amadeus: Improve and fix delay effect processing (#3205)
* amadeus: Improve and fix delay effect processing This rework the delay effect processing by representing calculation with the appropriate matrix and by unrolling some loop in the code. This allows better optimization by the JIT while making it more readeable. Also fix a bug in the Surround code path found while looking back at my notes. * Remove useless GetHashCode * Address gdkchan's comments
This commit is contained in:
Mary
GitHub
parent
55ee261363
commit
d04ba51bb0
6 changed files with 363 additions and 66 deletions
|
|
@ -18,8 +18,10 @@
|
|||
using Ryujinx.Audio.Renderer.Dsp.State;
|
||||
using Ryujinx.Audio.Renderer.Parameter.Effect;
|
||||
using Ryujinx.Audio.Renderer.Server.Effect;
|
||||
using Ryujinx.Audio.Renderer.Utils.Math;
|
||||
using System;
|
||||
using System.Diagnostics;
|
||||
using System.Numerics;
|
||||
using System.Runtime.CompilerServices;
|
||||
|
||||
namespace Ryujinx.Audio.Renderer.Dsp.Command
|
||||
|
|
@ -70,7 +72,7 @@ namespace Ryujinx.Audio.Renderer.Dsp.Command
|
|||
DataSourceHelper.RemapChannelResourceMappingToLegacy(newEffectChannelMappingSupported, OutputBufferIndices);
|
||||
}
|
||||
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining)]
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
|
||||
private unsafe void ProcessDelayMono(ref DelayState state, float* outputBuffer, float* inputBuffer, uint sampleCount)
|
||||
{
|
||||
float feedbackGain = FixedPointHelper.ToFloat(Parameter.FeedbackGain, FixedPointPrecision);
|
||||
|
|
@ -83,133 +85,148 @@ namespace Ryujinx.Audio.Renderer.Dsp.Command
|
|||
float input = inputBuffer[i] * 64;
|
||||
float delayLineValue = state.DelayLines[0].Read();
|
||||
|
||||
float lowPassResult = (input * inGain + delayLineValue * feedbackGain) * state.LowPassBaseGain + state.LowPassZ[0] * state.LowPassFeedbackGain;
|
||||
float temp = input * inGain + delayLineValue * feedbackGain;
|
||||
|
||||
state.LowPassZ[0] = lowPassResult;
|
||||
|
||||
state.DelayLines[0].Update(lowPassResult);
|
||||
state.UpdateLowPassFilter(ref temp, 1);
|
||||
|
||||
outputBuffer[i] = (input * dryGain + delayLineValue * outGain) / 64;
|
||||
}
|
||||
}
|
||||
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining)]
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
|
||||
private unsafe void ProcessDelayStereo(ref DelayState state, Span<IntPtr> outputBuffers, ReadOnlySpan<IntPtr> inputBuffers, uint sampleCount)
|
||||
{
|
||||
const ushort channelCount = 2;
|
||||
|
||||
Span<float> channelInput = stackalloc float[channelCount];
|
||||
Span<float> delayLineValues = stackalloc float[channelCount];
|
||||
Span<float> temp = stackalloc float[channelCount];
|
||||
|
||||
float delayFeedbackBaseGain = state.DelayFeedbackBaseGain;
|
||||
float delayFeedbackCrossGain = state.DelayFeedbackCrossGain;
|
||||
float inGain = FixedPointHelper.ToFloat(Parameter.InGain, FixedPointPrecision);
|
||||
float dryGain = FixedPointHelper.ToFloat(Parameter.DryGain, FixedPointPrecision);
|
||||
float outGain = FixedPointHelper.ToFloat(Parameter.OutGain, FixedPointPrecision);
|
||||
|
||||
Matrix2x2 delayFeedback = new Matrix2x2(delayFeedbackBaseGain , delayFeedbackCrossGain,
|
||||
delayFeedbackCrossGain, delayFeedbackBaseGain);
|
||||
|
||||
