public static void ApplyPrecompile(CudaCompiler compiler, DeviceKernelTemplate template, int tensorCount)
{
foreach (ApplySpecialization spec in ApplySpecialization.AllSpecializations(tensorCount))
{
template.PtxForConfig(compiler, spec.GetConfig());
}
}
public byte[] GetPtx(CudaCompiler compiler)
{
if (this.ptx == null)
{
this.Precompile(compiler);
}
return(this.ptx);
}
/// <summary>
/// Gets the PTX.
/// </summary>
/// <param name="compiler">The compiler.</param>
/// <returns>System.Byte[].</returns>
public byte[] GetPtx(CudaCompiler compiler)
{
if (ptx == null)
{
Precompile(compiler);
}
return(ptx);
}
public static void PrecompileAllFields(object instance, CudaCompiler compiler)
{
var type = instance.GetType();
foreach (var field in type.GetFields())
{
if (typeof(IPrecompilable).IsAssignableFrom(field.FieldType))
{
var precompilableField = (IPrecompilable)field.GetValue(instance);
Console.WriteLine("Compiling field " + field.Name);
precompilableField.Precompile(compiler);
}
}
}
public void Precompile(CudaCompiler compiler)
{
this.ptx = compiler.CompileToPtx(this.code, this.requiredHeaders);
}
/// <summary>
/// Precompiles the specified compiler.
/// </summary>
/// <param name="compiler">The compiler.</param>
public void Precompile(CudaCompiler compiler)
{
ptx = compiler.CompileToPtx(code, requiredHeaders);
}
