static void Main()
{
using (var context = new Context())
{
// For each available accelerator...
foreach (var acceleratorId in Accelerator.Accelerators)
{
// Create default accelerator for the given accelerator id
using (var accelerator = Accelerator.Create(context, acceleratorId))
{
Console.WriteLine($"Performing operations on {accelerator}");
int groupSize = accelerator.MaxNumThreadsPerGroup;
// Scenario 1: simple version
using (var buffer = accelerator.Allocate <int>(groupSize))
{
var kernel = accelerator.LoadStreamKernel <
ArrayView <int>,
SpecializedValue <int> >(SpecializedKernel);
kernel((1, groupSize), buffer.View, SpecializedValue.New(2));
kernel((1, groupSize), buffer.View, SpecializedValue.New(23));
kernel((1, groupSize), buffer.View, SpecializedValue.New(42));
}
// Scenario 2: custom structure
using (var buffer = accelerator.Allocate <int>(groupSize))
{
var kernel = accelerator.LoadStreamKernel <
ArrayView <int>,
SpecializedValue <CustomStruct> >(SpecializedCustomStructKernel);
kernel(
(1, groupSize),
buffer.View,
SpecializedValue.New(
new CustomStruct(1, 7)));
kernel(
(1, groupSize),
buffer.View,
SpecializedValue.New(
new CustomStruct(23, 42)));
}
// Scenario 3: generic kernel
using (var buffer = accelerator.Allocate <long>(groupSize))
{
var kernel = accelerator.LoadStreamKernel <
ArrayView <long>,
SpecializedValue <long> >(SpecializedGenericKernel);
kernel((1, groupSize), buffer.View, SpecializedValue.New(23L));
kernel((1, groupSize), buffer.View, SpecializedValue.New(42L));
}
}
}
}
}
/// <summary>
/// Demonstrates using the mul.hi.u64 and mul.lo.u64 inline PTX instructions to
/// multiply two UInt64 values to produce a UInt128 value.
/// </summary>
static void MultiplyUInt128(CudaAccelerator accelerator)
{
using var buffer = accelerator.Allocate1D <UInt128>(1024);
var kernel = accelerator.LoadAutoGroupedStreamKernel <Index1D, ArrayView <UInt128>, SpecializedValue <ulong> >(MultiplyUInt128Kernel);
kernel(
(int)buffer.Length,
buffer.View,
SpecializedValue.New(ulong.MaxValue));
var results = buffer.GetAsArray1D();
for (var i = 0; i < results.Length; i++)
{
Console.WriteLine($"[{i}] = {results[i]}");
}
}
private static void IlGpuOptimisedImpl(
CudaAccelerator gpu,
Real[] mSquaredDistances,
Real[] mCoordinates,
int c,
int n,
string name,
Action <ArrayView2D <Real>, ArrayView <Real>, SpecializedValue <int>, SpecializedValue <int>, int> kernelFunc)
{
using var cudaSquaredDistance = gpu.Allocate <Real>(n, n);
using var cudaCoordinates = gpu.Allocate(mCoordinates);
var timer = Stopwatch.StartNew();
const int blockSize = 128;
var gridSize = Util.DivUp(n, blockSize);
var lp = ((gridSize, gridSize, 1), (blockSize, 1, 1));
gpu.Launch(kernelFunc, gpu.DefaultStream, lp, cudaSquaredDistance.View, cudaCoordinates.View, SpecializedValue.New(blockSize), SpecializedValue.New(c), n);
gpu.Synchronize();
Util.PrintPerformance(timer, name, n, c, n);
cudaSquaredDistance.CopyTo(mSquaredDistances, (0, 0), 0, (n, n));
}