public TileInfo(ArrayView <T> input, Index1 numIterationsPerGroup)
{
TileSize = Group.DimX * numIterationsPerGroup;
StartIndex = Grid.IdxX * TileSize + Group.IdxX;
EndIndex = (Grid.IdxX + Index1.One) * TileSize;
MaxLength = XMath.Min(input.Length, EndIndex);
}
public TileInfo(int inputLength, Index1D numIterationsPerGroup)
{
TileSize = Group.DimX * numIterationsPerGroup;
StartIndex = Grid.IdxX * TileSize + Group.IdxX;
EndIndex = (Grid.IdxX + Index1D.One) * TileSize;
MaxLength = XMath.Min(inputLength, EndIndex);
}
/// <summary>
/// The actual unique kernel implementation.
/// </summary>
/// <typeparam name="T">The element type.</typeparam>
/// <typeparam name="TComparisonOperation">The comparison operation.</typeparam>
/// <param name="input">The input view.</param>
/// <param name="output">The output view to store the new length.</param>
/// <param name="sequentialGroupExecutor">
/// The sequential group executor to use.
/// </param>
/// <param name="tileSize">The tile size.</param>
/// <param name="numIterationsPerGroup">
/// The number of iterations per group.
/// </param>
internal static void UniqueKernel <T, TComparisonOperation>(
ArrayView <T> input,
ArrayView <long> output,
SequentialGroupExecutor sequentialGroupExecutor,
SpecializedValue <int> tileSize,
Index1D numIterationsPerGroup)
where T : unmanaged
where TComparisonOperation : struct, IComparisonOperation <T>
{
TComparisonOperation comparison = default;
var isFirstGrid = Grid.IdxX == 0;
var tileInfo = new TileInfo(input.IntLength, numIterationsPerGroup);
// Sync groups and wait for the current one to become active
sequentialGroupExecutor.Wait();
var temp = SharedMemory.Allocate <bool>(tileSize);
var startIdx = Grid.ComputeGlobalIndex(Grid.IdxX, 0);
for (
int i = tileInfo.StartIndex;
i < tileInfo.MaxLength;
i += Group.DimX)
{
if (Group.IsFirstThread && i == tileInfo.StartIndex && isFirstGrid)
{
temp[0] = true;
}
else
{
var currIdx = i;
var prevIdx = Group.IsFirstThread && i == tileInfo.StartIndex
? output[0] - 1
: currIdx - 1;
temp[currIdx - startIdx] =
comparison.Compare(input[currIdx], input[prevIdx]) != 0;
}
}
Group.Barrier();
if (Group.IsFirstThread)
{
var offset = isFirstGrid ? 0 : output[0];
var maxLength =
XMath.Min(startIdx + temp.IntLength, tileInfo.MaxLength) - startIdx;
for (var i = 0; i < maxLength; i++)
{
if (temp[i])
{
input[offset++] = input[startIdx + i];
}
}
output[0] = offset;
}
MemoryFence.DeviceLevel();
Group.Barrier();
sequentialGroupExecutor.Release();
}
