINT message, WPARAM wparam, LPARAM lparam)
{
    if (message == WM_ERASEBKGND) {
        return 1; // we draw all of it
    }
    return DefWindowProcW(window, message, wparam, lparam);
}

// A child of the client's window. Disabled, so mouse input goes to the client's window under it.
HWND create_child_window(HWND parent)
{
    WNDCLASSEXW windowClass{.cbSize = sizeof(WNDCLASSEXW),
                            .lpfnWndProc = window_procedure,
                            .hInstance = GetModuleHandleW(nullptr),
                            .hCursor = LoadCursorW(nullptr, IDC_ARROW),
                            .lpszClassName = L"T5DERenderer"};
    RegisterClassExW(&windowClass);
    return CreateWindowExW(0, windowClass.lpszClassName, L"", WS_CHILD | WS_VISIBLE | WS_DISABLED | WS_CLIPSIBLINGS,
                           0, 0, 1, 1, parent, nullptr, windowClass.hInstance, nullptr);
}

// The view, as the adapter set it, or as %TEMP%\t5de-renderer-view.txt says while the client runs (off, split, full
// or compare): for comparing the renderers without restarting the client. Read every second or so.
View current_view(View adapter)
{
    static View view = adapter;
    static ULONGLONG checked = 0;
    if (GetTickCount64() - checked < 1000) {
        return view;
    }
    checked = GetTickCount64();
    view = adapter;
    wchar_t temp[MAX_PATH + 1];
    const DWORD length = GetTempPathW(MAX_PATH + 1, temp);
    if (length == 0 || length > MAX_PATH) {
        return view;
    }
    const HANDLE file = CreateFileW((std::filesystem::path(temp) / L"t5de-renderer-view.txt").c_str(), GENERIC_READ,
                                    FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE, nullptr, OPEN_EXISTING,
                                    FILE_ATTRIBUTE_NORMAL, nullptr);
    if (file == INVALID_HANDLE_VALUE) {
        return view;
    }
    char text[16] = {};
    DWORD read = 0;
    ReadFile(file, text, sizeof(text) - 1, &read, nullptr);
    CloseHandle(file);
    const std::string_view setting(text, std::strcspn(text, "\r\n "));
    view = setting == "off"       ? View::Off
           : setting == "split"   ? View::Split
           : setting == "full"    ? View::Full
           : setting == "compare" ? View::Compare
                                  : adapter;
    return view;
}

// Shader changes to try while the client runs: %TEMP%\t5de-renderer-shaders\fixed_function.hlsl, if there is
// one, replaces the built-in shaders, and is read again when it changes (checked every second or so).
void reload_shaders(t5de::Scene& scene)
{
    static ULONGLONG checked = 0;
    static std::filesystem::file_time_type loaded{};
    if (GetTickCount64() - checked < 1000) {
        return;
    }
    checked = GetTickCount64();
    std::error_code error;
    const auto path = std::filesystem::temp_directory_path(error) / L"t5de-renderer-shaders" / L"fixed_function.hlsl";
    const auto modified = std::filesystem::last_write_time(path, error);
    if (error || modified == loaded) {
        return;
    }
    const HANDLE file = CreateFileW(path.c_str(), GENERIC_READ, FILE_SHARE_READ | FILE_SHARE_WRITE, nullptr,
                                    OPEN_EXISTING, FILE_ATTRIBUTE_NORMAL, nullptr);
    if (file == INVALID_HANDLE_VALUE) {
        return;
    }
    std::string source(static_cast<std::size_t>(std::filesystem::file_size(path, error)), ' ');
    DWORD read = 0;
    ReadFile(file, source.data(), static_cast<DWORD>(source.size()), &read, nullptr);
    CloseHandle(file);
    source.resize(read);
    loaded = modified;
    scene.setShaderSource(std::move(source));
    info("shaders reloaded from {}", path.string());
}

