gistfile1.cs

using System;
using System.Diagnostics;
using Mono.Simd;
using System.Runtime.CompilerServices;

namespace VectorizedQueryTest
{
    class Program
    {
        static bool MinimalSimdSupport()
        {
            return SimdRuntime.IsMethodAccelerated (typeof (Vector4f), "op_Addition") &&
                   SimdRuntime.IsMethodAccelerated (typeof (Vector4f), "op_Multiply") &&
                   SimdRuntime.IsMethodAccelerated (typeof (VectorOperations), "Shuffle", typeof (Vector4f), typeof (ShuffleSel));
        }

        static bool EnhancedSimdSupport()
        {
            return SimdRuntime.IsMethodAccelerated (typeof (VectorOperations), "HorizontalAdd", typeof (Vector4f), typeof (Vector4f)) &&
                   SimdRuntime.IsMethodAccelerated (typeof (Vector4f), "op_Multiply");
        }

        static void Main(string[] args)
        {
           // Console.WriteLine("Minimal SIMD Support: " + MinimalSimdSupport());
           // Console.WriteLine("Enhanched SIMD Support: " + EnhancedSimdSupport());


            // SELECT * FROM Table WHERE ColumnA < ColumnB
            int rowCount = 10000000; // 10M
            var rand = new Random();
            var rowids = new int[rowCount];
            var columna = new int[rowids.Length];
            var columnb = new int[rowids.Length];
            var result = new int[rowids.Length + 1];
            var pos = 0;

            for (int i = 0; i < rowids.Length; i++)
            {
                rowids[i] = i;
                columna[i] = rand.Next();
                columnb[i] = rand.Next();
            }

            var columnbVector = new Vector4i();
            var columnaVector = new Vector4i();
            var rowIdBlock = new int[4];

            // Start of performance measuring.
            var watch = new Stopwatch();
            watch.Start();


            for (int i = 0; i < rowids.Length; i += 4)
            {
//                columnaVector.X = columna[i];
//                columnaVector.Y = columna[i + 1];
//                columnaVector.Z = columna[i + 2];
//                columnaVector.W = columna[i + 3];
                  columnaVector = columna.GetVector(i);

//                columnbVector.X = columnb[i];
//                columnbVector.Y = columnb[i + 1];
//                columnbVector.Z = columnb[i + 2];
//                columnbVector.W = columnb[i + 3];
                  columnbVector = columnb.GetVector(i);

                var mask = columnbVector.CompareGreaterThan(columnaVector);
                // To compare greater than or equals, compare equals, compare greater than, and then and the two masks together. Use bitwise OR.

//                rowIdBlock[0] = rowids[i];
//                rowIdBlock[1] = rowids[i + 1];
//                rowIdBlock[2] = rowids[i + 2];
//                rowIdBlock[3] = rowids[i + 3];
//                SIMD_Process(mask, rowids, i, result, ref pos);
            if ((mask.X | mask.Y | mask.Z | mask.W) < 0) // Profiler says 4.8% spent here.
            {
//                for (int j = 0; j < 4; j++)
//                {
                    var tmp = mask[0] & 1; // Profiler says 10.3% spent here.
                    result[pos] = rowids[i]; // Profiler says 8.7% spent here.
                    pos += tmp; // Profiler says 5.6% spent here.

                    tmp = mask[1] & 1; // Profiler says 10.3% spent here.
                    result[pos] = rowids[i + 1]; // Profiler says 8.7% spent here.
                    pos += tmp; // Profiler says 5.6% spent here.

                    tmp = mask[2] & 1; // Profiler says 10.3% spent here.
                    result[pos] = rowids[i + 2]; // Profiler says 8.7% spent here.
                    pos += tmp; // Profiler says 5.6% spent here.

                    tmp = mask[3] & 1; // Profiler says 10.3% spent here.
                    result[pos] = rowids[i + 3]; // Profiler says 8.7% spent here.
                    pos += tmp; // Profiler says 5.6% spent here.
//                }
            } 



            }
            result[pos] = 0;
            watch.Stop(); // 114 milliseconds to process 10M rows.

//            Debug.WriteLine(watch.Elapsed);
            Console.WriteLine(watch.Elapsed);
//            Console.ReadKey();
        }

        [MethodImpl(MethodImplOptions.AggressiveInlining)]
        public static void SIMD_Process(Vector4i mask, int[] rowids, int rowidsStart, int[] result, ref int pos)
        {
            if ((mask.X | mask.Y | mask.Z | mask.W) < 0) // Profiler says 4.8% spent here.
            {
//                for (int j = 0; j < 4; j++)
//                {
                    var tmp = mask[0] & 1; // Profiler says 10.3% spent here.
                    result[pos] = rowids[rowidsStart]; // Profiler says 8.7% spent here.
                    pos += tmp; // Profiler says 5.6% spent here.

                    tmp = mask[1] & 1; // Profiler says 10.3% spent here.
                    result[pos] = rowids[rowidsStart + 1]; // Profiler says 8.7% spent here.
                    pos += tmp; // Profiler says 5.6% spent here.

                    tmp = mask[2] & 1; // Profiler says 10.3% spent here.
                    result[pos] = rowids[rowidsStart + 2]; // Profiler says 8.7% spent here.
                    pos += tmp; // Profiler says 5.6% spent here.

                    tmp = mask[3] & 1; // Profiler says 10.3% spent here.
                    result[pos] = rowids[rowidsStart + 3]; // Profiler says 8.7% spent here.
                    pos += tmp; // Profiler says 5.6% spent here.
//                }
            } 
        }
    }
}