//--------------------------------------------------------------------------
//
// Copyright (c) Microsoft Corporation. All rights reserved.
//
// File: ParallelAlgorithms_Scan.cs
//
//--------------------------------------------------------------------------
using System.Collections.Concurrent;
using System.Collections.Generic;
using System.Linq;
using System.Threading.Tasks;
namespace System.Threading.Algorithms
{
public static partial class ParallelAlgorithms
{
/// <summary>Computes a parallel prefix scan over the source enumerable using the specified function.</summary>
/// <typeparam name="T">The type of the data in the source.</typeparam>
/// <param name="source">The source data over which a prefix scan should be computed.</param>
/// <param name="function">The function to use for the scan.</param>
/// <returns>The results of the scan operation.</returns>
/// <remarks>
/// For very small functions, such as additions, an implementation targeted
/// at the relevant type and operation will perform significantly better than
/// this generalized implementation.
/// </remarks>
public static T[] Scan<T>(IEnumerable<T> source, Func<T, T, T> function)
{
return Scan(source, function, loadBalance: false);
}
/// <summary>Computes a parallel prefix scan over the source enumerable using the specified function.</summary>
/// <typeparam name="T">The type of the data in the source.</typeparam>
/// <param name="source">The source data over which a prefix scan should be computed.</param>
/// <param name="function">The function to use for the scan.</param>
/// <param name="loadBalance">Whether to load-balance during process.</param>
/// <returns>The results of the scan operation.</returns>
/// <remarks>
/// For very small functions, such as additions, an implementation targeted
/// at the relevant type and operation will perform significantly better than
/// this generalized implementation.
/// </remarks>
public static T[] Scan<T>(IEnumerable<T> source, Func<T, T, T> function, bool loadBalance)
{
// Validate arguments
if (source == null) throw new ArgumentNullException("source");
// Create output copy
var output = source.ToArray();
// Do the prefix scan in-place on the copy and return the results
ScanInPlace(output, function, loadBalance);
return output;
}
/// <summary>Computes a parallel prefix scan in-place on an array using the specified function.</summary>
/// <typeparam name="T">The type of the data in the source.</typeparam>
/// <param name="data">The data over which a prefix scan should be computed. Upon exit, stores the results.</param>
/// <param name="function">The function to use for the scan.</param>
/// <returns>The results of the scan operation.</returns>
/// <remarks>
/// For very small functions, such as additions, an implementation targeted
/// at the relevant type and operation will perform significantly better than
/// this generalized implementation.
/// </remarks>
public static void ScanInPlace<T>(T[] data, Func<T, T, T> function)
{
ScanInPlace(data, function, loadBalance:false);
}
/// <summary>Computes a parallel prefix scan in-place on an array using the specified function.</summary>
/// <typeparam name="T">The type of the data in the source.</typeparam>
/// <param name="data">The data over which a prefix scan should be computed. Upon exit, stores the results.</param>
/// <param name="function">The function to use for the scan.</param>
/// <param name="loadBalance">Whether to load-balance during process.</param>
/// <returns>The results of the scan operation.</returns>
/// <remarks>
/// For very small functions, such as additions, an implementation targeted
/// at the relevant type and operation will perform significantly better than
/// this generalized implementation.
/// </remarks>
public static void ScanInPlace<T>(T [] data, Func<T, T, T> function, bool loadBalance)
{
// Validate arguments
if (data == null) throw new ArgumentNullException("data");
if (function == null) throw new ArgumentNullException("function");
// Do the prefix scan in-place and return the results. This implementation
// of parallel prefix scan ends up executing the function twice as many
// times as the sequential implementation. Thus, only if we have more than 2 cores
// will the parallel version have a chance of running faster than sequential.
