public void Dgemv(bool transposeA, double alpha, DoubleMatrix2D A, DoubleMatrix1D x, double beta, DoubleMatrix1D y)
{
/*
* split A, as follows:
*
* x x
* x
* x
* A
* xxx x y
* xxx x
* --- -
* xxx x
* xxx x
* --- -
* xxx x
*
*/
if (transposeA)
{
Dgemv(false, alpha, A.ViewDice(), x, beta, y);
return;
}
int m = A.Rows;
int n = A.Columns;
long flops = 2L * m * n;
int noOfTasks = (int)System.Math.Min(flops / 30000, this.maxThreads); // each thread should process at least 30000 flops
int width = A.Rows;
noOfTasks = System.Math.Min(width, noOfTasks);
if (noOfTasks < 2)
{ // parallelization doesn't pay off (too much start up overhead)
seqBlas.Dgemv(transposeA, alpha, A, x, beta, y);
return;
}
// set up concurrent tasks
int span = width / noOfTasks;
//FJTask[] subTasks = new FJTask[noOfTasks];
for (int i = 0; i < noOfTasks; i++)
{
int offset = i * span;
if (i == noOfTasks - 1)
{
span = width - span * i; // last span may be a bit larger
}
// split A along rows into blocks
DoubleMatrix2D AA = A.ViewPart(offset, 0, span, n);
DoubleMatrix1D yy = y.ViewPart(offset, span);
// subTasks[i] = new FJTask()
// {
// public void run()
// {
// seqBlas.Dgemv(transposeA, alpha, AA, x, beta, yy);
// //Console.WriteLine("Hello "+offset);
// }
//};
Action task = (() =>
{
seqBlas.Dgemv(transposeA, alpha, AA, x, beta, yy);
});
// run tasks and wait for completion
try
{
this.smp.TaskGroup.QueueTask(() => task());
}
catch (TaskCanceledException exc) { }
}
}
