perf: ⚡envelope caching with MemoryCache and waveform utils improve…

…ments (#24) * perf: ⚡ more specialized waveform mixing * perf: ⚡ envelope caching with `MemoryCache`
kahojyun · Jul 1, 2023 · 5070ce2 · 5070ce2
1 parent 2c28940
commit 5070ce2
Show file tree

Hide file tree

Showing 21 changed files with 947 additions and 201 deletions.
diff --git a/Qynit.Pulsewave.sln b/Qynit.Pulsewave.sln
@@ -18,7 +18,9 @@ Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "tests", "tests", "{3B8B8B3E
 EndProject
 Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "examples", "examples", "{6741804E-962A-4762-B289-CD4899A62AD8}"
 EndProject
-Project("{FAE04EC0-301F-11D3-BF4B-00C04F79EFBC}") = "WaveGenDemo", "examples\WaveGenDemo\WaveGenDemo.csproj", "{B83CB224-4C00-4011-BF6A-C76C7D786908}"
+Project("{9A19103F-16F7-4668-BE54-9A1E7A4F7556}") = "WaveGenDemo", "examples\WaveGenDemo\WaveGenDemo.csproj", "{B83CB224-4C00-4011-BF6A-C76C7D786908}"
+EndProject
+Project("{FAE04EC0-301F-11D3-BF4B-00C04F79EFBC}") = "WaveGenBenchmarks", "examples\WaveGenBenchmarks\WaveGenBenchmarks.csproj", "{ECE2579D-700C-4BB3-9009-F35395177B7A}"
 EndProject
 Global
 	GlobalSection(SolutionConfigurationPlatforms) = preSolution
@@ -38,6 +40,10 @@ Global
 		{B83CB224-4C00-4011-BF6A-C76C7D786908}.Debug|Any CPU.Build.0 = Debug|Any CPU
 		{B83CB224-4C00-4011-BF6A-C76C7D786908}.Release|Any CPU.ActiveCfg = Release|Any CPU
 		{B83CB224-4C00-4011-BF6A-C76C7D786908}.Release|Any CPU.Build.0 = Release|Any CPU
+		{ECE2579D-700C-4BB3-9009-F35395177B7A}.Debug|Any CPU.ActiveCfg = Debug|Any CPU
+		{ECE2579D-700C-4BB3-9009-F35395177B7A}.Debug|Any CPU.Build.0 = Debug|Any CPU
+		{ECE2579D-700C-4BB3-9009-F35395177B7A}.Release|Any CPU.ActiveCfg = Release|Any CPU
+		{ECE2579D-700C-4BB3-9009-F35395177B7A}.Release|Any CPU.Build.0 = Release|Any CPU
 	EndGlobalSection
 	GlobalSection(SolutionProperties) = preSolution
 		HideSolutionNode = FALSE
@@ -46,6 +52,7 @@ Global
 		{8A71F7BA-E700-4DA3-BBE7-5E7C866466B8} = {3EBAD22F-230A-4BCB-B16C-A9063E2A4E9A}
 		{33659EDC-29F9-45DB-A9FA-E01E648BE4B9} = {3B8B8B3E-9095-4A7E-9309-661439C30D9A}
 		{B83CB224-4C00-4011-BF6A-C76C7D786908} = {6741804E-962A-4762-B289-CD4899A62AD8}
+		{ECE2579D-700C-4BB3-9009-F35395177B7A} = {6741804E-962A-4762-B289-CD4899A62AD8}
 	EndGlobalSection
 	GlobalSection(ExtensibilityGlobals) = postSolution
 		SolutionGuid = {1DC52AEE-D9FC-4B2F-9492-0C92B0479A3C}

diff --git a/WaveGenBenchmarks.WaveformUtilsBench-report-github.md b/WaveGenBenchmarks.WaveformUtilsBench-report-github.md
@@ -0,0 +1,43 @@
+``` ini
+
+BenchmarkDotNet=v0.13.5, OS=Windows 11 (10.0.22621.1848/22H2/2022Update/SunValley2)
+AMD Ryzen 5 5600, 1 CPU, 12 logical and 6 physical cores
+.NET SDK=7.0.304
+  [Host]     : .NET 7.0.7 (7.0.723.27404), X64 RyuJIT AVX2
+  DefaultJob : .NET 7.0.7 (7.0.723.27404), X64 RyuJIT AVX2
+
+
+```
+|                  Method | Length |        Mean |     Error |    StdDev |      Median | Ratio | RatioSD |
