/// <summary>
/// Generates a dynamic kernel-launcher method that will be just-in-time compiled
/// during the first invocation. Using the generated launcher lowers the overhead
/// for kernel launching dramatically, since unnecessary operations (like boxing)
/// can be avoided.
/// </summary>
/// <param name="kernel">The kernel to generate a launcher for.</param>
/// <param name="customGroupSize">The custom group size for the launching operation.</param>
/// <returns>The generated launcher method.</returns>
private MethodInfo GenerateKernelLauncherMethod(ILCompiledKernel kernel, int customGroupSize)
{
var entryPoint = kernel.EntryPoint;
AdjustAndVerifyKernelGroupSize(ref customGroupSize, entryPoint);
var launcher = entryPoint.CreateLauncherMethod(Context);
var emitter = new ILEmitter(launcher.ILGenerator);
var cpuKernel = emitter.DeclareLocal(typeof(CPUKernel));
KernelLauncherBuilder.EmitLoadKernelArgument <CPUKernel, ILEmitter>(
Kernel.KernelInstanceParamIdx, emitter);
emitter.Emit(LocalOperation.Store, cpuKernel);
// Create an instance of the custom task type
var task = emitter.DeclareLocal(kernel.TaskType);
{
var sharedMemSize = KernelLauncherBuilder.EmitSharedMemorySizeComputation(entryPoint, emitter);
emitter.Emit(LocalOperation.Load, cpuKernel);
emitter.EmitCall(
typeof(CPUKernel).GetProperty(
nameof(CPUKernel.KernelExecutionDelegate),
BindingFlags.NonPublic | BindingFlags.Public | BindingFlags.Instance).GetGetMethod(true));
// Load custom user dimension
KernelLauncherBuilder.EmitLoadDimensions(
entryPoint,
emitter,
Kernel.KernelParamDimensionIdx,
() => emitter.EmitNewObject(
typeof(Index3).GetConstructor(
new Type[] { typeof(int), typeof(int), typeof(int) })));
// Load dimensions as index3 arguments
KernelLauncherBuilder.EmitLoadDimensions(
entryPoint,
emitter,
Kernel.KernelParamDimensionIdx,
() => emitter.EmitNewObject(
typeof(Index3).GetConstructor(
new Type[] { typeof(int), typeof(int), typeof(int) })),
customGroupSize);
// Load shared-memory size
emitter.Emit(LocalOperation.Load, sharedMemSize);
// Create new task object
emitter.EmitNewObject(kernel.TaskConstructor);
// Store task
emitter.Emit(LocalOperation.Store, task);
}
// Assign parameters
var parameters = entryPoint.Parameters;
for (int i = 0, e = parameters.NumParameters; i < e; ++i)
{
emitter.Emit(LocalOperation.Load, task);
emitter.Emit(ArgumentOperation.Load, i + Kernel.KernelParameterOffset);
if (parameters.IsByRef(i))
{
emitter.Emit(OpCodes.Ldobj, parameters[i]);
}
emitter.Emit(OpCodes.Stfld, kernel.TaskArgumentMapping[i]);
}
// Launch task: ((CPUKernel)kernel).CPUAccelerator.Launch(task);
emitter.Emit(LocalOperation.Load, cpuKernel);
emitter.EmitCall(
typeof(CPUKernel).GetProperty(
nameof(CPUKernel.CPUAccelerator)).GetGetMethod(false));
emitter.Emit(LocalOperation.Load, task);
emitter.EmitCall(
typeof(CPUAccelerator).GetMethod(
nameof(CPUAccelerator.Launch),
BindingFlags.NonPublic | BindingFlags.Instance));
// End of launch method
emitter.Emit(OpCodes.Ret);
emitter.Finish();
return(launcher.Finish());
}
/// <summary>
/// Generates a dynamic kernel-launcher method that will be just-in-time compiled
/// during the first invocation. Using the generated launcher lowers the overhead
/// for kernel launching dramatically, since unnecessary operations (like boxing)
/// can be avoided.
/// </summary>
/// <param name="kernel">The kernel to generate a launcher for.</param>
/// <param name="customGroupSize">
/// The custom group size for the launching operation.
/// </param>
/// <returns>The generated launcher method.</returns>
private MethodInfo GenerateKernelLauncherMethod(
ILCompiledKernel kernel,
int customGroupSize)
{
var entryPoint = kernel.EntryPoint;
AdjustAndVerifyKernelGroupSize(ref customGroupSize, entryPoint);
// Add support for by ref parameters
if (entryPoint.HasByRefParameters)
{
throw new NotSupportedException(
ErrorMessages.NotSupportedByRefKernelParameters);
}
using var scopedLock = entryPoint.CreateLauncherMethod(
Context.RuntimeSystem,
out var launcher);
var emitter = new ILEmitter(launcher.ILGenerator);
// Pretend to map kernel arguments (like a GPU accelerator would perform).
var argumentMapper = Backend.ArgumentMapper;
argumentMapper.Map(entryPoint);
var cpuKernel = emitter.DeclareLocal(typeof(CPUKernel));
KernelLauncherBuilder.EmitLoadKernelArgument <CPUKernel, ILEmitter>(
Kernel.KernelInstanceParamIdx, emitter);
emitter.Emit(LocalOperation.Store, cpuKernel);
// Create an instance of the custom task type
var task = emitter.DeclareLocal(kernel.TaskType);
{
emitter.Emit(LocalOperation.Load, cpuKernel);
emitter.EmitCall(CPUKernel.GetKernelExecutionDelegate);
// Load custom user dimension
KernelLauncherBuilder.EmitLoadKernelConfig(
entryPoint,
emitter,
Kernel.KernelParamDimensionIdx,
MaxGridSize,
MaxGroupSize);
// Load dimensions
KernelLauncherBuilder.EmitLoadRuntimeKernelConfig(
entryPoint,
emitter,
Kernel.KernelParamDimensionIdx,
MaxGridSize,
MaxGroupSize,
customGroupSize);
// Create new task object
emitter.EmitNewObject(kernel.TaskConstructor);
// Store task
emitter.Emit(LocalOperation.Store, task);
}
// Assign parameters
var parameters = entryPoint.Parameters;
for (int i = 0, e = parameters.Count; i < e; ++i)
{
emitter.Emit(LocalOperation.Load, task);
emitter.Emit(ArgumentOperation.Load, i + Kernel.KernelParameterOffset);
if (parameters.IsByRef(i))
{
emitter.Emit(OpCodes.Ldobj, parameters[i]);
}
emitter.Emit(OpCodes.Stfld, kernel.TaskArgumentMapping[i]);
}
// Launch task: ((CPUKernel)kernel).CPUAccelerator.Launch(task);
emitter.Emit(LocalOperation.Load, cpuKernel);
emitter.EmitCall(
typeof(CPUKernel).GetProperty(
nameof(CPUKernel.CPUAccelerator)).GetGetMethod(false));
emitter.Emit(LocalOperation.Load, task);
emitter.EmitCall(
typeof(CPUAccelerator).GetMethod(
nameof(CPUAccelerator.Launch),
BindingFlags.NonPublic | BindingFlags.Instance));
// End of launch method
emitter.Emit(OpCodes.Ret);
emitter.Finish();
return(launcher.Finish());
}