Implement HwOpus multistream functions (#3275)

* Implement HwOpus multistream functions

* Avoid one copy

Ryujinx.HLE/HOS/Services/Audio/HardwareOpusDecoderManager/Decoder.cs

@@ -0,0 +1,27 @@
+using Concentus.Structs;
+
+namespace Ryujinx.HLE.HOS.Services.Audio.HardwareOpusDecoderManager
+{
+    class Decoder : IDecoder
+    {
+        private readonly OpusDecoder _decoder;
+
+        public int SampleRate => _decoder.SampleRate;
+        public int ChannelsCount => _decoder.NumChannels;
+
+        public Decoder(int sampleRate, int channelsCount)
+        {
+            _decoder = new OpusDecoder(sampleRate, channelsCount);
+        }
+
+        public int Decode(byte[] inData, int inDataOffset, int len, short[] outPcm, int outPcmOffset, int frameSize)
+        {
+            return _decoder.Decode(inData, inDataOffset, len, outPcm, outPcmOffset, frameSize);
+        }
+
+        public void ResetState()
+        {
+            _decoder.ResetState();
+        }
+    }
+}

Ryujinx.HLE/HOS/Services/Audio/HardwareOpusDecoderManager/DecoderCommon.cs

@@ -0,0 +1,92 @@
+using Concentus;
+using Concentus.Enums;
+using Concentus.Structs;
+using Ryujinx.HLE.HOS.Services.Audio.Types;
+using System;
+using System.Runtime.CompilerServices;
+
+namespace Ryujinx.HLE.HOS.Services.Audio.HardwareOpusDecoderManager
+{
+    static class DecoderCommon
+    {
+        private static ResultCode GetPacketNumSamples(this IDecoder decoder, out int numSamples, byte[] packet)
+        {
+            int result = OpusPacketInfo.GetNumSamples(packet, 0, packet.Length, decoder.SampleRate);
+
+            numSamples = result;
+
+            if (result == OpusError.OPUS_INVALID_PACKET)
+            {
+                return ResultCode.OpusInvalidInput;
+            }
+            else if (result == OpusError.OPUS_BAD_ARG)
+            {
+                return ResultCode.OpusInvalidInput;
+            }
+
+            return ResultCode.Success;
+        }
+
+        public static ResultCode DecodeInterleaved(
+            this IDecoder decoder,
+            bool reset,
+            ReadOnlySpan<byte> input,
+            out short[] outPcmData,
+            ulong outputSize,
+            out uint outConsumed,
+            out int outSamples)
+        {
+            outPcmData = null;
+            outConsumed = 0;
+            outSamples = 0;
+
+            int streamSize = input.Length;
+
+            if (streamSize < Unsafe.SizeOf<OpusPacketHeader>())
+            {
+                return ResultCode.OpusInvalidInput;
+            }
+
+            OpusPacketHeader header = OpusPacketHeader.FromSpan(input);
+            int headerSize = Unsafe.SizeOf<OpusPacketHeader>();
+            uint totalSize = header.length + (uint)headerSize;
+
+            if (totalSize > streamSize)
+            {
+                return ResultCode.OpusInvalidInput;
+            }
+
+            byte[] opusData = input.Slice(headerSize, (int)header.length).ToArray();
+
+            ResultCode result = decoder.GetPacketNumSamples(out int numSamples, opusData);
+
+            if (result == ResultCode.Success)
+            {
+                if ((uint)numSamples * (uint)decoder.ChannelsCount * sizeof(short) > outputSize)
+                {
+                    return ResultCode.OpusInvalidInput;
+                }
+
+                outPcmData = new short[numSamples * decoder.ChannelsCount];
+
+                if (reset)
+                {
+                    decoder.ResetState();
+                }
+
+                try
+                {
+                    outSamples = decoder.Decode(opusData, 0, opusData.Length, outPcmData, 0, outPcmData.Length / decoder.ChannelsCount);
+                    outConsumed = totalSize;
+                }
+                catch (OpusException)
+                {
+                    // TODO: as OpusException doesn't provide us the exact error code, this is kind of inaccurate in some cases...
+                    return ResultCode.OpusInvalidInput;
+                }
+            }
+
+            return ResultCode.Success;
+        }
+    }
+}