if (Environment.ProcessorCount <= 2)
{
InclusiveScanInPlaceSerial(data, function, 0, data.Length, 1);
}
else if (loadBalance)
{
InclusiveScanInPlaceWithLoadBalancingParallel(data, function, 0, data.Length, 1);
}
else // parallel, non-loadbalance
{
InclusiveScanInPlaceParallel(data, function);
}
}
/// <summary>Computes a sequential prefix scan over the array using the specified function.</summary>
/// <typeparam name="T">The type of the data in the array.</typeparam>
/// <param name="arr">The data, which will be overwritten with the computed prefix scan.</param>
/// <param name="function">The function to use for the scan.</param>
/// <param name="arrStart">The start of the data in arr over which the scan is being computed.</param>
/// <param name="arrLength">The length of the data in arr over which the scan is being computed.</param>
/// <param name="skip">The inclusive distance between elements over which the scan is being computed.</param>
/// <remarks>No parameter validation is performed.</remarks>
private static void InclusiveScanInPlaceSerial<T>(T[] arr, Func<T, T, T> function, int arrStart, int arrLength, int skip)
{
int count = 0;
for (int i = arrStart; count < arrLength - 1; i += skip, count++)
{
arr[i + skip] = function(arr[i], arr[i + skip]);
}
}
/// <summary>Computes a sequential exclusive prefix scan over the array using the specified function.</summary>
/// <param name="arr">The data, which will be overwritten with the computed prefix scan.</param>
/// <param name="function">The function to use for the scan.</param>
/// <param name="lowerBoundInclusive">The inclusive lower bound of the array at which to start the scan.</param>
/// <param name="upperBoundExclusive">The exclusive upper bound of the array at which to end the scan.</param>
public static void ExclusiveScanInPlaceSerial<T>(T[] arr, Func<T, T, T> function, int lowerBoundInclusive, int upperBoundExclusive)
{
T total = arr[lowerBoundInclusive];
arr[lowerBoundInclusive] = default(T);
for (int i = lowerBoundInclusive + 1; i < upperBoundExclusive; i++)
{
T prevTotal = total;
total = function(total, arr[i]);
arr[i] = prevTotal;
}
}
/// <summary>Computes a parallel prefix scan over the array using the specified function.</summary>
/// <typeparam name="T">The type of the data in the array.</typeparam>
/// <param name="arr">The data, which will be overwritten with the computed prefix scan.</param>
/// <param name="function">The function to use for the scan.</param>
/// <param name="arrStart">The start of the data in arr over which the scan is being computed.</param>
/// <param name="arrLength">The length of the data in arr over which the scan is being computed.</param>
/// <param name="skip">The inclusive distance between elements over which the scan is being computed.</param>
/// <remarks>No parameter validation is performed.</remarks>
private static void InclusiveScanInPlaceWithLoadBalancingParallel<T>(T[] arr, Func<T, T, T> function,
int arrStart, int arrLength, int skip)
{
// If the length is 0 or 1, just return a copy of the original array.
if (arrLength <= 1) return;
int halfInputLength = arrLength / 2;
// Pairwise combine. Use static partitioning, as the function
// is likely to be very small.
Parallel.For(0, halfInputLength, i =>
{
int loc = arrStart + (i * 2 * skip);
arr[loc + skip] = function(arr[loc], arr[loc + skip]);
});
// Recursively prefix scan the pairwise computations.
InclusiveScanInPlaceWithLoadBalancingParallel(arr, function, arrStart + skip, halfInputLength, skip * 2);
// Generate output. As before, use static partitioning.
Parallel.For(0, (arrLength % 2) == 0 ? halfInputLength - 1 : halfInputLength, i =>
{
int loc = arrStart + (i * 2 * skip) + skip;
arr[loc + skip] = function(arr[loc], arr[loc + skip]);
});
}
/// <summary>Computes a parallel inclusive prefix scan over the array using the specified function.</summary>
public static void InclusiveScanInPlaceParallel<T>(T[] arr, Func<T, T, T> function)
{
int procCount = Environment.ProcessorCount;
T[] intermediatePartials = new T[procCount];
using (var phaseBarrier = new Barrier(procCount,
_ => ExclusiveScanInPlaceSerial(intermediatePartials, function, 0, intermediatePartials.Length)))
{
// Compute the size of each range
int rangeSize = arr.Length / procCount;
int nextRangeStart = 0;
// Create, store, and wait on all of the tasks
var tasks = new Task[procCount];
for (int i = 0; i < procCount; i++, nextRangeStart += rangeSize)
{
// Get the range for each task, then start it
int rangeNum = i;
int lowerRangeInclusive = nextRangeStart;
int upperRangeExclusive = i < procCount - 1 ? nextRangeStart + rangeSize : arr.Length;
tasks[rangeNum] = Task.Factory.StartNew(() =>
{
// Phase 1: Prefix scan assigned range, and copy upper bound to intermediate partials
InclusiveScanInPlaceSerial(arr, function, lowerRangeInclusive, upperRangeExclusive, 1);
intermediatePartials[rangeNum] = arr[upperRangeExclusive - 1];
// Phase 2: One thread only should prefix scan the intermediaries... done implicitly by the barrier
phaseBarrier.SignalAndWait();
// Phase 3: Incorporate partials
if (rangeNum != 0)
{
for (int j = lowerRangeInclusive; j < upperRangeExclusive; j++)
{
arr[j] = function(intermediatePartials[rangeNum], arr[j]);
}
}
});
}
// Wait for all of the tasks to complete
Task.WaitAll(tasks);
}
}
}
}
Submit an article or tip