using System.Diagnostics;
using Microsoft.Xna.Framework;
using Microsoft.Xna.Framework.Graphics;

namespace MrGameEng.Graphics;

/// <summary>
/// The engine's 2D renderer: a sprite batcher over dynamic vertex buffers.
/// Per frame: <see cref="BeginFrame"/> (camera) → <see cref="Submit"/> per sprite (with culling)
/// → <see cref="EndFrame"/> (stable sort layer → depth → texture, build vertices, issue draw calls).
/// Above <see cref="Renderer2DOptions.ParallelThreshold"/> sprites, submission and vertex
/// building run on all cores; the vertex buffer is ring-written (NoOverwrite) to avoid GPU stalls.
/// Registered as a service; scenes attach it via <c>scene.UseRenderer2D()</c>.
/// </summary>
public sealed class Renderer2D : IDisposable
{
    private const int MaxQuadsPerDraw = 8192;
    private const int ParallelBlock = 4096;

    private static readonly int VertexStride = VertexPositionColorTexture.VertexDeclaration.VertexStride;

    /// <summary>Render layer registry. Register layers before the first frame.</summary>
    public LayerRegistry Layers { get; } = new();

    /// <summary>Camera state of the current frame. Valid between BeginFrame and the next BeginFrame.</summary>
    public CameraState Camera { get; private set; }

    /// <summary>Draw calls issued by the last <see cref="EndFrame"/>.</summary>
    public int DrawCalls { get; private set; }

    /// <summary>Sprites accepted by <see cref="Submit"/> this frame.</summary>
    public int SubmittedSprites { get; private set; }

    /// <summary>Sprites rejected by culling this frame.</summary>
    public int CulledSprites { get; private set; }

    /// <summary>Milliseconds spent submitting sprites (BeginFrame → EndFrame) last frame.</summary>
    public float SubmitMs { get; private set; }

    /// <summary>Milliseconds spent sorting last frame.</summary>
    public float SortMs { get; private set; }

    /// <summary>Milliseconds spent building vertices last frame.</summary>
    public float BuildMs { get; private set; }

    /// <summary>Milliseconds spent uploading vertices to the GPU last frame.</summary>
    public float UploadMs { get; private set; }

    /// <summary>Milliseconds spent issuing draw calls last frame.</summary>
    public float DrawMs { get; private set; }

    private readonly GraphicsDevice _device;
    private readonly Renderer2DOptions _options;
    private readonly SpriteBatcher _batcher;
    private readonly BasicEffect _effect;
    private readonly IndexBuffer _indexBuffer;
    private DynamicVertexBuffer _vertexBuffer;
    private VertexPositionColorTexture[] _vertices;
    private CameraState _screenCamera;
    private bool _begun;
    private long _submitStartTimestamp;
    private int _ringCursor;
    private int _ringBaseVertex;

    /// <summary>Creates the renderer. One instance per game is enough.</summary>
    public Renderer2D(GraphicsDevice device, Renderer2DOptions? options = null)
    {
        _device = device;
        _options = options ?? new Renderer2DOptions();
        ArgumentOutOfRangeException.ThrowIfLessThan(_options.InitialCapacity, 1, nameof(options));
        _batcher = new SpriteBatcher(_options.InitialCapacity);
        _vertices = new VertexPositionColorTexture[_options.InitialCapacity * 4];
        _vertexBuffer = new DynamicVertexBuffer(
            device, VertexPositionColorTexture.VertexDeclaration, _vertices.Length * 2, BufferUsage.WriteOnly);

        _effect = new BasicEffect(device)
        {
            TextureEnabled = true,
            VertexColorEnabled = true,
            World = Matrix.Identity,
        };

        _indexBuffer = CreateQuadIndexBuffer(device);
    }

    internal int ParallelThreshold => _options.ParallelThreshold;

    /// <summary>Begins a frame with the given camera. Called by <see cref="CameraSystem"/>.</summary>
    public void BeginFrame(in Camera camera)
    {
        var (virtualW, virtualH, mapping) = ResolveVirtualResolution();
        Camera = CameraMath.Compute(camera, virtualW, virtualH, mapping);
        _screenCamera = CameraMath.Compute(
            new Camera(new Vector2(virtualW / 2f, virtualH / 2f)), virtualW, virtualH, mapping);

        _batcher.Clear();
        SubmittedSprites = 0;
        CulledSprites = 0;
        _begun = true;
        _submitStartTimestamp = Stopwatch.GetTimestamp();
    }