for (int i = 0; i < sampleCount; i++)
|
||||
{
|
||||
for (int j = 0; j < channelCount; j++)
|
||||
Vector2 channelInput = new Vector2
|
||||
{
|
||||
channelInput[j] = *((float*)inputBuffers[j] + i) * 64;
|
||||
delayLineValues[j] = state.DelayLines[j].Read();
|
||||
}
|
||||
X = *((float*)inputBuffers[0] + i) * 64,
|
||||
Y = *((float*)inputBuffers[1] + i) * 64,
|
||||
};
|
||||
|
||||
temp[0] = channelInput[0] * inGain + delayLineValues[1] * delayFeedbackCrossGain + delayLineValues[0] * delayFeedbackBaseGain;
|
||||
temp[1] = channelInput[1] * inGain + delayLineValues[0] * delayFeedbackCrossGain + delayLineValues[1] * delayFeedbackBaseGain;
|
||||
|
||||
for (int j = 0; j < channelCount; j++)
|
||||
Vector2 delayLineValues = new Vector2()
|
||||
{
|
||||
float lowPassResult = state.LowPassFeedbackGain * state.LowPassZ[j] + temp[j] * state.LowPassBaseGain;
|
||||
X = state.DelayLines[0].Read(),
|
||||
Y = state.DelayLines[1].Read(),
|
||||
};
|
||||
|
||||
state.LowPassZ[j] = lowPassResult;
|
||||
state.DelayLines[j].Update(lowPassResult);
|
||||
Vector2 temp = MatrixHelper.Transform(ref channelInput, ref delayFeedback) + channelInput * inGain;
|
||||
|
||||
*((float*)outputBuffers[j] + i) = (channelInput[j] * dryGain + delayLineValues[j] * outGain) / 64;
|
||||
}
|
||||
state.UpdateLowPassFilter(ref Unsafe.As<Vector2, float>(ref temp), channelCount);
|
||||
|
||||
*((float*)outputBuffers[0] + i) = (channelInput.X * dryGain + delayLineValues.X * outGain) / 64;
|
||||
*((float*)outputBuffers[1] + i) = (channelInput.Y * dryGain + delayLineValues.Y * outGain) / 64;
|
||||
}
|
||||
}
|
||||
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining)]
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
|
||||
private unsafe void ProcessDelayQuadraphonic(ref DelayState state, Span<IntPtr> outputBuffers, ReadOnlySpan<IntPtr> inputBuffers, uint sampleCount)
|
||||
{
|
||||
const ushort channelCount = 4;
|
||||
|
||||
Span<float> channelInput = stackalloc float[channelCount];
|
||||
Span<float> delayLineValues = stackalloc float[channelCount];
|
||||
Span<float> temp = stackalloc float[channelCount];
|
||||
|
||||
float delayFeedbackBaseGain = state.DelayFeedbackBaseGain;
|
||||
float delayFeedbackCrossGain = state.DelayFeedbackCrossGain;
|
||||
float inGain = FixedPointHelper.ToFloat(Parameter.InGain, FixedPointPrecision);
|
||||
float dryGain = FixedPointHelper.ToFloat(Parameter.DryGain, FixedPointPrecision);
|
||||
float outGain = FixedPointHelper.ToFloat(Parameter.OutGain, FixedPointPrecision);
|
||||
|
||||
Matrix4x4 delayFeedback = new Matrix4x4(delayFeedbackBaseGain , delayFeedbackCrossGain, delayFeedbackCrossGain, 0.0f,
|
||||
delayFeedbackCrossGain, delayFeedbackBaseGain , 0.0f , delayFeedbackCrossGain,
|
||||
delayFeedbackCrossGain, 0.0f , delayFeedbackBaseGain , delayFeedbackCrossGain,
|
||||
0.0f , delayFeedbackCrossGain, delayFeedbackCrossGain, delayFeedbackBaseGain);
|
||||
|
||||
|
||||
for (int i = 0; i < sampleCount; i++)
|
||||
{
|
||||
for (int j = 0; j < channelCount; j++)
|
||||
Vector4 channelInput = new Vector4
|
||||
{
|
||||
channelInput[j] = *((float*)inputBuffers[j] + i) * 64;
|
||||
delayLineValues[j] = state.DelayLines[j].Read();
|
||||
}
|
||||
X = *((float*)inputBuffers[0] + i) * 64,
|
||||
Y = *((float*)inputBuffers[1] + i) * 64,
|
||||
