/// <summary>
/// Multiplies two dense matrices and returns the resultant matrix (using tiling).
/// </summary>
/// <param name="accelerator">The Accelerator to run the multiplication on</param>
/// <param name="a">A dense MxK matrix</param>
/// <param name="b">A dense KxN matrix</param>
/// <returns>A dense MxN matrix</returns>
static float[,] MatrixMultiplyTiled(Accelerator accelerator, float[,] a, float[,] b)
{
var m = a.GetLength(0);
var ka = a.GetLength(1);
var kb = b.GetLength(0);
var n = b.GetLength(1);
if (ka != kb)
{
throw new ArgumentException($"Cannot multiply {m}x{ka} matrix by {n}x{kb} matrix", nameof(b));
}
var kernel = accelerator.LoadStreamKernel <
ArrayView2D <float, Stride2D.DenseX>,
ArrayView2D <float, Stride2D.DenseX>,
ArrayView2D <float, Stride2D.DenseX> >(
MatrixMultiplyTiledKernel);
var groupSize = new Index2D(TILE_SIZE, TILE_SIZE);
var numGroups = new Index2D((m + TILE_SIZE - 1) / TILE_SIZE, (n + TILE_SIZE - 1) / TILE_SIZE);
using (var aBuffer = accelerator.Allocate2DDenseX <float>(new Index2D(m, ka)))
using (var bBuffer = accelerator.Allocate2DDenseX <float>(new Index2D(ka, n)))
using (var cBuffer = accelerator.Allocate2DDenseX <float>(new Index2D(m, n)))
{
aBuffer.CopyFromCPU(a);
bBuffer.CopyFromCPU(b);
kernel((numGroups, groupSize), aBuffer, bBuffer, cBuffer);
accelerator.Synchronize();
return(cBuffer.GetAsArray2D());
}
}
/// <summary>
/// Multiplies two dense matrices and returns the resultant matrix.
/// </summary>
/// <param name="accelerator">The Accelerator to run the multiplication on</param>
/// <param name="a">A dense MxK matrix</param>
/// <param name="b">A dense KxN matrix</param>
/// <returns>A dense MxN matrix</returns>
static float[,] MatrixMultiplyAccelerated(Accelerator accelerator, float[,] a, float[,] b)
{
var m = a.GetLength(0);
var ka = a.GetLength(1);
var kb = b.GetLength(0);
var n = b.GetLength(1);
if (ka != kb)
{
throw new ArgumentException($"Cannot multiply {m}x{ka} matrix by {n}x{kb} matrix", nameof(b));
}
var kernel = accelerator.LoadAutoGroupedStreamKernel <
Index2D,
ArrayView2D <float, Stride2D.DenseX>,
ArrayView2D <float, Stride2D.DenseX>,
ArrayView2D <float, Stride2D.DenseX> >(
MatrixMultiplyAcceleratedKernel);
using (var aBuffer = accelerator.Allocate2DDenseX <float>(new Index2D(m, ka)))
using (var bBuffer = accelerator.Allocate2DDenseX <float>(new Index2D(ka, n)))
using (var cBuffer = accelerator.Allocate2DDenseX <float>(new Index2D(m, n)))
{
aBuffer.CopyFromCPU(a);
bBuffer.CopyFromCPU(b);
kernel(cBuffer.Extent.ToIntIndex(), aBuffer.View, bBuffer.View, cBuffer.View);
accelerator.Synchronize();
return(cBuffer.GetAsArray2D());
}
}
