From af7d96d21462e6e080488f7d1933d1430834fadd Mon Sep 17 00:00:00 2001
From: Anton Firszov <antonfir@gmail.com>
Date: Mon, 15 Oct 2018 01:11:59 +0200
Subject: [PATCH] SIMD byte -> float conversion:
 BulkConvertByteToNormalizedFloatFast

---
 src/ImageSharp/Common/Extensions/SimdUtils.cs | 45 +++++++++++++++++--
 .../Color/Bulk/ToVector4.cs                   | 16 +++++--
 .../ImageSharp.Benchmarks.csproj              |  2 +-
 .../PixelFormats/PixelOperationsTests.cs      | 24 ++++++++++
 .../Tests/TestEnvironmentTests.cs             |  2 +
 5 files changed, 81 insertions(+), 8 deletions(-)

diff --git a/src/ImageSharp/Common/Extensions/SimdUtils.cs b/src/ImageSharp/Common/Extensions/SimdUtils.cs
index db1e80dda..56118a764 100644
--- a/src/ImageSharp/Common/Extensions/SimdUtils.cs
+++ b/src/ImageSharp/Common/Extensions/SimdUtils.cs
@@ -105,13 +105,50 @@ namespace SixLabors.ImageSharp
             }
         }
 
-        internal static void BulkConvertByteToNormalizedFloat(ReadOnlySpan<byte> source, Span<float> dest)
+        /// <summary>
+        /// Fast <see cref="byte"/> -> <see cref="float"/> conversion for RyuJIT runtimes having dotnet/coreclr#10662 merged.
+        /// <see>
+        ///     <cref>https://github.com/dotnet/coreclr/pull/10662</cref>
+        /// </see>
+        /// </summary>
+        internal static void BulkConvertByteToNormalizedFloatFast(ReadOnlySpan<byte> source, Span<float> dest)
         {
-            if (!Vector.IsHardwareAccelerated)
+            Guard.IsTrue(
+                source.Length % Vector<byte>.Count == 0,
+                nameof(source),
+                "dest.Length should be divisable by Vector<byte>.Count!");
+
+            int n = source.Length / Vector<byte>.Count;
+
+            ref Vector<byte> sourceBase = ref Unsafe.As<byte, Vector<byte>>(ref MemoryMarshal.GetReference(source));
+            ref Vector<float> destBase = ref Unsafe.As<float, Vector<float>>(ref MemoryMarshal.GetReference(dest));
+
+            var scale = new Vector<float>(1f / 255f);
+
+            for (int i = 0; i < n; i++)
             {
-                throw new InvalidOperationException(
-                    "Rgba32.PixelOperations.ToVector4SimdAligned() should not be called when Vector.IsHardwareAccelerated == false!");
+                Vector<byte> b = Unsafe.Add(ref sourceBase, i);
+
+                Vector.Widen(b, out Vector<ushort> s0, out Vector<ushort> s1);
+                Vector.Widen(s0, out Vector<uint> w0, out Vector<uint> w1);
+                Vector.Widen(s1, out Vector<uint> w2, out Vector<uint> w3);
+
+                Vector<float> f0 = Vector.ConvertToSingle(w0) * scale;
+                Vector<float> f1 = Vector.ConvertToSingle(w1) * scale;
+                Vector<float> f2 = Vector.ConvertToSingle(w2) * scale;
+                Vector<float> f3 = Vector.ConvertToSingle(w3) * scale;
+
+                ref Vector<float> d = ref Unsafe.Add(ref destBase, i * 4);
+                d = f0;
+                Unsafe.Add(ref d, 1) = f1;
+                Unsafe.Add(ref d, 2) = f2;
+                Unsafe.Add(ref d, 3) = f3;
             }
+        }
+
+        internal static void BulkConvertByteToNormalizedFloat(ReadOnlySpan<byte> source, Span<float> dest)
+        {
+            GuardAvx2(nameof(BulkConvertByteToNormalizedFloat));
 
             DebugGuard.IsTrue((dest.Length % Vector<float>.Count) == 0, nameof(source), "dest.Length should be divisable by Vector<float>.Count!");
 
diff --git a/tests/ImageSharp.Benchmarks/Color/Bulk/ToVector4.cs b/tests/ImageSharp.Benchmarks/Color/Bulk/ToVector4.cs
index 7b6f902d8..0e5e9d94f 100644
--- a/tests/ImageSharp.Benchmarks/Color/Bulk/ToVector4.cs
+++ b/tests/ImageSharp.Benchmarks/Color/Bulk/ToVector4.cs
@@ -28,7 +28,7 @@ namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk
 
