NVDEC (H264): Use separate contexts per channel and decode frames in DTS order (#2671)

* Use separate NVDEC contexts per channel (for FFMPEG)

* Remove NVDEC -> VIC frame override hack

* Add missing bottom_field_pic_order_in_frame_present_flag

* Make FFMPEG logging static

* nit: Remove empty lines

* New FFMPEG decoding approach -- call h264_decode_frame directly, trim surface cache to reduce memory usage

* Fix case

* Silence warnings

* PR feedback

* Per-decoder rather than per-codec ownership of surfaces on the cache

18 files changed, 354 insertions, 196 deletions

diff --git a/Ryujinx.Graphics.Device/IDeviceStateWithContext.cs b/Ryujinx.Graphics.Device/IDeviceStateWithContext.cs
new file mode 100644
index 00000000..17b2fa21
--- /dev/null
+++ b/Ryujinx.Graphics.Device/IDeviceStateWithContext.cs
@@ -0,0 +1,9 @@
+namespace Ryujinx.Graphics.Device
+{
+    public interface IDeviceStateWithContext : IDeviceState
+    {
+        long CreateContext();
+        void DestroyContext(long id);
+        void BindContext(long id);
+    }
+}
diff --git a/Ryujinx.Graphics.Host1x/Host1xDevice.cs b/Ryujinx.Graphics.Host1x/Host1xDevice.cs
index 7f6cef62..61408fc4 100644
--- a/Ryujinx.Graphics.Host1x/Host1xDevice.cs
+++ b/Ryujinx.Graphics.Host1x/Host1xDevice.cs
@@ -9,8 +9,20 @@ namespace Ryujinx.Graphics.Host1x
 {
     public sealed class Host1xDevice : IDisposable
     {
+        private struct Command
+        {
+            public int[] Buffer { get; }
+            public long ContextId { get; }
+
+            public Command(int[] buffer, long contextId)
+            {
+                Buffer = buffer;
+                ContextId = contextId;
+            }
+        }
+
         private readonly SyncptIncrManager _syncptIncrMgr;
-        private readonly AsyncWorkQueue<int[]> _commandQueue;
+        private readonly AsyncWorkQueue<Command> _commandQueue;
 
         private readonly Devices _devices = new Devices();
 
@@ -26,7 +38,7 @@ namespace Ryujinx.Graphics.Host1x
         public Host1xDevice(SynchronizationManager syncMgr)
         {
             _syncptIncrMgr = new SyncptIncrManager(syncMgr);
-            _commandQueue = new AsyncWorkQueue<int[]>(Process, "Ryujinx.Host1xProcessor");
+            _commandQueue = new AsyncWorkQueue<Command>(Process, "Ryujinx.Host1xProcessor");
 
             Class = new Host1xClass(syncMgr);
 
@@ -39,13 +51,52 @@ namespace Ryujinx.Graphics.Host1x
             _devices.RegisterDevice(classId, thi);
         }
 
-        public void Submit(ReadOnlySpan<int> commandBuffer)
+        public long CreateContext()
         {
-            _commandQueue.Add(commandBuffer.ToArray());
+            if (_devices.GetDevice(ClassId.Nvdec) is IDeviceStateWithContext nvdec)
+            {
+                return nvdec.CreateContext();
+            }
+
+            return -1;
         }
 
-        private void Process(int[] commandBuffer)
+        public void DestroyContext(long id)
         {
+            if (id == -1)
+            {
+                return;
+            }
+
+            if (_devices.GetDevice(ClassId.Nvdec) is IDeviceStateWithContext nvdec)
+            {
+                nvdec.DestroyContext(id);
+            }
+        }
+
+        private void SetNvdecContext(long id)
+        {
+            if (id == -1)
+            {
+                return;
+            }
+
+            if (_devices.GetDevice(ClassId.Nvdec) is IDeviceStateWithContext nvdec)
+            {
+                nvdec.BindContext(id);
+            }
+        }
+
+        public void Submit(ReadOnlySpan<int> commandBuffer, long contextId)
+        {
+            _commandQueue.Add(new Command(commandBuffer.ToArray(), contextId));
+        }
+
+        private void Process(Command command)
+        {
+            SetNvdecContext(command.ContextId);
+            int[] commandBuffer = command.Buffer;
+
             for (int index = 0; index < commandBuffer.Length; index++)
             {
                 Step(commandBuffer[index]);
diff --git a/Ryujinx.Graphics.Host1x/ThiDevice.cs b/Ryujinx.Graphics.Host1x/ThiDevice.cs
index 8e3e11b0..114ee26e 100644
--- a/Ryujinx.Graphics.Host1x/ThiDevice.cs
+++ b/Ryujinx.Graphics.Host1x/ThiDevice.cs
@@ -5,19 +5,24 @@ using System.Collections.Generic;
 
 namespace Ryujinx.Graphics.Host1x
 {
-    class ThiDevice : IDeviceState, IDisposable
+    class ThiDevice : IDeviceStateWithContext, IDisposable
     {
         private readonly ClassId _classId;
         private readonly IDeviceState _device;
 
         private readonly SyncptIncrManager _syncptIncrMgr;
 
+        private long _currentContextId;
+        private long _previousContextId;
+
         private class CommandAction
         {
+            public long ContextId { get; }
             public int Data { get; }
 
