protected T[] InternalExecuteMPI <T>(
byte[] kernelBinary,
String function,
int bufferSize,
ParallelTaskParams loaderParams,
params Object[] kernelParams) where T : struct
{
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointStart);
String binaryPath = Path.GetTempFileName();
File.WriteAllBytes(binaryPath, kernelBinary);
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointPlatformInit);
String mpiDirectory = SystemArchitecture.ProgramFolder(ArchitectureType.x86, @"OpenMPI*");
Process mpirunProcess = new Process();
mpirunProcess.StartInfo.CreateNoWindow = true;
mpirunProcess.StartInfo.UseShellExecute = false;
mpirunProcess.StartInfo.RedirectStandardOutput = true;
mpirunProcess.StartInfo.EnvironmentVariables["PATH"] += @";" + mpiDirectory + @"\bin";
mpirunProcess.StartInfo.FileName = mpiDirectory + @"\bin\mpirun.exe";
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointKernelBuild);
mpirunProcess.StartInfo.Arguments = String.Format("-n {0} \"{1}\" {2} {3} {4}",
loaderParams.ProcessCount,
ShortPath(binaryPath),
TypeName(typeof(T)),
function,
bufferSize);
foreach (Object param in kernelParams)
{
mpirunProcess.StartInfo.Arguments += " " + param.ToString();
}
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointDeviceWrite);
mpirunProcess.Start();
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointKernelExecute);
using (MemoryStream resultStream = new MemoryStream())
{
mpirunProcess.StandardOutput.BaseStream.CopyTo(resultStream);
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointDeviceRead);
byte[] processOutput = resultStream.ToArray();
T[] result = new T[(int)Math.Ceiling((float)processOutput.Length / Marshal.SizeOf(typeof(T)))];
Buffer.BlockCopy(processOutput, 0, result, 0, processOutput.Length);
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointPlatformDeinit);
return(result);
}
}
protected T[] InternalExecuteOpencl <T>(
String source,
String function,
int bufferSize,
ParallelTaskParams loaderParams,
params Object[] kernelParams)
where T : struct
{
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointStart);
ComputeCommandQueue queue = QueueWithDevice(loaderParams.OpenCLDevice);
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointPlatformInit);
String updatedSource = "#define OpenCL\r\n" + source;
ComputeProgram program = new ComputeProgram(queue.Context, updatedSource);
program.Build(new ComputeDevice[] { queue.Device }, null, null, IntPtr.Zero);
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointKernelBuild);
T[] resultBuffer = new T[bufferSize];
ComputeBuffer <T> resultBufferVar = new ComputeBuffer <T>(queue.Context, ComputeMemoryFlags.WriteOnly, bufferSize);
List <ComputeMemory> vars = new List <ComputeMemory>();
vars.Add(resultBufferVar);
vars.AddRange(WrapDeviceVariables(kernelParams, queue.Context));
ComputeKernel kernel = program.CreateKernel(function);
for (int i = 0; i < vars.Count; i++)
{
kernel.SetMemoryArgument(i, vars[i]);
}
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointDeviceWrite);
long[] workersGlobal = new long[2] {
loaderParams.GlobalWorkers.Width, loaderParams.GlobalWorkers.Height
};
queue.Execute(kernel, null, workersGlobal, null, null);
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointKernelExecute);
queue.ReadFromBuffer <T>(resultBufferVar, ref resultBuffer, false, null);
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointDeviceRead);
queue.Finish();
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointPlatformDeinit);
return(resultBuffer);
}
protected T[] InternalExecuteCuda <T>(
byte[] kernelBinary,
String function,
int bufferSize,
ParallelTaskParams loaderParams,
params Object[] kernelParams) where T : struct
{
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointStart);
CudaContext context = ContextWithDevice(loaderParams.CudaDevice);
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointPlatformInit);
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointKernelBuild);
CudaDeviceVariable <T> resultBufferVar = new CudaDeviceVariable <T>(bufferSize);
resultBufferVar.Memset(0);
List <Tuple <Object, IDisposable> > vars = new List <Tuple <Object, IDisposable> >();
vars.Add(new Tuple <Object, IDisposable>(resultBufferVar.DevicePointer, resultBufferVar));
vars.AddRange(WrapDeviceVariables(kernelParams, true));
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointDeviceWrite);
CudaKernel kernel = context.LoadKernelPTX(kernelBinary, function);
kernel.BlockDimensions = new dim3(loaderParams.BlockSize.Width, loaderParams.BlockSize.Height);
kernel.GridDimensions = new dim3(loaderParams.GridSize.Width, loaderParams.GridSize.Height);
kernel.Run(vars.Select(tuple => tuple.Item1).ToArray());
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointKernelExecute);
T[] resultBuffer = resultBufferVar;
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointDeviceRead);
vars.Where(tuple => tuple.Item2 != null).ToList().ForEach(tuple => tuple.Item2.Dispose());
TriggerCheckpoint(ParallelExecutionCheckpointType.CheckpointPlatformDeinit);
return(resultBuffer);
}