    /// <summary>
    /// Begins a frame with a default camera that shows the world origin at the top-left
    /// corner of the screen. Used when the scene has no camera entity.
    /// </summary>
    public void BeginFrameWithDefaultCamera()
    {
        var (virtualW, virtualH, _) = ResolveVirtualResolution();
        var camera = new Camera(new Vector2(virtualW / 2f, virtualH / 2f));
        BeginFrame(in camera);
    }

    /// <summary>Submits one sprite. Invisible sprites (outside the camera) are culled here.</summary>
    public void Submit(in Transform2D transform, in Sprite sprite)
    {
        EnsureBegun();
        switch (TryBuildInstance(in transform, in sprite, out var instance, out var key))
        {
            case SubmitResult.Visible:
                _batcher.Submit(in instance, key);
                SubmittedSprites++;
                break;
            case SubmitResult.Culled:
                CulledSprites++;
                break;
        }
    }

    /// <summary>Sorts, builds vertices and issues draw calls. Called by <see cref="RenderFlushSystem"/>.</summary>
    public void EndFrame()
    {
        EnsureBegun();
        _begun = false;
        DrawCalls = 0;

        var submitEnd = Stopwatch.GetTimestamp();
        SubmitMs = ToMs(submitEnd - _submitStartTimestamp);
        SortMs = 0f;
        BuildMs = 0f;
        UploadMs = 0f;
        DrawMs = 0f;

        _batcher.Sort();
        var count = _batcher.Count;
        var sortEnd = Stopwatch.GetTimestamp();
        SortMs = ToMs(sortEnd - submitEnd);
        if (count == 0)
        {
            return;
        }

        var order = _batcher.SortedOrder;
        EnsureVertexCapacity(count * 4);
        if (count >= _options.ParallelThreshold)
        {
            var blocks = (count + ParallelBlock - 1) / ParallelBlock;
            Parallel.For(0, blocks, block =>
            {
                var end = Math.Min((block + 1) * ParallelBlock, count);
                for (var i = block * ParallelBlock; i < end; i++)
                {
                    BuildVertex(order, i);
                }
            });
        }
        else
        {
            for (var i = 0; i < count; i++)
            {
                BuildVertex(order, i);
            }
        }

        var buildEnd = Stopwatch.GetTimestamp();
        BuildMs = ToMs(buildEnd - sortEnd);

        // Кольцевая запись: NoOverwrite не заставляет GPU ждать предыдущий кадр;
        // Discard только на перемотке кольца.
        var vertexCount = count * 4;
        SetDataOptions hint;
        if (_ringCursor + vertexCount <= _vertexBuffer.VertexCount)
        {
            _ringBaseVertex = _ringCursor;
            hint = SetDataOptions.NoOverwrite;
        }
        else
        {
            _ringBaseVertex = 0;
            hint = SetDataOptions.Discard;
        }

        _vertexBuffer.SetData(_ringBaseVertex * VertexStride, _vertices, 0, vertexCount, VertexStride, hint);
        _ringCursor = _ringBaseVertex + vertexCount;
        var uploadEnd = Stopwatch.GetTimestamp();
        UploadMs = ToMs(uploadEnd - buildEnd);

        _device.BlendState = BlendState.AlphaBlend;
        _device.SamplerStates[0] = _options.Sampler;
        _device.DepthStencilState = DepthStencilState.None;
        _device.RasterizerState = RasterizerState.CullNone;
        _device.SetVertexBuffer(_vertexBuffer);
        _device.Indices = _indexBuffer;

        // Виртуальное разрешение рисуется в letterbox-прямоугольник — тот же, по которому
        // считают ScreenToWorld/WorldToScreen; иначе картинка растягивается мимо маппинга.
        var previousViewport = _device.Viewport;
        var letterboxed = _options.VirtualResolution is not null;
        if (letterboxed)
        {
            var mapping = Camera.Mapping;
            _device.Viewport = new Viewport(
                (int)MathF.Round(mapping.Offset.X),
                (int)MathF.Round(mapping.Offset.Y),
                (int)MathF.Round(Camera.VirtualWidth * mapping.Scale),
                (int)MathF.Round(Camera.VirtualHeight * mapping.Scale));
        }

        DrawBatches(order, count);
        if (letterboxed)
        {
            _device.Viewport = previousViewport;
        }

        DrawMs = ToMs(Stopwatch.GetTimestamp() - uploadEnd);
    }

    /// <summary>Converts a physical screen point to world coordinates using the current camera.</summary>
    public Vector2 ScreenToWorld(Vector2 screen) => Camera.ScreenToWorld(screen);