Z = *((float*)inputBuffers[2] + i) * 64,
|
||||
W = *((float*)inputBuffers[3] + i) * 64
|
||||
};
|
||||
|
||||
temp[0] = channelInput[0] * inGain + (delayLineValues[2] + delayLineValues[1]) * delayFeedbackCrossGain + delayLineValues[0] * delayFeedbackBaseGain;
|
||||
temp[1] = channelInput[1] * inGain + (delayLineValues[0] + delayLineValues[3]) * delayFeedbackCrossGain + delayLineValues[1] * delayFeedbackBaseGain;
|
||||
temp[2] = channelInput[2] * inGain + (delayLineValues[3] + delayLineValues[0]) * delayFeedbackCrossGain + delayLineValues[2] * delayFeedbackBaseGain;
|
||||
temp[3] = channelInput[3] * inGain + (delayLineValues[1] + delayLineValues[2]) * delayFeedbackCrossGain + delayLineValues[3] * delayFeedbackBaseGain;
|
||||
|
||||
for (int j = 0; j < channelCount; j++)
|
||||
Vector4 delayLineValues = new Vector4()
|
||||
{
|
||||
float lowPassResult = state.LowPassFeedbackGain * state.LowPassZ[j] + temp[j] * state.LowPassBaseGain;
|
||||
X = state.DelayLines[0].Read(),
|
||||
Y = state.DelayLines[1].Read(),
|
||||
Z = state.DelayLines[2].Read(),
|
||||
W = state.DelayLines[3].Read()
|
||||
};
|
||||
|
||||
state.LowPassZ[j] = lowPassResult;
|
||||
state.DelayLines[j].Update(lowPassResult);
|
||||
Vector4 temp = MatrixHelper.Transform(ref channelInput, ref delayFeedback) + channelInput * inGain;
|
||||
|
||||
state.UpdateLowPassFilter(ref Unsafe.As<Vector4, float>(ref temp), channelCount);
|
||||
|
||||
*((float*)outputBuffers[j] + i) = (channelInput[j] * dryGain + delayLineValues[j] * outGain) / 64;
|
||||
}
|
||||
*((float*)outputBuffers[0] + i) = (channelInput.X * dryGain + delayLineValues.X * outGain) / 64;
|
||||
*((float*)outputBuffers[1] + i) = (channelInput.Y * dryGain + delayLineValues.Y * outGain) / 64;
|
||||
*((float*)outputBuffers[2] + i) = (channelInput.Z * dryGain + delayLineValues.Z * outGain) / 64;
|
||||
*((float*)outputBuffers[3] + i) = (channelInput.W * dryGain + delayLineValues.W * outGain) / 64;
|
||||
}
|
||||
}
|
||||
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining)]
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
|
||||
private unsafe void ProcessDelaySurround(ref DelayState state, Span<IntPtr> outputBuffers, ReadOnlySpan<IntPtr> inputBuffers, uint sampleCount)
|
||||
{
|
||||
const ushort channelCount = 6;
|
||||
|
||||
Span<float> channelInput = stackalloc float[channelCount];
|
||||
Span<float> delayLineValues = stackalloc float[channelCount];
|
||||
Span<float> temp = stackalloc float[channelCount];
|
||||
|
||||
float feedbackGain = FixedPointHelper.ToFloat(Parameter.FeedbackGain, FixedPointPrecision);
|
||||
float delayFeedbackBaseGain = state.DelayFeedbackBaseGain;
|
||||
float delayFeedbackCrossGain = state.DelayFeedbackCrossGain;
|
||||
float inGain = FixedPointHelper.ToFloat(Parameter.InGain, FixedPointPrecision);
|
||||
float dryGain = FixedPointHelper.ToFloat(Parameter.DryGain, FixedPointPrecision);
|
||||
float outGain = FixedPointHelper.ToFloat(Parameter.OutGain, FixedPointPrecision);
|
||||
|
||||
Matrix6x6 delayFeedback = new Matrix6x6(delayFeedbackBaseGain , 0.0f , 0.0f , 0.0f , delayFeedbackCrossGain, delayFeedbackCrossGain,
|
||||
0.0f , delayFeedbackBaseGain , 0.0f , delayFeedbackCrossGain, delayFeedbackCrossGain, 0.0f ,
|
||||
delayFeedbackCrossGain, 0.0f , delayFeedbackBaseGain , delayFeedbackCrossGain, 0.0f , 0.0f ,