Ryujinx.HLE/HOS/Services/Audio/HardwareOpusDecoderManager/IDecoder.cs

@@ -0,0 +1,11 @@
+namespace Ryujinx.HLE.HOS.Services.Audio.HardwareOpusDecoderManager
+{
+    interface IDecoder
+    {
+        int SampleRate { get; }
+        int ChannelsCount { get; }
+
+        int Decode(byte[] inData, int inDataOffset, int len, short[] outPcm, int outPcmOffset, int frameSize);
+        void ResetState();
+    }
+}

Ryujinx.HLE/HOS/Services/Audio/HardwareOpusDecoderManager/IHardwareOpusDecoder.cs

@@ -1,244 +1,112 @@
-using Concentus;
-using Concentus.Enums;
-using Concentus.Structs;
 using Ryujinx.HLE.HOS.Services.Audio.Types;
 using System;
-using System.IO;
 using System.Runtime.InteropServices;
 
 namespace Ryujinx.HLE.HOS.Services.Audio.HardwareOpusDecoderManager
 {
     class IHardwareOpusDecoder : IpcService
     {
-        private int  _sampleRate;
-        private int  _channelsCount;
-        private bool _reset;
+        private readonly IDecoder _decoder;
+        private readonly OpusDecoderFlags _flags;
 
-        private OpusDecoder _decoder;
-
-        public IHardwareOpusDecoder(int sampleRate, int channelsCount)
+        public IHardwareOpusDecoder(int sampleRate, int channelsCount, OpusDecoderFlags flags)
         {
-            _sampleRate    = sampleRate;
-            _channelsCount = channelsCount;
-            _reset         = false;
-
-            _decoder = new OpusDecoder(sampleRate, channelsCount);
+            _decoder = new Decoder(sampleRate, channelsCount);
+            _flags = flags;
         }
 
-        private ResultCode GetPacketNumSamples(out int numSamples, byte[] packet)
+        public IHardwareOpusDecoder(int sampleRate, int channelsCount, int streams, int coupledStreams, OpusDecoderFlags flags, byte[] mapping)
         {
-            int result = OpusPacketInfo.GetNumSamples(_decoder, packet, 0, packet.Length);
-
-            numSamples = result;
-
-            if (result == OpusError.OPUS_INVALID_PACKET)
-            {
-                return ResultCode.OpusInvalidInput;
-            }
-            else if (result == OpusError.OPUS_BAD_ARG)
-            {
-                return ResultCode.OpusInvalidInput;
-            }
-
-            return ResultCode.Success;
+            _decoder = new MultiSampleDecoder(sampleRate, channelsCount, streams, coupledStreams, mapping);
+            _flags = flags;
         }
 
-        private ResultCode DecodeInterleavedInternal(BinaryReader input, out short[] outPcmData, long outputSize, out uint outConsumed, out int outSamples)
+        [CommandHipc(0)]
+        // DecodeInterleavedOld(buffer<unknown, 5>) -> (u32, u32, buffer<unknown, 6>)
+        public ResultCode DecodeInterleavedOld(ServiceCtx context)
         {
-            outPcmData  = null;
-            outConsumed = 0;
-            outSamples  = 0;
-
-            long streamSize = input.BaseStream.Length;
-
-            if (streamSize < Marshal.SizeOf<OpusPacketHeader>())
-            {
-                return ResultCode.OpusInvalidInput;
-            }
-
-            OpusPacketHeader header = OpusPacketHeader.FromStream(input);
-
-            uint totalSize = header.length + (uint)Marshal.SizeOf<OpusPacketHeader>();
-
-            if (totalSize > streamSize)
-            {
-                return ResultCode.OpusInvalidInput;
-            }
-
-            byte[] opusData = input.ReadBytes((int)header.length);
-
-            ResultCode result = GetPacketNumSamples(out int numSamples, opusData);
-
-            if (result == ResultCode.Success)
-            {
-                if ((uint)numSamples * (uint)_channelsCount * sizeof(short) > outputSize)
-                {
-                    return ResultCode.OpusInvalidInput;
-                }
-
-                outPcmData = new short[numSamples * _channelsCount];
-
-                if (_reset)
-                {
-                    _reset = false;
-
-                    _decoder.ResetState();
-                }
-
-                try
-                {
-                    outSamples  = _decoder.Decode(opusData, 0, opusData.Length, outPcmData, 0, outPcmData.Length / _channelsCount);
-                    outConsumed = totalSize;
-                }
-                catch (OpusException)
-                {
-                    // TODO: as OpusException doesn't provide us the exact error code, this is kind of inaccurate in some cases...
-                    return ResultCode.OpusInvalidInput;
-                }
-            }
-
-            return ResultCode.Success;
+            return DecodeInterleavedInternal(context, OpusDecoderFlags.None, reset: false, withPerf: false);
         }
 
