private static DirtyImage ForwardCalculateB(Intracommunicator comm, GriddingConstants c, List <List <Subgrid> > metadata, Complex[,,] visibilities, double[,,] uvw, double[] frequencies, Complex[,] PsfCorrelation, float[,] psfCut, float maxSidelobe, Stopwatch watchIdg)
{
Stopwatch another = new Stopwatch();
comm.Barrier();
if (comm.Rank == 0)
{
watchIdg.Start();
}
var localGrid = IDG.Grid(c, metadata, visibilities, uvw, frequencies);
float[,] image = null;
float maxSideLobeLevel = 0.0f;
var grid_total = comm.Reduce <Complex[, ]>(localGrid, SequentialSum, 0);
if (comm.Rank == 0)
{
var dirtyImage = FFT.BackwardFloat(grid_total, c.VisibilitiesCount);
FFT.Shift(dirtyImage);
if (comm.Rank == 0)
{
FitsIO.Write(dirtyImage, "dirtyImage.fits");
}
maxSideLobeLevel = maxSidelobe * Residuals.GetMax(dirtyImage);
//remove spheroidal
image = Residuals.CalcGradientMap(dirtyImage, PsfCorrelation, new Rectangle(0, 0, psfCut.GetLength(0), psfCut.GetLength(1)));
watchIdg.Stop();
}
comm.Broadcast(ref maxSideLobeLevel, 0);
comm.Broadcast(ref image, 0);
return(new DirtyImage(image, maxSideLobeLevel));
}
// End StopSubTimer
// Synchronize a boolean across MPI processes
public static void SynchronizeMPIvariable(ref bool cosmicboolean)
{
if (MPI_Size > 1)
{
StartSubTimer(MPIBROADCASTTiming);
MPI_communicator.Broadcast(ref cosmicboolean, 0);
StopSubTimer(MPIBROADCASTTiming);
}
return;
}
private static int determineLatencyReps()
{
double t0, duration = 0;
int reps = 1, prev_reps = 0;
int i;
simpleAllreduce();
simpleAllreduce();
simpleAllreduce();
t0 = when();
t0 = when();
t0 = when();
while ((duration < 1) || (duration < 3 && reps < 1000))
{
t0 = when();
for (i = 0; i < reps - prev_reps; i++)
{
simpleAllreduce();
}
duration += when() - t0;
prev_reps = reps;
reps = reps * 2;
comm.Broadcast(ref duration, 0);
}
return(reps);
}
public static void Floyd(int[] local_mat, int n, int procid, int numproc, Intracommunicator comm)
{
int global_k, local_i, global_j, temp;
int root;
int[] row_k = new int[40];
for (global_k = 0; global_k < n; global_k++)
{
root = global_k / (n / numproc);
//root = Owner(global_k, numproc, n);
if (procid == root)
{
Copy_row(local_mat, n, numproc, row_k, global_k);
}
comm.Broadcast <int>(ref row_k, root);
for (local_i = 0; local_i < n / numproc; local_i++)
{
for (global_j = 0; global_j < n; global_j++)
{
temp = local_mat[local_i * n + global_k] + row_k[global_j];
if (temp < local_mat[local_i * n + global_j])
{
local_mat[local_i * n + global_j] = temp;
}
}
}
}
}
private void SendMatrixToWorkers(Intracommunicator communicator)
{
using (var perfCounter = new PerformanceCounter("Master. Broadcasting Jacoby: "))
{
var jacoby = this;
communicator.Broadcast <Jacoby>(ref jacoby, 0);
}
}
static void Main(string[] args)
{
using (new MPI.Environment(ref args))
{
int root = 0;
int arraySize = 100;
if (args.Length == 1)
{
arraySize = Convert.ToInt32(args[0]);
}
Intracommunicator comm = Communicator.world;
int[] lotsOfNumbers = new int[arraySize];
if (comm.Rank == 0)
{
for (int i = 0; i < lotsOfNumbers.Length; i++)
{
lotsOfNumbers[i] = i;
}
}
int sum = 0;
comm.Broadcast(ref lotsOfNumbers, 0);
//divides up the work
int x = arraySize / comm.Size;
int startingIndex = comm.Rank * x;
int endingIndex = startingIndex + x;
if (comm.Rank == comm.Size - 1)
{
endingIndex = lotsOfNumbers.Length;
}
for (int i = startingIndex; i < endingIndex; i++)
{
sum += lotsOfNumbers[i];
}
Console.WriteLine("Rank " + comm.Rank + ": " + "summed the numbers from index " + startingIndex + " to index " + (endingIndex - 1) + " and got " + sum + ".");
int totalSum = comm.Reduce(sum, Operation <int> .Add, root);
if (comm.Rank == root)
{
Console.WriteLine("The total sum is: " + totalSum);
}
}
}
private static float[,] CalculatePSF(Intracommunicator comm, GriddingConstants c, List <List <Subgrid> > metadata, double[,,] uvw, bool[,,] flags, double[] frequencies)
{
float[,] psf = null;
var localGrid = IDG.GridPSF(c, metadata, uvw, flags, frequencies);
var psf_total = comm.Reduce(localGrid, SequentialSum, 0);
if (comm.Rank == 0)
{
psf = FFT.BackwardFloat(psf_total, c.VisibilitiesCount);
FFT.Shift(psf);
}
comm.Broadcast(ref psf, 0);
return(psf);
}
public static void MPIWorker(Intracommunicator communicator)
{
Jacoby jacoby = null;
ReceiveRequest request;
using (var perf = new PerformanceCounter($"Comm {communicator.Rank} waiting for Jacoby duration: "))
{
communicator.Broadcast <Jacoby>(ref jacoby, 0);
}
bool inProcess = true;
using (var totalPerformance = new PerformanceCounter($"Comm {communicator.Rank} total work duration: "))