    /// <summary>Converts a world point to physical screen coordinates using the current camera.</summary>
    public Vector2 WorldToScreen(Vector2 world) => Camera.WorldToScreen(world);

    /// <inheritdoc />
    public void Dispose()
    {
        _effect.Dispose();
        _vertexBuffer.Dispose();
        _indexBuffer.Dispose();
    }

    // --- Параллельная по-чанковая подача (используется SpriteRenderSystem выше порога) ----

    internal void BeginChunkedSubmit(ReadOnlySpan<int> chunkLengths)
    {
        EnsureBegun();
        _batcher.BeginChunks(chunkLengths);
    }

    internal SpriteChunkWriter GetChunkWriter(int chunkIndex) => _batcher.GetChunkWriter(chunkIndex);

    internal void SubmitInto(ref SpriteChunkWriter writer, in Transform2D transform, in Sprite sprite)
    {
        switch (TryBuildInstance(in transform, in sprite, out var instance, out var key))
        {
            case SubmitResult.Visible:
                writer.Add(in instance, key);
                break;
            case SubmitResult.Culled:
                writer.AddCulled();
                break;
        }
    }

    internal void EndChunk(int chunkIndex, in SpriteChunkWriter writer) => _batcher.EndChunk(chunkIndex, in writer);

    internal void CommitChunkedSubmit()
    {
        SubmittedSprites += _batcher.CommitChunks();
        CulledSprites += _batcher.LastChunkCulled;
    }

    private enum SubmitResult
    {
        Skipped,
        Culled,
        Visible,
    }

    private SubmitResult TryBuildInstance(
        in Transform2D transform, in Sprite sprite, out SpriteInstance instance, out ulong key)
    {
        instance = default;
        key = 0;
        if (sprite.Region is not { } region)
        {
            return SubmitResult.Skipped;
        }

        var layer = Layers[sprite.Layer];
        var (center, radius) = CullingMath.SpriteBoundingCircle(in transform, region, sprite.Origin);

        if (layer.Space == LayerSpace.World &&
            !CullingMath.CircleIntersectsRect(center, radius, Camera.CullRect))
        {
            return SubmitResult.Culled;
        }

        // Y-sort по пивоту (Transform2D.Position), а не по центру квада: спрайты разной
        // высоты с origin «в ногах» сортируются по ногам, как принято в top-down.
        var depth = layer.SortMode == LayerSortMode.YSort ? transform.Position.Y : sprite.Depth;

        instance = new SpriteInstance
        {
            Region = region,
            Center = center,
            HalfSize = new Vector2(region.Width * transform.Scale.X, region.Height * transform.Scale.Y) / 2f,
            Rotation = transform.Rotation,
            Color = sprite.Color,
            Flip = sprite.Flip,
            Layer = sprite.Layer.Value,
        };
        key = SpriteSortKey.Make(sprite.Layer.Value, depth, region.TextureSortKey);
        return SubmitResult.Visible;
    }

    private void EnsureBegun()
    {
        if (!_begun)
        {
            throw new InvalidOperationException(
                "Renderer used outside BeginFrame/EndFrame (is CameraSystem registered first?).");
        }
    }

    private (int Width, int Height, ViewportMapping Mapping) ResolveVirtualResolution()
    {
        var viewport = _device.Viewport;
        if (_options.VirtualResolution is not { } virtualSize)
        {
            return (viewport.Width, viewport.Height, ViewportMapping.Identity);
        }

        return (virtualSize.X, virtualSize.Y,
            CameraMath.ComputeMapping(viewport.Width, viewport.Height, virtualSize.X, virtualSize.Y));
    }

    private void BuildVertex(int[] order, int i)
    {
        ref readonly var instance = ref _batcher[order[i]];
        var region = instance.Region;

        var u0 = region.U0;
        var v0 = region.V0;
        var u1 = region.U1;
        var v1 = region.V1;

        if ((instance.Flip & SpriteFlip.X) != 0)
        {
            (u0, u1) = (u1, u0);
        }

        if ((instance.Flip & SpriteFlip.Y) != 0)
        {
            (v0, v1) = (v1, v0);
        }