+|------------------------ |------- |------------:|----------:|----------:|------------:|------:|--------:|
+|           **MixAddPlateau** |     **16** |    **12.94 ns** |  **0.027 ns** |  **0.023 ns** |    **12.95 ns** |  **0.25** |    **0.00** |
+|  MixAddPlateauFrequency |     16 |    28.13 ns |  0.090 ns |  0.080 ns |    28.13 ns |  0.54 |    0.00 |
+|                  MixAdd |     16 |    20.65 ns |  0.047 ns |  0.042 ns |    20.66 ns |  0.40 |    0.00 |
+|         MixAddFrequency |     16 |    38.46 ns |  0.113 ns |  0.088 ns |    38.49 ns |  0.74 |    0.00 |
+|          MixAddWithDrag |     16 |    29.16 ns |  0.143 ns |  0.134 ns |    29.16 ns |  0.56 |    0.00 |
+| MixAddFrequencyWithDrag |     16 |    51.62 ns |  0.134 ns |  0.119 ns |    51.61 ns |  1.00 |    0.00 |
+|                  Simple |     16 |    63.72 ns |  0.322 ns |  0.301 ns |    63.59 ns |  1.23 |    0.01 |
+|                         |        |             |           |           |             |       |         |
+|           **MixAddPlateau** |     **64** |    **42.69 ns** |  **0.113 ns** |  **0.105 ns** |    **42.71 ns** |  **0.53** |    **0.01** |
+|  MixAddPlateauFrequency |     64 |    46.73 ns |  0.218 ns |  0.193 ns |    46.74 ns |  0.58 |    0.01 |
+|                  MixAdd |     64 |    37.26 ns |  0.101 ns |  0.095 ns |    37.27 ns |  0.46 |    0.00 |
+|         MixAddFrequency |     64 |    62.61 ns |  0.192 ns |  0.160 ns |    62.61 ns |  0.78 |    0.01 |
+|          MixAddWithDrag |     64 |    51.32 ns |  0.466 ns |  0.363 ns |    51.19 ns |  0.64 |    0.01 |
+| MixAddFrequencyWithDrag |     64 |    80.17 ns |  0.779 ns |  0.691 ns |    80.01 ns |  1.00 |    0.00 |
+|                  Simple |     64 |   217.39 ns |  4.377 ns |  7.892 ns |   213.14 ns |  2.78 |    0.13 |
+|                         |        |             |           |           |             |       |         |
+|           **MixAddPlateau** |    **256** |    **72.68 ns** |  **0.321 ns** |  **0.300 ns** |    **72.73 ns** |  **0.37** |    **0.00** |
+|  MixAddPlateauFrequency |    256 |   122.84 ns |  0.210 ns |  0.197 ns |   122.80 ns |  0.62 |    0.00 |
+|                  MixAdd |    256 |    74.62 ns |  0.266 ns |  0.249 ns |    74.60 ns |  0.38 |    0.00 |
+|         MixAddFrequency |    256 |   163.94 ns |  0.327 ns |  0.306 ns |   164.00 ns |  0.83 |    0.01 |
+|          MixAddWithDrag |    256 |   141.93 ns |  0.305 ns |  0.285 ns |   141.89 ns |  0.72 |    0.00 |
+| MixAddFrequencyWithDrag |    256 |   196.62 ns |  1.133 ns |  1.060 ns |   196.16 ns |  1.00 |    0.00 |
+|                  Simple |    256 |   782.64 ns |  4.104 ns |  3.839 ns |   781.45 ns |  3.98 |    0.03 |
+|                         |        |             |           |           |             |       |         |
+|           **MixAddPlateau** |   **1024** |   **277.91 ns** |  **0.331 ns** |  **0.294 ns** |   **277.91 ns** |  **0.40** |    **0.00** |
