PR #611: Bzip input stream simple vectorization

* Added benchmark for BZip2 decompression.
* Updated benchmarks to be run also on .NET Core 3.1.
* Simple automatic vectorization of the rotation loop.
* Added comment describing vectorization.

Author: konrad-kruczynski

benchmark/ICSharpCode.SharpZipLib.Benchmark/BZip2/BZip2InputStream.cs

@@ -0,0 +1,37 @@
+using System;
+using System.IO;
+using BenchmarkDotNet.Attributes;
+
+namespace ICSharpCode.SharpZipLib.Benchmark.BZip2
+{
+	[Config(typeof(MultipleRuntimes))]
+	public class BZip2InputStream
+	{
+		private byte[] compressedData;
+
+		public BZip2InputStream()
+		{
+			var outputMemoryStream = new MemoryStream();
+			using (var outputStream = new SharpZipLib.BZip2.BZip2OutputStream(outputMemoryStream))
+			{
+				var random = new Random(1234);
+				var inputData = new byte[1024 * 1024 * 30];
+				random.NextBytes(inputData);
+				var inputMemoryStream = new MemoryStream(inputData);
+				inputMemoryStream.CopyTo(outputStream);
+			}
+
+			compressedData = outputMemoryStream.ToArray();
+		}
+
+		[Benchmark]
+		public void DecompressData()
+		{
+			var memoryStream = new MemoryStream(compressedData);
+			using (var inputStream = new SharpZipLib.BZip2.BZip2InputStream(memoryStream))
+			{
+				inputStream.CopyTo(Stream.Null);
+			}
+		}
+	}
+}

benchmark/ICSharpCode.SharpZipLib.Benchmark/ICSharpCode.SharpZipLib.Benchmark.csproj

@@ -2,7 +2,7 @@
 
   <PropertyGroup>
     <OutputType>Exe</OutputType>
-    <TargetFrameworks>netcoreapp2.1;net461</TargetFrameworks>
+    <TargetFrameworks>netcoreapp2.1;netcoreapp3.1;net461</TargetFrameworks>
   </PropertyGroup>
 
 	<ItemGroup>

benchmark/ICSharpCode.SharpZipLib.Benchmark/Program.cs

@@ -13,7 +13,7 @@ public MultipleRuntimes()
 		{
 			AddJob(Job.Default.WithToolchain(CsProjClassicNetToolchain.Net461).AsBaseline()); // NET 4.6.1
 			AddJob(Job.Default.WithToolchain(CsProjCoreToolchain.NetCoreApp21)); // .NET Core 2.1
-			//Add(Job.Default.With(CsProjCoreToolchain.NetCoreApp30)); // .NET Core 3.0
+			AddJob(Job.Default.WithToolchain(CsProjCoreToolchain.NetCoreApp31)); // .NET Core 3.1
 		}
 	}
 

src/ICSharpCode.SharpZipLib/BZip2/BZip2InputStream.cs

@@ -19,7 +19,11 @@ public class BZip2InputStream : Stream
 		private const int NO_RAND_PART_B_STATE = 6;
 		private const int NO_RAND_PART_C_STATE = 7;
 
-		#endregion Constants
+#if NETSTANDARD2_1
+		private static readonly int VectorSize = System.Numerics.Vector<byte>.Count;
+#endif
+
+#endregion Constants
 
 		#region Instance Fields
 
@@ -711,10 +715,27 @@ cache misses.
 					unzftab[seqToUnseq[tmp]]++;
 					ll8[last] = seqToUnseq[tmp];
 
-					for (int j = nextSym - 1; j > 0; --j)
+					var j = nextSym - 1;
+
+#if !NETSTANDARD2_0 && !NETFRAMEWORK
+					// This is vectorized memory move. Going from the back, we're taking chunks of array
+					// and write them at the new location shifted by one. Since chunks are VectorSize long,
+					// at the end we have to move "tail" (or head actually) of the array using a plain loop.
+					// If System.Numerics.Vector API is not available, the plain loop is used to do the whole copying.
+
+					while(j >= VectorSize)
 					{
-						yy[j] = yy[j - 1];
+						var arrayPart = new System.Numerics.Vector<byte>(yy, j - VectorSize);
+						arrayPart.CopyTo(yy, j - VectorSize + 1);
+						j -= VectorSize;
 					}
+#endif
+
+					while(j > 0)
+					{
+						yy[j] = yy[--j];
+					}
+
 					yy[0] = tmp;
 
 					if (groupPos == 0)