Prefer QSV System Memory Input With GPU Frame Pool Fallback

Util/FFmpeg/FFmpegNativeDecoder.cs

@@ -206,6 +206,12 @@ internal sealed unsafe class FFmpegNativeDecoder : IDisposable
     private int _srcWidth;
     private int _srcHeight;
 
+    /// <summary>YUV→RGB matrix id (<c>SWS_CS_*</c>) applied to the context. 0 when no matrix is set (non-RGBA output).</summary>
+    private int _srcSwsColorspace;
+
+    /// <summary>Whether the source range applied to the context is the full range.</summary>
+    private bool _srcFullRange;
+
     // ──────────────────────────────────────────────────────────────
     // Public Properties
     // ──────────────────────────────────────────────────────────────
@@ -296,6 +302,59 @@ internal sealed unsafe class FFmpegNativeDecoder : IDisposable
         }
     }
 
+    /// <summary>
+    /// Color matrix id used to convert the YUV frames of this stream to RGB, selected from the colorspace tag of the stream.
+    /// The ids are shared with the <c>ColorMatrix</c> uniform of ColorConversion.compute:
+    /// 0 = BT.709 (also used when the stream has no usable tag), 1 = BT.601, 2 = BT.2020 (non-constant luminance).
+    /// </summary>
+    public int ColorMatrix
+    {
+        get
+        {
+            lock (_lock)
+            {
+                if (_pCodecContext != null && _pCodecContext->colorspace != AVColorSpace.AVCOL_SPC_UNSPECIFIED)
+                {
+                    return GetColorMatrixId(_pCodecContext->colorspace);
+                }
+                if (_pFormatContext != null && _videoStreamIndex >= 0)
+                {
+                    return GetColorMatrixId(_pFormatContext->streams[_videoStreamIndex]->codecpar->color_space);
+                }
+                return ColorMatrixBT709;
+            }
+        }
+    }
+
+    /// <summary>Color matrix id of BT.709.</summary>
+    // ReSharper disable InconsistentNaming
+    public const int ColorMatrixBT709 = 0;
+
+    /// <summary>Color matrix id of BT.601 (SMPTE 170M / BT.470 BG).</summary>
+    public const int ColorMatrixBT601 = 1;
+
+    /// <summary>Color matrix id of BT.2020 (non-constant luminance).</summary>
+    public const int ColorMatrixBT2020 = 2;
+    // ReSharper restore InconsistentNaming
+
+    /// <summary>
+    /// Converts a colorspace tag to a color matrix id (see <see cref="ColorMatrix"/>).
+    /// Tags that have no matrix of their own (unspecified, RGB, other standards) use BT.709.
+    /// </summary>
+    private static int GetColorMatrixId(AVColorSpace colorspace)
+    {
+        switch (colorspace)
+        {
+            case AVColorSpace.AVCOL_SPC_SMPTE170M:
+            case AVColorSpace.AVCOL_SPC_BT470BG:
+                return ColorMatrixBT601;
+            case AVColorSpace.AVCOL_SPC_BT2020_NCL:
+                return ColorMatrixBT2020;
+            default:
+                return ColorMatrixBT709;
+        }
+    }
+
     /// <summary>
     /// Gets a value indicating whether hardware-accelerated decoding is currently active.
     /// </summary>
@@ -778,12 +837,67 @@ internal sealed unsafe class FFmpegNativeDecoder : IDisposable
     }
 
     /// <summary>
-    /// Lazily initializes or updates the <see cref="SwsContext"/> when source/destination formats or dimensions change.
+    /// Selects the YUV→RGB conversion matrix (libswscale <c>SWS_CS_*</c>) from the colorspace tag of a frame.
+    /// The matrix set is the same as that of the compute shader path (ColorConversion.compute, <c>ColorMatrix</c> uniform):
+    /// BT.709 (also for frames without a usable tag), BT.601 and BT.2020, so both paths give the same colors.
+    /// Reference: https://ffmpeg.org/doxygen/trunk/group__libsws.html (sws_getCoefficients)
+    /// </summary>
+    private static int GetSwsColorspace(AVColorSpace colorspace)
+    {
+        switch (GetColorMatrixId(colorspace))
+        {
+            case ColorMatrixBT601:
+                return ffmpeg.SWS_CS_ITU601;
+            case ColorMatrixBT2020:
+                return ffmpeg.SWS_CS_BT2020;
+            default:
+                return ffmpeg.SWS_CS_ITU709;
+        }
+    }
+
+    /// <summary>
+    /// Returns true when the Y/U/V samples of the frame use the full range (0-255) instead of the limited range (16-235 / 16-240).
+    /// The range is given by the frame's color_range tag or by a deprecated full range pixel format (YUVJ*).
+    /// </summary>
+    private static bool IsFullRange(AVFrame* frame)
+    {
+        if (frame->color_range == AVColorRange.AVCOL_RANGE_JPEG) return true;
+
+        switch ((AVPixelFormat)frame->format)
+        {
+            case AVPixelFormat.AV_PIX_FMT_YUVJ420P:
+            case AVPixelFormat.AV_PIX_FMT_YUVJ422P:
+            case AVPixelFormat.AV_PIX_FMT_YUVJ444P:
+            case AVPixelFormat.AV_PIX_FMT_YUVJ440P:
+                return true;
+            default:
+                return false;
+        }
+    }
+
+    /// <summary>
+    /// Lazily initializes or updates the <see cref="SwsContext"/> when source/destination formats, dimensions,
+    /// or (for RGBA output) the colorspace / range of the source frame change.
+    /// For RGBA output the YUV→RGB matrix and the source range are taken from <paramref name="colorFrame"/>.
     /// Reference: https://ffmpeg.org/doxygen/trunk/group__libsws.html
     /// </summary>
-    private void InitOrUpdateSwsContext(AVPixelFormat srcFormat, int srcWidth, int srcHeight, AVPixelFormat dstFormat)
+    /// <param name="srcFormat">Pixel format of the source frame.</param>
+    /// <param name="srcWidth">Width of the source frame in pixels.</param>
+    /// <param name="srcHeight">Height of the source frame in pixels.</param>
+    /// <param name="dstFormat">Pixel format of the output.</param>
+    /// <param name="colorFrame">Source frame whose colorspace / range tags are applied. Required for RGBA output.</param>
+    private void InitOrUpdateSwsContext(AVPixelFormat srcFormat, int srcWidth, int srcHeight, AVPixelFormat dstFormat, AVFrame* colorFrame = null)
     {
-        if (_pSwsContext != null && _srcPixFmt == srcFormat && _srcWidth == srcWidth && _srcHeight == srcHeight && _dstPixFmt == dstFormat)
+        var swsColorspace = 0;
+        var fullRange = false;
+        if (colorFrame != null && dstFormat == AVPixelFormat.AV_PIX_FMT_RGBA)
+        {
+            swsColorspace = GetSwsColorspace(colorFrame->colorspace);
+            fullRange = IsFullRange(colorFrame);
+        }
+
+        if (_pSwsContext != null && _srcPixFmt == srcFormat && _srcWidth == srcWidth && _srcHeight == srcHeight && _dstPixFmt == dstFormat
+            && _srcSwsColorspace == swsColorspace && _srcFullRange == fullRange)
         {
             return;
         }
@@ -798,6 +912,8 @@ internal sealed unsafe class FFmpegNativeDecoder : IDisposable
         _srcWidth = srcWidth;
         _srcHeight = srcHeight;
         _dstPixFmt = dstFormat;
+        _srcSwsColorspace = swsColorspace;
+        _srcFullRange = fullRange;
 
         _pSwsContext = ffmpeg.sws_getContext(
             srcWidth, srcHeight, srcFormat,
@@ -809,6 +925,44 @@ internal sealed unsafe class FFmpegNativeDecoder : IDisposable
         {
             throw new Exception($"Failed to initialize SwsContext (src: {srcFormat} {srcWidth}x{srcHeight} -> dst: {dstFormat} {Width}x{Height}).");
         }
+
+        if (swsColorspace != 0)
+        {
+            ApplySwsColorspace(swsColorspace, fullRange);
+        }
+    }
+
+    /// <summary>
+    /// Sets the YUV→RGB matrix and the source range of <see cref="_pSwsContext"/>.
+    /// The destination range, brightness, contrast and saturation keep the values of the context.
+    /// Reference: https://ffmpeg.org/doxygen/trunk/group__libsws.html (sws_setColorspaceDetails)
+    /// </summary>
+    /// <param name="swsColorspace">Matrix id (<c>SWS_CS_*</c>) of the source.</param>
+    /// <param name="fullRange">True when the source uses the full range.</param>
+    private void ApplySwsColorspace(int swsColorspace, bool fullRange)
+    {
+        int* currentInvTable;
+        int* currentTable;
+        int srcRange, dstRange, brightness, contrast, saturation;
+
+        // Returns a negative value when the context cannot change its colorspace details (e.g. YUV to YUV conversion).
+        if (ffmpeg.sws_getColorspaceDetails(_pSwsContext, &currentInvTable, &srcRange, &currentTable, &dstRange,
+                &brightness, &contrast, &saturation) < 0)
+        {
+            return;
+        }
+
+        var coefficients = ffmpeg.sws_getCoefficients(swsColorspace);
+        var invTable = new int_array4();
+        var table = new int_array4();
+        for (uint i = 0; i < 4; i++)
+        {
+            invTable[i] = coefficients[i];
+            table[i] = currentTable[i];
+        }
+
+        ffmpeg.sws_setColorspaceDetails(_pSwsContext, invTable, fullRange ? 1 : 0, table, dstRange,
+            brightness, contrast, saturation);
     }
 
     /// <summary>
@@ -900,7 +1054,7 @@ internal sealed unsafe class FFmpegNativeDecoder : IDisposable
             if (frame == null) return false;
 
             timestamp = GetTimestampAndUpdateIndex(frame);
-            InitOrUpdateSwsContext((AVPixelFormat)frame->format, frame->width, frame->height, AVPixelFormat.AV_PIX_FMT_RGBA);
+            InitOrUpdateSwsContext((AVPixelFormat)frame->format, frame->width, frame->height, AVPixelFormat.AV_PIX_FMT_RGBA, frame);
 
             ClearConvertedFramePointers();
             if (vflip)

Util/FFmpeg/FFmpegNativeEncoder.cs

@@ -63,7 +63,7 @@ namespace ADOFAIRenderer.Util.FFmpeg
             public AVCodecID AudioCodecId;
 
             /// <summary>
-            /// Audio bitrate (in bps). If default (0), 128000 bps is used.
+            /// Audio bitrate (in bps). If default (0), 256000 bps is used.
             /// For AAC, 128000–320000 bps is common.
             /// </summary>
             public int Bitrate;
@@ -299,6 +299,23 @@ namespace ADOFAIRenderer.Util.FFmpeg
             return GPUUtil.GetPrioritizedEncoders(_settings.Codec, _settings.ForceSoftware).ToArray();
         }
 
+        /// <summary>
+        /// Expands the candidate list into (encoder, input mode) attempts in order.
+        /// A QSV encoder is attempted twice: system memory input mode (useFramePool = false) first,
+        /// then GPU frame pool input mode (useFramePool = true). Other encoders are attempted once (useFramePool = false).
+        /// </summary>
+        private static System.Collections.Generic.IEnumerable<(VideoEncoder Encoder, bool UseFramePool)> ExpandQsvInputModes(VideoEncoder[] candidates)
+        {
+            foreach (var encoder in candidates)
+            {
+                yield return (encoder, false);
+                if (encoder.Kind == EncoderKind.QSV)
+                {
+                    yield return (encoder, true);
+                }
+            }
+        }
+
         /// <summary>
         /// Converts an FFmpeg error code (negative integer) to a human-readable string.
         /// e.g., -1 → "Operation not permitted"
@@ -367,7 +384,7 @@ namespace ADOFAIRenderer.Util.FFmpeg
                     var candidates = GetEncoderCandidates();
                     var openedSuccessfully = false;
 
