/// <summary>
/// Constructs a new kernel key.
/// </summary>
/// <param name="entry">The entry point description.</param>
/// <param name="specialization">The kernel specialization.</param>
public CachedCompiledKernelKey(
EntryPointDescription entry,
KernelSpecialization specialization)
{
Entry = entry;
Specialization = specialization;
}
/// <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();
}
internal static Kernel CreateKernel(MethodInfo kernelFunction)
{
var entryPointDescCalculate = EntryPointDescription.FromImplicitlyGroupedKernel(kernelFunction);
var compiledKernel = GPUHelper.accelerator.Backend.Compile(entryPointDescCalculate, KernelSpecialization.Empty);
var kernel = GPUHelper.accelerator.LoadAutoGroupedKernel(compiledKernel);
kernels.Add(kernel);
return(kernel);
}
/// <summary>
/// Constructs a new specialization cache.
/// </summary>
/// <param name="accelerator">The parent accelerator.</param>
/// <param name="kernelMethod">The IR kernel method.</param>
/// <param name="loader">The loader instance.</param>
/// <param name="entry">The associated entry point.</param>
/// <param name="specialization">The kernel specialization.</param>
public SpecializationCache(
Accelerator accelerator,
Method kernelMethod,
TLoader loader,
EntryPointDescription entry,
KernelSpecialization specialization)
{
Accelerator = accelerator;
KernelMethod = kernelMethod;
Loader = loader;
Entry = entry;
KernelSpecialization = specialization;
}
/// <summary>
/// Compiles and launches an auto-grouped implicitly-grouped kernel.
/// </summary>
static void CompileAndLaunchAutoGroupedKernel(Accelerator accelerator)
{
// Access the current backend for this device
var backend = accelerator.Backend;
// Resolve and compile method into a kernel
var method = typeof(Program).GetMethod(nameof(MyKernel), BindingFlags.NonPublic | BindingFlags.Static);
var entryPointDesc = EntryPointDescription.FromImplicitlyGroupedKernel(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 implicitly grouped kernel with an automatically determined group size.
// Note that the kernel has to be disposed manually.
using (var kernel = accelerator.LoadAutoGroupedKernel(compiledKernel))
{
var launcher = kernel.CreateLauncherDelegate <Action <AcceleratorStream, Index1, ArrayView <int>, int> >();
// -------------------------------------------------------------------------------
using (var buffer = accelerator.Allocate <int>(1024))
{
// Launch buffer.Length many threads and pass a view to buffer.
// You can also use kernel.Launch; however, the generic launch method involves boxing.
launcher(
accelerator.DefaultStream,
buffer.Length,
buffer.View,
42);
// Wait for the kernel to finish...
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");
}
}
}
accelerator.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($"Error at element location {i}: {data[i]} found");
}
}
}
}
}
/// <summary>
/// Compiles and launches an implicitly-grouped kernel.
/// </summary>
static void CompileAndLaunchImplicitlyGroupedKernel(Accelerator accelerator, int groupSize)
{
// Access the current backend for this device
var backend = accelerator.GetBackend();
// Resolve and compile method into a kernel
var method = typeof(Program).GetMethod(nameof(MyKernel), BindingFlags.NonPublic | BindingFlags.Static);
var entryPointDesc = EntryPointDescription.FromImplicitlyGroupedKernel(method);
var compiledKernel = backend.Compile(entryPointDesc, default);
// Info: If the current accelerator is a CudaAccelerator, we can cast the compiled kernel to a
// PTXCompiledKernel in order to extract the PTX assembly code.
// -------------------------------------------------------------------------------
// Load the implicitly grouped kernel with the custom group size
// Note that the kernel has to be disposed manually.
using var kernel = accelerator.LoadImplicitlyGroupedKernel(compiledKernel, groupSize);
var launcher = kernel.CreateLauncherDelegate <Action <AcceleratorStream, Index1D, ArrayView <int>, int> >();
// -------------------------------------------------------------------------------
using var buffer = accelerator.Allocate1D <int>(1024);
// Launch buffer.Length many threads and pass a view to buffer.