{
int cycleIndex = 0;
while (inProcess)
{
using (var perf = new PerformanceCounter($"Comm {communicator.Rank} cycle {cycleIndex} duration: "))
{
JobToDo jobToDo = null;
using (var perfRecv = new PerformanceCounter($"Comm {communicator.Rank} waiting for job-to-do (cycle {cycleIndex}): "))
{
request = communicator.ImmediateReceive <JobToDo>(0, (int)JacobyMessageType.CalculatePart);
request.Wait();
jobToDo = request.GetValue() as JobToDo;
if (jobToDo.finish)
{
Console.WriteLine($"Comm {communicator.Rank} receive FINISH signal");
return;
}
}
using (var perfWork = new PerformanceCounter($"Comm {communicator.Rank} work duration (cycle {cycleIndex}): "))
{
Console.WriteLine($"Comm {communicator.Rank} start={jobToDo.start} end={jobToDo.end}");
var result = jacoby.PartialCalculation(jobToDo.start, jobToDo.end, jobToDo.initial);
var srequest = communicator.ImmediateSend <JobDone>(new JobDone(result), 0, (int)JacobyMessageType.JobDone);
}
}
cycleIndex++;
}
}
}
static void Main(string[] args)
{
using (new MPI.Environment(ref args))
{
Intracommunicator comm = Communicator.world;
int Id = comm.Rank;
Connect4 game = new Connect4(comm, comm.Size, Id);
if (Id == 0)
{
Board board = new Board();
//game.CalculateNextMove(board);
Stopwatch timer = new Stopwatch();
timer.Start();
Console.WriteLine("Igra krece!");
while (true)
{
board.Draw();
bool kraj = false;
while (true)
{
Console.WriteLine("Daj potez:");
try
{
int column = Int32.Parse(Console.ReadLine());
if (board.Put(column, 1))
{
kraj = true;
}
break;
}
catch (Exception)
{
}
}
if (kraj)
{
Console.WriteLine("Pobijedio je igrac!");
break;
}
timer.Reset();
timer.Start();
int nextMove = game.CalculateNextMove(board.Duplicate());
Console.WriteLine("Potrebno vrijeme za izračun: " + timer.ElapsedMilliseconds);
Console.WriteLine("Računalo igra potez " + nextMove);
if (board.Put(nextMove, 2))
{
Console.WriteLine("Pobijedilo je računalo!");
break;
}
}
board.Draw();
Message exitMsg = new Message(0, EXIT);
comm.Broadcast <Message>(ref exitMsg, 0);
}
else
{
while (true)
{
Message msg = null;
comm.Broadcast <Message>(ref msg, 0);
if (msg.MessageType == EXIT)
{
break;
}
comm.Send <Message>(new Message(Id, NEXT_TASK), 0, 0);
while (true)
{
msg = comm.Receive <Message>(0, 0);
if (msg.MessageType == STOP)
{
break;
}
else if (msg.MessageType == TASK)
{
Task task = msg.task;
task.value = game.CalculateStateValue(task.b, game.Reverse(task.nextPlayer), -1, 0);
Message answer = new Message(Id, RESULT);
answer.task = task;
comm.Send <Message>(answer, 0, 0);
}
}
}
}
}
}
static void Main(string[] args)
{
// By default, test with primitives types and the predefined MPI_SUM
Test test = testPrimitiveAndPredefined;
using (new MPI.Environment(ref args))
{
if (args.Length > 0 && args[0] == "/direct")
{
test = testDirect;
System.Console.WriteLine("Using direct MPI interface.");
}
else if (args.Length > 0 && args[0] == "/user")
{
test = testPrimitiveAndMethod;
Operation <double> .UseGeneratedUserOps = true;
System.Console.WriteLine("Using primitive type (double) with user-defined sum and run-time code generation");
}
else if (args.Length > 0 && args[0] == "/marshaluser")
{
test = testPrimitiveAndMethod;
Operation <double> .UseGeneratedUserOps = false;
System.Console.WriteLine("Using primitive type (double) with user-defined sum and marshalling");
}
else if (args.Length > 0 && args[0] == "/valuetype")
{
test = testValueType;
System.Console.WriteLine("Using value types with user-defined sum");
}
else if (args.Length > 0 && args[0] == "/reftype")
{
test = testReferenceType;
System.Console.WriteLine("Using reference types with user-defined sum");
}
else
{
System.Console.WriteLine("Using MPI.NET interface.");
}
comm = MPI.Communicator.world;
self = comm.Rank;
System.Console.WriteLine(comm.Rank + ": " + MPI.Environment.ProcessorName);
bwstats = new Stats[nSamp];
testLatency();
testSyncTime();
comm.Broadcast(ref latency, 0);
if (self == 0)
{
System.Console.WriteLine("Latency: {0:F9}", latency);
System.Console.WriteLine("Sync Time: {0:F9}", synctime);
System.Console.WriteLine("Now starting main loop");
}
int n, nq;
int inc = 1, len;
int start = 0, end = 1024 * 1024 * 1024;
int bufflen = start;
double tlast = latency;
for (n = nq = 0, len = start; tlast < stopTime && len <= end; len += inc, nq++)
{
if (nq > 2 && (nq % 2 != 0))
{
inc *= 2;
}
int ipert, pert;
for (ipert = 0, pert = (inc > PERT + 1) ? -PERT : 0;
pert <= PERT;
ipert++, n++, pert += (inc > PERT + 1) ? PERT : PERT + 1)
{
int nRepeat = bufflen == 0 ?
