private static void IlGpuKernelConstants(
ArrayView2D <Real> mSquaredDistances,
ArrayView <Real> mCoordinates,
SpecializedValue <int> c,
int n)
{
// Same as CudaKernelOptimised2, but the number of coordinates is given as a meta-constant.
// Also, we write the results as float2.
var shared = SharedMemory.GetDynamic <Real>();
var coordinatesI = shared.GetSubView(0, c * Group.DimX);
var coordinatesJ = shared.GetSubView(c * Group.DimX);
var bI = Grid.IdxY * Group.DimX;
var bJ = Grid.IdxX * Group.DimX;
for (int k = 0; k != c; ++k)
{
if (bI + Group.IdxX < n)
{
coordinatesI[k * Group.DimX + Group.IdxX] = mCoordinates[k * n + bI + Group.IdxX];
}
if (bJ + Group.IdxX < n)
{
coordinatesJ[k * Group.DimX + Group.IdxX] = mCoordinates[k * n + bJ + Group.IdxX];
}
}
Group.Barrier();
var line = Group.IdxX / (Group.DimX / 2);
var tid = Group.IdxX % (Group.DimX / 2);
if (bJ + tid * 2 < n)
{
var coordinatesJ2 = coordinatesJ.Cast <IlReal2>();
for (int i = line; i < Group.DimX & bI + i < n; i += 2)
{
var dist = default(IlReal2);
for (int k = 0; k != c; ++k)
{
var coord1 = coordinatesI[k * Group.DimX + i];
var coord2 = coordinatesJ2[(k * Group.DimX / 2) + tid];
var diff = new IlReal2(coord1 - coord2.X, coord1 - coord2.Y);
dist += diff * diff;
}
var dst = mSquaredDistances.Cast <IlReal2>();
dst[bJ / 2 + tid, bI + i] = dist;
}
}
}
private static void IlGpuKernelFloat2(
ArrayView2D <Real> mSquaredDistances,
ArrayView <Real> mCoordinates,
int c,
int n)
{
// Same as KernelSharedMemory, but one thread does two element in one by using float2 reads.
var shared = SharedMemory.GetDynamic <Real>();
var coordinatesI = shared.GetSubView(0, c * Group.DimX);
var coordinatesJ = shared.GetSubView(c * Group.DimX);
var bI = Grid.IdxY * Group.DimX;
var bJ = Grid.IdxX * Group.DimX;
for (int k = 0; k != c; ++k)
{
if (bI + Group.IdxX < n)
{
coordinatesI[k * Group.DimX + Group.IdxX] = mCoordinates[k * n + bI + Group.IdxX];
}
if (bJ + Group.IdxX < n)
{
coordinatesJ[k * Group.DimX + Group.IdxX] = mCoordinates[k * n + bJ + Group.IdxX];
}
}
Group.Barrier();
var line = Group.IdxX / (Group.DimX / 2);
var tid = Group.IdxX % (Group.DimX / 2);
if (bJ + tid * 2 < n)
{
var coordinatesJ2 = coordinatesJ.Cast <IlReal2>();
for (int i = line; i < Group.DimX && bI + i < n; i += 2)
{
var dist = default(IlReal2);
for (int k = 0; k != c; ++k)
{
var coord1 = coordinatesI[k * Group.DimX + i];
var coord2 = coordinatesJ2[(k * Group.DimX / 2) + tid];
var diff = new IlReal2(coord1 - coord2.X, coord1 - coord2.Y);
dist += diff * diff;
}
mSquaredDistances[bJ + 2 * tid + 0, bI + i] = dist.X;
mSquaredDistances[bJ + 2 * tid + 1, bI + i] = dist.Y;
}
}
}
private static void IlGpuKernelLocalMemory(
ArrayView2D <Real> mSquaredDistances,
ArrayView <Real> mCoordinates,
SpecializedValue <int> dimX,
SpecializedValue <int> c,
int n)
{
// Same as KernelConstants, but use both local and shared memory to increase the effective shared memory.
var coordinatesI = SharedMemory.Allocate <Real>(c * dimX);
var coordinatesJ = new IlReal2[c.Value];
var bI = Grid.IdxY * dimX;
var bJ = Grid.IdxX * dimX;
var line = Group.IdxX / (dimX / 2);
var tid = Group.IdxX % (dimX / 2);
var isActive = bJ + tid * 2 < n;
for (int k = 0; k != c.Value; ++k)
{
if (bI + Group.IdxX < n)
{
coordinatesI[k * dimX + Group.IdxX] = mCoordinates[k * n + bI + Group.IdxX];
}
if (isActive)
{
var mCoordinates2 = mCoordinates.Cast <IlReal2>();
coordinatesJ[k] = mCoordinates2[(k * n + bJ) / 2 + tid];
}
}
Group.Barrier();
if (isActive)
{
for (int i = line; i < dimX && bI + i < n; i += 2)
{
var dist = default(IlReal2);
for (int k = 0; k != c.Value; ++k)
{
var coord1 = coordinatesI[k * dimX + i];
var coord2 = coordinatesJ[k];
var diff = new IlReal2(coord1 - coord2.X, coord1 - coord2.Y);
dist += diff * diff;
}
var dst = mSquaredDistances.Cast <IlReal2>();
dst[bJ / 2 + tid, bI + i] = dist;
}
}
}