// You can also use kernel.Launch; however, the generic launch method involves boxing.
launcher(
accelerator.DefaultStream,
(int)buffer.Length,
buffer.View,
42);
// Reads data from the GPU buffer into a new CPU array.
// Implicitly calls accelerator.DefaultStream.Synchronize() to ensure
// that the kernel and memory copy are completed first.
var data = buffer.GetAsArray1D();
for (int i = 0, e = data.Length; i < e; ++i)
{
if (data[i] != 42 + i)
{
Console.WriteLine($"Error at element location {i}: {data[i]} found");
}
}
}
public static void Main2()
{
using (var context = new Context())
{
foreach (var acceleratorId in Accelerator.Accelerators)
{
if (acceleratorId.AcceleratorType == AcceleratorType.CPU)
{
continue;
}
using (var accelerator = Accelerator.Create(context, acceleratorId))
{
CompiledKernel compiledKernel;
using (Backend b = new CLBackend(context, ILGPU.Runtime.OpenCL.CLAcceleratorVendor.AMD))
{
MethodInfo methodInfo = typeof(GPU).GetMethod("PixelKernel");
KernelSpecialization spec = KernelSpecialization.Empty;
compiledKernel = b.Compile(EntryPointDescription.FromImplicitlyGroupedKernel(methodInfo), spec);
// debug: check kernel.Source for source text
}
var kernel = accelerator.LoadAutoGroupedKernel(compiledKernel);
// var kernel = accelerator.LoadAutoGroupedStreamKernel<Index2, ArrayView2D<FSMUnit>>(MathKernel);
// kernel = accelerator.LoadAutoGroupedStreamKernel<Index2, ArrayView2D<Color3>, ArrayView<byte>, ArrayView2D<Neuron>>(PixelKernel);
MemoryBuffer2D <Color3> buffer = accelerator.Allocate <Color3>(pixelMap.GetLength(0), pixelMap.GetLength(1));
MemoryBuffer <byte> buffer2 = accelerator.Allocate <byte>(imageBytes.Length);
MemoryBuffer2D <Neuron> buffer3 = accelerator.Allocate <Neuron>(nrn.GetLength(0), nrn.GetLength(1));
buffer3.CopyFrom(nrn, new LongIndex2(0, 0), new LongIndex2(0, 0), new LongIndex2(nrn.GetLength(0), nrn.GetLength(1)));
while (running == true)
{
Stopwatch sw = new Stopwatch();
sw.Start();
Index2 gridSize = new Index2(pixelMap.GetLength(0), pixelMap.GetLength(1));
//kernel(gridSize, buffer.View, buffer2.View, buffer3.View);
sw.OutputDelta("Kernel");
accelerator.Synchronize();
sw.OutputDelta("Sync");
// imageBytes = buffer2.GetAsArray();
buffer2.CopyTo(imageBytes, 0, 0, imageBytes.Length);
sw.OutputDelta("Copy ImageBytes");
// Resolve and verify data
//pixelMap = buffer.GetAs2DArray();
// buffer.CopyTo(pixelMap, new LongIndex2(0, 0), new LongIndex2(0, 0), new LongIndex2(pixelMap.GetLength(0), pixelMap.GetLength(1)));
// Color3[] pixelMap1D = buffer.GetAsArray();
//Copy1DTo2DArray(pixelMap1D, pixelMap); // ~36ms, a bit faster
//Array.Copy(pixelMap1D, imageBytes, pixelMap1D.Length); // fails
//Buffer.BlockCopy(pixelMap1D, 0, pixelMap, 0, pixelMap1D.Length * Marshal.SizeOf(typeof(Color3))); // fails
// pixelMap = Make2DArray(pixelMap1D, pixelMap.GetLength(0), pixelMap.GetLength(1)); // still slow
//sw.OutputDelta("Copy PixelMap");
// MainForm.form.DrawPixels(pixelMap);
MainForm.form.DrawPixels(imageBytes, pixelMap.GetLength(0), pixelMap.GetLength(1));
Application.DoEvents();
//Debugger.Break();
sw.OutputDelta("DrawPixels");
}
buffer.Dispose();
buffer2.Dispose();
buffer3.Dispose();
Application.Exit();
//Debugger.Break();
}
}
}
}