/// <summary>
/// Loads an implicitly grouped kernel on the current accelerator.
/// </summary>
/// <param name="kernel">The compiled kernel to load.</param>
/// <param name="customGroupSize">The user-defined group size.</param>
/// <param name="kernelInfo">The resolved kernel information.</param>
/// <returns>The loaded kernel.</returns>
protected sealed override Kernel LoadImplicitlyGroupedKernelInternal(
CompiledKernel kernel,
int customGroupSize,
out KernelInfo kernelInfo)
{
if (kernel == null)
{
throw new ArgumentNullException(nameof(kernel));
}
if (customGroupSize < 0 || customGroupSize > MaxNumThreadsPerGroup)
{
throw new ArgumentOutOfRangeException(nameof(customGroupSize));
}
if (!(kernel is TCompiledKernel compiledKernel))
{
throw new NotSupportedException(RuntimeErrorMessages.NotSupportedKernel);
}
if (kernel.EntryPoint.IsExplicitlyGrouped)
{
throw new NotSupportedException(
RuntimeErrorMessages.NotSupportedExplicitlyGroupedKernel);
}
kernelInfo = KernelInfo.CreateFrom(
kernel.Info,
customGroupSize,
null);
return(CreateKernel(
compiledKernel,
GenerateKernelLauncherMethod(compiledKernel, customGroupSize)));
}
/// <summary>
/// Loads the given implicitly-grouped kernel while using an automatically
/// computed grouping configuration.
/// </summary>
/// <param name="kernel">The kernel to load.</param>
/// <param name="kernelInfo">
/// Detailed kernel information about the loaded kernel.
/// </param>
/// <returns>The loaded kernel.</returns>
/// <remarks>
/// Note that the returned kernel will not be managed by the kernel cache.
/// </remarks>
public Kernel LoadAutoGroupedKernel(
CompiledKernel kernel,
out KernelInfo kernelInfo)
{
Bind();
return(LoadAutoGroupedKernelInternal(kernel, out kernelInfo));
}
/// <summary>
/// Loads an auto grouped kernel on the current accelerator.
/// </summary>
/// <param name="kernel">The compiled kernel to load.</param>
/// <param name="kernelInfo">The resolved kernel information.</param>
/// <returns>The loaded kernel.</returns>
protected sealed override Kernel LoadAutoGroupedKernelInternal(
CompiledKernel kernel,
out KernelInfo kernelInfo)
{
if (kernel == null)
{
throw new ArgumentNullException(nameof(kernel));
}
if (!(kernel is TCompiledKernel compiledKernel))
{
throw new NotSupportedException(RuntimeErrorMessages.NotSupportedKernel);
}
if (kernel.EntryPoint.IsExplicitlyGrouped)
{
throw new NotSupportedException(
RuntimeErrorMessages.NotSupportedExplicitlyGroupedKernel);
}
var result = CreateKernel(compiledKernel);
int groupSize = EstimateGroupSizeInternal(result, 0, 0, out int minGridSize);
result.Launcher = GenerateKernelLauncherMethod(compiledKernel, groupSize);
kernelInfo = KernelInfo.CreateFrom(
kernel.Info,
groupSize,
minGridSize);
return(result);
}
/// <summary>
/// Loads the given implicitly-grouped kernel while using an automatically
/// computed grouping configuration.
/// </summary>
/// <param name="kernel">The kernel to load.</param>
/// <param name="groupSize">The estimated group size to gain maximum occupancy on this device.</param>
/// <param name="minGridSize">The minimum grid size to gain maximum occupancy on this device.</param>
/// <returns>The loaded kernel.</returns>
/// <remarks>
/// Note that the returned kernel will not be managed by the kernel cache.
/// </remarks>
public Kernel LoadAutoGroupedKernel(
CompiledKernel kernel,
out int groupSize,
out int minGridSize)
{
Bind(); return(LoadAutoGroupedKernelInternal(kernel, out groupSize, out minGridSize));
}
/// <summary>
/// Loads an automatically grouped kernel.