-        [CommandHipc(0)]
-        // DecodeInterleaved(buffer<unknown, 5>) -> (u32, u32, buffer<unknown, 6>)
-        public ResultCode DecodeInterleavedOriginal(ServiceCtx context)
+        [CommandHipc(2)]
+        // DecodeInterleavedForMultiStreamOld(buffer<unknown, 5>) -> (u32, u32, buffer<unknown, 6>)
+        public ResultCode DecodeInterleavedForMultiStreamOld(ServiceCtx context)
         {
-            ResultCode result;
-
-            ulong inPosition     = context.Request.SendBuff[0].Position;
-            ulong inSize         = context.Request.SendBuff[0].Size;
-            ulong outputPosition = context.Request.ReceiveBuff[0].Position;
-            ulong outputSize     = context.Request.ReceiveBuff[0].Size;
-
-            byte[] buffer = new byte[inSize];
-
-            context.Memory.Read(inPosition, buffer);
-
-            using (BinaryReader inputStream = new BinaryReader(new MemoryStream(buffer)))
-            {
-                result = DecodeInterleavedInternal(inputStream, out short[] outPcmData, (long)outputSize, out uint outConsumed, out int outSamples);
-
-                if (result == ResultCode.Success)
-                {
-                    byte[] pcmDataBytes = new byte[outPcmData.Length * sizeof(short)];
-                    Buffer.BlockCopy(outPcmData, 0, pcmDataBytes, 0, pcmDataBytes.Length);
-                    context.Memory.Write(outputPosition, pcmDataBytes);
-
-                    context.ResponseData.Write(outConsumed);
-                    context.ResponseData.Write(outSamples);
-                }
-            }
-
-            return result;
+            return DecodeInterleavedInternal(context, OpusDecoderFlags.None, reset: false, withPerf: false);
         }
 
         [CommandHipc(4)] // 6.0.0+
         // DecodeInterleavedWithPerfOld(buffer<unknown, 5>) -> (u32, u32, u64, buffer<unknown, 0x46>)
         public ResultCode DecodeInterleavedWithPerfOld(ServiceCtx context)
         {
-            ResultCode result;
-
-            ulong inPosition     = context.Request.SendBuff[0].Position;
-            ulong inSize         = context.Request.SendBuff[0].Size;
-            ulong outputPosition = context.Request.ReceiveBuff[0].Position;
-            ulong outputSize     = context.Request.ReceiveBuff[0].Size;
-
-            byte[] buffer = new byte[inSize];
-
-            context.Memory.Read(inPosition, buffer);
-
-            using (BinaryReader inputStream = new BinaryReader(new MemoryStream(buffer)))
-            {
-                result = DecodeInterleavedInternal(inputStream, out short[] outPcmData, (long)outputSize, out uint outConsumed, out int outSamples);
-
-                if (result == ResultCode.Success)
-                {
-                    byte[] pcmDataBytes = new byte[outPcmData.Length * sizeof(short)];
-                    Buffer.BlockCopy(outPcmData, 0, pcmDataBytes, 0, pcmDataBytes.Length);
-                    context.Memory.Write(outputPosition, pcmDataBytes);
-
-                    context.ResponseData.Write(outConsumed);
-                    context.ResponseData.Write(outSamples);
-
-                    // This is the time the DSP took to process the request, TODO: fill this.
-                    context.ResponseData.Write(0);
-                }
-            }
+            return DecodeInterleavedInternal(context, OpusDecoderFlags.None, reset: false, withPerf: true);
+        }
 