-            public CommandAction(int data)
+            public CommandAction(long contextId, int data)
             {
+                ContextId = contextId;
                 Data = data;
             }
         }
@@ -26,7 +31,7 @@ namespace Ryujinx.Graphics.Host1x
         {
             public int Method { get; }
 
-            public MethodCallAction(int method, int data) : base(data)
+            public MethodCallAction(long contextId, int method, int data) : base(contextId, data)
             {
                 Method = method;
             }
@@ -34,7 +39,7 @@ namespace Ryujinx.Graphics.Host1x
 
         private class SyncptIncrAction : CommandAction
         {
-            public SyncptIncrAction(uint syncptIncrHandle) : base((int)syncptIncrHandle)
+            public SyncptIncrAction(long contextId, uint syncptIncrHandle) : base(contextId, (int)syncptIncrHandle)
             {
             }
         }
@@ -54,6 +59,31 @@ namespace Ryujinx.Graphics.Host1x
                 { nameof(ThiRegisters.IncrSyncpt), new RwCallback(IncrSyncpt, null) },
                 { nameof(ThiRegisters.Method1), new RwCallback(Method1, null) }
             });
+
+            _previousContextId = -1;
+        }
+
+        public long CreateContext()
+        {
+            if (_device is IDeviceStateWithContext deviceWithContext)
+            {
+                return deviceWithContext.CreateContext();
+            }
+
+            return -1;
+        }
+
+        public void DestroyContext(long id)
+        {
+            if (_device is IDeviceStateWithContext deviceWithContext)
+            {
+                deviceWithContext.DestroyContext(id);
+            }
+        }
+
+        public void BindContext(long id)
+        {
+            _currentContextId = id;
         }
 
         public int Read(int offset) => _state.Read(offset);
@@ -70,17 +100,28 @@ namespace Ryujinx.Graphics.Host1x
             }
             else
             {
-                _commandQueue.Add(new SyncptIncrAction(_syncptIncrMgr.IncrementWhenDone(_classId, syncpointId)));
+                _commandQueue.Add(new SyncptIncrAction(_currentContextId, _syncptIncrMgr.IncrementWhenDone(_classId, syncpointId)));
             }
         }
 
         private void Method1(int data)
         {
-            _commandQueue.Add(new MethodCallAction((int)_state.State.Method0 * 4, data));
+            _commandQueue.Add(new MethodCallAction(_currentContextId, (int)_state.State.Method0 * 4, data));
         }
 
         private void Process(CommandAction cmdAction)
         {
+            long contextId = cmdAction.ContextId;
+            if (contextId != _previousContextId)
+            {
+                _previousContextId = contextId;
+
+                if (_device is IDeviceStateWithContext deviceWithContext)
+                {
+                    deviceWithContext.BindContext(contextId);
+                }
+            }
+
             if (cmdAction is SyncptIncrAction syncptIncrAction)
             {
                 _syncptIncrMgr.SignalDone((uint)syncptIncrAction.Data);
diff --git a/Ryujinx.Graphics.Nvdec.H264/FFmpegContext.cs b/Ryujinx.Graphics.Nvdec.H264/FFmpegContext.cs
index 17de1a03..7676093f 100644
--- a/Ryujinx.Graphics.Nvdec.H264/FFmpegContext.cs
+++ b/Ryujinx.Graphics.Nvdec.H264/FFmpegContext.cs
@@ -9,30 +9,34 @@ namespace Ryujinx.Graphics.Nvdec.H264
 {
     unsafe class FFmpegContext : IDisposable
     {
-        private readonly av_log_set_callback_callback _logFunc;
+        private readonly AVCodec_decode _h264Decode;
+        private static readonly av_log_set_callback_callback _logFunc;
         private readonly AVCodec* _codec;
         private AVPacket* _packet;
         private AVCodecContext* _context;
 
         public FFmpegContext()
         {
-            _logFunc = Log;
-
-            // Redirect log output
-            ffmpeg.av_log_set_level(ffmpeg.AV_LOG_MAX_OFFSET);
-            ffmpeg.av_log_set_callback(_logFunc);
-
             _codec = ffmpeg.avcodec_find_decoder(AVCodecID.AV_CODEC_ID_H264);
             _context = ffmpeg.avcodec_alloc_context3(_codec);
+            _context->debug |= ffmpeg.FF_DEBUG_MMCO;
 
             ffmpeg.avcodec_open2(_context, _codec, null);
 
             _packet = ffmpeg.av_packet_alloc();
+
+            _h264Decode = Marshal.GetDelegateForFunctionPointer<AVCodec_decode>(_codec->decode.Pointer);
         }
 
         static FFmpegContext()
         {
             SetRootPath();
+
+            _logFunc = Log;
+
+            // Redirect log output.
+            ffmpeg.av_log_set_level(ffmpeg.AV_LOG_MAX_OFFSET);
+            ffmpeg.av_log_set_callback(_logFunc);
         }
 
         private static void SetRootPath()
@@ -64,7 +68,7 @@ namespace Ryujinx.Graphics.Nvdec.H264
             }
         }
 
