Add Compute Shader Color Conversion

This commit is contained in:
copecone committed 2026-06-16 11:44:34 +09:00
1 parent 7df15eae08
commit b23d168ed7
7 files changed
+369 -32

No files matched your search

+147 -22
View File
@@ -3,6 +3,8 @@ using System.Buffers;
using System.Collections.Generic;
using System.Linq;
using ADOFAIRenderer.Core;
using ADOFAIRenderer.Util.FFmpeg;
using FFmpeg.AutoGen;
using MelonLoader;
using Unity.Collections;
using Unity.Collections.LowLevel.Unsafe;
@@ -13,6 +15,10 @@ namespace ADOFAIRenderer.Util
{
public 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();
@@ -42,6 +48,24 @@ namespace ADOFAIRenderer.Util
}
}
private static unsafe AVPixelFormat PredictPixelFormat(Codec codec)
{
var candidates = GPUUtil.GetPrioritizedEncoderCandidates(codec);
// ReSharper disable once ForeachCanBePartlyConvertedToQueryUsingAnotherGetEnumerator
foreach (var encoderName in candidates)
{
var pCodec = ffmpeg.avcodec_find_encoder_by_name(encoderName);
if (pCodec == null) continue;
if (encoderName.Contains("qsv") || encoderName.Contains("nvenc") || encoderName.Contains("amf"))
{
return AVPixelFormat.AV_PIX_FMT_NV12;
}
return encoderName == "libvvenc" ? AVPixelFormat.AV_PIX_FMT_YUV420P10LE : AVPixelFormat.AV_PIX_FMT_YUV420P;
}
return AVPixelFormat.AV_PIX_FMT_YUV420P;
}
public void RequestFrameCapture(Texture texture, long currentFrame)
{
if (!RenderManager.IsCapturingActive) return;
@@ -62,15 +86,74 @@ namespace ADOFAIRenderer.Util
state.IsInFlight = true;
}
var bufferSize = texture.width * texture.height * 4; // RGBA = 4 channel
state.Setup(currentFrame, bufferSize);
// GPU 컬러 변환 활성화 여부 검사
bool useGPUConversion = RenderMod.ColorConversionShader;
AVPixelFormat expectedFmt = AVPixelFormat.AV_PIX_FMT_RGBA;
if (useGPUConversion && RenderManager.Instance)
{
expectedFmt = PredictPixelFormat(RenderManager.Instance.Configuration.Codec);
// Compute Shader는 NV12와 YUV420P만 지원하므로 그 외 포맷은 CPU 폴백 처리
if (expectedFmt != AVPixelFormat.AV_PIX_FMT_NV12 && expectedFmt != AVPixelFormat.AV_PIX_FMT_YUV420P)
{
useGPUConversion = false;
}
}
AsyncGPUReadback.RequestIntoNativeArray(
ref state.NativeBuffer,
texture,
0,
state.CachedCallback
);
var bufferSize = texture.width * texture.height * 4; // RGBA = 4 channel
state.Setup(currentFrame, bufferSize, useGPUConversion, expectedFmt, texture.width, texture.height);
if (useGPUConversion && state.GPUBuffer != null && RenderMod.ColorConversionShader)
{
int kernelIndex = -1;
if (expectedFmt == AVPixelFormat.AV_PIX_FMT_NV12)
{
kernelIndex = RenderMod.ColorConversionShader.FindKernel("CSMain_NV12");
}
else if (expectedFmt == AVPixelFormat.AV_PIX_FMT_YUV420P)
{
kernelIndex = RenderMod.ColorConversionShader.FindKernel("CSMain_YUV420P");
}
if (kernelIndex >= 0)
{
RenderMod.ColorConversionShader.SetTexture(kernelIndex, SourceTexture, texture);
RenderMod.ColorConversionShader.SetBuffer(kernelIndex, DestinationBuffer, state.GPUBuffer);
RenderMod.ColorConversionShader.SetInt(Width, texture.width);
