// note. cannot use TimeSpan here because it ain't work with fractions of milliseconds
public static ElapsedTime Benchmark(Action action)
{
CudaDriver.Ensure();
var before = CUevent.Null;
var after = CUevent.Null;
try
{
before = nvcuda.cuEventCreate(CUevent_flags.Default);
after = nvcuda.cuEventCreate(CUevent_flags.Default);
nvcuda.cuEventRecord(before);
action();
nvcuda.cuEventRecord(after);
nvcuda.cuEventSynchronize(after);
return(nvcuda.cuEventElapsedTime(before, after));
}
finally
{
// todo. there's a slight possibility of resource leak here
if (before.IsNotNull)
{
nvcuda.cuEventDestroy(before);
}
if (after.IsNotNull)
{
nvcuda.cuEventDestroy(after);
}
}
}
public static Object Invoke(this JittedKernel kernel, dim3 gridDim, dim3 blockDim, IEnumerable <KernelArgument> args)
{
kernel.AssertNotNull();
args = args ?? Seq.Empty <KernelArgument>().ToArray();
CudaDriver.Ensure();
return(kernel.Function.Invoke(gridDim, blockDim, args));
}
public static Object Invoke(this JittedFunction function, dim3 gridDim, dim3 blockDim, params KernelArgument[] args)
{
function.AssertNotNull();
args = args ?? Seq.Empty <KernelArgument>().ToArray();
CudaDriver.Ensure();
return(function.Invoke(gridDim, blockDim, (IEnumerable <KernelArgument>)args));
}
public static KernelResult Run(this CUfunction function, dim3 gridDim, dim3 blockDim, IEnumerable <KernelArgument> args)
{
function.IsNotNull.AssertTrue();
args = args ?? Seq.Empty <KernelArgument>().ToArray();
CudaDriver.Ensure();
return(new JittedFunction(function).Run(gridDim, blockDim, args));
}
public static KernelResult Run(this JittedKernel kernel, dim3 gridDim, dim3 blockDim, params KernelArgument[] args)
{
kernel.AssertNotNull();
args = args ?? Seq.Empty <KernelArgument>().ToArray();
CudaDriver.Ensure();
return(kernel.Function.Run(gridDim, blockDim, args));
}
public static Object Invoke(this CUfunction function, dim3 gridDim, dim3 blockDim, params KernelArgument[] args)
{
function.IsNotNull.AssertTrue();
args = args ?? Seq.Empty <KernelArgument>().ToArray();
CudaDriver.Ensure();
return(new JittedFunction(function).Invoke(gridDim, blockDim, args));
}
public static KernelResult Run(this JittedFunction function, dim3 gridDim, dim3 blockDim, IEnumerable <KernelArgument> args)
{
function.AssertNotNull();
args = args ?? Seq.Empty <KernelArgument>().ToArray();
CudaDriver.Ensure();
var invocation = new KernelInvocation(function, args);
return(invocation.Launch(gridDim, blockDim));
}
public static JittedKernel JitKernel(this String ptx, dim3 reqntid, HardwareIsa target)
{
ptx.AssertNotNull();
CudaDriver.Ensure();
var tuning = new JitTuning {
Reqntid = reqntid
};
return(ptx.JitKernel(tuning, target));
}
public static Object Invoke(this JittedFunction function, dim3 gridDim, dim3 blockDim, IEnumerable <KernelArgument> args)
{
function.AssertNotNull();
args = args ?? Seq.Empty <KernelArgument>().ToArray();
CudaDriver.Ensure();
var invocation = new KernelInvocation(function, args);
var invocation_result = invocation.Launch(gridDim, blockDim);
var result = invocation_result.Result;
args.ForEach(arg => arg.Dispose());
return(result);
}
public static JittedKernel JitKernel(this String ptx, JitTuning tuning, HardwareIsa target)
{
ptx.AssertNotNull();
CudaDriver.Ensure();
var compiler = new JitCompiler();
compiler.Target = target;
compiler.Tuning = tuning;
var result = compiler.Compile(ptx);
return(new JittedKernel(result));
}
public static JittedKernel JitKernel(this String ptx, JitTuning tuning)
{
ptx.AssertNotNull();
CudaDriver.Ensure();
var compiler = new JitCompiler();
compiler.TargetFromContext = true;
compiler.Tuning = tuning;
var result = compiler.Compile(ptx);
return(new JittedKernel(result));
}