|
||||
0.0f , delayFeedbackCrossGain, delayFeedbackCrossGain, delayFeedbackBaseGain , 0.0f , 0.0f ,
|
||||
delayFeedbackCrossGain, delayFeedbackCrossGain, 0.0f , 0.0f , delayFeedbackBaseGain , 0.0f ,
|
||||
0.0f , 0.0f , 0.0f , 0.0f , 0.0f , feedbackGain);
|
||||
|
||||
for (int i = 0; i < sampleCount; i++)
|
||||
{
|
||||
for (int j = 0; j < channelCount; j++)
|
||||
Vector6 channelInput = new Vector6
|
||||
{
|
||||
channelInput[j] = *((float*)inputBuffers[j] + i) * 64;
|
||||
delayLineValues[j] = state.DelayLines[j].Read();
|
||||
}
|
||||
X = *((float*)inputBuffers[0] + i) * 64,
|
||||
Y = *((float*)inputBuffers[1] + i) * 64,
|
||||
Z = *((float*)inputBuffers[2] + i) * 64,
|
||||
W = *((float*)inputBuffers[3] + i) * 64,
|
||||
V = *((float*)inputBuffers[4] + i) * 64,
|
||||
U = *((float*)inputBuffers[5] + i) * 64
|
||||
};
|
||||
|
||||
temp[0] = channelInput[0] * inGain + (delayLineValues[2] + delayLineValues[4]) * delayFeedbackCrossGain + delayLineValues[0] * delayFeedbackBaseGain;
|
||||
temp[1] = channelInput[1] * inGain + (delayLineValues[4] + delayLineValues[3]) * delayFeedbackCrossGain + delayLineValues[1] * delayFeedbackBaseGain;
|
||||
temp[2] = channelInput[2] * inGain + (delayLineValues[3] + delayLineValues[0]) * delayFeedbackCrossGain + delayLineValues[2] * delayFeedbackBaseGain;
|
||||
temp[3] = channelInput[3] * inGain + (delayLineValues[1] + delayLineValues[2]) * delayFeedbackCrossGain + delayLineValues[3] * delayFeedbackBaseGain;
|
||||
temp[4] = channelInput[4] * inGain + (delayLineValues[0] + delayLineValues[1]) * delayFeedbackCrossGain + delayLineValues[4] * delayFeedbackBaseGain;
|
||||
temp[5] = channelInput[5] * inGain + delayLineValues[5] * delayFeedbackBaseGain;
|
||||
|
||||
for (int j = 0; j < channelCount; j++)
|
||||
Vector6 delayLineValues = new Vector6
|
||||
{
|
||||
float lowPassResult = state.LowPassFeedbackGain * state.LowPassZ[j] + temp[j] * state.LowPassBaseGain;
|
||||
X = state.DelayLines[0].Read(),
|
||||
Y = state.DelayLines[1].Read(),
|
||||
Z = state.DelayLines[2].Read(),
|
||||
W = state.DelayLines[3].Read(),
|
||||
V = state.DelayLines[4].Read(),
|
||||
U = state.DelayLines[5].Read()
|
||||
};
|
||||
|
||||
state.LowPassZ[j] = lowPassResult;
|
||||
state.DelayLines[j].Update(lowPassResult);
|
||||
Vector6 temp = MatrixHelper.Transform(ref channelInput, ref delayFeedback) + channelInput * inGain;
|
||||
|
||||
*((float*)outputBuffers[j] + i) = (channelInput[j] * dryGain + delayLineValues[j] * outGain) / 64;
|
||||
}
|
||||
state.UpdateLowPassFilter(ref Unsafe.As<Vector6, float>(ref temp), channelCount);
|
||||
|
||||
*((float*)outputBuffers[0] + i) = (channelInput.X * dryGain + delayLineValues.X * outGain) / 64;
|
||||
*((float*)outputBuffers[1] + i) = (channelInput.Y * dryGain + delayLineValues.Y * outGain) / 64;
|
||||
*((float*)outputBuffers[2] + i) = (channelInput.Z * dryGain + delayLineValues.Z * outGain) / 64;
|
||||
*((float*)outputBuffers[3] + i) = (channelInput.W * dryGain + delayLineValues.W * outGain) / 64;
|
||||
*((float*)outputBuffers[4] + i) = (channelInput.V * dryGain + delayLineValues.V * outGain) / 64;
|
||||
*((float*)outputBuffers[5] + i) = (channelInput.U * dryGain + delayLineValues.U * outGain) / 64;
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue