Misaki.HighPerformance/Misaki.HighPerformance.Mathematics.SPMD/Templates/IJobSPMD.tt

<#@ template debug="false" hostspecific="false" language="C#" #>
<#@ assembly name="System.Core" #>
<#@ import namespace="System.Linq" #>
<#@ import namespace="System.Text" #>
<#@ import namespace="System.Collections.Generic" #>
<#@ output extension="gen.cs" #>
using Misaki.HighPerformance.Jobs;
using System.Numerics;

namespace Misaki.HighPerformance.Mathematics.SPMD;

<#
const string TLane = "TLane";
const string TNumber = "TNumber";
const string GenericParameters = $"{TLane}, {TNumber}";

var TLaneRestrictions = $@"where {TLane} : ISPMDLane<{TLane}, {TNumber}>";
var TNumberRestrictions = $@"where {TNumber} : unmanaged, INumber<{TNumber}>, IBinaryNumber<{TNumber}>, IMinMaxValue<{TNumber}>, IBitwiseOperators<{TNumber}, {TNumber}, {TNumber}>";

for (var i = 0; i < 8; i++) { #>
/// <summary>
/// A job interface for Single Program Multiple Data (SPMD) execution, allowing for efficient parallel processing of data across multiple lanes.
/// </summary>
/// <remarks>
/// Always use TNumber0 as the primary type for determining lane width and job scheduling, even if it's not used in the job execution.
/// </remarks>
<#= ForEachDimension(i + 1, j => @$"/// <typeparam name=""TNumber{j}"">The first numeric type used in the SPMD job.</typeparam>", Environment.NewLine) #>
public interface IJobSPMD<<#= ForEachDimension(i + 1, j => $"TNumber{j}") #>>
<#= GetTNumberRestrictions(i + 1) #>
{
    void Execute<<#= ForEachDimension(i + 1, j => $"TLane{j}") #>>(TLane0 indices, TLane0 mask, ref readonly JobExecutionContext ctx)
<#= GetTLaneRestrictions(i + 1, "        ") #>;
}

internal struct SPMDJobWrapper<T, <#= ForEachDimension(i + 1, j => $"TNumber{j}") #>> : IJobParallelFor
    where T : IJobSPMD<<#= ForEachDimension(i + 1, j => $"TNumber{j}") #>>
<#= GetTNumberRestrictions(i + 1) #>
{
    public T innerJob;
    public int totalIteration;

    public void Execute(int loopIndex, ref readonly JobExecutionContext ctx)
    {
        var baseIndex = loopIndex * WideLane<TNumber0>.LaneWidth;
        var indices = WideLane<TNumber0>.Sequence(TNumber0.CreateTruncating(baseIndex), TNumber0.One);
        var mask = indices < TNumber0.CreateTruncating(totalIteration);

        innerJob.Execute<<#= ForEachDimension(i + 1, j => $"WideLane<TNumber{j}>") #>>(indices, mask, in ctx);
    }
}

internal struct SPMDScalerJobWrapper<T, <#= ForEachDimension(i + 1, j => $"TNumber{j}") #>> : IJobParallelFor
    where T : IJobSPMD<<#= ForEachDimension(i + 1, j => $"TNumber{j}") #>>
<#= GetTNumberRestrictions(i + 1) #>
{
    public T innerJob;
    public int totalIteration;

    public void Execute(int loopIndex, ref readonly JobExecutionContext ctx)
    {
        innerJob.Execute<<#= ForEachDimension(i + 1, j => $"ScalarLane<TNumber{j}>") #>>(TNumber0.CreateTruncating(loopIndex), ScalarLane<TNumber0>.AllBitsSet, in ctx);
    }
}

<# } #>

public static class IJobParallelForSPMDExtensions
{
<# for (var i = 0; i < 8; i++) { #>
    /// <summary>
    /// Run the SPMD job with the specified total count and job execution context directly on the calling thread.
    /// </summary>
    /// <remarks>
    /// Always use TNumber0 as the primary type for determining lane width and job scheduling, even if it's not used in the job execution.
    /// </remarks>
<#= ForEachDimension(i + 1, j => @$"    /// <typeparam name=""TNumber{j}"">The first numeric type used in the SPMD job.</typeparam>", Environment.NewLine) #>
    /// <param name="job">The SPMD job to run.</param>
    /// <param name="totalIteration">The total number of iterations to execute across all lanes.</param>
    /// <param name="ctx">The job execution context providing information about the current execution environment.</param>
    public static void Run<T, <#= ForEachDimension(i + 1, j => $"TNumber{j}") #>>(this T job, int totalIteration, ref readonly JobExecutionContext ctx)
        where T : IJobSPMD<<#= ForEachDimension(i + 1, j => $"TNumber{j}") #>>
<#= GetTNumberRestrictions(i + 1, "        ") #>
    {
        if (WideLane.IsSupported)
        {
            var iterations = (totalIteration + WideLane<TNumber0>.LaneWidth - 1) / WideLane<TNumber0>.LaneWidth;
            for (var loopIndex = 0; loopIndex < iterations; loopIndex++)
            {
                var baseIndex = loopIndex * WideLane<TNumber0>.LaneWidth;
                var indices = WideLane<TNumber0>.Sequence(TNumber0.CreateTruncating(baseIndex), TNumber0.One);
                var mask = indices < TNumber0.CreateTruncating(totalIteration);