latencyReps :
(int)Math.Max((RUNTM / ((double)bufflen / (bufflen - inc + 1.0) * tlast)),
TRIALS);
comm.Broadcast(ref nRepeat, 0);
bufflen = len + pert;
if (self == 0)
{
System.Console.Write("{0,3:D}: {1,9:D} doubles {2,7:D} times ---> ", n, bufflen, nRepeat);
}
GC.Collect();
test(bufflen, n, nRepeat, ref tlast);
if (self == 0)
{
System.Console.WriteLine("{0,9:F2} Mbps in {1:F9} sec", bwstats[n].bps, tlast);
}
}
}
}
}
// return true if it converges. Output: solution matrix, errors, loops it took
public static Boolean solve(Matrix A, Matrix b, out Matrix x, out Matrix err, out int loops, Intracommunicator comm)
{
// check sanity. rank 0 only
if (comm.Rank == 0 && (!A.isSquare || !b.isColumn || (A.Height != b.Height)))
{
Exception e = new Exception("Matrix A must be square! Matrix b must be a column matrix with the same height as matrix A!");
throw e;
}
// follow samples in Wikipedia step by step https://en.wikipedia.org/wiki/Gauss%E2%80%93Seidel_method
benchmark bm = new benchmark(), bm2 = new benchmark(), bm3 = new benchmark();
double sequential = 0, parallel = 0, communication = 0;
bm.start();
bm2.start();
// decompose A into the sum of a lower triangular component L* and a strict upper triangular component U
int size = 0; Matrix L = null, U = null, L_1;
if (comm.Rank == 0)
{
size = A.Height;
Matrix.Decompose(A, out L, out U);
}
bm2.pause();
sequential += bm2.getElapsedSeconds();
bm2.start();
comm.Broadcast(ref size, 0);
comm.Broadcast(ref U, 0);
comm.Broadcast(ref b, 0);
bm2.pause();
communication += bm2.getElapsedSeconds();
// Inverse matrix L*
comm.Barrier();
L_1 = MatrixParallel.Inverse(L, comm, ref sequential, ref parallel, ref communication);
// Main iteration: x (at step k+1) = T * x (at step k) + C
// where T = - (inverse of L*) * U, and C = (inverse of L*) * b
// split T & C into groups of rows, each for one slave, according to the nature of this algorithm
// each slave will have one piece of T & one piece of C stored locally. the rest of T & C is not needed
// there might be cases where jobs > slaves, so some might get no job at all
// Changes: only split L_1. Slaves will calculate T & C (pieces) themselves
bm2.start();
Matrix jobDistro = Utils.splitJob(size, comm.Size);
int startRow = 0, endRow = 0, myJobSize = (int)jobDistro[0, comm.Rank];
for (int p = 0; p < comm.Size; p++)
{
if (p != comm.Rank)
{
startRow += (int)jobDistro[0, p];
}
else
{
endRow = startRow + (int)jobDistro[0, p] - 1;
break;
}
}
Matrix[] L_1Ps = new Matrix[comm.Size];
if (comm.Rank == 0)
{
int slaveStart = 0;
for (int p = 0; p < comm.Size; p++)
{
L_1Ps[p] = Matrix.extractRows(L_1, slaveStart, slaveStart + (int)jobDistro[0, p] - 1);
slaveStart += (int)jobDistro[0, p];
}
}
bm2.pause();
sequential += bm2.getElapsedSeconds();
bm2.start();
Matrix L_1P = comm.Scatter(L_1Ps, 0);
bm2.pause();
communication += bm2.getElapsedSeconds();
bm2.start();
Matrix T = -L_1P * U; Matrix C = L_1P * b;
bm2.pause();
parallel += bm2.getElapsedSeconds();
// the actual iteration
// if it still doesn't converge after this many loops, assume it won't converge and give up
Boolean converge = false;
int loopLimit = 100;
x = Matrix.zeroLike(b); // at step k
for (loops = 0; loops < loopLimit; loops++)
{
bm3.start();
// (re-)distributing x vector. Must be done every single loop
// this loop needs x from the previous loop
comm.Broadcast(ref x, 0);
bm3.pause();
communication += bm3.getElapsedSeconds();
// calculation step
bm3.start();
comm.Barrier();
Matrix new_x = T * x + C;
// check convergence
converge = Matrix.SomeClose(new_x, x, 1e-15, startRow);
// collect result x
comm.Barrier();
x = comm.Reduce(new_x, Matrix.Concatenate, 0);
// collect convergence. consider converged if ALL slaves claim so
converge = comm.Reduce(converge, bothTrue, 0);
comm.Broadcast(ref converge, 0); // make sure EVERYONE breaks/coninues
bm3.pause();
parallel += bm3.getElapsedSeconds();
if (converge)
{
loops++;
break;
}
}
bm2.start();
// round the result slightly
err = null;
if (comm.Rank == 0)
{
x.Round(1e-14);
err = A * x - b;
err.Round(1e-14);
}
bm2.pause();
sequential += bm2.getElapsedSeconds();
bm.pause();
if (showBenchmark)
{
Console.WriteLine("Sequential part took " + sequential + " secs.");
Console.WriteLine("Parallel part took " + parallel + " secs.");
Console.WriteLine("Communication took " + communication + " secs.");
Console.WriteLine("Total: " + bm.getResult() + " (" + bm.getElapsedSeconds() + " secs). Seq + Parallel: " + (sequential + parallel));
}
return(converge);
}
public static float[,] Reconstruct(Intracommunicator comm, DistributedData.LocalDataset local, GriddingConstants c, int maxCycle, float lambda, float alpha, int iterPerCycle = 1000, bool usePathDeconvolution = false)
{
var watchTotal = new Stopwatch();
var watchForward = new Stopwatch();