+|  MixAddPlateauFrequency |   1024 |   424.67 ns |  0.439 ns |  0.389 ns |   424.60 ns |  0.61 |    0.00 |
+|                  MixAdd |   1024 |   274.94 ns |  1.182 ns |  1.048 ns |   275.11 ns |  0.39 |    0.00 |
+|         MixAddFrequency |   1024 |   566.09 ns |  1.301 ns |  1.086 ns |   566.40 ns |  0.81 |    0.00 |
+|          MixAddWithDrag |   1024 |   534.09 ns |  2.571 ns |  2.405 ns |   533.32 ns |  0.76 |    0.00 |
+| MixAddFrequencyWithDrag |   1024 |   701.37 ns |  2.718 ns |  2.270 ns |   701.43 ns |  1.00 |    0.00 |
+|                  Simple |   1024 | 3,061.59 ns | 12.202 ns | 10.189 ns | 3,060.17 ns |  4.37 |    0.02 |
diff --git a/examples/WaveGenBenchmarks/Program.cs b/examples/WaveGenBenchmarks/Program.cs
@@ -0,0 +1,5 @@
+using BenchmarkDotNet.Running;
+
+using WaveGenBenchmarks;
+
+BenchmarkRunner.Run<WaveformUtilsBench>();
diff --git a/examples/WaveGenBenchmarks/WaveGenBenchmarks.csproj b/examples/WaveGenBenchmarks/WaveGenBenchmarks.csproj
@@ -0,0 +1,18 @@
+<Project Sdk="Microsoft.NET.Sdk">
+
+  <PropertyGroup>
+    <OutputType>Exe</OutputType>
+    <TargetFramework>net7.0</TargetFramework>
+    <ImplicitUsings>enable</ImplicitUsings>
+    <Nullable>enable</Nullable>
+  </PropertyGroup>
+
+  <ItemGroup>
+    <PackageReference Include="BenchmarkDotNet" Version="0.13.5" />
+  </ItemGroup>
+
+  <ItemGroup>
+    <ProjectReference Include="..\..\src\Qynit.Pulsewave\Qynit.Pulsewave.csproj" />
+  </ItemGroup>
+
+</Project>
diff --git a/examples/WaveGenBenchmarks/WaveformUtilsBench.cs b/examples/WaveGenBenchmarks/WaveformUtilsBench.cs
@@ -0,0 +1,97 @@
+using System.Numerics;
+
+using BenchmarkDotNet.Attributes;
+
+using Qynit.Pulsewave;
+
+namespace WaveGenBenchmarks;
+public class WaveformUtilsBench
+{
+    [Params(16, 64, 256, 1024)]
+    public int Length { get; set; }
+    private PooledComplexArray<double>? Source { get; set; }
+    private PooledComplexArray<double>? Target { get; set; }
+    private IqPair<double> Amplitude { get; } = new IqPair<double>(1, 1);
+    private IqPair<double> DragAmplitude { get; } = new IqPair<double>(1, 1);
+    private double DPhase { get; } = Math.PI / 11;
+
+    [GlobalSetup]
+    public void Init()
+    {
+        Source = new PooledComplexArray<double>(Length, true);
+        Source.DataI.Fill(0.125);
+        Source.DataQ.Fill(-0.125);
+        Target = new PooledComplexArray<double>(Length, true);
+    }
+
+    [Benchmark]
+    public void MixAddPlateau()
+    {
+        WaveformUtils.MixAddPlateau(Target!, Amplitude);
+    }
+
+    [Benchmark]
+    public void MixAddPlateauFrequency()
+    {
+        WaveformUtils.MixAddPlateauFrequency(Target!, Amplitude, DPhase);
+    }
+
+    [Benchmark]
+    public void MixAdd()
+    {
+        WaveformUtils.MixAdd(Target!, Source!, Amplitude);
+    }
+
+    [Benchmark]
+    public void MixAddFrequency()
+    {
+        WaveformUtils.MixAddFrequency(Target!, Source!, Amplitude, DPhase);
+    }
+