-                    foreach (var encoder in candidates)
+                    foreach (var (encoder, useFramePool) in ExpandQsvInputModes(candidates))
                     {
                         // avcodec_find_encoder_by_name: Searches for an encoder by name (e.g., "h264_nvenc").
                         // Returns null if the encoder is not included in the build.
@@ -514,9 +531,10 @@ namespace ADOFAIRenderer.Util.FFmpeg
                         //   Strategy 3) Create QSV device directly (legacy method)
                         //               → Fallback on older drivers/systems
                         //
-                        // After obtaining device, attempt using GPU frame pool:
-                        //   hw_frames_ctx: Pre-allocate frame pool in GPU VRAM (zero-copy, highest performance)
-                        //   hw_device_ctx: Encoder reads frames directly from system memory (better compatibility)
+                        // After obtaining device, the input mode is configured (see ExpandQsvInputModes):
+                        //   Input mode 1, hw_device_ctx: Encoder receives NV12 frames from system memory (attempted first)
+                        //   Input mode 2, hw_frames_ctx: Frames are uploaded to a QSV surface pool in GPU memory
+                        //                                (attempted when the encoder cannot be opened in input mode 1)
                         //
                         // Reference: https://ffmpeg.org/doxygen/trunk/group__lavu__hwcontext.html
                         if (encoder.Kind == EncoderKind.QSV)
@@ -585,30 +603,64 @@ namespace ADOFAIRenderer.Util.FFmpeg
                                 continue;
                             }
 