-        private void Log(void* p0, int level, string format, byte* vl)
+        private static void Log(void* p0, int level, string format, byte* vl)
         {
             if (level > ffmpeg.av_log_get_level())
             {
@@ -102,23 +106,40 @@ namespace Ryujinx.Graphics.Nvdec.H264
 
         public int DecodeFrame(Surface output, ReadOnlySpan<byte> bitstream)
         {
-            // Ensure the packet is clean before proceeding
-            ffmpeg.av_packet_unref(_packet);
+            int result;
+            int gotFrame;
 
             fixed (byte* ptr = bitstream)
             {
                 _packet->data = ptr;
                 _packet->size = bitstream.Length;
+                result = _h264Decode(_context, output.Frame, &gotFrame, _packet);
+            }
+
+            if (gotFrame == 0)
+            {
+                ffmpeg.av_frame_unref(output.Frame);
 
-                int rc = ffmpeg.avcodec_send_packet(_context, _packet);
+                // If the frame was not delivered, it was probably delayed.
+                // Get the next delayed frame by passing a 0 length packet.
+                _packet->data = null;
+                _packet->size = 0;
+                result = _h264Decode(_context, output.Frame, &gotFrame, _packet);
 
-                if (rc != 0)
-                {
-                    return rc;
-                }
+                // We need to set B frames to 0 as we already consumed all delayed frames.
+                // This prevents the decoder from trying to return a delayed frame next time.
+                _context->has_b_frames = 0;
+            }
+
+            ffmpeg.av_packet_unref(_packet);
+
+            if (gotFrame == 0)
+            {
+                ffmpeg.av_frame_unref(output.Frame);
+                return -1;
             }
 
-            return ffmpeg.avcodec_receive_frame(_context, output.Frame);
+            return result < 0 ? result : 0;
         }
 
         public void Dispose()
diff --git a/Ryujinx.Graphics.Nvdec.H264/SpsAndPpsReconstruction.cs b/Ryujinx.Graphics.Nvdec.H264/SpsAndPpsReconstruction.cs
index e75c555e..6fd1ce79 100644
--- a/Ryujinx.Graphics.Nvdec.H264/SpsAndPpsReconstruction.cs
+++ b/Ryujinx.Graphics.Nvdec.H264/SpsAndPpsReconstruction.cs
@@ -74,7 +74,7 @@ namespace Ryujinx.Graphics.Nvdec.H264
             writer.WriteUe(0); // Seq parameter set id
 
             writer.WriteBit(pictureInfo.EntropyCodingModeFlag);
-            writer.WriteBit(false); // Bottom field pic order in frame present flag
+            writer.WriteBit(pictureInfo.PicOrderPresentFlag);
             writer.WriteUe(0); // Num slice groups minus 1
             writer.WriteUe(pictureInfo.NumRefIdxL0ActiveMinus1);
             writer.WriteUe(pictureInfo.NumRefIdxL1ActiveMinus1);
diff --git a/Ryujinx.Graphics.Nvdec/H264Decoder.cs b/Ryujinx.Graphics.Nvdec/H264Decoder.cs
index 57ce12d0..1ee3997b 100644
--- a/Ryujinx.Graphics.Nvdec/H264Decoder.cs
+++ b/Ryujinx.Graphics.Nvdec/H264Decoder.cs
@@ -10,9 +10,7 @@ namespace Ryujinx.Graphics.Nvdec
     {
         private const int MbSizeInPixels = 16;
 
-        private static readonly Decoder _decoder = new Decoder();
-
-        public unsafe static void Decode(NvdecDevice device, ResourceManager rm, ref NvdecRegisters state)
+        public unsafe static void Decode(NvdecDecoderContext context, ResourceManager rm, ref NvdecRegisters state)
         {
             PictureInfo pictureInfo = rm.Gmm.DeviceRead<PictureInfo>(state.SetPictureInfoOffset);
             H264PictureInfo info = pictureInfo.Convert();
@@ -22,19 +20,18 @@ namespace Ryujinx.Graphics.Nvdec
             int width  = (int)pictureInfo.PicWidthInMbs * MbSizeInPixels;
             int height = (int)pictureInfo.PicHeightInMbs * MbSizeInPixels;
 
-            ISurface outputSurface = rm.Cache.Get(_decoder, CodecId.H264, 0, 0, width, height);
+            int surfaceIndex = (int)pictureInfo.OutputSurfaceIndex;
 
-            if (_decoder.Decode(ref info, outputSurface, bitstream))
-            {
-                int li = (int)pictureInfo.LumaOutputSurfaceIndex;
-                int ci = (int)pictureInfo.ChromaOutputSurfaceIndex;
+            uint lumaOffset   = state.SetSurfaceLumaOffset[surfaceIndex];
+            uint chromaOffset = state.SetSurfaceChromaOffset[surfaceIndex];
 
-                uint lumaOffset   = state.SetSurfaceLumaOffset[li];
-                uint chromaOffset = state.SetSurfaceChromaOffset[ci];
+            Decoder decoder = context.GetDecoder();
 
-                SurfaceWriter.Write(rm.Gmm, outputSurface, lumaOffset, chromaOffset);
+            ISurface outputSurface = rm.Cache.Get(decoder, 0, 0, width, height);
 