        Vector2 rx, ry;
        if (instance.Rotation == 0f)
        {
            rx = new Vector2(instance.HalfSize.X, 0f);
            ry = new Vector2(0f, instance.HalfSize.Y);
        }
        else
        {
            var (sin, cos) = MathF.SinCos(instance.Rotation);
            rx = new Vector2(instance.HalfSize.X * cos, instance.HalfSize.X * sin);
            ry = new Vector2(-instance.HalfSize.Y * sin, instance.HalfSize.Y * cos);
        }

        var center = instance.Center;
        var vertex = i * 4;
        _vertices[vertex + 0] = Vertex(center - rx - ry, instance.Color, u0, v0);
        _vertices[vertex + 1] = Vertex(center + rx - ry, instance.Color, u1, v0);
        _vertices[vertex + 2] = Vertex(center - rx + ry, instance.Color, u0, v1);
        _vertices[vertex + 3] = Vertex(center + rx + ry, instance.Color, u1, v1);
    }

    private void DrawBatches(int[] order, int count)
    {
        var batchStart = 0;
        ref readonly var first = ref _batcher[order[0]];
        var currentTexture = first.Region.Texture;
        var currentLayer = first.Layer;
        ApplyLayerMatrices(currentLayer);

        for (var i = 1; i <= count; i++)
        {
            Texture2D? texture = null;
            byte layer = 0;
            if (i < count)
            {
                ref readonly var instance = ref _batcher[order[i]];
                texture = instance.Region.Texture;
                layer = instance.Layer;
                if (ReferenceEquals(texture, currentTexture) && layer == currentLayer)
                {
                    continue;
                }
            }

            DrawRange(currentTexture, batchStart, i - batchStart);
            batchStart = i;

            if (i < count)
            {
                currentTexture = texture!;
                if (layer != currentLayer)
                {
                    currentLayer = layer;
                    ApplyLayerMatrices(currentLayer);
                }
            }
        }
    }

    private void ApplyLayerMatrices(byte layer)
    {
        var state = Layers[new LayerId(layer)].Space == LayerSpace.Screen ? _screenCamera : Camera;
        _effect.View = state.View;
        _effect.Projection = state.Projection;
    }

    private void DrawRange(Texture2D texture, int firstQuad, int quadCount)
    {
        _effect.Texture = texture;

        while (quadCount > 0)
        {
            var quads = Math.Min(quadCount, MaxQuadsPerDraw);
            foreach (var pass in _effect.CurrentTechnique.Passes)
            {
                pass.Apply();
                _device.DrawIndexedPrimitives(
                    PrimitiveType.TriangleList, _ringBaseVertex + firstQuad * 4, 0, quads * 2);
                DrawCalls++;
            }

            firstQuad += quads;
            quadCount -= quads;
        }
    }

    private void EnsureVertexCapacity(int vertexCount)
    {
        if (_vertices.Length < vertexCount)
        {
            var capacity = _vertices.Length;
            while (capacity < vertexCount)
            {
                capacity *= 2;
            }

            _vertices = new VertexPositionColorTexture[capacity];
        }

        // GPU-буфер держим вдвое больше CPU-массива — кольцу нужен запас,
        // чтобы NoOverwrite срабатывал чаще, чем Discard.
        var wantedBuffer = _vertices.Length * 2;
        if (_vertexBuffer.VertexCount < wantedBuffer)
        {
            _vertexBuffer.Dispose();
            _vertexBuffer = new DynamicVertexBuffer(
                _device, VertexPositionColorTexture.VertexDeclaration, wantedBuffer, BufferUsage.WriteOnly);
            _ringCursor = 0;
        }
    }

    private static float ToMs(long timestampDelta) => (float)timestampDelta * 1000f / Stopwatch.Frequency;

    private static VertexPositionColorTexture Vertex(Vector2 position, Color color, float u, float v) =>
        new(new Vector3(position, 0f), color, new Vector2(u, v));

    private static IndexBuffer CreateQuadIndexBuffer(GraphicsDevice device)
    {
        var indices = new ushort[MaxQuadsPerDraw * 6];
        for (var quad = 0; quad < MaxQuadsPerDraw; quad++)
        {
            var vertex = quad * 4;
            var index = quad * 6;
            indices[index + 0] = (ushort)(vertex + 0);
            indices[index + 1] = (ushort)(vertex + 1);
            indices[index + 2] = (ushort)(vertex + 2);
            indices[index + 3] = (ushort)(vertex + 2);
            indices[index + 4] = (ushort)(vertex + 1);
            indices[index + 5] = (ushort)(vertex + 3);
        }

        var buffer = new IndexBuffer(device, IndexElementSize.SixteenBits, indices.Length, BufferUsage.WriteOnly);
        buffer.SetData(indices);
        return buffer;
    }
}