+    [Benchmark]
+    public void MixAddWithDrag()
+    {
+        WaveformUtils.MixAddWithDrag(Target!, Source!, Amplitude, DragAmplitude);
+    }
+
+    [Benchmark(Baseline = true)]
+    public void MixAddFrequencyWithDrag()
+    {
+        WaveformUtils.MixAddFrequencyWithDrag(Target!, Source!, Amplitude, DragAmplitude, DPhase);
+    }
+
+    [Benchmark]
+    public void Simple()
+    {
+        MixAddWithDragSimple(Target!, Source!, Amplitude, DragAmplitude, DPhase);
+    }
+
+    private static void MixAddWithDragSimple<T>(ComplexSpan<T> target, ComplexReadOnlySpan<T> source, IqPair<T> amplitude, IqPair<T> dragAmplitude, T dPhase)
+       where T : unmanaged, IFloatingPointIeee754<T>
+    {
+        var length = source.Length;
+        var sourceI = source.DataI;
+        var sourceQ = source.DataQ;
+        var targetI = target.DataI;
+        var targetQ = target.DataQ;
+
+        var carrier = amplitude;
+        var dragCarrier = dragAmplitude;
+        var phaser = IqPair<T>.FromPolarCoordinates(T.One, dPhase);
+        for (var i = 0; i < length; i++)
+        {
+            var diff = i switch
+            {
+                0 => new IqPair<T>(sourceI[i + 1], sourceQ[i + 1]) - new IqPair<T>(sourceI[i], sourceQ[i]),
+                _ when i == length - 1 => new IqPair<T>(sourceI[i], sourceQ[i]) - new IqPair<T>(sourceI[i - 1], sourceQ[i - 1]),
+                _ => (new IqPair<T>(sourceI[i + 1], sourceQ[i + 1]) - new IqPair<T>(sourceI[i - 1], sourceQ[i - 1])) * T.CreateChecked(0.5),
+            };
+            var sourceIq = new IqPair<T>(sourceI[i], sourceQ[i]);
+            var totalIq = sourceIq * carrier + diff * dragCarrier;
+            targetI[i] += totalIq.I;
+            targetQ[i] += totalIq.Q;
+            carrier *= phaser;
+            dragCarrier *= phaser;
+        }
+    }
+}
diff --git a/src/Qynit.Pulsewave/ComplexArrayReadOnlySpan.cs b/src/Qynit.Pulsewave/ComplexArrayReadOnlySpan.cs
diff --git a/src/Qynit.Pulsewave/ComplexArraySpan.cs b/src/Qynit.Pulsewave/ComplexArraySpan.cs
diff --git a/src/Qynit.Pulsewave/ComplexReadOnlySpan.cs b/src/Qynit.Pulsewave/ComplexReadOnlySpan.cs
@@ -0,0 +1,44 @@
+using System.Diagnostics;
+using System.Runtime.CompilerServices;
+
+namespace Qynit.Pulsewave;
+public readonly ref struct ComplexReadOnlySpan<T>
+    where T : unmanaged
+{
+    public ReadOnlySpan<T> DataI { get; }
+    public ReadOnlySpan<T> DataQ { get; }
+    public int Length => DataI.Length;
+    public bool IsEmpty => Length == 0;
+    internal ComplexReadOnlySpan(ReadOnlySpan<T> dataI, ReadOnlySpan<T> dataQ)
+    {
+        Debug.Assert(dataI.Length == dataQ.Length);
+        DataI = dataI;
+        DataQ = dataQ;
+    }
+    public static implicit operator ComplexReadOnlySpan<T>(PooledComplexArray<T> source)
+    {
+        return new ComplexReadOnlySpan<T>(source.DataI, source.DataQ);
+    }
+    public static implicit operator ComplexReadOnlySpan<T>(ComplexSpan<T> source)