-                device.OnFrameDecoded(CodecId.H264, lumaOffset, chromaOffset);
+            if (decoder.Decode(ref info, outputSurface, bitstream))
+            {
+                SurfaceWriter.Write(rm.Gmm, outputSurface, lumaOffset, chromaOffset);
             }
 
             rm.Cache.Put(outputSurface);
diff --git a/Ryujinx.Graphics.Nvdec/Image/SurfaceCache.cs b/Ryujinx.Graphics.Nvdec/Image/SurfaceCache.cs
index c362185f..dc119673 100644
--- a/Ryujinx.Graphics.Nvdec/Image/SurfaceCache.cs
+++ b/Ryujinx.Graphics.Nvdec/Image/SurfaceCache.cs
@@ -21,7 +21,7 @@ namespace Ryujinx.Graphics.Nvdec.Image
             public uint ChromaOffset;
             public int Width;
             public int Height;
-            public CodecId CodecId;
+            public IDecoder Owner;
             public ISurface Surface;
         }
 
@@ -34,104 +34,110 @@ namespace Ryujinx.Graphics.Nvdec.Image
             _gmm = gmm;
         }
 
-        public ISurface Get(IDecoder decoder, CodecId codecId, uint lumaOffset, uint chromaOffset, int width, int height)
+        public ISurface Get(IDecoder decoder, uint lumaOffset, uint chromaOffset, int width, int height)
         {
-            ISurface surface = null;
-
-            // Try to find a compatible surface with same parameters, and same offsets.
-            for (int i = 0; i < MaxItems; i++)
+            lock (_pool)
             {
-                ref CacheItem item = ref _pool[i];
+                ISurface surface = null;
 
-                if (item.LumaOffset == lumaOffset &&
-                    item.ChromaOffset == chromaOffset &&
-                    item.CodecId == codecId &&
-                    item.Width == width &&
-                    item.Height == height)
-                {
-                    item.ReferenceCount++;
-                    surface = item.Surface;
-                    MoveToFront(i);
-                    break;
-                }
-            }
-
-            // If we failed to find a perfect match, now ignore the offsets.
-            // Search backwards to replace the oldest compatible surface,
-            // this avoids thrashing frquently used surfaces.
-            // Now we need to ensure that the surface is not in use, as we'll change the data.
-            if (surface == null)
-            {
-                for (int i = MaxItems - 1; i >= 0; i--)
+                // Try to find a compatible surface with same parameters, and same offsets.
+                for (int i = 0; i < MaxItems; i++)
                 {
                     ref CacheItem item = ref _pool[i];
 
-                    if (item.ReferenceCount == 0 && item.CodecId == codecId && item.Width == width && item.Height == height)
+                    if (item.LumaOffset == lumaOffset &&
+                        item.ChromaOffset == chromaOffset &&
+                        item.Owner == decoder &&
+                        item.Width == width &&
+                        item.Height == height)
                     {
-                        item.ReferenceCount = 1;
-                        item.LumaOffset = lumaOffset;
-                        item.ChromaOffset = chromaOffset;
+                        item.ReferenceCount++;
                         surface = item.Surface;
-
-                        if ((lumaOffset | chromaOffset) != 0)
-                        {
-                            SurfaceReader.Read(_gmm, surface, lumaOffset, chromaOffset);
-                        }
-
                         MoveToFront(i);
                         break;
                     }
                 }
-            }
 
-            // If everything else failed, we try to create a new surface,
-            // and insert it on the pool. We replace the oldest item on the
-            // pool to avoid thrashing frequently used surfaces.
-            // If even the oldest item is in use, that means that the entire pool
-            // is in use, in that case we throw as there's no place to insert
-            // the new surface.
-            if (surface == null)
-            {
-                if (_pool[MaxItems - 1].ReferenceCount == 0)
+                // If we failed to find a perfect match, now ignore the offsets.
+                // Search backwards to replace the oldest compatible surface,
+                // this avoids thrashing frequently used surfaces.
+                // Now we need to ensure that the surface is not in use, as we'll change the data.
+                if (surface == null)
                 {
-                    surface = decoder.CreateSurface(width, height);
-
-                    if ((lumaOffset | chromaOffset) != 0)
+                    for (int i = MaxItems - 1; i >= 0; i--)
                     {
-                        SurfaceReader.Read(_gmm, surface, lumaOffset, chromaOffset);
-                    }
+                        ref CacheItem item = ref _pool[i];
 
