Amadeus: Add ARM SIMD fast path (#2069)

Add fast paths in the audio renderer for AArch64 in all current fast paths.

2 files changed, 51 insertions, 0 deletions

diff --git a/Ryujinx.Audio/Renderer/Dsp/Command/MixCommand.cs b/Ryujinx.Audio/Renderer/Dsp/Command/MixCommand.cs
index 566fea92..06968871 100644
--- a/Ryujinx.Audio/Renderer/Dsp/Command/MixCommand.cs
+++ b/Ryujinx.Audio/Renderer/Dsp/Command/MixCommand.cs
@@ -19,6 +19,7 @@ using System;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
 using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.Arm;
 using System.Runtime.Intrinsics.X86;
 
 namespace Ryujinx.Audio.Renderer.Dsp.Command
@@ -90,6 +91,27 @@ namespace Ryujinx.Audio.Renderer.Dsp.Command
         }
 
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        private void ProcessMixAdvSimd(Span<float> outputMix, ReadOnlySpan<float> inputMix)
+        {
+            Vector128<float> volumeVec = Vector128.Create(Volume);
+
+            ReadOnlySpan<Vector128<float>> inputVec = MemoryMarshal.Cast<float, Vector128<float>>(inputMix);
+            Span<Vector128<float>> outputVec = MemoryMarshal.Cast<float, Vector128<float>>(outputMix);
+
+            int sisdStart = inputVec.Length * 4;
+
+            for (int i = 0; i < inputVec.Length; i++)
+            {
+                outputVec[i] = AdvSimd.Add(outputVec[i], AdvSimd.Ceiling(AdvSimd.Multiply(inputVec[i], volumeVec)));
+            }
+
+            for (int i = sisdStart; i < inputMix.Length; i++)
+            {
+                outputMix[i] += FloatingPointHelper.MultiplyRoundUp(inputMix[i], Volume);
+            }
+        }
+
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
         private void ProcessMixSlowPath(Span<float> outputMix, ReadOnlySpan<float> inputMix)
         {
             for (int i = 0; i < inputMix.Length; i++)
@@ -108,6 +130,10 @@ namespace Ryujinx.Audio.Renderer.Dsp.Command
             {
                 ProcessMixSse41(outputMix, inputMix);
             }
+            else if (AdvSimd.IsSupported)
+            {
+                ProcessMixAdvSimd(outputMix, inputMix);
+            }
             else
             {
                 ProcessMixSlowPath(outputMix, inputMix);
diff --git a/Ryujinx.Audio/Renderer/Dsp/Command/VolumeCommand.cs b/Ryujinx.Audio/Renderer/Dsp/Command/VolumeCommand.cs
index b58ae1f8..217d51c9 100644
--- a/Ryujinx.Audio/Renderer/Dsp/Command/VolumeCommand.cs
+++ b/Ryujinx.Audio/Renderer/Dsp/Command/VolumeCommand.cs
@@ -19,6 +19,7 @@ using System;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
 using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.Arm;
 using System.Runtime.Intrinsics.X86;
 
 namespace Ryujinx.Audio.Renderer.Dsp.Command
@@ -89,6 +90,26 @@ namespace Ryujinx.Audio.Renderer.Dsp.Command
             }
         }
 
+        private void ProcessVolumeAdvSimd(Span<float> outputBuffer, ReadOnlySpan<float> inputBuffer)
+        {
+            Vector128<float> volumeVec = Vector128.Create(Volume);
+
+            ReadOnlySpan<Vector128<float>> inputVec = MemoryMarshal.Cast<float, Vector128<float>>(inputBuffer);
+            Span<Vector128<float>> outputVec = MemoryMarshal.Cast<float, Vector128<float>>(outputBuffer);
+
+            int sisdStart = inputVec.Length * 4;
+
+            for (int i = 0; i < inputVec.Length; i++)
+            {
+                outputVec[i] = AdvSimd.Ceiling(AdvSimd.Multiply(inputVec[i], volumeVec));
+            }
+
+            for (int i = sisdStart; i < inputBuffer.Length; i++)
+            {
+                outputBuffer[i] = FloatingPointHelper.MultiplyRoundUp(inputBuffer[i], Volume);
+            }
+        }
+
         private void ProcessVolume(Span<float> outputBuffer, ReadOnlySpan<float> inputBuffer)
         {
             if (Avx.IsSupported)
@@ -99,6 +120,10 @@ namespace Ryujinx.Audio.Renderer.Dsp.Command
             {
                 ProcessVolumeSse41(outputBuffer, inputBuffer);
             }
+            else if (AdvSimd.IsSupported)
+            {
+                ProcessVolumeAdvSimd(outputBuffer, inputBuffer);
+            }
             else
             {
                 ProcessVolumeSlowPath(outputBuffer, inputBuffer);