/// </summary>
public Kernel LoadKernel(
Accelerator accelerator,
CompiledKernel compiledKernel,
out KernelInfo kernelInfo) =>
accelerator.LoadAutoGroupedKernel(
compiledKernel,
out kernelInfo);
/// <summary>
/// Loads an implicitly grouped kernel.
/// </summary>
public Kernel LoadKernel(
Accelerator accelerator,
CompiledKernel compiledKernel,
out KernelInfo kernelInfo) =>
accelerator.LoadImplicitlyGroupedKernel(
compiledKernel,
GroupSize,
out kernelInfo);
/// <summary cref="Accelerator.LoadAutoGroupedKernelInternal(CompiledKernel, out int, out int)"/>
protected override Kernel LoadAutoGroupedKernelInternal(
CompiledKernel kernel,
out int groupSize,
out int minGridSize)
{
groupSize = WarpSize;
minGridSize = NumThreads / WarpSize;
return(LoadKernel(kernel, groupSize));
}
internal CudaKernel(
CudaAccelerator accelerator,
CompiledKernel kernel,
MethodInfo launcher)
: base(accelerator, kernel, launcher)
{
CudaException.ThrowIfFailed(CudaAPI.Current.LoadModule(out modulePtr, kernel.GetBuffer()));
CudaException.ThrowIfFailed(CudaAPI.Current.GetModuleFunction(out functionPtr, modulePtr, kernel.EntryName));
}
/// <summary>
/// Constructs a new kernel.
/// </summary>
/// <param name="accelerator">The associated accelerator.</param>
/// <param name="compiledKernel">The source kernel.</param>
/// <param name="launcher">The launcher method for the given kernel.</param>
protected Kernel(
Accelerator accelerator,
CompiledKernel compiledKernel,
MethodInfo launcher)
: base(accelerator)
{
CompiledKernel = compiledKernel ?? throw new ArgumentNullException(nameof(compiledKernel));
Launcher = launcher;
}
/// <summary>
/// Loads a compiled kernel into the given Cuda context as kernel program.
/// </summary>
/// <param name="accelerator">The associated accelerator.</param>
/// <param name="kernel">The source kernel.</param>
/// <param name="launcher">The launcher method for the given kernel.</param>
/// <param name="kernelExecutionDelegate">The execution method.</param>
internal CPUKernel(
CPUAccelerator accelerator,
CompiledKernel kernel,
MethodInfo launcher,
CPUKernelExecutionHandler kernelExecutionDelegate)
: base(accelerator, kernel, launcher)
{
KernelExecutionDelegate = kernelExecutionDelegate ?? throw new ArgumentNullException(nameof(kernelExecutionDelegate));
}
/// <summary>
/// Loads an auto grouped kernel.
/// </summary>
protected override Kernel LoadAutoGroupedKernelInternal(
CompiledKernel kernel,
out KernelInfo kernelInfo)
{
var result = LoadKernel(kernel, WarpSize);
kernelInfo = new KernelInfo(WarpSize, NumThreads / WarpSize);
return(result);
}
/// <summary>
/// Loads the given implicitly-grouped kernel.
/// </summary>
/// <param name="kernel">The kernel to load.</param>
/// <param name="customGroupSize">The custom group size to use.</param>
/// <returns>The loaded kernel.</returns>
/// <remarks>
/// Note that implicitly-grouped kernel will be launched with the given group
/// size.
/// </remarks>
/// <remarks>
/// Note that the returned kernel will not be managed by the kernel cache.
/// </remarks>
public Kernel LoadImplicitlyGroupedKernel(
CompiledKernel kernel,
int customGroupSize)
{
Bind();
return(LoadImplicitlyGroupedKernelInternal(
kernel,
customGroupSize,
out var _));
}
/// <summary cref="Accelerator.LoadImplicitlyGroupedKernelInternal(CompiledKernel, int)"/>
protected override Kernel LoadImplicitlyGroupedKernelInternal(
CompiledKernel kernel,
int customGroupSize)
{
if (customGroupSize < 0)
{
throw new ArgumentOutOfRangeException(nameof(customGroupSize));
}
return(LoadKernel(kernel, customGroupSize));
}
/// <summary>
/// Constructs a new kernel.