-                    MoveToFront(MaxItems - 1);
-                    ref CacheItem item = ref _pool[0];
-                    item.Surface?.Dispose();
-                    item.ReferenceCount = 1;
-                    item.LumaOffset = lumaOffset;
-                    item.ChromaOffset = chromaOffset;
-                    item.Width = width;
-                    item.Height = height;
-                    item.CodecId = codecId;
-                    item.Surface = surface;
+                        if (item.ReferenceCount == 0 && item.Owner == decoder && item.Width == width && item.Height == height)
+                        {
+                            item.ReferenceCount = 1;
+                            item.LumaOffset = lumaOffset;
+                            item.ChromaOffset = chromaOffset;
+                            surface = item.Surface;
+
+                            if ((lumaOffset | chromaOffset) != 0)
+                            {
+                                SurfaceReader.Read(_gmm, surface, lumaOffset, chromaOffset);
+                            }
+
+                            MoveToFront(i);
+                            break;
+                        }
+                    }
                 }
-                else
+
+                // If everything else failed, we try to create a new surface,
+                // and insert it on the pool. We replace the oldest item on the
+                // pool to avoid thrashing frequently used surfaces.
+                // If even the oldest item is in use, that means that the entire pool
+                // is in use, in that case we throw as there's no place to insert
+                // the new surface.
+                if (surface == null)
                 {
-                    throw new InvalidOperationException("No free slot on the surface pool.");
+                    if (_pool[MaxItems - 1].ReferenceCount == 0)
+                    {
+                        surface = decoder.CreateSurface(width, height);
+
+                        if ((lumaOffset | chromaOffset) != 0)
+                        {
+                            SurfaceReader.Read(_gmm, surface, lumaOffset, chromaOffset);
+                        }
+
+                        MoveToFront(MaxItems - 1);
+                        ref CacheItem item = ref _pool[0];
+                        item.Surface?.Dispose();
+                        item.ReferenceCount = 1;
+                        item.LumaOffset = lumaOffset;
+                        item.ChromaOffset = chromaOffset;
+                        item.Width = width;
+                        item.Height = height;
+                        item.Owner = decoder;
+                        item.Surface = surface;
+                    }
+                    else
+                    {
+                        throw new InvalidOperationException("No free slot on the surface pool.");
+                    }
                 }
-            }
 
-            return surface;
+                return surface;
+            }
         }
 
         public void Put(ISurface surface)
         {
-            for (int i = 0; i < MaxItems; i++)
+            lock (_pool)
             {
-                ref CacheItem item = ref _pool[i];
-
-                if (item.Surface == surface)
+                for (int i = 0; i < MaxItems; i++)
                 {
-                    item.ReferenceCount--;
-                    Debug.Assert(item.ReferenceCount >= 0);
-                    break;
+                    ref CacheItem item = ref _pool[i];
+
+                    if (item.Surface == surface)
+                    {
+                        item.ReferenceCount--;
+                        Debug.Assert(item.ReferenceCount >= 0);
+                        break;
+                    }
                 }
             }
         }
@@ -147,5 +153,22 @@ namespace Ryujinx.Graphics.Nvdec.Image
                 _pool[0] = temp;
             }
         }
+
+        public void Trim()
+        {
+            lock (_pool)
+            {
+                for (int i = 0; i < MaxItems; i++)
+                {
+                    ref CacheItem item = ref _pool[i];
+
+                    if (item.ReferenceCount == 0)
+                    {
+                        item.Surface?.Dispose();
+                        item = default;
+                    }
+                }
+            }
+        }
     }
 }
diff --git a/Ryujinx.Graphics.Nvdec/NvdecDecoderContext.cs b/Ryujinx.Graphics.Nvdec/NvdecDecoderContext.cs
new file mode 100644
index 00000000..90da0bee
--- /dev/null
+++ b/Ryujinx.Graphics.Nvdec/NvdecDecoderContext.cs
@@ -0,0 +1,21 @@
+using Ryujinx.Graphics.Nvdec.H264;
+using System;
+
+namespace Ryujinx.Graphics.Nvdec
+{
+    class NvdecDecoderContext : IDisposable
+    {
+        private Decoder _decoder;
+
+        public Decoder GetDecoder()
+        {
+            return _decoder ??= new Decoder();
+        }
+
+        public void Dispose()
+        {
+            _decoder?.Dispose();
+            _decoder = null;
+        }
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Nvdec/NvdecDevice.cs b/Ryujinx.Graphics.Nvdec/NvdecDevice.cs
index 08f802a1..5319429b 100644
--- a/Ryujinx.Graphics.Nvdec/NvdecDevice.cs
+++ b/Ryujinx.Graphics.Nvdec/NvdecDevice.cs
@@ -2,17 +2,20 @@
 using Ryujinx.Graphics.Device;
 using Ryujinx.Graphics.Gpu.Memory;
 using Ryujinx.Graphics.Nvdec.Image;
-using System;
+using System.Collections.Concurrent;
 using System.Collections.Generic;
+using System.Threading;
 
 namespace Ryujinx.Graphics.Nvdec
 {
-    public class NvdecDevice : IDeviceState
+    public class NvdecDevice : IDeviceStateWithContext
     {
         private readonly ResourceManager _rm;
         private readonly DeviceState<NvdecRegisters> _state;
 
-        public event Action<FrameDecodedEventArgs> FrameDecoded;
+        private long _currentId;
+        private ConcurrentDictionary<long, NvdecDecoderContext> _contexts;
+        private NvdecDecoderContext _currentContext;
 
         public NvdecDevice(MemoryManager gmm)
         {
@@ -21,6 +24,33 @@ namespace Ryujinx.Graphics.Nvdec
             {
                 { nameof(NvdecRegisters.Execute), new RwCallback(Execute, null) }
             });
+            _contexts = new ConcurrentDictionary<long, NvdecDecoderContext>();
+        }
+
+        public long CreateContext()
+        {
+            long id = Interlocked.Increment(ref _currentId);
+            _contexts.TryAdd(id, new NvdecDecoderContext());
+
+            return id;
+        }
+
+        public void DestroyContext(long id)
+        {
+            if (_contexts.TryRemove(id, out var context))
+            {
+                context.Dispose();
+            }
+
+            _rm.Cache.Trim();
+        }
+
+        public void BindContext(long id)
+        {
+            if (_contexts.TryGetValue(id, out var context))
+            {
+                _currentContext = context;
+            }
         }
 
         public int Read(int offset) => _state.Read(offset);
@@ -36,20 +66,15 @@ namespace Ryujinx.Graphics.Nvdec
             switch (codecId)
             {
                 case CodecId.H264:
-                    H264Decoder.Decode(this, _rm, ref _state.State);
+                    H264Decoder.Decode(_currentContext, _rm, ref _state.State);
                     break;
                 case CodecId.Vp9:
-                    Vp9Decoder.Decode(this, _rm, ref _state.State);
+                    Vp9Decoder.Decode(_rm, ref _state.State);
                     break;
                 default:
                     Logger.Error?.Print(LogClass.Nvdec, $"Unsupported codec \"{codecId}\".");
                     break;
             }
         }
-
-        internal void OnFrameDecoded(CodecId codecId, uint lumaOffset, uint chromaOffset)
-        {
-            FrameDecoded?.Invoke(new FrameDecodedEventArgs(codecId, lumaOffset, chromaOffset));
-        }
     }
 }
diff --git a/Ryujinx.Graphics.Nvdec/NvdecStatus.cs b/Ryujinx.Graphics.Nvdec/NvdecStatus.cs
new file mode 100644
index 00000000..0712af88
--- /dev/null
+++ b/Ryujinx.Graphics.Nvdec/NvdecStatus.cs
@@ -0,0 +1,16 @@
+using Ryujinx.Graphics.Nvdec.Types.Vp9;
+
+namespace Ryujinx.Graphics.Nvdec
+{
+    struct NvdecStatus
+    {
+#pragma warning disable CS0649
+        public uint MbsCorrectlyDecoded;
+        public uint MbsInError;
+        public uint Reserved;
+        public uint ErrorStatus;
+        public FrameStats Stats;
+        public uint SliceHeaderErrorCode;
+#pragma warning restore CS0649
+    }
+}
\ No newline at end of file
diff --git a/Ryujinx.Graphics.Nvdec/Types/H264/PictureInfo.cs b/Ryujinx.Graphics.Nvdec/Types/H264/PictureInfo.cs
index 86570342..326c40ae 100644
--- a/Ryujinx.Graphics.Nvdec/Types/H264/PictureInfo.cs
+++ b/Ryujinx.Graphics.Nvdec/Types/H264/PictureInfo.cs
@@ -26,13 +26,13 @@ namespace Ryujinx.Graphics.Nvdec.Types.H264
         public uint Transform8x8ModeFlag;
         public uint LumaPitch;
         public uint ChromaPitch;
-        public uint Unknown94;
-        public uint LumaSecondFieldOffset;
-        public uint Unknown9C;
-        public uint UnknownA0;
-        public uint ChromaSecondFieldOffset;
-        public uint UnknownA8;
-        public uint UnknownAC;
+        public uint LumaTopOffset;
+        public uint LumaBottomOffset;
+        public uint LumaFrameOffset;
+        public uint ChromaTopOffset;
+        public uint ChromaBottomFieldOffset;
+        public uint ChromaFrameOffset;
+        public uint HistBufferSize;
         public ulong Flags;
         public Array2<int> FieldOrderCnt;
         public Array16<ReferenceFrame> RefFrames;
@@ -64,8 +64,8 @@ namespace Ryujinx.Graphics.Nvdec.Types.H264
         public int ChromaQpIndexOffset => ExtractSx(Flags, 22, 5);
         public int SecondChromaQpIndexOffset => ExtractSx(Flags, 27, 5);
         public uint WeightedBipredIdc => (uint)(Flags >> 32) & 3;
-        public uint LumaOutputSurfaceIndex => (uint)(Flags >> 34) & 0x7f;
-        public uint ChromaOutputSurfaceIndex => (uint)(Flags >> 41) & 0x1f;
+        public uint OutputSurfaceIndex => (uint)(Flags >> 34) & 0x7f;
+        public uint ColIndex => (uint)(Flags >> 41) & 0x1f;
         public ushort FrameNum => (ushort)(Flags >> 46);
         public bool QpprimeYZeroTransformBypassFlag => (Flags2 & (1 << 1)) != 0;
 
diff --git a/Ryujinx.Graphics.Nvdec/Types/H264/ReferenceFrame.cs b/Ryujinx.Graphics.Nvdec/Types/H264/ReferenceFrame.cs
index 163a8783..d205a47a 100644
--- a/Ryujinx.Graphics.Nvdec/Types/H264/ReferenceFrame.cs
+++ b/Ryujinx.Graphics.Nvdec/Types/H264/ReferenceFrame.cs
@@ -1,12 +1,15 @@
-namespace Ryujinx.Graphics.Nvdec.Types.H264
+using Ryujinx.Common.Memory;
+
+namespace Ryujinx.Graphics.Nvdec.Types.H264
 {
     struct ReferenceFrame
     {
 #pragma warning disable CS0649
-        public uint Unknown0;
-        public uint Unknown4;
-        public uint Unknown8;
-        public uint UnknownC;
+        public uint Flags;
+        public Array2<uint> FieldOrderCnt;
+        public uint FrameNum;
 #pragma warning restore CS0649
+
+        public uint OutputSurfaceIndex => (uint)Flags & 0x7f;
     }
 }
diff --git a/Ryujinx.Graphics.Nvdec/Types/Vp9/FrameStats.cs b/Ryujinx.Graphics.Nvdec/Types/Vp9/FrameStats.cs
index c360d378..26aab506 100644
--- a/Ryujinx.Graphics.Nvdec/Types/Vp9/FrameStats.cs
+++ b/Ryujinx.Graphics.Nvdec/Types/Vp9/FrameStats.cs
@@ -16,7 +16,6 @@
         public uint FrameStatusBwdMvyCnt;
         public uint ErrorCtbPos;
         public uint ErrorSlicePos;
-        public uint Unknown34;
 #pragma warning restore CS0649
     }
 }
diff --git a/Ryujinx.Graphics.Nvdec/Vp9Decoder.cs b/Ryujinx.Graphics.Nvdec/Vp9Decoder.cs
index f05555c6..b56dc56e 100644
--- a/Ryujinx.Graphics.Nvdec/Vp9Decoder.cs
+++ b/Ryujinx.Graphics.Nvdec/Vp9Decoder.cs
@@ -15,14 +15,14 @@ namespace Ryujinx.Graphics.Nvdec
     {
         private static Decoder _decoder = new Decoder();
 
-        public unsafe static void Decode(NvdecDevice device, ResourceManager rm, ref NvdecRegisters state)
+        public unsafe static void Decode(ResourceManager rm, ref NvdecRegisters state)
         {
             PictureInfo pictureInfo = rm.Gmm.DeviceRead<PictureInfo>(state.SetPictureInfoOffset);
             EntropyProbs entropy = rm.Gmm.DeviceRead<EntropyProbs>(state.SetVp9EntropyProbsOffset);
 
             ISurface Rent(uint lumaOffset, uint chromaOffset, FrameSize size)
             {
-                return rm.Cache.Get(_decoder, CodecId.Vp9, lumaOffset, chromaOffset, size.Width, size.Height);
+                return rm.Cache.Get(_decoder, lumaOffset, chromaOffset, size.Width, size.Height);
             }
 
             ISurface lastSurface    = Rent(state.SetSurfaceLumaOffset[0], state.SetSurfaceChromaOffset[0], pictureInfo.LastFrameSize);
@@ -60,8 +60,6 @@ namespace Ryujinx.Graphics.Nvdec
             if (_decoder.Decode(ref info, currentSurface, bitstream, mvsIn, mvsOut))
             {
                 SurfaceWriter.Write(rm.Gmm, currentSurface, lumaOffset, chromaOffset);
-
-                device.OnFrameDecoded(CodecId.Vp9, lumaOffset, chromaOffset);
             }
 
             WriteBackwardUpdates(rm.Gmm, state.SetVp9BackwardUpdatesOffset, ref info.BackwardUpdateCounts);
diff --git a/Ryujinx.Graphics.Vic/VicDevice.cs b/Ryujinx.Graphics.Vic/VicDevice.cs
index db4fe513..21021c58 100644
--- a/Ryujinx.Graphics.Vic/VicDevice.cs
+++ b/Ryujinx.Graphics.Vic/VicDevice.cs
@@ -1,9 +1,7 @@
-using Ryujinx.Common.Logging;
-using Ryujinx.Graphics.Device;
+using Ryujinx.Graphics.Device;
 using Ryujinx.Graphics.Gpu.Memory;
 using Ryujinx.Graphics.Vic.Image;
 using Ryujinx.Graphics.Vic.Types;
-using System;
 using System.Collections.Generic;
 
 namespace Ryujinx.Graphics.Vic
@@ -14,9 +12,6 @@ namespace Ryujinx.Graphics.Vic
         private readonly ResourceManager _rm;
         private readonly DeviceState<VicRegisters> _state;
 
-        private PlaneOffsets _overrideOffsets;
-        private bool _hasOverride;
-
         public VicDevice(MemoryManager gmm)
         {
             _gmm = gmm;
@@ -27,32 +22,6 @@ namespace Ryujinx.Graphics.Vic
             });
         }
 
-        /// <summary>
-        /// Overrides all input surfaces with a custom surface.
-        /// </summary>
-        /// <param name="lumaOffset">Offset of the luma plane or packed data for this surface</param>
-        /// <param name="chromaUOffset">Offset of the U chroma plane (for planar formats) or both chroma planes (for semiplanar formats)</param>
-        /// <param name="chromaVOffset">Offset of the V chroma plane for planar formats</param>
-        public void SetSurfaceOverride(uint lumaOffset, uint chromaUOffset, uint chromaVOffset)
-        {
-            _overrideOffsets.LumaOffset = lumaOffset;
-            _overrideOffsets.ChromaUOffset = chromaUOffset;
-            _overrideOffsets.ChromaVOffset = chromaVOffset;
-            _hasOverride = true;
-        }
-
-        /// <summary>
-        /// Disables overriding input surfaces.
-        /// </summary>
-        /// <remarks>
-        /// Surface overrides are disabled by default.
-        /// Call this if you previously called <see cref="SetSurfaceOverride(uint, uint, uint)"/> and which to disable it.
-        /// </remarks>
-        public void DisableSurfaceOverride()
-        {
-            _hasOverride = false;
-        }
-
         public int Read(int offset) => _state.Read(offset);
         public void Write(int offset, int data) => _state.Write(offset, data);
 
@@ -76,11 +45,6 @@ namespace Ryujinx.Graphics.Vic
 
                 var offsets = _state.State.SetSurfacexSlotx[i][0];
 
-                if (_hasOverride)
-                {
-                    offsets = _overrideOffsets;
-                }
-
                 using Surface src = SurfaceReader.Read(_rm, ref slot.SlotSurfaceConfig, ref offsets);
 
                 Blender.BlendOne(output, src, ref slot);
diff --git a/Ryujinx.Graphics.Vic/VicRegisters.cs b/Ryujinx.Graphics.Vic/VicRegisters.cs
index 74d41ade..1c11b554 100644
--- a/Ryujinx.Graphics.Vic/VicRegisters.cs
+++ b/Ryujinx.Graphics.Vic/VicRegisters.cs
@@ -4,9 +4,11 @@ namespace Ryujinx.Graphics.Vic
 {
     struct PlaneOffsets
     {
+#pragma warning disable CS0649
         public uint LumaOffset;
         public uint ChromaUOffset;
         public uint ChromaVOffset;
+#pragma warning restore CS0649
     }
 
     struct VicRegisters
diff --git a/Ryujinx.HLE/HOS/Services/Nv/Host1xContext.cs b/Ryujinx.HLE/HOS/Services/Nv/Host1xContext.cs
index 024b0bb4..14f3389b 100644
--- a/Ryujinx.HLE/HOS/Services/Nv/Host1xContext.cs
+++ b/Ryujinx.HLE/HOS/Services/Nv/Host1xContext.cs
@@ -22,21 +22,6 @@ namespace Ryujinx.HLE.HOS.Services.Nv
             var vic = new VicDevice(Smmu);
             Host1x.RegisterDevice(ClassId.Nvdec, nvdec);
             Host1x.RegisterDevice(ClassId.Vic, vic);
-
-            nvdec.FrameDecoded += (FrameDecodedEventArgs e) =>
-            {
-                // FIXME:
-                // Figure out what is causing frame ordering issues on H264.
-                // For now this is needed as workaround.
-                if (e.CodecId == CodecId.H264)
-                {
-                    vic.SetSurfaceOverride(e.LumaOffset, e.ChromaOffset, 0);
-                }
-                else
-                {
-                    vic.DisableSurfaceOverride();
-                }
-            };
         }
 
         public void Dispose()
diff --git a/Ryujinx.HLE/HOS/Services/Nv/NvDrvServices/NvHostChannel/NvHostChannelDeviceFile.cs b/Ryujinx.HLE/HOS/Services/Nv/NvDrvServices/NvHostChannel/NvHostChannelDeviceFile.cs
index a85fd44c..e11949be 100644
--- a/Ryujinx.HLE/HOS/Services/Nv/NvDrvServices/NvHostChannel/NvHostChannelDeviceFile.cs
+++ b/Ryujinx.HLE/HOS/Services/Nv/NvDrvServices/NvHostChannel/NvHostChannelDeviceFile.cs
@@ -27,6 +27,7 @@ namespace Ryujinx.HLE.HOS.Services.Nv.NvDrvServices.NvHostChannel
 
         private readonly IVirtualMemoryManager _memory;
         private readonly Host1xContext _host1xContext;
+        private readonly long _contextId;
 
         public GpuChannel Channel { get; }
 
@@ -52,6 +53,7 @@ namespace Ryujinx.HLE.HOS.Services.Nv.NvDrvServices.NvHostChannel
             _submitTimeout = 0;
             _timeslice     = 0;
             _host1xContext = GetHost1XContext(context.Device.Gpu, owner);
+            _contextId     = _host1xContext.Host1x.CreateContext();
             Channel        = _device.Gpu.CreateChannel();
 
             ChannelInitialization.InitializeState(Channel);
@@ -167,7 +169,7 @@ namespace Ryujinx.HLE.HOS.Services.Nv.NvDrvServices.NvHostChannel
 
                     var data = _memory.GetSpan(map.Address + commandBuffer.Offset, commandBuffer.WordsCount * 4);
 
-                    _host1xContext.Host1x.Submit(MemoryMarshal.Cast<byte, int>(data));
+                    _host1xContext.Host1x.Submit(MemoryMarshal.Cast<byte, int>(data), _contextId);
                 }
             }
 
@@ -177,7 +179,7 @@ namespace Ryujinx.HLE.HOS.Services.Nv.NvDrvServices.NvHostChannel
 
             tmpCmdBuff[0] = (4 << 28) | (int)fences[0].Id;
 
-            _host1xContext.Host1x.Submit(tmpCmdBuff);
+            _host1xContext.Host1x.Submit(tmpCmdBuff, _contextId);
 
             return NvInternalResult.Success;
         }
@@ -548,6 +550,7 @@ namespace Ryujinx.HLE.HOS.Services.Nv.NvDrvServices.NvHostChannel
 
         public override void Close()
         {
+            _host1xContext.Host1x.DestroyContext(_contextId);
             Channel.Dispose();
         }