                job.Execute<<#= ForEachDimension(i + 1, j => $"WideLane<TNumber{j}>") #>>(indices, mask, in ctx);
            }
        }
        else
        {
            for (var loopIndex = 0; loopIndex < totalIteration; loopIndex++)
            {
                job.Execute<<#= ForEachDimension(i + 1, j => $"ScalarLane<TNumber{j}>") #>>(TNumber0.CreateTruncating(loopIndex), ScalarLane<TNumber0>.AllBitsSet, in ctx);
            }
        }
    }

    /// <summary>
    /// Schedule the SPMD job for parallel execution across multiple threads, with the specified total count, batch size, and job execution context.
    /// </summary>
<#= ForEachDimension(i + 1, j => @$"    /// <typeparam name=""TNumber{j}"">The first numeric type used in the SPMD job.</typeparam>", Environment.NewLine) #>
    /// <remarks>
    /// Always use TNumber0 as the primary type for determining lane width and job scheduling, even if it's not used in the job execution.
    /// </remarks>
    /// <param name="jobScheduler">The job scheduler to use for scheduling the job.</param>
    /// <param name="job">The SPMD job to schedule.</param>
    /// <param name="totalIteration">The total number of iterations to execute across all lanes.</param>
    /// <param name="batchSize">The number of iterations to execute in each batch for parallel execution.</param>
    /// <param name="preferLocal">Whether to prefer scheduling the job on the local thread for better cache locality.</param>
    /// <param name="priority">The priority of the job.</param>
    /// <param name="dependencies">Any job handles that this job depends on, which must complete before this job can start.</param>
    public static JobHandle ScheduleParallelSPDM<T, <#= ForEachDimension(i + 1, j => $"TNumber{j}") #>>(this JobScheduler jobScheduler, ref T job, int totalIteration, int batchSize, bool preferLocal, JobPriority priority, params ReadOnlySpan<JobHandle> dependencies)
        where T : IJobSPMD<<#= ForEachDimension(i + 1, j => $"TNumber{j}") #>>
<#= GetTNumberRestrictions(i + 1, "        ") #>
    {
        if (WideLane.IsSupported)
        {
            var warper = new SPMDJobWrapper<T, <#= ForEachDimension(i + 1, j => $"TNumber{j}") #>>
            {
                innerJob = job,
                totalIteration = totalIteration,
            };

            var iterations = (totalIteration + WideLane<TNumber0>.LaneWidth - 1) / WideLane<TNumber0>.LaneWidth;
            return jobScheduler.ScheduleParallelFor(ref warper, iterations, batchSize, preferLocal, priority, dependencies);
        }
        else
        {
            var warper = new SPMDScalerJobWrapper<T, <#= ForEachDimension(i + 1, j => $"TNumber{j}") #>>
            {
                innerJob = job,
                totalIteration = totalIteration,
            };

            return jobScheduler.ScheduleParallelFor(ref warper, totalIteration, batchSize, preferLocal, priority, dependencies);
        }
    }

<# } #>
}

<#+
public string ForEachDimension(int dimension, Func<int, string> action, string spliter = ", ")
{
    return string.Join(spliter, Enumerable.Range(0, dimension).Select(i => action(i)));
}

public string GetTNumberRestrictions(int dimension, string space = "    ")
{
    var sb = new StringBuilder();
    for (var i = 0; i < dimension; i++)
    {
        sb.Append(space + $@"where TNumber{i} : unmanaged, INumber<TNumber{i}>, IBinaryNumber<TNumber{i}>, IMinMaxValue<TNumber{i}>, IBitwiseOperators<TNumber{i}, TNumber{i}, TNumber{i}>");
        if (i < dimension - 1)
        {
            sb.AppendLine();
        }
    }

    return sb.ToString();
}

public string GetTLaneRestrictions(int dimension, string space = "    ")
{
    var sb = new StringBuilder();
    for (var i = 0; i < dimension; i++)
    {
        sb.Append(space + $@"where TLane{i} : unmanaged, ISPMDLane<TLane{i}, TNumber{i}>");
        if (i < dimension - 1)
        {
            sb.AppendLine();
        }
    }

    return sb.ToString();
}
#>