/// </summary>
/// <param name="accelerator">The associated accelerator.</param>
/// <param name="compiledKernel">The source kernel.</param>
/// <param name="launcher">The launcher method for the given kernel.</param>
protected Kernel(
Accelerator accelerator,
CompiledKernel compiledKernel,
MethodInfo launcher)
: base(accelerator)
{
Debug.Assert(compiledKernel != null, "Invalid compiled kernel");
Specialization = compiledKernel.Specialization;
Launcher = launcher;
NumParameters = compiledKernel.EntryPoint.Parameters.NumParameters;
}
/// <summary cref="IKernelLoader.LoadKernel(Accelerator, CompiledKernel)"/>
public Kernel LoadKernel(Accelerator accelerator, CompiledKernel compiledKernel)
{
var result = accelerator.LoadAutoGroupedKernel(
compiledKernel,
out int groupSize,
out int minGridSize);
GroupSize = groupSize;
MinGridSize = minGridSize;
return(result);
}
/// <summary>
/// Loads an explicitly grouped kernel.
/// </summary>
public Kernel LoadKernel(
Accelerator accelerator,
CompiledKernel compiledKernel,
out KernelInfo kernelInfo)
{
var kernel = accelerator.LoadKernel(compiledKernel);
kernelInfo = KernelInfo.CreateFrom(
compiledKernel.Info,
null,
null);
return(kernel);
}
/// <summary cref="Accelerator.LoadKernelInternal(CompiledKernel)"/>
protected sealed override Kernel LoadKernelInternal(CompiledKernel kernel)
{
if (kernel == null)
{
throw new ArgumentNullException(nameof(kernel));
}
if (!(kernel is TCompiledKernel compiledKernel))
{
throw new NotSupportedException(RuntimeErrorMessages.NotSupportedKernel);
}
return(CreateKernel(
compiledKernel,
GenerateKernelLauncherMethod(compiledKernel, 0)));
}
/// <summary>
/// Loads an implicitly grouped kernel.
/// </summary>
protected override Kernel LoadImplicitlyGroupedKernelInternal(
CompiledKernel kernel,
int customGroupSize,
out KernelInfo kernelInfo)
{
if (customGroupSize < 0)
{
throw new ArgumentOutOfRangeException(nameof(customGroupSize));
}
kernelInfo = KernelInfo.CreateFrom(
kernel.Info,
customGroupSize,
null);
return(LoadKernel(kernel, customGroupSize));
}
/// <summary>
/// Loads the given kernel.
/// </summary>
/// <param name="kernel">The kernel to load.</param>
/// <param name="customGroupSize">The custom group size.</param>
/// <returns>The loaded kernel</returns>
private Kernel LoadKernel(CompiledKernel kernel, int customGroupSize)
{
if (kernel == null)
{
throw new ArgumentNullException(nameof(kernel));
}
if (!(kernel is ILCompiledKernel ilKernel))
{
throw new NotSupportedException(RuntimeErrorMessages.NotSupportedKernel);
}
var launcherMethod = GenerateKernelLauncherMethod(ilKernel, customGroupSize);
return(new CPUKernel(
this,
kernel,
launcherMethod,
ilKernel.ExecutionHandler));
}
/// <summary>
/// Loads the given kernel.
/// </summary>
/// <param name="kernel">The kernel to load.</param>
/// <param name="customGroupSize">The custom group size.</param>
/// <returns>The loaded kernel</returns>
private Kernel LoadKernel(CompiledKernel kernel, int customGroupSize)
{
if (kernel == null)
{
throw new ArgumentNullException(nameof(kernel));
}
var launcherMethod = GenerateKernelLauncherMethod(
kernel,
out Type taskType,
out FieldInfo[] taskArgumentMapping,
customGroupSize);
var executionMethod = GenerateKernelExecutionMethod(
kernel,
taskType,
taskArgumentMapping);
return(new CPUKernel(
this,
kernel,
launcherMethod,
(CPUKernelExecutionHandler)executionMethod.CreateDelegate(
typeof(CPUKernelExecutionHandler))));
}
/// <summary> /// Loads the given kernel. /// Note that implictly-grouped kernels will be launched with a group size /// of the current warp size of the accelerator. /// </summary> /// <param name="kernel">The kernel to load.</param> /// <returns>The loaded kernel.</returns> /// <remarks> /// Note that the returned kernel will not be managed by the kernel cache. /// </remarks> protected abstract Kernel LoadKernelInternal(CompiledKernel kernel);
/// <summary>
/// Loads the given kernel.