-            return result;
+        [CommandHipc(5)] // 6.0.0+
+        // DecodeInterleavedForMultiStreamWithPerfOld(buffer<unknown, 5>) -> (u32, u32, u64, buffer<unknown, 0x46>)
+        public ResultCode DecodeInterleavedForMultiStreamWithPerfOld(ServiceCtx context)
+        {
+            return DecodeInterleavedInternal(context, OpusDecoderFlags.None, reset: false, withPerf: true);
         }
 
         [CommandHipc(6)] // 6.0.0+
-        // DecodeInterleavedOld(bool reset, buffer<unknown, 5>) -> (u32, u32, u64, buffer<unknown, 0x46>)
-        public ResultCode DecodeInterleavedOld(ServiceCtx context)
+        // DecodeInterleavedWithPerfAndResetOld(bool reset, buffer<unknown, 5>) -> (u32, u32, u64, buffer<unknown, 0x46>)
+        public ResultCode DecodeInterleavedWithPerfAndResetOld(ServiceCtx context)
         {
-            ResultCode result;
-
-            _reset = context.RequestData.ReadBoolean();
-
-            ulong inPosition     = context.Request.SendBuff[0].Position;
-            ulong inSize         = context.Request.SendBuff[0].Size;
-            ulong outputPosition = context.Request.ReceiveBuff[0].Position;
-            ulong outputSize     = context.Request.ReceiveBuff[0].Size;
+            bool reset = context.RequestData.ReadBoolean();
 
-            byte[] buffer = new byte[inSize];
-
-            context.Memory.Read(inPosition, buffer);
-
-            using (BinaryReader inputStream = new BinaryReader(new MemoryStream(buffer)))
-            {
-                result = DecodeInterleavedInternal(inputStream, out short[] outPcmData, (long)outputSize, out uint outConsumed, out int outSamples);
-
-                if (result == ResultCode.Success)
-                {
-                    byte[] pcmDataBytes = new byte[outPcmData.Length * sizeof(short)];
-                    Buffer.BlockCopy(outPcmData, 0, pcmDataBytes, 0, pcmDataBytes.Length);
-                    context.Memory.Write(outputPosition, pcmDataBytes);
-
-                    context.ResponseData.Write(outConsumed);
-                    context.ResponseData.Write(outSamples);
+            return DecodeInterleavedInternal(context, OpusDecoderFlags.None, reset, withPerf: true);
+        }
 
-                    // This is the time the DSP took to process the request, TODO: fill this.
-                    context.ResponseData.Write(0);
-                }
-            }
+        [CommandHipc(7)] // 6.0.0+
+        // DecodeInterleavedForMultiStreamWithPerfAndResetOld(bool reset, buffer<unknown, 5>) -> (u32, u32, u64, buffer<unknown, 0x46>)
+        public ResultCode DecodeInterleavedForMultiStreamWithPerfAndResetOld(ServiceCtx context)
+        {
+            bool reset = context.RequestData.ReadBoolean();
 
-            return result;
+            return DecodeInterleavedInternal(context, OpusDecoderFlags.None, reset, withPerf: true);
         }
 
         [CommandHipc(8)] // 7.0.0+
         // DecodeInterleaved(bool reset, buffer<unknown, 0x45>) -> (u32, u32, u64, buffer<unknown, 0x46>)
         public ResultCode DecodeInterleaved(ServiceCtx context)
         {
-            ResultCode result;
+            bool reset = context.RequestData.ReadBoolean();
 
-            _reset = context.RequestData.ReadBoolean();
+            return DecodeInterleavedInternal(context, _flags, reset, withPerf: true);
+        }
 
