/// <summary>Executes a loop in chunks, using up to the given number of threads.</summary>
/// <param name="maxParallelism">The maximum number of threads to use to execute the loop.</param>
/// <include file="documentation.xml" path="/Utilities/Tasks/ParallelForChunked/*"/>
public static void ParallelFor(int start, int endExclusive, Action <int, int, LoopThreadInfo> body, int maxParallelism)
{
if (start > endExclusive || maxParallelism <= 0)
{
throw new ArgumentOutOfRangeException();
}
if (body == null)
{
throw new ArgumentNullException();
}
uint iterations = (uint)(endExclusive - start), parallelism = Math.Min((uint)maxParallelism, iterations);
if (iterations == 0)
{
return;
}
if (parallelism == 1)
{
body(start, endExclusive, new LoopThreadInfo(0));
}
else
{
WorkItemTask[] tasks = new WorkItemTask[parallelism];
for (uint basicChunkSize = iterations / parallelism, errorInc = iterations % parallelism, error = 0, i = 0; i < (uint)tasks.Length; i++)
{
int threadNumber = (int)i, chunkStart = start; // make a copy of these so they can be properly captured by the closure
uint chunkSize = basicChunkSize;
// say we're dividing 100 iterations over 6 threads. then each chunk should be 16.666... iterations in length. since we
// can't have a fractional iteration, we'll truncate the chunk size (to 16) and keep track of an error value. the error
// will be incremented by the fractional part of the ideal chunk size (0.666...). if the error is greater than or equal
// to 0.5, then we'll add one to the chunk size and subtract 1 from the error. since we don't want to use floating point
// math, due to its inaccuracy, we'll do the same thing with integers. the error increment will be the integer remainder
// from the calculation of the chunk size (so 100/6 leaves a remainder of 4). and instead of comparing against 0.5, we'll
// compare double the error against the divisor (parallelism). this is the same as comparing the error against half the
// divisor, except that it avoids truncation error. unfortunately, i can't really prove that the algorithm works for all
// combinations of iterations and thread counts, but it seems to, and i'm showing my trust in it by not adding a special
// case for the last chunk that simply takes up the remaining iterations.
error += errorInc;
if (error * 2 >= parallelism)
{
chunkSize++;
error -= parallelism;
}
tasks[i] = new WorkItemTask(task =>
{
LoopThreadInfo info = new LoopThreadInfo(threadNumber);
body(chunkStart, chunkStart + (int)chunkSize, info);
});
start += (int)chunkSize;
}
new CompositeTask(tasks).Run();
}
}
/// <summary>Executes a loop using up to the given number of threads.</summary>
/// <param name="maxParallelism">The maximum number of threads to use to execute the loop.</param>
/// <include file="documentation.xml" path="/Utilities/Tasks/ParallelFor/*[not(@name='body')]"/>
/// <include file="documentation.xml" path="/Utilities/Tasks/ParallelForWithInit/*"/>
public static void ParallelFor <T>(int start, int endExclusive, Func <LoopThreadInfo, T> threadInitializer,
Action <int, T, LoopThreadInfo> body, int maxParallelism)
{
if (start > endExclusive || maxParallelism <= 0)
{
throw new ArgumentOutOfRangeException();
}
if (threadInitializer == null || body == null)
{
throw new ArgumentNullException();
}
int iterations = endExclusive - start;
if (iterations == 0)
{
return;
}
if (maxParallelism > iterations)
{
maxParallelism = iterations;
}
if (maxParallelism == 1)
{
LoopThreadInfo info = new LoopThreadInfo(0);
T value = threadInitializer(info);
for (; start < endExclusive; start++)
{
body(start, value, info);
}
}
else
{
WorkItemTask[] tasks = new WorkItemTask[maxParallelism];
for (int i = 0; i < tasks.Length; i++)
{
int threadNumber = i;
tasks[i] = new WorkItemTask(task =>
{
LoopThreadInfo info = new LoopThreadInfo(threadNumber);
T value = threadInitializer(info);
// get an available index
int currentIndex;
while (true)
{
do
{
currentIndex = start;
if (currentIndex == endExclusive)
{
goto done;
}
} while(Interlocked.CompareExchange(ref start, currentIndex + 1, currentIndex) != currentIndex);
body(currentIndex, value, info);
}
done:;
});
}
new CompositeTask(tasks).Run();
}
}