RenderMod.ColorConversionShader.SetInt(Height, texture.height);
var threadGroupsX = expectedFmt == AVPixelFormat.AV_PIX_FMT_NV12
? (texture.width + 31) / 32
: (texture.width + 63) / 64;
var threadGroupsY = (texture.height + 15) / 16;
RenderMod.ColorConversionShader.Dispatch(kernelIndex, threadGroupsX, threadGroupsY, 1);
AsyncGPUReadback.Request(
state.GPUBuffer,
state.CachedCallback
);
}
else
{
// 커널을 못 찾은 경우 기존 RGBA 텍스처 백업 기법으로 폴백
AsyncGPUReadback.RequestIntoNativeArray(
ref state.NativeBuffer,
texture,
0,
state.CachedCallback
);
}
}
else
{
AsyncGPUReadback.RequestIntoNativeArray(
ref state.NativeBuffer,
texture,
0,
state.CachedCallback
);
}
}
internal void OnReadbackComplete(ReadbackRequestState state, AsyncGPUReadbackRequest request)
@@ -101,7 +184,9 @@ namespace ADOFAIRenderer.Util
}
Melon<RenderMod>.Logger.Warning($"[OOM Prevent] GPU Readback failed for frame {frame}. Queue deadlock prevented by sending a black frame.");
var byteSize = state.NativeBuffer.Length;
var byteSize = state.IsGPUConversion
? state.Width * state.Height * 3 / 2
: state.Width * state.Height * 4;
var blackFrame = ArrayPool<byte>.Shared.Rent(byteSize);
Array.Clear(blackFrame, 0, byteSize);
@@ -117,14 +202,32 @@ namespace ADOFAIRenderer.Util
{
unsafe
{
var ptr = (IntPtr)state.NativeBuffer.GetUnsafeReadOnlyPtr();
var capFrame = new RenderManager.CapturedFrame
if (state.IsGPUConversion)
{
NativePointer = ptr,
Length = state.NativeBuffer.Length,
ReadbackState = state
};
PushToQueue(frame, capFrame);
var data = request.GetData<byte>();
// 비동기 요청 버퍼 데이터를 지속성(NativeBuffer) 메모리로 복사
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
};
PushToQueue(frame, capFrame);
}
else
{
var ptr = (IntPtr)state.NativeBuffer.GetUnsafeReadOnlyPtr();
var capFrame = new RenderManager.CapturedFrame
{
NativePointer = ptr,
Length = state.NativeBuffer.Length,
ReadbackState = state
};
PushToQueue(frame, capFrame);
}
}
}
}
@@ -178,6 +281,11 @@ namespace ADOFAIRenderer.Util
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)
@@ -187,19 +295,34 @@ namespace ADOFAIRenderer.Util
FrameIndex = -1;
}
public void Setup(long frameIndex, int bufferSize)
public void Setup(long frameIndex, int bufferSize, bool isGPUConversion, AVPixelFormat pixelFormat, int width, int height)
{
FrameIndex = frameIndex;
// 버퍼가 생성되지 않았거나 크기가 다른 경우 재생성
if (!NativeBuffer.IsCreated || NativeBuffer.Length != bufferSize)
IsGPUConversion = isGPUConversion;
Width = width; Height = height;
var readbackSize = isGPUConversion ? (width * height * 3 / 2) : bufferSize;
if (!NativeBuffer.IsCreated || NativeBuffer.Length != readbackSize)
{
// 이전 메모리 해제
if (NativeBuffer.IsCreated) NativeBuffer.Dispose();
NativeBuffer = new NativeArray<byte>(bufferSize, Allocator.Persistent,
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)
@@ -215,6 +338,8 @@ namespace ADOFAIRenderer.Util
public void Dispose()
{
if (NativeBuffer.IsCreated) NativeBuffer.Dispose();
GPUBuffer?.Dispose();
GPUBuffer = null;
}
}
}