ADOFAIRenderer/Util/FrameReader.cs

using System;
using System.Buffers;
using System.Collections.Generic;
using System.Linq;
using ADOFAIRenderer.Core;
using ADOFAIRenderer.UI;
using FFmpeg.AutoGen;
using MelonLoader;
using Unity.Collections;
using Unity.Collections.LowLevel.Unsafe;
using UnityEngine;
using UnityEngine.Rendering;

namespace ADOFAIRenderer.Util
{
    internal class FrameReader : IDisposable
    {
        private static readonly int Height = Shader.PropertyToID("Height");
        private static readonly int Width = Shader.PropertyToID("Width");
        private static readonly int DestinationBuffer = Shader.PropertyToID("DestinationBuffer");
        private static readonly int SourceTexture = Shader.PropertyToID("SourceTexture");
        private const int InitialBufferSize = 16;
        
        private readonly Stack<ReadbackRequestState> _statePool = new();
        private readonly List<ReadbackRequestState> _allocatedStates = new();
        private readonly object _poolLock = new();
        private bool _isInitialized;
        private volatile bool _isDisposed;

        private readonly ComputeShader? _computeShader;
        private readonly int _kernelIndex = -1;
        private readonly bool _useGPUConversion;
        private readonly AVPixelFormat _expectedFormat;

        public FrameReader(AVPixelFormat targetFormat)
        {
            _computeShader = ModAssets.ColorConversionShader;
            if (_computeShader)
            {
                var kernelName = targetFormat switch
                {
                    AVPixelFormat.AV_PIX_FMT_NV12 => "CSMain_NV12",
                    AVPixelFormat.AV_PIX_FMT_YUV420P => "CSMain_YUV420P",
                    _ => null
                };

                if (kernelName != null)
                {
                    try { _kernelIndex = _computeShader.FindKernel(kernelName); }
                    catch { _kernelIndex = -1; }
                }
            }

            _useGPUConversion = _kernelIndex >= 0;
            _expectedFormat = _useGPUConversion ? targetFormat : AVPixelFormat.AV_PIX_FMT_RGBA;

            InitializePool();
        }

        private void InitializePool()
        {
            lock (_poolLock)
            {
                if (_isInitialized) return;
                for (var i = 0; i < InitialBufferSize; i++)
                {
                    var state = new ReadbackRequestState(this);
                    _statePool.Push(state);
                    _allocatedStates.Add(state);
                }

                _isInitialized = true;
            }
        }

        /// <summary>
        /// Bytes of the readback buffer of one frame: Width x Height x 3 / 2 when the GPU conversion to NV12 / YUV420P
        /// is used, Width x Height x 4 (RGBA) otherwise. The same size as ReadbackRequestState.Setup uses.
        /// </summary>
        public long GetFrameBufferBytes(int width, int height)
        {
            var pixels = (long)width * height;
            return _useGPUConversion ? pixels * 3 / 2 : pixels * 4;
        }

        public void RequestFrameCapture(Texture texture, long currentFrame)
        {
            if (!RenderManager.IsCapturingActive) return;
            ReadbackRequestState state;

            lock (_poolLock)
            {
                if (_statePool.Count == 0)
                {
                    state = new ReadbackRequestState(this);
                    _allocatedStates.Add(state);
                }
                else
                {
                    state = _statePool.Pop();
                }
                
                state.IsInFlight = true;
            }

            var bufferSize = texture.width * texture.height * 4; // RGBA = 4 channel
            state.Setup(currentFrame, bufferSize, _useGPUConversion, _expectedFormat, texture.width, texture.height);

            if (_useGPUConversion && state.GPUBuffer != null && _computeShader)
            {
                _computeShader.SetTexture(_kernelIndex, SourceTexture, texture);
                _computeShader.SetBuffer(_kernelIndex, DestinationBuffer, state.GPUBuffer);
                _computeShader.SetInt(Width, texture.width);
                _computeShader.SetInt(Height, texture.height);

                var threadGroupsX = _expectedFormat == AVPixelFormat.AV_PIX_FMT_NV12
                    ? (texture.width + 31) / 32
                    : (texture.width + 63) / 64;
                var threadGroupsY = (texture.height + 15) / 16;

                _computeShader.Dispatch(_kernelIndex, threadGroupsX, threadGroupsY, 1);

                AsyncGPUReadback.RequestIntoNativeArray(
                    ref state.NativeBuffer, state.GPUBuffer, state.CachedCallback
                );
                /*AsyncGPUReadback.Request(
                    state.GPUBuffer,
                    state.CachedCallback
                );*/
            }
            else
            {
                AsyncGPUReadback.RequestIntoNativeArray(
                    ref state.NativeBuffer, texture, 0, state.CachedCallback
                );
            }
        }

        internal void OnReadbackComplete(ReadbackRequestState state, AsyncGPUReadbackRequest request)
        {
            var frame = state.FrameIndex;

            lock (_poolLock)
            {
                state.IsInFlight = false;
                if (_isDisposed)
                {
                    state.Dispose();
                    return;
                }
            }

            if (RenderManager.Instance && RenderManager.Instance.IsRenderAborted)
            {
                RecycleState(state);
                return;
            }

            if (request.hasError)
            {
                if (RenderManager.Instance)
                {
                    RenderManager.Instance.readbackFailureCount++;
                }
                Melon<RenderMod>.Logger.Warning($"[OOM Prevent] GPU Readback failed for frame {frame}. Queue deadlock prevented by sending a black frame.");