var watchBackward = new Stopwatch();
var watchDeconv = new Stopwatch();
watchTotal.Start();
var metadata = Partitioner.CreatePartition(c, local.UVW, local.Frequencies);
var patchSize = CalculateLocalImageSection(comm.Rank, comm.Size, c.GridSize, c.GridSize);
var totalSize = new Rectangle(0, 0, c.GridSize, c.GridSize);
//calculate psf and prepare for correlation in the Fourier space
var psf = CalculatePSF(comm, c, metadata, local.UVW, local.Flags, local.Frequencies);
Complex[,] PsfCorrelation = null;
var maxSidelobe = PSF.CalcMaxSidelobe(psf);
lambda = (float)(lambda * PSF.CalcMaxLipschitz(psf));
StreamWriter writer = null;
if (comm.Rank == 0)
{
FitsIO.Write(psf, "psf.fits");
Console.WriteLine("done PSF gridding ");
PsfCorrelation = PSF.CalcPaddedFourierCorrelation(psf, totalSize);
writer = new StreamWriter(comm.Size + "runtimestats.txt");
}
var deconvovler = new MPIGreedyCD(comm, totalSize, patchSize, psf);
var residualVis = local.Visibilities;
var xLocal = new float[patchSize.YEnd - patchSize.Y, patchSize.XEnd - patchSize.X];
for (int cycle = 0; cycle < maxCycle; cycle++)
{
if (comm.Rank == 0)
{
Console.WriteLine("cycle " + cycle);
}
var dirtyImage = ForwardCalculateB(comm, c, metadata, residualVis, local.UVW, local.Frequencies, PsfCorrelation, psf, maxSidelobe, watchForward);
var bLocal = GetImgSection(dirtyImage.Image, patchSize);
MPIGreedyCD.Statistics lastRun;
if (usePathDeconvolution)
{
var currentLambda = Math.Max(1.0f / alpha * dirtyImage.MaxSidelobeLevel, lambda);
lastRun = deconvovler.DeconvolvePath(xLocal, bLocal, currentLambda, 4.0f, alpha, 5, iterPerCycle, 2e-5f);
}
else
{
lastRun = deconvovler.Deconvolve(xLocal, bLocal, lambda, alpha, iterPerCycle, 1e-5f);
}
if (comm.Rank == 0)
{
WriteToFile(cycle, lastRun, writer);
if (lastRun.Converged)
{
Console.WriteLine("-----------------------------CONVERGED!!!!------------------------");
}
else
{
Console.WriteLine("-------------------------------not converged----------------------");
}
}
comm.Barrier();
if (comm.Rank == 0)
{
watchDeconv.Stop();
}
float[][,] totalX = null;
comm.Gather(xLocal, 0, ref totalX);
Complex[,] modelGrid = null;
if (comm.Rank == 0)
{
watchBackward.Start();
var x = new float[c.GridSize, c.GridSize];
StitchImage(totalX, x, comm.Size);
FitsIO.Write(x, "xImage_" + cycle + ".fits");
FFT.Shift(x);
modelGrid = FFT.Forward(x);
}
comm.Broadcast(ref modelGrid, 0);
var modelVis = IDG.DeGrid(c, metadata, modelGrid, local.UVW, local.Frequencies);
residualVis = Visibilities.Substract(local.Visibilities, modelVis, local.Flags);
}
writer.Close();
float[][,] gatherX = null;
comm.Gather(xLocal, 0, ref gatherX);
float[,] reconstructed = null;
if (comm.Rank == 0)
{
reconstructed = new float[c.GridSize, c.GridSize];;
StitchImage(gatherX, reconstructed, comm.Size);
}
return(reconstructed);
}
static void Main(string[] args)
{
// By default, test with primitives types and the predefined MPI_SUM
Test test = testPrimitiveAndPredefined;
using (new MPI.Environment(ref args))
{
if (args.Length > 0 && args[0] == "/direct")
{
test = testDirect;
System.Console.WriteLine("Using direct MPI interface.");
}
else if (args.Length > 0 && args[0] == "/user")
{
test = testPrimitiveAndMethod;
Operation<double>.UseGeneratedUserOps = true;
System.Console.WriteLine("Using primitive type (double) with user-defined sum and run-time code generation");
}
else if (args.Length > 0 && args[0] == "/marshaluser")
{
test = testPrimitiveAndMethod;
Operation<double>.UseGeneratedUserOps = false;
System.Console.WriteLine("Using primitive type (double) with user-defined sum and marshalling");
}
else if (args.Length > 0 && args[0] == "/valuetype")
{
test = testValueType;
System.Console.WriteLine("Using value types with user-defined sum");
}
else if (args.Length > 0 && args[0] == "/reftype")
{
test = testReferenceType;
System.Console.WriteLine("Using reference types with user-defined sum");
}
else
System.Console.WriteLine("Using MPI.NET interface.");
comm = MPI.Communicator.world;
self = comm.Rank;
System.Console.WriteLine(comm.Rank + ": " + MPI.Environment.ProcessorName);
bwstats = new Stats[nSamp];
testLatency();
testSyncTime();
comm.Broadcast(ref latency, 0);
if (self == 0)
{
System.Console.WriteLine("Latency: {0:F9}", latency);
System.Console.WriteLine("Sync Time: {0:F9}", synctime);
System.Console.WriteLine("Now starting main loop");
}
int n, nq;
int inc = 1, len;
int start = 0, end = 1024 * 1024 * 1024;
int bufflen = start;
double tlast = latency;
for (n = nq = 0, len = start; tlast < stopTime && len <= end; len += inc, nq++)
{
if (nq > 2 && (nq % 2 != 0)) inc *= 2;
int ipert, pert;
for (ipert = 0, pert = (inc > PERT + 1) ? -PERT : 0;
pert <= PERT;
ipert++, n++, pert += (inc > PERT + 1) ? PERT : PERT + 1)
{
int nRepeat = bufflen == 0 ?