+        [CommandHipc(9)] // 7.0.0+
+        // DecodeInterleavedForMultiStream(bool reset, buffer<unknown, 0x45>) -> (u32, u32, u64, buffer<unknown, 0x46>)
+        public ResultCode DecodeInterleavedForMultiStream(ServiceCtx context)
+        {
+            bool reset = context.RequestData.ReadBoolean();
+
+            return DecodeInterleavedInternal(context, _flags, reset, withPerf: true);
+        }
+
+        private ResultCode DecodeInterleavedInternal(ServiceCtx context, OpusDecoderFlags flags, bool reset, bool withPerf)
+        {
             ulong inPosition     = context.Request.SendBuff[0].Position;
             ulong inSize         = context.Request.SendBuff[0].Size;
             ulong outputPosition = context.Request.ReceiveBuff[0].Position;
             ulong outputSize     = context.Request.ReceiveBuff[0].Size;
 
-            byte[] buffer = new byte[inSize];
+            ReadOnlySpan<byte> input = context.Memory.GetSpan(inPosition, (int)inSize);
 
-            context.Memory.Read(inPosition, buffer);
+            ResultCode result = _decoder.DecodeInterleaved(reset, input, out short[] outPcmData, outputSize, out uint outConsumed, out int outSamples);
 
-            using (BinaryReader inputStream = new BinaryReader(new MemoryStream(buffer)))
+            if (result == ResultCode.Success)
             {
-                result = DecodeInterleavedInternal(inputStream, out short[] outPcmData, (long)outputSize, out uint outConsumed, out int outSamples);
+                context.Memory.Write(outputPosition, MemoryMarshal.Cast<short, byte>(outPcmData.AsSpan()));
 
-                if (result == ResultCode.Success)
-                {
-                    byte[] pcmDataBytes = new byte[outPcmData.Length * sizeof(short)];
-                    Buffer.BlockCopy(outPcmData, 0, pcmDataBytes, 0, pcmDataBytes.Length);
-                    context.Memory.Write(outputPosition, pcmDataBytes);
-
-                    context.ResponseData.Write(outConsumed);
-                    context.ResponseData.Write(outSamples);
+                context.ResponseData.Write(outConsumed);
+                context.ResponseData.Write(outSamples);
 
+                if (withPerf)
+                {
                     // This is the time the DSP took to process the request, TODO: fill this.
-                    context.ResponseData.Write(0);
+                    context.ResponseData.Write(0UL);
                 }
             }
 

Ryujinx.HLE/HOS/Services/Audio/HardwareOpusDecoderManager/MultiSampleDecoder.cs

@@ -0,0 +1,28 @@
+using Concentus.Structs;
+
+namespace Ryujinx.HLE.HOS.Services.Audio.HardwareOpusDecoderManager
+{
+    class MultiSampleDecoder : IDecoder
+    {
+        private readonly OpusMSDecoder _decoder;
+
+        public int SampleRate => _decoder.SampleRate;
+        public int ChannelsCount { get; }
+
+        public MultiSampleDecoder(int sampleRate, int channelsCount, int streams, int coupledStreams, byte[] mapping)
+        {
+            ChannelsCount = channelsCount;
+            _decoder = new OpusMSDecoder(sampleRate, channelsCount, streams, coupledStreams, mapping);
+        }
+
+        public int Decode(byte[] inData, int inDataOffset, int len, short[] outPcm, int outPcmOffset, int frameSize)
+        {
+            return _decoder.DecodeMultistream(inData, inDataOffset, len, outPcm, outPcmOffset, frameSize, 0);
+        }
+
+        public void ResetState()
+        {
+            _decoder.ResetState();
+        }
+    }
+}

Ryujinx.HLE/HOS/Services/Audio/IHardwareOpusDecoderManager.cs

@@ -1,6 +1,7 @@
 using Ryujinx.Common;
 using Ryujinx.HLE.HOS.Services.Audio.HardwareOpusDecoderManager;
 using Ryujinx.HLE.HOS.Services.Audio.Types;
+using System.Runtime.InteropServices;
 
 namespace Ryujinx.HLE.HOS.Services.Audio
 {
@@ -16,7 +17,7 @@ namespace Ryujinx.HLE.HOS.Services.Audio
             int sampleRate    = context.RequestData.ReadInt32();
             int channelsCount = context.RequestData.ReadInt32();
 
-            MakeObject(context, new IHardwareOpusDecoder(sampleRate, channelsCount));
+            MakeObject(context, new IHardwareOpusDecoder(sampleRate, channelsCount, OpusDecoderFlags.None));
 