-                            // ── Attempt hw_frames_ctx: Pre-create GPU frame pool ──
-                            // AVHWFramesContext pre-allocates frame pool in GPU VRAM,
-                            // eliminating memory allocation/deallocation overhead for each frame.
+                            if (!useFramePool)
+                            {
+                                // ── Input mode 1 (primary): hw_device_ctx + NV12 system memory ──
+                                // The encoder receives NV12 frames in system memory (RAM) and the QSV runtime uploads them to the GPU.
                                 //
-                            // Pros: Zero-copy capability, highest performance
-                            // Cons: May fail depending on GPU/driver resolution/format constraints
-                            //       (e.g. specific texture size constraints on Intel Arc → E_INVALIDARG)
+                                // pix_fmt is the software format NV12.
+                                // _isQsvEncoder = false: WriteFrame() passes the NV12 frame to the encoder without uploading it.
+                                _pCodecContext->pix_fmt = AVPixelFormat.AV_PIX_FMT_NV12;
+                                _pCodecContext->hw_device_ctx = ffmpeg.av_buffer_ref(hwDeviceCtx);
+                                _isQsvEncoder = false;
+                                Melon<RenderMod>.Logger.Msg($"{encoder.Name}: Using system memory input mode (hw_device_ctx)");
+                            }
+                            else
+                            {
+                                // ── Input mode 2 (fallback): hw_frames_ctx GPU frame pool ──
+                                // Used when the encoder cannot be opened in input mode 1.
+                                // AVHWFramesContext pre-allocates a pool of QSV surfaces in GPU memory (pix_fmt AV_PIX_FMT_QSV),
+                                // and WriteFrame() uploads each NV12 frame to a surface of the pool (av_hwframe_transfer_data).
                                 //
-                            // initial_pool_size: Number of frames to pre-allocate. Matches encoder pipeline depth.
-                            // Align resolution to multiples of 32: Texture alignment condition required by QSV hardware.
+                                // Pool creation depends on the texture types the GPU/driver accepts
+                                // (e.g. Intel Arc A770 rejects NV12 textures with D3D11_BIND_RENDER_TARGET → E_INVALIDARG 0x80070057).
+                                //
+                                // initial_pool_size: Number of surfaces in the pool (QsvFramePoolSize).
+                                //   The encoder keeps submitted surfaces until they are encoded, so the pool has to hold
+                                //   the frames in the encoder pipeline plus the frame being uploaded.
+                                // Width/height are aligned to multiples of 32: texture alignment condition required by QSV hardware.
+                                //
+                                // Frame pool memory type (AVQSVFramesContext.frame_type, a combination of MFX_MEMTYPE_* flags):
+                                //   Windows (D3D11): MFX_MEMTYPE_VIDEO_MEMORY_DECODER_TARGET (0x0010) is tried first, then 0.
+                                //     0x0010 creates the D3D11 textures with D3D11_BIND_DECODER.
+                                //     0 lets FFmpeg select MFX_MEMTYPE_VIDEO_MEMORY_PROCESSOR_TARGET (0x0020),
+                                //     which creates the textures with D3D11_BIND_RENDER_TARGET.
+                                //   Other platforms (VAAPI): 0. FFmpeg sets the memory type of a VAAPI child pool itself.
+                                // Both types are accepted by the QSV encoder as video memory input.
+                                // When every type fails, the next candidate encoder is tested.
                                 //
                                 // Reference:
                                 //   AVHWFramesContext: https://ffmpeg.org/doxygen/trunk/structAVHWFramesContext.html
                                 //   av_hwframe_ctx_alloc: https://ffmpeg.org/doxygen/trunk/group__lavu__hwcontext.html
+                                //   AVQSVFramesContext.frame_type: https://github.com/FFmpeg/FFmpeg/blob/release/8.0/libavutil/hwcontext_qsv.h
+                                //   frame_type → D3D11 bind flags (qsv_init_child_ctx, qsv_get_d3d11va_bind_flags): https://github.com/FFmpeg/FFmpeg/blob/release/8.0/libavutil/hwcontext_qsv.c
+                                //   MFX_MEMTYPE_* values: https://github.com/intel/libvpl/blob/main/api/vpl/mfxstructures.h
                                 bool framesCtxReady = false;
-                            AVBufferRef* hwFramesRef = ffmpeg.av_hwframe_ctx_alloc(hwDeviceCtx);
-                            if (hwFramesRef != null)
+                                int[] frameTypes = isWindows
+                                    ? new[] { QsvMemTypeVideoMemoryDecoderTarget, 0 }
+                                    : new[] { 0 };
+                                foreach (var frameType in frameTypes)
                                 {
+                                    AVBufferRef* hwFramesRef = ffmpeg.av_hwframe_ctx_alloc(hwDeviceCtx);
+                                    if (hwFramesRef == null) break;
+
                                     var framesCtx = (AVHWFramesContext*)hwFramesRef->data;
                                     framesCtx->format = AVPixelFormat.AV_PIX_FMT_QSV;       // Hardware surface format
                                     framesCtx->sw_format = AVPixelFormat.AV_PIX_FMT_NV12;    // Software-side format
                                     framesCtx->width = (_settings.Width + 31) & ~31;          // 32-multiple alignment (bitmask)
                                     framesCtx->height = (_settings.Height + 31) & ~31;        // 32-multiple alignment (bitmask)
-                                framesCtx->initial_pool_size = 4;                         // Frame pool size
+                                    framesCtx->initial_pool_size = QsvFramePoolSize;
+                                    SetQsvFrameType(framesCtx, frameType);
 
                                     err = ffmpeg.av_hwframe_ctx_init(hwFramesRef);
                                     if (err >= 0)
@@ -618,31 +670,30 @@ namespace ADOFAIRenderer.Util.FFmpeg
                                         _pCodecContext->hw_frames_ctx = ffmpeg.av_buffer_ref(hwFramesRef);
                                         _isQsvEncoder = true;
                                         framesCtxReady = true;
-                                    Melon<RenderMod>.Logger.Msg($"{encoder.Name}: hw_frames_ctx initialization succeeded (GPU frame pool mode)");
+                                        Melon<RenderMod>.Logger.Msg($"{encoder.Name}: Using GPU frame pool input mode (hw_frames_ctx, frame_type=0x{frameType:X4}, pool={QsvFramePoolSize})");
+                                        ffmpeg.av_buffer_unref(&hwFramesRef);
 