latencyReps :
(int)Math.Max((RUNTM / ((double)bufflen / (bufflen - inc + 1.0) * tlast)),
TRIALS);
comm.Broadcast(ref nRepeat, 0);
bufflen = len + pert;
if (self == 0)
System.Console.Write("{0,3:D}: {1,9:D} doubles {2,7:D} times ---> ", n, bufflen, nRepeat);
GC.Collect();
test(bufflen, n, nRepeat, ref tlast);
if (self == 0)
System.Console.WriteLine("{0,9:F2} Mbps in {1:F9} sec", bwstats[n].bps, tlast);
}
}
}
}
public int CalculateNextMove(Board b)
{
List <Task> tasks = GenerateTasks(new Board(), "", PLAYER, -1, 0, new List <Task>());
var map = new Dictionary <string, double>();
int stopped = 0;
if (N == 1)
{
foreach (Task task in tasks)
{
map[task.key] = CalculateStateValue(task.b, Reverse(task.nextPlayer), -1, 0);
}
}
else
{
Message startMsg = new Message(0, START);
comm.Broadcast <Message>(ref startMsg, 0);
Message msg;
while (tasks.Any() || stopped < N - 1)
{
msg = comm.Receive <Message>(MPI.Unsafe.MPI_ANY_SOURCE, 0);
if (msg.MessageType == RESULT)
{
map[msg.task.key] = msg.task.value;
}
if (tasks.Any())
{
Message newTask = new Message(0, TASK);
newTask.task = tasks[0];
tasks.RemoveAt(0);
comm.Send <Message>(newTask, msg.SourceId, 0);
}
else
{
stopped++;
comm.Send <Message>(new Message(0, STOP), msg.SourceId, 0);
}
}
}
double bestSol = Double.MinValue, currSol;
int bestMove = -1;
for (int move = 0; move < C; move++)
{
if (b.columns[move].LastPosition() < R - 1)
{
currSol = CalculateMoveValue(b, "", COMPUTER, move, 1, map);
Console.WriteLine(move + " " + currSol);
if (currSol > bestSol)
{
bestSol = currSol;
bestMove = move;
}
}
}
return(bestMove);
}
public int initialConfig()
{
int fstatus = 0, niter = 0;
if (node == 0)
{
Console.WriteLine(" NAS Parallel Benchmarks " + "3.3" + " -- FT Benchmark ");
try {
Console.Write("Trying Read from input file inputft.data: ");
int[] conf = { 1, 1, 2 };//Line 1: 1 var; Line 2: 1 var; Line 3: 2 vars;
string[] vetTemp = readFileData("inputft.data", conf);
niter = int.Parse(vetTemp[0]);
layout_type = int.Parse(vetTemp[1]);
np1 = int.Parse(vetTemp[2]);
np2 = int.Parse(vetTemp[3]);
}
catch {
Console.WriteLine("inputft.data not found");
fstatus = 1;
}
if (fstatus == 0)
{
Console.WriteLine("inputft.data found");
if (np1 * np2 != np)
{
Console.WriteLine(" np1 and np2 given in input file are not valid.");
Console.WriteLine("Product is " + np1 * np2 + " and should be " + np);
System.Environment.Exit(0);
}
if (layout_type != layout_0D && layout_type != layout_1D && layout_type != layout_2D)
{
Console.WriteLine(" Layout type specified in inputft.data is invalid ");
System.Environment.Exit(0);
}
if (layout_type == layout_0D && (np1 != 1 || np2 != 1))
{
Console.WriteLine(" For 0D layout, both np1 and np2 must be 1 ");
System.Environment.Exit(0);
}
if (layout_type == layout_1D && np1 != 1)
{
Console.WriteLine(" For 1D layout, np1 must be 1 ");
System.Environment.Exit(0);
}
}
else
{
Console.WriteLine(" No input file inputft.data. Using compiled defaults");
niter = niter_default;
if (np == 1)
{
np1 = 1;
np2 = 1;
layout_type = layout_0D;
}
else if (np <= nz)
{
np1 = 1;
np2 = np;
layout_type = layout_1D;
}
else
{
np1 = nz;
np2 = np / nz;
layout_type = layout_2D;
}
}
Console.WriteLine(" Size: " + nx + "x" + ny + "x" + nz);
Console.WriteLine(" Iterations: " + niter);
Console.WriteLine(" Number of processes : " + np);
Console.WriteLine(" Processor array : " + np1 + "x" + np2);
if (layout_type == layout_0D)
{
Console.WriteLine(" Layout type: OD");
}
else if (layout_type == layout_1D)
{
Console.WriteLine(" Layout type: 1D");
}
else
{
Console.WriteLine(" Layout type: 2D");
}
}
worldcomm.Broadcast <int>(ref np1, root);
worldcomm.Broadcast <int>(ref niter, root);
worldcomm.Broadcast <int>(ref np2, root);
if (np1 == 1 && np2 == 1)
{
layout_type = layout_0D;
}
else if (np1 == 1)
{
layout_type = layout_1D;
}
else
{
layout_type = layout_2D;
}
return(niter);
}
public void Broadcast <T>(ref T message, int rank)
{
communicator.Broadcast(ref message, rank);
}
static void Main(String[] args)
{
using (new MPI.Environment(ref args))
{
Intracommunicator comm = MPI.Communicator.world;
int p = comm.Size;
int rank = comm.Rank;
int range = 0, rem = 0, tag = 0;
int N = 0; // the number of subparts of an image
int noOfComponent = 0;
int noOfPersons = 0; // the no of persons in the training data (here it's 5)
string minLabel = "";
double minDistance = 0;
//this is used to get the distance of the training data and the predicted image
EuclideanDistance ed = new EuclideanDistance();
//this list will containt the parts of image after deviding it to N parts
List <double[][]> testSubImgs = new List <double[][]>();
//this will contain the weighs of each persono (N parts for each person)
List <double[][]> weights = new List <double[][]>();
//the label associated with each weight
List <string> labels = new List <string>();
if (rank == 0) // It's the root
