//[/GenericScanDownsweepKernel]
public T[] Apply(T[] input, bool inclusive)
{
var n = input.Length;
var numSm = GPUWorker.Device.Attributes.MULTIPROCESSOR_COUNT;
var tup = Plan.BlockRanges(numSm, n);
var ranges = tup.Item1;
var numRanges = tup.Item2;
var lpUpsweep = new LaunchParam(numRanges, Plan.NumThreads);
var lpReduce = new LaunchParam(1, Plan.NumThreadsReduction);
var lpDownsweep = new LaunchParam(numRanges, Plan.NumThreads);
var _inclusive = inclusive ? 1 : 0;
using (var dRanges = GPUWorker.Malloc(ranges))
using (var dRangeTotals = GPUWorker.Malloc <T>(numRanges + 1))
using (var dInput = GPUWorker.Malloc(input))
using (var dOutput = GPUWorker.Malloc(input))
{
_reduceModule.Upsweep(lpUpsweep, dInput.Ptr, dRanges.Ptr, dRangeTotals.Ptr);
GPULaunch(ScanReduce, lpReduce, numRanges, dRangeTotals.Ptr);
GPULaunch(Downsweep, lpDownsweep, dInput.Ptr, dOutput.Ptr, dRangeTotals.Ptr, dRanges.Ptr, _inclusive);
return(dOutput.Gather());
}
}