                                         _pHwFrame = ffmpeg.av_frame_alloc();
                                         if (_pHwFrame == null)
                                             throw new Exception("Failed to allocate QSV hardware frame structure.");
+                                        break;
                                     }
-                                else
-                                {
-                                    Melon<RenderMod>.Logger.Warning($"{encoder.Name}: hw_frames_ctx failed — {GetFFmpegErrorString(err)} (Code: {err}). Falling back to hw_device_ctx mode.");
-                                }
+
+                                    Melon<RenderMod>.Logger.Warning($"{encoder.Name}: hw_frames_ctx failed with frame_type=0x{frameType:X4} — {GetFFmpegErrorString(err)} (Code: {err})");
                                     ffmpeg.av_buffer_unref(&hwFramesRef);
                                 }
 
-                            // ── When hw_frames_ctx fails: Fallback to hw_device_ctx + NV12 system memory ──
-                            // In hw_device_ctx mode, the encoder receives NV12 frames from system memory (RAM)
-                            // directly and uploads to GPU internally.
-                            // Slightly slower than hw_frames_ctx, but compatibility is much better.
-                            //
-                            // Change pix_fmt to NV12: Use software format directly instead of hardware surface (QSV).
-                            // _isQsvEncoder = false: Skip manual upload in WriteFrame().
                                 if (!framesCtxReady)
                                 {
-                                _pCodecContext->pix_fmt = AVPixelFormat.AV_PIX_FMT_NV12;
-                                _pCodecContext->hw_device_ctx = ffmpeg.av_buffer_ref(hwDeviceCtx);
-                                _isQsvEncoder = false; // System memory mode, no manual frame upload required
+                                    Melon<RenderMod>.Logger.Warning($"{encoder.Name}: GPU frame pool input mode is unavailable. Testing next candidate codec.");
+                                    ffmpeg.av_buffer_unref(&hwDeviceCtx);
+                                    {
+                                        var pCtx = _pCodecContext;
+                                        ffmpeg.avcodec_free_context(&pCtx);
+                                    }
+                                    _pCodecContext = null;
+                                    continue;
+                                }
                             }
 
                             // Transfer ownership of hwDeviceCtx to codec context,
@@ -704,11 +755,19 @@ namespace ADOFAIRenderer.Util.FFmpeg
                             break;
                         }
 
-                        // If open fails, free codec context and try next encoder.
-                        Melon<RenderMod>.Logger.Warning($"Failed to initialize hardware encoder {encoder.Name}: {GetFFmpegErrorString(openResult)} (Code: {openResult}). Testing next candidate codec.");
+                        // If open fails, free codec context and QSV frame state, then try next candidate.
+                        var inputMode = encoder.Kind == EncoderKind.QSV ? (useFramePool ? " (GPU frame pool input mode)" : " (system memory input mode)") : "";
+                        Melon<RenderMod>.Logger.Warning($"Failed to initialize encoder {encoder.Name}{inputMode}: {GetFFmpegErrorString(openResult)} (Code: {openResult}). Testing next candidate.");
                         var pCodecCtx = _pCodecContext;
                         ffmpeg.avcodec_free_context(&pCodecCtx);
                         _pCodecContext = null;
+                        _isQsvEncoder = false;
+                        if (_pHwFrame != null)
+                        {
+                            var pHwFrame = _pHwFrame;
+                            ffmpeg.av_frame_free(&pHwFrame);
+                            _pHwFrame = null;
+                        }
                     }
 
                     if (!openedSuccessfully || _pCodecContext == null)
@@ -977,6 +1036,7 @@ namespace ADOFAIRenderer.Util.FFmpeg
         /// Sets a private option (priv_data) of the encoder.
         /// Returns true when FFmpeg accepts the option. Otherwise logs a warning with the FFmpeg error and returns false.
         /// </summary>
+        // ReSharper disable once UnusedMethodReturnValue.Local
         private bool TrySetOption(string name, string value)
         {
             var ret = ffmpeg.av_opt_set(_pCodecContext->priv_data, name, value, 0);
@@ -986,6 +1046,32 @@ namespace ADOFAIRenderer.Util.FFmpeg
             return false;
         }
 