                var blackFrame = CreateBlackFrame(state, out var byteSize);
                var capFrame = new RenderManager.CapturedFrame
                {
                    ManagedBuffer = blackFrame, Length = byteSize,
                    IsColorConverted = state.IsGPUConversion
                };
                PushToQueue(frame, capFrame);
                RecycleState(state);
            }
            else
            {
                unsafe
                {
                    /*
                    if (state.IsGPUConversion)
                    {
                        var data = request.GetData<byte>();
                        // Copy async request buffer data to persistent NativeBuffer memory
                        NativeArray<byte>.Copy(data, state.NativeBuffer, data.Length);
                    }
                    */
                    
                    var ptr = (IntPtr)state.NativeBuffer.GetUnsafeReadOnlyPtr();
                    var capFrame = new RenderManager.CapturedFrame
                    {
                        NativePointer = ptr, Length = state.NativeBuffer.Length,
                        ReadbackState = state, IsColorConverted = state.IsGPUConversion
                    };
                    PushToQueue(frame, capFrame);
                }
            }
        }

        /// <summary>
        /// Creates a black frame in the same format as the frames of this reader, from ArrayPool&lt;byte&gt;.Shared
        /// (returned to the pool by the render thread after encoding).
        ///
        ///   GPU conversion (NV12 / YUV420P, full range): Width × Height × 3 / 2 bytes.
        ///     Y plane (Width × Height bytes) = 0, U/V data (the rest) = 128.
        ///     128 is the zero chroma value of the 8-bit full range: the U/V offset 0.5 of RGB2YUV_BT709_Full
        ///     in ColorConversion.compute, which the encoder input of this path uses.
        ///     Both NV12 (interleaved UV) and YUV420P (U plane, then V plane) store all chroma after the Y plane,
        ///     so the same fill applies to both.
        ///   No GPU conversion (RGBA input, converted by the encoder): Width × Height × 4 bytes, all 0.
        /// </summary>
        private static byte[] CreateBlackFrame(ReadbackRequestState state, out int byteSize)
        {
            var lumaSize = state.Width * state.Height;
            byteSize = state.IsGPUConversion ? lumaSize * 3 / 2 : lumaSize * 4;

            var frame = ArrayPool<byte>.Shared.Rent(byteSize);
            if (state.IsGPUConversion)
            {
                Array.Clear(frame, 0, lumaSize);
                frame.AsSpan(lumaSize, byteSize - lumaSize).Fill(128);
            }
            else
            {
                Array.Clear(frame, 0, byteSize);
            }
            return frame;
        }

        internal void RecycleState(ReadbackRequestState state)
        {
            lock (_poolLock)
            {
                if (_isDisposed || !_isInitialized)
                {
                    state.Dispose();
                    return;
                }
                
                state.Recycle();
                _statePool.Push(state);
            }
        }
        
        private static void PushToQueue(long frame, RenderManager.CapturedFrame capturedFrame)
        {
            if (!RenderManager.Instance) return;
            lock (RenderManager.Instance.ImageQueue)
            {
                RenderManager.Instance.ImageQueue[frame] = capturedFrame;
            }
            RenderManager.Instance.FrameAvailableEvent.Set();
        }
        
        public void Dispose()
        {
            lock (_poolLock)
            {
                if (_isDisposed) return;
                _isDisposed = true;
                foreach (var state in _allocatedStates.Where(state => !state.IsInFlight))
                {
                    state.Dispose();
                }
                _statePool.Clear();
                _allocatedStates.Clear();
                _isInitialized = false;
            }
        }
    }

    internal class ReadbackRequestState : IDisposable
    {
        private readonly FrameReader _owner;
        public long FrameIndex { get; private set; }
        public NativeArray<byte> NativeBuffer;
        public bool IsInFlight { get; internal set; }

        public ComputeBuffer? GPUBuffer;
        public bool IsGPUConversion { get; private set; }
        public int Width { get; private set; }
        public int Height { get; private set; }

        public readonly Action<AsyncGPUReadbackRequest> CachedCallback;

        public ReadbackRequestState(FrameReader owner)
        {
            _owner = owner;
            CachedCallback = OnComplete;
            FrameIndex = -1;
        }

        public void Setup(long frameIndex, int bufferSize, bool isGPUConversion, AVPixelFormat pixelFormat, int width, int height)
        {
            FrameIndex = frameIndex;
            IsGPUConversion = isGPUConversion;
            Width = width; Height = height;

            var readbackSize = isGPUConversion ? (width * height * 3 / 2) : bufferSize;
            if (!NativeBuffer.IsCreated || NativeBuffer.Length != readbackSize)
            {
                if (NativeBuffer.IsCreated) NativeBuffer.Dispose();

                NativeBuffer = new NativeArray<byte>(readbackSize, Allocator.Persistent,
                    NativeArrayOptions.UninitializedMemory);
            }

            if (isGPUConversion)
            {
                var computeBufferSize = readbackSize / 4;
                if (GPUBuffer != null && GPUBuffer.count == computeBufferSize) return;
                
                GPUBuffer?.Dispose();
                GPUBuffer = new ComputeBuffer(computeBufferSize, 4, ComputeBufferType.Raw);
            }
            else
            {
                GPUBuffer?.Dispose();
                GPUBuffer = null;
            }
        }

        private void OnComplete(AsyncGPUReadbackRequest request)
        {
            _owner.OnReadbackComplete(this, request);
        }

        public void Recycle()
        {
            FrameIndex = -1;
        }

        public void Dispose()
        {
            if (NativeBuffer.IsCreated) NativeBuffer.Dispose();
            GPUBuffer?.Dispose();
            GPUBuffer = null;
        }
    }
}