             // Close transfer memory immediately as we don't use it.
             context.Device.System.KernelContext.Syscall.CloseHandle(context.Request.HandleDesc.ToCopy[0]);
@@ -28,11 +29,50 @@ namespace Ryujinx.HLE.HOS.Services.Audio
         // GetWorkBufferSize(bytes<8, 4>) -> u32
         public ResultCode GetWorkBufferSize(ServiceCtx context)
         {
-            // NOTE: The sample rate is ignored because it is fixed to 48KHz.
             int sampleRate    = context.RequestData.ReadInt32();
             int channelsCount = context.RequestData.ReadInt32();
 
-            context.ResponseData.Write(GetOpusDecoderSize(channelsCount));
+            int opusDecoderSize = GetOpusDecoderSize(channelsCount);
+
+            int frameSize = BitUtils.AlignUp(channelsCount * 1920 / (48000 / sampleRate), 64);
+            int totalSize = opusDecoderSize + 1536 + frameSize;
+
+            context.ResponseData.Write(totalSize);
+
+            return ResultCode.Success;
+        }
+
+        [CommandHipc(2)] // 3.0.0+
+        // InitializeForMultiStream(u32, handle<copy>, buffer<unknown<0x110>, 0x19>) -> object<nn::codec::detail::IHardwareOpusDecoder>
+        public ResultCode InitializeForMultiStream(ServiceCtx context)
+        {
+            ulong parametersAddress = context.Request.PtrBuff[0].Position;
+
+            OpusMultiStreamParameters parameters = context.Memory.Read<OpusMultiStreamParameters>(parametersAddress);
+
+            MakeObject(context, new IHardwareOpusDecoder(parameters.SampleRate, parameters.ChannelsCount, OpusDecoderFlags.None));
+
+            // Close transfer memory immediately as we don't use it.
+            context.Device.System.KernelContext.Syscall.CloseHandle(context.Request.HandleDesc.ToCopy[0]);
+
+            return ResultCode.Success;
+        }
+
+        [CommandHipc(3)] // 3.0.0+
+        // GetWorkBufferSizeForMultiStream(buffer<unknown<0x110>, 0x19>) -> u32
+        public ResultCode GetWorkBufferSizeForMultiStream(ServiceCtx context)
+        {
+            ulong parametersAddress = context.Request.PtrBuff[0].Position;
+
+            OpusMultiStreamParameters parameters = context.Memory.Read<OpusMultiStreamParameters>(parametersAddress);
+
+            int opusDecoderSize = GetOpusMultistreamDecoderSize(parameters.NumberOfStreams, parameters.NumberOfStereoStreams);
+
+            int streamSize = BitUtils.AlignUp(parameters.NumberOfStreams * 1500, 64);
+            int frameSize = BitUtils.AlignUp(parameters.ChannelsCount * 1920 / (48000 / parameters.SampleRate), 64);
+            int totalSize = opusDecoderSize + streamSize + frameSize;
+
+            context.ResponseData.Write(totalSize);
 
             return ResultCode.Success;
         }
@@ -44,7 +84,7 @@ namespace Ryujinx.HLE.HOS.Services.Audio
             OpusParametersEx parameters = context.RequestData.ReadStruct<OpusParametersEx>();
 
             // UseLargeFrameSize can be ignored due to not relying on fixed size buffers for storing the decoded result.
-            MakeObject(context, new IHardwareOpusDecoder(parameters.SampleRate, parameters.ChannelCount));
+            MakeObject(context, new IHardwareOpusDecoder(parameters.SampleRate, parameters.ChannelsCount, parameters.Flags));
 
             // Close transfer memory immediately as we don't use it.
             context.Device.System.KernelContext.Syscall.CloseHandle(context.Request.HandleDesc.ToCopy[0]);
@@ -58,15 +98,84 @@ namespace Ryujinx.HLE.HOS.Services.Audio
         {
             OpusParametersEx parameters = context.RequestData.ReadStruct<OpusParametersEx>();
 