+        /// <summary>
+        /// MFX_MEMTYPE_VIDEO_MEMORY_DECODER_TARGET (0x0010): frames are in video memory and belong to video decoder render targets.
+        /// Reference: https://github.com/intel/libvpl/blob/main/api/vpl/mfxstructures.h
+        /// </summary>
+        private const int QsvMemTypeVideoMemoryDecoderTarget = 0x0010;
+
+        /// <summary>
+        /// Number of surfaces in the QSV GPU frame pool (AVHWFramesContext.initial_pool_size) of GPU frame pool input mode.
+        /// Same value as the encoder frame pool of the FFmpeg hardware encoding example (initial_pool_size = 20).
+        /// Reference: https://github.com/FFmpeg/FFmpeg/blob/release/8.0/doc/examples/vaapi_encode.c
+        /// </summary>
+        private const int QsvFramePoolSize = 20;
+
+        /// <summary>
+        /// Writes AVQSVFramesContext.frame_type of a QSV frames context (AVHWFramesContext.hwctx).
+        /// FFmpeg.AutoGen has no AVQSVFramesContext type, so the field is written at its offset.
+        /// Field order of AVQSVFramesContext: mfxFrameSurface1* surfaces, int nb_surfaces, int frame_type, mfxFrameInfo* info
+        /// → frame_type is at offset sizeof(pointer) + sizeof(int) (12 bytes on 64-bit).
+        /// Reference: https://github.com/FFmpeg/FFmpeg/blob/release/8.0/libavutil/hwcontext_qsv.h
+        /// </summary>
+        private static void SetQsvFrameType(AVHWFramesContext* framesCtx, int frameType)
+        {
+            if (framesCtx->hwctx == null) return;
+            *(int*)((byte*)framesCtx->hwctx + IntPtr.Size + sizeof(int)) = frameType;
+        }
+
         /// <summary>
         /// Largest declared maximum that is treated as an actual quality range.
         /// A declared maximum above this value is treated as unbounded, and the fallback range 0~51 is used.

Util/VideoBGPlayer.cs

@@ -20,6 +20,7 @@ internal class VideoBGPlayer : MonoBehaviour
     private static readonly int UTex = Shader.PropertyToID("_UTex");
     private static readonly int VTex = Shader.PropertyToID("_VTex");
     private static readonly int ColorRange = Shader.PropertyToID("ColorRange");
+    private static readonly int ColorMatrix = Shader.PropertyToID("ColorMatrix");
 
     [SuppressMessage("ReSharper", "InconsistentNaming")]
     private enum GPUMode
@@ -189,6 +190,14 @@ internal class VideoBGPlayer : MonoBehaviour
         var vidOffset = scrVfxPlus.instance ? scrVfxPlus.instance.vidOffset : 0.0;
         var targetVideoTime = songTime + vidOffset;
 
+        // When looping is enabled, the playback position is the elapsed time wrapped into [0, Duration).
+        // When it is disabled, the position keeps increasing and the last frame stays on screen after the end.
+        var videoDuration = _decoder.Duration;
+        if (_isLooping && targetVideoTime >= 0 && videoDuration > 0)
+        {
+            targetVideoTime %= videoDuration;
+        }
+
         if (targetVideoTime < 0)
         {
             if (_targetTexture)
@@ -307,6 +316,7 @@ internal class VideoBGPlayer : MonoBehaviour
             _computeShader.SetInt(Width, w);
             _computeShader.SetInt(Height, h);
             _computeShader.SetInt(ColorRange, _decoder.ColorRange == AVColorRange.AVCOL_RANGE_JPEG ? 1 : 0);
+            _computeShader.SetInt(ColorMatrix, _decoder.ColorMatrix);
 
             int threadGroupsX = (w + 7) / 8;
             int threadGroupsY = (h + 7) / 8;
@@ -341,6 +351,7 @@ internal class VideoBGPlayer : MonoBehaviour
             _computeShader.SetInt(Width, w);
             _computeShader.SetInt(Height, h);
             _computeShader.SetInt(ColorRange, _decoder.ColorRange == AVColorRange.AVCOL_RANGE_JPEG ? 1 : 0);
+            _computeShader.SetInt(ColorMatrix, _decoder.ColorMatrix);
 
             int threadGroupsX = (w + 7) / 8;
             int threadGroupsY = (h + 7) / 8;