{
//create an object of the algorithm
ModularFaceRecognitionAlgorithms mpca = new ModularFaceRecognitionAlgorithms();
//load the training data from file
mpca.loadTrainingData("C:/train.txt");
//prepare the image for testing
// u can change s1 to s2,s3,s4 ... s5 and watch out the result
String filePath = "C:/test/s4/10.bmp";
Matrix test = FileManager.GetBitMapColorMatrix(filePath);
//divide the image into N parts
testSubImgs = testSubImgs = mpca.devideImageToN(test, mpca.N);
//prepare local variables
noOfPersons = mpca.weights.Count / mpca.N;
N = mpca.N;
noOfComponent = mpca.numOfComponents;
weights = mpca.weights;
labels = mpca.labels;
if (p > 1) //this cond. to handle the exception of a single master process
{
//compute the no. of persons checked per process
//each process will be resposible for a nubmber of persons
//the process returns the dist. and label of the min distance of its persons
range = noOfPersons / (p - 1);
rem = noOfPersons % (p - 1);
if (range == 0) // in case for ex. we have 5 persons and 6 slaves
{
range = 1;
rem = 0;
}
}
else
{
Console.WriteLine("There's only a master process");
}
//broadcast the needed variables
comm.Broadcast(ref N, 0);
comm.Broadcast(ref noOfComponent, 0);
comm.Broadcast(ref range, 0);
comm.Broadcast(ref rem, 0);
comm.Broadcast(ref testSubImgs, 0);
comm.Broadcast(ref weights, 0);
comm.Broadcast(ref labels, 0);
comm.Broadcast(ref noOfPersons, 0);
string resLabel = ""; //the final resulted label
double resDistance = 0; // the final resulted distance
minLabel = ""; //used to receive the min label of each slave
minDistance = 0; //used to reciec the min distance of each slave
//in the following for loop we are receiving the min distance and label
//resulted from each slave and then get the min of them all
// the resulted resLabel and resDistance is the final result
//these line is used to handle if we have processes more than the noOfpersons
int endLoop = p - 1;
if (noOfPersons < (p - 1))
{
endLoop = noOfPersons;
}
for (int src = 1; src <= endLoop; src++)
{
comm.Receive(src, tag, out minDistance);
comm.Receive(src, tag, out minLabel);
if (src == 1 || minDistance < resDistance)
{
resLabel = minLabel;
resDistance = minDistance;
}
}
Console.WriteLine("resLabel = " + resLabel);
Console.WriteLine("resDistance = " + resDistance);
}
else
{
comm.Broadcast(ref N, 0);
comm.Broadcast(ref noOfComponent, 0);
comm.Broadcast(ref range, 0);
comm.Broadcast(ref rem, 0);
comm.Broadcast(ref testSubImgs, 0);
comm.Broadcast(ref weights, 0);
comm.Broadcast(ref labels, 0);
comm.Broadcast(ref noOfPersons, 0);
if (rank <= noOfPersons) //other wise do nothing
{
if (rank <= rem)
{
range++;
}
int start = 0;
if (rank <= rem)
{
start = (rank - 1) * N + (rank - 1) * N;
}
else
{
start = (rank - 1) * N + rem * N;
}
//As we mentioned before the range is the number of personse per process
// so in this for loop we are calculating the distance of each person
// and eventually send the min distance and lable to the master process
for (int i = 0; i < range; i++)
{
double dpj = 0;
int begin = i * N + start;
for (int j = begin, m = 0; m < N; j++, m++)
{
double dist = ed.getDistance(weights[j], testSubImgs[m]);
dpj += ((double)1 / noOfComponent) * dist;
}
double dp = ((double)1 / N) * dpj;
if (i == 0 || dp < minDistance)
{
minLabel = labels[begin];
minDistance = dp;
}
}
comm.Send(minDistance, 0, 0);
comm.Send(minLabel, 0, 0);
}
}
}
}
static void Main(string[] args)
{
// Whether we should use the unsafe, Direct interface to MPI.
// When false, use the normal MPI.NET interface.
bool useDirectInterface = false;
using (MPI.Environment env = new MPI.Environment(ref args))
{
if (args.Length > 0 && args[0] == "/direct")
{
useDirectInterface = true;
System.Console.WriteLine("Using direct MPI interface.");
}
else
{
System.Console.WriteLine("Using MPI.NET interface.");
}
comm = MPI.Communicator.world;
if (comm.Size != 2)
{
if (comm.Rank == 0)
{
System.Console.WriteLine("Only two processes allowed. Rerun with -np 2");
}
return;
}
else
{
self = comm.Rank;
other = (comm.Rank + 1) % 2;
}
System.Console.WriteLine(comm.Rank + ": " + MPI.Environment.ProcessorName);
bwstats = new Stats[nSamp];
testLatency();
testSyncTime();
comm.Broadcast(ref latency, 0);
if (self == 0)
{
System.Console.WriteLine("Latency: {0:F9}", latency);
System.Console.WriteLine("Sync Time: {0:F9}", synctime);
System.Console.WriteLine("Now starting main loop");
}
int i, j, n, nq;
int inc = 1, len;
int start = 0, end = 1024 * 1024 * 1024;
int bufflen = start;
double tlast = latency;
for (n = nq = 0, len = start; tlast < stopTime && len <= end; len += inc, nq++)
{
if (nq > 2 && (nq % 2 != 0))
{
inc *= 2;
}
int ipert, pert;
for (ipert = 0, pert = (inc > PERT + 1) ? -PERT : 0;
pert <= PERT;
ipert++, n++, pert += (inc > PERT + 1) ? PERT : PERT + 1)
{
int nRepeat = bufflen == 0 ?