         [Params(
             //64, 
-            1024)]
+            2048)]
         public int Count { get; set; }
 
         [GlobalSetup]
@@ -72,14 +72,14 @@ namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk
     }
 
     [CoreJob]
-    //[ClrJob]
+    [ClrJob]
     public class ToVector4_Rgba32 : ToVector4<Rgba32>
     {
         class Config : ManualConfig
         {
         }
 
-        [Benchmark]
+        //[Benchmark]
         public void BulkConvertByteToNormalizedFloat()
         {
             Span<byte> sBytes = MemoryMarshal.Cast<Rgba32, byte>(this.source.GetSpan());
@@ -87,5 +87,15 @@ namespace SixLabors.ImageSharp.Benchmarks.ColorSpaces.Bulk
 
             SimdUtils.BulkConvertByteToNormalizedFloat(sBytes, dFloats);
         }
+
+        [Benchmark]
+        public void BulkConvertByteToNormalizedFloatFast()
+        {
+            Span<byte> sBytes = MemoryMarshal.Cast<Rgba32, byte>(this.source.GetSpan());
+            Span<float> dFloats = MemoryMarshal.Cast<Vector4, float>(this.destination.GetSpan());
+
+            SimdUtils.BulkConvertByteToNormalizedFloatFast(sBytes, dFloats);
+        }
+
     }
 }
\ No newline at end of file
diff --git a/tests/ImageSharp.Benchmarks/ImageSharp.Benchmarks.csproj b/tests/ImageSharp.Benchmarks/ImageSharp.Benchmarks.csproj
index 36b7d4db4..e470e7821 100644
--- a/tests/ImageSharp.Benchmarks/ImageSharp.Benchmarks.csproj
+++ b/tests/ImageSharp.Benchmarks/ImageSharp.Benchmarks.csproj
@@ -1,6 +1,6 @@
 <Project Sdk="Microsoft.NET.Sdk">
   <PropertyGroup>
-    <TargetFrameworks>netcoreapp2.0;net461</TargetFrameworks>
+    <TargetFrameworks>netcoreapp2.1;net461</TargetFrameworks>
     <OutputType>Exe</OutputType>
     <AllowUnsafeBlocks>True</AllowUnsafeBlocks>
     <RootNamespace>SixLabors.ImageSharp.Benchmarks</RootNamespace>
diff --git a/tests/ImageSharp.Tests/PixelFormats/PixelOperationsTests.cs b/tests/ImageSharp.Tests/PixelFormats/PixelOperationsTests.cs
index a96da03e7..2e84886c0 100644
--- a/tests/ImageSharp.Tests/PixelFormats/PixelOperationsTests.cs
+++ b/tests/ImageSharp.Tests/PixelFormats/PixelOperationsTests.cs
@@ -73,6 +73,30 @@ namespace SixLabors.ImageSharp.Tests.PixelFormats
                     );
             }
 
+            [Fact]
+            public void BulkConvertByteToNormalizedFloatFast()
+            {
+                if (!Vector.IsHardwareAccelerated)
+                {
+                    return;
+                }
+
+                ImageSharp.PixelFormats.Rgba32[] source = CreatePixelTestData(128);
+                Vector4[] expected = CreateExpectedVector4Data(source);
+
+                TestOperation(
+                    source,
+                    expected,
+                    (s, d) =>
+                        {
+                            ReadOnlySpan<byte> sBytes = MemoryMarshal.Cast<ImageSharp.PixelFormats.Rgba32, byte>(s);
+                            Span<float> dFloats = MemoryMarshal.Cast<Vector4, float>(d.Memory.Span);
+
+                            SimdUtils.BulkConvertByteToNormalizedFloatFast(sBytes, dFloats);
+                        }
+                );
+            }
+
 
             // [Fact] // Profiling benchmark - enable manually!
 #pragma warning disable xUnit1013 // Public method should be marked as test
diff --git a/tests/ImageSharp.Tests/TestUtilities/Tests/TestEnvironmentTests.cs b/tests/ImageSharp.Tests/TestUtilities/Tests/TestEnvironmentTests.cs
index 8a3e69059..30bb16c2a 100644
--- a/tests/ImageSharp.Tests/TestUtilities/Tests/TestEnvironmentTests.cs
+++ b/tests/ImageSharp.Tests/TestUtilities/Tests/TestEnvironmentTests.cs
@@ -3,6 +3,8 @@
 
 using System;
 using System.IO;
+using System.Reflection;
+using System.Runtime.InteropServices;
 
 using SixLabors.ImageSharp.Common.Helpers;
 using SixLabors.ImageSharp.Formats;