// Where the child window goes in the client's 3D view, and where in the 3D view its scene starts (offset).
struct Placement {
    RECT window{};
    int offset = 0;
};

Placement placement(HWND parent, View view)
{
    Placement place;
    GetClientRect(parent, &place.window);
    if (view == View::Split || view == View::Compare) {
        place.window.left = place.window.right / 2;
    }
    place.offset = view == View::Compare ? 0 : place.window.left;
    if (view == View::Off) {
        place.window = {};
    }
    return place;
}

} // namespace

int WINAPI wWinMain(HINSTANCE, HINSTANCE, PWSTR, int)
{
    const Arguments arguments = parse_arguments();
    if (arguments.channel.empty()) {
        return 2;
    }
    HANDLE mapping = OpenFileMappingW(FILE_MAP_ALL_ACCESS, FALSE, arguments.channel.c_str());
    auto* channel = mapping == nullptr ? nullptr
                                       : static_cast<Channel*>(MapViewOfFile(mapping, FILE_MAP_ALL_ACCESS, 0, 0, 0));
    HANDLE ready = OpenEventW(SYNCHRONIZE, FALSE, (arguments.channel + t5de::protocol::readyEventSuffix).c_str());
    HANDLE consumed =
        OpenEventW(EVENT_MODIFY_STATE, FALSE, (arguments.channel + t5de::protocol::consumedEventSuffix).c_str());
    HANDLE parent = OpenProcess(SYNCHRONIZE, FALSE, arguments.parent);
    if (channel == nullptr || ready == nullptr || consumed == nullptr || parent == nullptr
        || channel->version != t5de::protocol::version) {
        return 3;
    }
    const auto* data = reinterpret_cast<const std::byte*>(channel + 1);

    const HWND clientWindow = reinterpret_cast<HWND>(static_cast<std::uintptr_t>(channel->window));
    const HWND window = create_child_window(clientWindow);
    if (window == nullptr) {
        return 4;
    }

    info("started for window {:#x}, view {}", channel->window, static_cast<int>(channel->view));
    IEngineFactoryD3D11* factory = GetEngineFactoryD3D11();
    factory->SetMessageCallback(diligent_message);
    RefCntAutoPtr<IRenderDevice> device;
    RefCntAutoPtr<IDeviceContext> context;
    EngineD3D11CreateInfo engineInfo;
    factory->CreateDeviceAndContextsD3D11(engineInfo, &device, &context);
    if (!device) {
        info("can't create a Direct3D 11 device");
        return 5;
    }
    // Show each frame as soon as it's drawn: Direct3D 11 otherwise lets 3 frames queue, which the view would trail by.
    if (RefCntAutoPtr<IRenderDeviceD3D11> d3d11{device, IID_RenderDeviceD3D11}) {
        IDXGIDevice1* dxgi = nullptr;
        if (SUCCEEDED(d3d11->GetD3D11Device()->QueryInterface(__uuidof(IDXGIDevice1), reinterpret_cast<void**>(&dxgi)))) {
            dxgi->SetMaximumFrameLatency(1);
            dxgi->Release();
        }
    }
    RefCntAutoPtr<ISwapChain> swapChain;
    SwapChainDesc swapChainDesc;
    swapChainDesc.ColorBufferFormat = TEX_FORMAT_RGBA8_UNORM; // no sRGB conversion, like the client's renderer
    swapChainDesc.DepthBufferFormat = TEX_FORMAT_D32_FLOAT;
    factory->CreateSwapChainD3D11(device, context, swapChainDesc, FullScreenModeDesc{}, Win32NativeWindow{window},
                                  &swapChain);
    if (!swapChain) {
        info("can't create a swap chain");
        return 6;
    }
    t5de::Scene scene(device, context, swapChain);

    // Takes the chunk waiting in the channel (after "ready"), and lets the adapter send the next: appends it to
    // frame, and returns whether it ends a frame.
    std::vector<std::byte> frame;
    const auto take_chunk = [&] {
        const std::uint32_t size = channel->chunkSize <= channel->capacity ? channel->chunkSize : 0;
        const bool endsFrame = (channel->chunkFlags & t5de::protocol::chunkEndsFrame) != 0;
        frame.insert(frame.end(), data, data + size);
        SetEvent(consumed);
        return endsFrame;
    };

    LARGE_INTEGER frequency;
    QueryPerformanceFrequency(&frequency);
    double drawTime = 0;
    double latency = 0;
    int drawn = 0;
    int skipped = 0;
    RECT placed{};
    for (;;) {
        const HANDLE handles[] = {ready, parent};
        const DWORD result = MsgWaitForMultipleObjects(2, handles, FALSE, 1000, QS_ALLINPUT);
        MSG message;
        while (PeekMessageW(&message, nullptr, 0, 0, PM_REMOVE)) {
            DispatchMessageW(&message);
        }
        if (result == WAIT_OBJECT_0 + 1 || channel->closing != 0) {
            break; // the client exited, or the engine is gone
        }
        if (result != WAIT_OBJECT_0 || !take_chunk()) {
            continue;
        }

        // If the client has sent more frames meanwhile (drawing fell behind), catch up: carry out the older ones'
        // changes without drawing them, and draw only the newest. (An incomplete frame stays for the next round.)
        std::size_t complete = frame.size();
        while (WaitForSingleObject(ready, 0) == WAIT_OBJECT_0) {
            bool ended = take_chunk();
            while (!ended && WaitForSingleObject(ready, 1000) == WAIT_OBJECT_0) {
                ended = take_chunk();
            }
            if (!ended) {
                break;
            }
            scene.execute(std::span(frame).first(complete), 0, false);
            frame.erase(frame.begin(), frame.begin() + static_cast<std::ptrdiff_t>(complete));
            complete = frame.size();
            ++skipped;
        }

        // Follow the client's 3D view's size.
        const auto [place, offset] = placement(clientWindow, current_view(channel->view));
        const int width = place.right - place.left;
        const int height = place.bottom - place.top;
        if (std::memcmp(&place, &placed, sizeof(RECT)) != 0) {
            placed = place;
            if (width > 0 && height > 0) {
                SetWindowPos(window, nullptr, place.left, place.top, width, height,
                             SWP_NOZORDER | SWP_NOACTIVATE | SWP_SHOWWINDOW);
                swapChain->Resize(static_cast<Uint32>(width), static_cast<Uint32>(height));
                info("drawing {}x{} at {} in the client's 3D view, its scene from {}", width, height, place.left, offset);
            } else {
                ShowWindow(window, SW_HIDE);
                info("hidden");
            }
        }
        reload_shaders(scene);
        LARGE_INTEGER start, end;
        QueryPerformanceCounter(&start);
        if (scene.execute(std::span(frame).first(complete), offset, width > 0 && height > 0)) {
            swapChain->Present(0);
            QueryPerformanceCounter(&end);
            drawTime += static_cast<double>(end.QuadPart - start.QuadPart) / static_cast<double>(frequency.QuadPart);
            latency += static_cast<double>(end.QuadPart - static_cast<LONGLONG>(scene.lastSentAt())) /
                       static_cast<double>(frequency.QuadPart);
            if (++drawn == 300) {
                info("last 300 frames: {:.1f} ms each to draw and present, {:.1f} ms from the client's end of frame to "
                     "presented, {} skipped to catch up",
                     drawTime * 1000 / drawn, latency * 1000 / drawn, skipped);
                drawTime = 0;
                latency = 0;
                drawn = 0;
                skipped = 0;
            }
        }
        frame.erase(frame.begin(), frame.begin() + static_cast<std::ptrdiff_t>(complete));
    }

    info("exiting");
    DestroyWindow(window);
    return 0;
}