/// Note that implictly-grouped kernels will be launched with a group size
/// of the current warp size of the accelerator.
/// </summary>
/// <param name="kernel">The kernel to load.</param>
/// <returns>The loaded kernel.</returns>
/// <remarks>
/// Note that the returned kernel will not be managed by the kernel cache.
/// </remarks>
public Kernel LoadKernel(CompiledKernel kernel)
{
Bind(); return(LoadKernelInternal(kernel));
}
/// <summary cref="IKernelLoader.LoadKernel(Accelerator, CompiledKernel)"/>
public Kernel LoadKernel(Accelerator accelerator, CompiledKernel compiledKernel)
{
return(accelerator.LoadImplicitlyGroupedKernel(compiledKernel, GroupSize));
}
/// <summary>
/// Loads the given implicitly-grouped kernel while using an automatically
/// computed grouping configuration.
/// </summary>
/// <param name="kernel">The kernel to load.</param>
/// <param name="groupSize">The estimated group size to gain maximum occupancy on this device.</param>
/// <param name="minGridSize">The minimum grid size to gain maximum occupancy on this device.</param>
/// <returns>The loaded kernel.</returns>
/// <remarks>
/// Note that the returned kernel will not be managed by the kernel cache.
/// </remarks>
protected abstract Kernel LoadAutoGroupedKernelInternal(
CompiledKernel kernel,
out int groupSize,
out int minGridSize);
/// <summary cref="IKernelLoader.LoadKernel(Accelerator, CompiledKernel)"/>
public Kernel LoadKernel(Accelerator accelerator, CompiledKernel compiledKernel)
{
return(accelerator.LoadKernel(compiledKernel));
}
/// <summary> /// Loads the given implicitly-grouped kernel while using an automatically /// computed grouping configuration. /// </summary> /// <param name="kernel">The kernel to load.</param> /// <returns>The loaded kernel.</returns> /// <remarks> /// Note that the returned kernel will not be managed by the kernel cache. /// </remarks> public Kernel LoadAutoGroupedKernel(CompiledKernel kernel) => LoadAutoGroupedKernel(kernel, out var _, out var _);
/// <summary>
/// Loads the given implicitly-grouped kernel while using an automatically
/// computed grouping configuration.
/// </summary>
/// <param name="kernel">The kernel to load.</param>
/// <returns>The loaded kernel.</returns>
/// <remarks>
/// Note that the returned kernel will not be managed by the kernel cache.
/// </remarks>
public Kernel LoadAutoGroupedKernel(CompiledKernel kernel)
{
return(LoadAutoGroupedKernel(kernel, out int groupSize, out int minGridSize));
}
/// <summary>
/// Loads the given implicitly-grouped kernel.
/// </summary>
/// <param name="kernel">The kernel to load.</param>
/// <param name="customGroupSize">The custom group size to use.</param>
/// <returns>The loaded kernel.</returns>
/// <remarks>
/// Note that implictly-grouped kernel will be launched with the given
/// group size.
/// </remarks>
/// <remarks>
/// Note that the returned kernel will not be managed by the kernel cache.
/// </remarks>
protected abstract Kernel LoadImplicitlyGroupedKernelInternal(
CompiledKernel kernel,
int customGroupSize);
/// <summary cref="Accelerator.LoadKernelInternal(CompiledKernel)"/> protected override Kernel LoadKernelInternal(CompiledKernel kernel) => LoadKernel(kernel, 0);
/// <summary cref="Accelerator.LoadKernelInternal(CompiledKernel)"/>
protected override Kernel LoadKernelInternal(CompiledKernel kernel)
{
return(LoadKernel(kernel, 0));
}