+    {
+        return new ComplexReadOnlySpan<T>(source.DataI, source.DataQ);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public ComplexReadOnlySpan<T> Slice(int start)
+    {
+        return new ComplexReadOnlySpan<T>(DataI[start..], DataQ[start..]);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public ComplexReadOnlySpan<T> Slice(int start, int length)
+    {
+        return new ComplexReadOnlySpan<T>(DataI.Slice(start, length), DataQ.Slice(start, length));
+    }
+
+    public void CopyTo(ComplexSpan<T> destination)
+    {
+        DataI.CopyTo(destination.DataI);
+        DataQ.CopyTo(destination.DataQ);
+    }
+}
diff --git a/src/Qynit.Pulsewave/ComplexSpan.cs b/src/Qynit.Pulsewave/ComplexSpan.cs
@@ -0,0 +1,49 @@
+using System.Diagnostics;
+using System.Runtime.CompilerServices;
+
+namespace Qynit.Pulsewave;
+public readonly ref struct ComplexSpan<T>
+    where T : unmanaged
+{
+    public Span<T> DataI { get; }
+    public Span<T> DataQ { get; }
+    public int Length => DataI.Length;
+    public bool IsEmpty => Length == 0;
+    internal ComplexSpan(Span<T> dataI, Span<T> dataQ)
+    {
+        Debug.Assert(dataI.Length == dataQ.Length);
+        DataI = dataI;
+        DataQ = dataQ;
+    }
+    public static implicit operator ComplexSpan<T>(PooledComplexArray<T> source)
+    {
+        return new ComplexSpan<T>(source.DataI, source.DataQ);
+    }
+
+    public void Fill(T i, T q)
+    {
+        DataI.Fill(i);
+        DataQ.Fill(q);
+    }
+    public void Clear()
+    {
+        DataI.Clear();
+        DataQ.Clear();
+    }
+    public void CopyTo(ComplexSpan<T> destination)
+    {
+        ((ComplexReadOnlySpan<T>)this).CopyTo(destination);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public ComplexSpan<T> Slice(int start)
+    {
+        return new ComplexSpan<T>(DataI[start..], DataQ[start..]);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public ComplexSpan<T> Slice(int start, int length)
+    {
+        return new ComplexSpan<T>(DataI.Slice(start, length), DataQ.Slice(start, length));
+    }
+}
diff --git a/src/Qynit.Pulsewave/Envelope.cs b/src/Qynit.Pulsewave/Envelope.cs
@@ -6,6 +6,8 @@ public readonly record struct Envelope
     public IPulseShape? Shape { get; }
     public double Width { get; }
     public double Plateau { get; }
+    public bool IsRectangle => Shape is null;
+    public bool IsZero => Width == 0 && Plateau == 0;
     public Envelope(IPulseShape? shape, double width, double plateau)
     {
         Guard.IsGreaterThanOrEqualTo(width, 0);

diff --git a/src/Qynit.Pulsewave/EnvelopeCacheKey.cs b/src/Qynit.Pulsewave/EnvelopeCacheKey.cs
@@ -0,0 +1,2 @@
+namespace Qynit.Pulsewave;
+internal record EnvelopeCacheKey(EnvelopeInfo EnvelopeInfo, Envelope Envelope);
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1,2 @@
		namespace Qynit.Pulsewave;
		internal record EnvelopeCacheKey(EnvelopeInfo EnvelopeInfo, Envelope Envelope);