latencyReps :
(int)Math.Max((RUNTM / ((double)bufflen / (bufflen - inc + 1.0) * tlast)),
TRIALS);
comm.Broadcast(ref nRepeat, 0);
bufflen = len + pert;
byte[] sendBuffer = new byte[bufflen]; // Align the data? Some day. Maybe.
byte[] recvBuffer = new byte[bufflen];
if (self == 0)
{
System.Console.Write("{0,3:D}: {1,9:D} bytes {2,7:D} times ---> ", n, bufflen, nRepeat);
}
bwstats[n].t = 1e99;
double t1 = 0, t2 = 0;
for (i = 0; i < TRIALS; i++)
{
sync();
double t0 = when();
if (useDirectInterface)
{
// Use the unsafe, direct interface to MPI via P/Invoke
unsafe
{
fixed(byte *sendPtr = sendBuffer, recvPtr = recvBuffer)
{
for (j = 0; j < nRepeat; j++)
{
if (self == 0)
{
Unsafe.MPI_Send(new IntPtr(sendPtr), bufflen, Unsafe.MPI_BYTE, other, 142, Unsafe.MPI_COMM_WORLD);
Unsafe.MPI_Recv(new IntPtr(recvPtr), bufflen, Unsafe.MPI_BYTE, other, 242, Unsafe.MPI_COMM_WORLD, out *Unsafe.MPI_STATUS_IGNORE);
}
else
{
Unsafe.MPI_Recv(new IntPtr(recvPtr), bufflen, Unsafe.MPI_BYTE, other, 142, Unsafe.MPI_COMM_WORLD, out *Unsafe.MPI_STATUS_IGNORE);
Unsafe.MPI_Send(new IntPtr(sendPtr), bufflen, Unsafe.MPI_BYTE, other, 242, Unsafe.MPI_COMM_WORLD);
}
}
}
}
}
else
{
for (j = 0; j < nRepeat; j++)
{
if (self == 0)
{
comm.Send(sendBuffer, other, 142);
comm.Receive(other, 242, ref recvBuffer);
}
else
{
comm.Receive(other, 142, ref recvBuffer);
comm.Send(sendBuffer, other, 242);
}
}
}
double t = (when() - t0) / (2.0 * nRepeat);
t2 += t * t;
t1 += t;
bwstats[n].t = Math.Min(bwstats[n].t, t);
bwstats[n].variance = t2 / TRIALS - t1 / TRIALS * t1 / TRIALS;
tlast = bwstats[n].t;
bwstats[n].bits = bufflen * sizeof(byte) * 8;
bwstats[n].bps = bwstats[n].bits / (bwstats[n].t * 1024 * 1024);
bwstats[n].repeat = nRepeat;
}
if (self == 0)
{
System.Console.WriteLine("{0,9:F2} Mbps in {1:F9} sec", bwstats[n].bps, tlast);
}
}
}
}
}
static void Main(string[] args)
{
// Whether we should use the unsafe, Direct interface to MPI.
// When false, use the normal MPI.NET interface.
bool useDirectInterface = false;
using (MPI.Environment env = new MPI.Environment(ref args))
{
if (args.Length > 0 && args[0] == "/direct")
{
useDirectInterface = true;
System.Console.WriteLine("Using direct MPI interface.");
}
else
System.Console.WriteLine("Using MPI.NET interface.");
comm = MPI.Communicator.world;
if (comm.Size != 2)
{
if (comm.Rank == 0)
System.Console.WriteLine("Only two processes allowed. Rerun with -np 2");
return;
}
else
{
self = comm.Rank;
other = (comm.Rank + 1) % 2;
}
System.Console.WriteLine(comm.Rank + ": " + MPI.Environment.ProcessorName);
bwstats = new Stats[nSamp];
testLatency();
testSyncTime();
comm.Broadcast(ref latency, 0);
if (self == 0)
{
System.Console.WriteLine("Latency: {0:F9}", latency);
System.Console.WriteLine("Sync Time: {0:F9}", synctime);
System.Console.WriteLine("Now starting main loop");
}
int i, j, n, nq;
int inc = 1, len;
int start = 0, end = 1024 * 1024 * 1024;
int bufflen = start;
double tlast = latency;
for (n = nq = 0, len = start; tlast < stopTime && len <= end; len += inc, nq++)
{
if (nq > 2 && (nq % 2 != 0)) inc *= 2;
int ipert, pert;
for (ipert = 0, pert = (inc > PERT + 1) ? -PERT : 0;
pert <= PERT;
ipert++, n++, pert += (inc > PERT + 1) ? PERT : PERT + 1)
{
int nRepeat = bufflen == 0 ?
latencyReps :
(int)Math.Max((RUNTM / ((double)bufflen / (bufflen - inc + 1.0) * tlast)),
TRIALS);
comm.Broadcast(ref nRepeat, 0);
bufflen = len + pert;
byte[] sendBuffer = new byte[bufflen]; // Align the data? Some day. Maybe.