-            // NOTE: The sample rate is ignored because it is fixed to 48KHz.
-            context.ResponseData.Write(GetOpusDecoderSize(parameters.ChannelCount));
+            int opusDecoderSize = GetOpusDecoderSize(parameters.ChannelsCount);
+
+            int frameSizeMono48KHz = parameters.Flags.HasFlag(OpusDecoderFlags.LargeFrameSize) ? 5760 : 1920;
+            int frameSize = BitUtils.AlignUp(parameters.ChannelsCount * frameSizeMono48KHz / (48000 / parameters.SampleRate), 64);
+            int totalSize = opusDecoderSize + 1536 + frameSize;
+
+            context.ResponseData.Write(totalSize);
+
+            return ResultCode.Success;
+        }
+
+        [CommandHipc(6)] // 12.0.0+
+        // InitializeForMultiStreamEx(u32, handle<copy>, buffer<unknown<0x118>, 0x19>) -> object<nn::codec::detail::IHardwareOpusDecoder>
+        public ResultCode InitializeForMultiStreamEx(ServiceCtx context)
+        {
+            ulong parametersAddress = context.Request.PtrBuff[0].Position;
+
+            OpusMultiStreamParametersEx parameters = context.Memory.Read<OpusMultiStreamParametersEx>(parametersAddress);
+
+            byte[] mappings = MemoryMarshal.Cast<uint, byte>(parameters.ChannelMappings.ToSpan()).ToArray();
+
+            // UseLargeFrameSize can be ignored due to not relying on fixed size buffers for storing the decoded result.
+            MakeObject(context, new IHardwareOpusDecoder(
+                parameters.SampleRate,
+                parameters.ChannelsCount,
+                parameters.NumberOfStreams,
+                parameters.NumberOfStereoStreams,
+                parameters.Flags,
+                mappings));
+
+            // Close transfer memory immediately as we don't use it.
+            context.Device.System.KernelContext.Syscall.CloseHandle(context.Request.HandleDesc.ToCopy[0]);
+
+            return ResultCode.Success;
+        }
+
+        [CommandHipc(7)] // 12.0.0+
+        // GetWorkBufferSizeForMultiStreamEx(buffer<unknown<0x118>, 0x19>) -> u32
+        public ResultCode GetWorkBufferSizeForMultiStreamEx(ServiceCtx context)
+        {
+            ulong parametersAddress = context.Request.PtrBuff[0].Position;
+
+            OpusMultiStreamParametersEx parameters = context.Memory.Read<OpusMultiStreamParametersEx>(parametersAddress);
+
+            int opusDecoderSize = GetOpusMultistreamDecoderSize(parameters.NumberOfStreams, parameters.NumberOfStereoStreams);
+
+            int frameSizeMono48KHz = parameters.Flags.HasFlag(OpusDecoderFlags.LargeFrameSize) ? 5760 : 1920;
+            int streamSize = BitUtils.AlignUp(parameters.NumberOfStreams * 1500, 64);
+            int frameSize = BitUtils.AlignUp(parameters.ChannelsCount * frameSizeMono48KHz / (48000 / parameters.SampleRate), 64);
+            int totalSize = opusDecoderSize + streamSize + frameSize;
+
+            context.ResponseData.Write(totalSize);
 
             return ResultCode.Success;
         }
 
+        private static int GetOpusMultistreamDecoderSize(int streams, int coupledStreams)
+        {
+            if (streams < 1 || coupledStreams > streams || coupledStreams < 0)
+            {
+                return 0;
+            }
+
+            int coupledSize = GetOpusDecoderSize(2);
+            int monoSize = GetOpusDecoderSize(1);
+
+            return Align4(monoSize - GetOpusDecoderAllocSize(1)) * (streams - coupledStreams) +
+                Align4(coupledSize - GetOpusDecoderAllocSize(2)) * coupledStreams + 0xb90c;
+        }
+
+        private static int Align4(int value)
+        {
+            return BitUtils.AlignUp(value, 4);
+        }
+
         private static int GetOpusDecoderSize(int channelsCount)
         {
-            const int silkDecoderSize = 0x2198;
+            const int SilkDecoderSize = 0x2160;
 
             if (channelsCount < 1 || channelsCount > 2)
             {
@@ -74,24 +183,23 @@ namespace Ryujinx.HLE.HOS.Services.Audio
             }
 
             int celtDecoderSize = GetCeltDecoderSize(channelsCount);
+            int opusDecoderSize = GetOpusDecoderAllocSize(channelsCount) | 0x4c;
 
-            int opusDecoderSize = (channelsCount * 0x800 + 0x4807) & -0x800 | 0x50;
+            return opusDecoderSize + SilkDecoderSize + celtDecoderSize;
+        }
 
-            return opusDecoderSize + silkDecoderSize + celtDecoderSize;
+        private static int GetOpusDecoderAllocSize(int channelsCount)
+        {
+            return (channelsCount * 0x800 + 0x4803) & -0x800;
         }
 
         private static int GetCeltDecoderSize(int channelsCount)
         {
-            const int decodeBufferSize = 0x2030;
-            const int celtDecoderSize  = 0x58;
-            const int celtSigSize      = 0x4;
-            const int overlap          = 120;
-            const int eBandsCount      = 21;
-
-            return (decodeBufferSize + overlap * 4) * channelsCount +
-                    eBandsCount * 16 +
-                    celtDecoderSize +
-                    celtSigSize;
+            const int DecodeBufferSize = 0x2030;
+            const int Overlap          = 120;
+            const int EBandsCount      = 21;
+
+            return (DecodeBufferSize + Overlap * 4) * channelsCount + EBandsCount * 16 + 0x50;
         }
     }
 }

Ryujinx.HLE/HOS/Services/Audio/Types/OpusMultiStreamParameters.cs

@@ -0,0 +1,15 @@
+using Ryujinx.Common.Memory;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.HLE.HOS.Services.Audio.Types
+{
+    [StructLayout(LayoutKind.Sequential, Size = 0x110)]
+    struct OpusMultiStreamParameters
+    {
+        public int SampleRate;
+        public int ChannelsCount;
+        public int NumberOfStreams;
+        public int NumberOfStereoStreams;
+        public Array64<uint> ChannelMappings;
+    }
+}

Ryujinx.HLE/HOS/Services/Audio/Types/OpusMultiStreamParametersEx.cs

@@ -0,0 +1,19 @@
+using Ryujinx.Common.Memory;
+using System.Runtime.InteropServices;
+
+namespace Ryujinx.HLE.HOS.Services.Audio.Types
+{
+    [StructLayout(LayoutKind.Sequential, Size = 0x118)]
+    struct OpusMultiStreamParametersEx
+    {
+        public int SampleRate;
+        public int ChannelsCount;
+        public int NumberOfStreams;
+        public int NumberOfStereoStreams;
+        public OpusDecoderFlags Flags;
+
+        Array4<byte> Padding1;
+
+        public Array64<uint> ChannelMappings;
+    }
+}

Ryujinx.HLE/HOS/Services/Audio/Types/OpusPacketHeader.cs

@@ -12,9 +12,9 @@ namespace Ryujinx.HLE.HOS.Services.Audio.Types
         public uint length;
         public uint finalRange;
 
-        public static OpusPacketHeader FromStream(BinaryReader reader)
+        public static OpusPacketHeader FromSpan(ReadOnlySpan<byte> data)
         {
-            OpusPacketHeader header = reader.ReadStruct<OpusPacketHeader>();
+            OpusPacketHeader header = MemoryMarshal.Cast<byte, OpusPacketHeader>(data)[0];
 
             header.length     = BitConverter.IsLittleEndian ? BinaryPrimitives.ReverseEndianness(header.length)     : header.length;
             header.finalRange = BitConverter.IsLittleEndian ? BinaryPrimitives.ReverseEndianness(header.finalRange) : header.finalRange;

Ryujinx.HLE/HOS/Services/Audio/Types/OpusParametersEx.cs

@@ -7,8 +7,8 @@ namespace Ryujinx.HLE.HOS.Services.Audio.Types
     struct OpusParametersEx
     {
         public int SampleRate;
-        public int ChannelCount;
-        public OpusDecoderFlags UseLargeFrameSize;
+        public int ChannelsCount;
+        public OpusDecoderFlags Flags;
 
         Array4<byte> Padding1;
     }