/// <summary>
/// Executes the specified kernel with the given arguments.
/// </summary>
/// <typeparam name="TIndex">The index type.</typeparam>
/// <param name="kernel">The kernel method.</param>
/// <param name="dimension">The dimension.</param>
/// <param name="arguments">The arguments.</param>
public void Execute <TIndex>(
MethodInfo kernel,
TIndex dimension,
params object[] arguments)
where TIndex : struct, IIndex
{
using var stream = Accelerator.CreateStream();
// Compile kernel manually and load the compiled kernel into the accelerator
var backend = Accelerator.Backend;
Output.WriteLine($"Compiling '{kernel.Name}'");
var entryPoint = typeof(TIndex) == typeof(KernelConfig)
? EntryPointDescription.FromExplicitlyGroupedKernel(kernel)
: EntryPointDescription.FromImplicitlyGroupedKernel(kernel);
var compiled = backend.Compile(entryPoint, new KernelSpecialization());
// Load the compiled kernel
Output.WriteLine($"Loading '{kernel.Name}'");
using var acceleratorKernel = Accelerator.LoadKernel(compiled);
// Launch the kernel
Output.WriteLine($"Launching '{kernel.Name}'");
acceleratorKernel.Launch(stream, dimension, arguments);
stream.Synchronize();
}
/// <summary>
/// Compiles and launches an explicitly grouped kernel.
/// </summary>
static void CompileAndLaunchKernel(Accelerator accelerator, int groupSize)
{
// Access the current backend for this device
var backend = accelerator.Backend;
// Resolve and compile method into a kernel
var method = typeof(Program).GetMethod(nameof(GroupedKernel), BindingFlags.NonPublic | BindingFlags.Static);
var entryPointDesc = EntryPointDescription.FromExplicitlyGroupedKernel(method);
var compiledKernel = backend.Compile(entryPointDesc, default);
// Info: If the current accelerator is a CudaAccelerator, we can cast the compiled kernel to a
// PTXCompiledKernel in order to extract the PTX assembly code.
// -------------------------------------------------------------------------------
// Load the explicitly grouped kernel
// Note that the kernel has to be disposed manually.
using (var kernel = accelerator.LoadKernel(compiledKernel))
{
var launcher = kernel.CreateLauncherDelegate <Action <AcceleratorStream, KernelConfig, ArrayView <int>, int> >();
// -------------------------------------------------------------------------------
using (var buffer = accelerator.Allocate <int>(1024))
{
// You can also use kernel.Launch; however, the generic launch method involves boxing.
launcher(
accelerator.DefaultStream,
((buffer.Length + groupSize - 1) / groupSize, // Compute the number of groups (round up)
groupSize), // Use the given group size
buffer.View,
42);
accelerator.Synchronize();
// Resolve and verify data
var data = buffer.GetAsArray();
for (int i = 0, e = data.Length; i < e; ++i)
{
if (data[i] != 42 + i)
{
Console.WriteLine($"Error at element location {i}: {data[i]} found");
}
}
}
}
}