byte[] recvBuffer = new byte[bufflen];
if (self == 0)
System.Console.Write("{0,3:D}: {1,9:D} bytes {2,7:D} times ---> ", n, bufflen, nRepeat);
bwstats[n].t = 1e99;
double t1 = 0, t2 = 0;
for (i = 0; i < TRIALS; i++)
{
sync();
double t0 = when();
if (useDirectInterface)
{
// Use the unsafe, direct interface to MPI via P/Invoke
unsafe
{
fixed (byte* sendPtr = sendBuffer, recvPtr = recvBuffer)
{
for (j = 0; j < nRepeat; j++)
{
if (self == 0)
{
Unsafe.MPI_Send(new IntPtr(sendPtr), bufflen, Unsafe.MPI_BYTE, other, 142, Unsafe.MPI_COMM_WORLD);
Unsafe.MPI_Recv(new IntPtr(recvPtr), bufflen, Unsafe.MPI_BYTE, other, 242, Unsafe.MPI_COMM_WORLD, out *Unsafe.MPI_STATUS_IGNORE);
}
else
{
Unsafe.MPI_Recv(new IntPtr(recvPtr), bufflen, Unsafe.MPI_BYTE, other, 142, Unsafe.MPI_COMM_WORLD, out *Unsafe.MPI_STATUS_IGNORE);
Unsafe.MPI_Send(new IntPtr(sendPtr), bufflen, Unsafe.MPI_BYTE, other, 242, Unsafe.MPI_COMM_WORLD);
}
}
}
}
}
else
{
for (j = 0; j < nRepeat; j++)
{
if (self == 0)
{
comm.Send(sendBuffer, other, 142);
comm.Receive(other, 242, ref recvBuffer);
}
else
{
comm.Receive(other, 142, ref recvBuffer);
comm.Send(sendBuffer, other, 242);
}
}
}
double t = (when() - t0) / (2.0 * nRepeat);
t2 += t*t;
t1 += t;
bwstats[n].t = Math.Min(bwstats[n].t, t);
bwstats[n].variance = t2 / TRIALS - t1 / TRIALS * t1 / TRIALS;
tlast = bwstats[n].t;
bwstats[n].bits = bufflen * sizeof(byte)*8;
bwstats[n].bps = bwstats[n].bits / (bwstats[n].t * 1024 * 1024);
bwstats[n].repeat = nRepeat;
}
if (self == 0)
System.Console.WriteLine("{0,9:F2} Mbps in {1:F9} sec", bwstats[n].bps, tlast);
}
}
}
}
private static void Test(int root)
{
Intracommunicator world = Communicator.world;
// Broadcast an integer
int intValue = default(int);
if (world.Rank == root)
{
intValue = 17;
System.Console.Write("Broadcasting integer from root " + root + "...");
}
world.Broadcast(ref intValue, root);
MPIDebug.Assert(intValue == 17);
if (world.Rank == root)
{
System.Console.WriteLine(" done.");
}
// Broadcast a string
string strValue = "";
if (world.Rank == root)
{
strValue = "Hello, World!";
System.Console.Write("Broadcasting string from root " + root + "...");
}
world.Broadcast(ref strValue, root);
MPIDebug.Assert(strValue == "Hello, World!");
if (world.Rank == root)
{
System.Console.WriteLine(" done.");
}
// Broadcast an array of integers
int[] intArray = new int[7];
if (world.Rank == root)
{
intArray = new int[] { 1, 1, 2, 3, 5, 8, 13 };
System.Console.Write("Broadcasting integer array from root " + root + "...");
}
world.Broadcast(ref intArray, root);
MPIDebug.Assert(intArray[3] == 3);
if (world.Rank == root)
{
System.Console.WriteLine(" done.");
}
// Broadcast an array of strings
string[] strArray = new string[2];
if (world.Rank == root)
{
strArray = new string[] { "Hello", "World" };
System.Console.Write("Broadcasting string array from root " + root + "...");
}
world.Broadcast(ref strArray, root);
MPIDebug.Assert(strArray[0] == "Hello" && strArray[1] == "World");
if (world.Rank == root)
{
System.Console.WriteLine(" done.");
}
}
private const int ROOT_PROCESS_ID = 0; // Define Root Process ID
public static int Main(string[] args)
{
using (new MPI.Environment(ref args)) // Initialize MPI Environment
{
Intracommunicator comm = Communicator.world; // Define a shortcut to world communicator
string[] contents = null; // Right here, we will store the lines of the text file
char what = '\0'; // The character which the user enters
int totalCharacters = 0; // Total characters per process
int totalOccurrences = 0; // Total occurrences per process
double totalTime = 0; // Total time per process
bool verbose = false; // Verbose output on command line
if (args.Length > 0 && args[0].ToLower() == "--verbose") // If the flag "--verbose" exists, set verbose = true
{
verbose = true;
}
int allCharacters; // All characters by all processes
int allOccurrences; // All occurrencess by all processes
// If we're in the root process
if (comm.Rank == ROOT_PROCESS_ID)
{
// Ask the user to enter the character he want to search for
Console.Write("Enter the character you want to search for: ");
what = (char)Console.Read();
contents = File.ReadAllLines(PATH); // Load contents of file inside the array
}
comm.Broadcast(ref contents, ROOT_PROCESS_ID); // Broadcast contents from root
comm.Broadcast(ref what, ROOT_PROCESS_ID); // Broadcast character from root
// All processes must reach this point before going on
// This is important for time measurement
comm.Barrier();
double start = MPI.Environment.Time; // The start of time measuring
// Let's divide the load for processes, each process is responsible for a specific number of lines
for (int i = comm.Rank; i < contents.Length; i += comm.Size)
{
if (verbose)
{
Console.WriteLine("Process ID = {0}, Line number = {1}", comm.Rank, i);
}
for (int j = 0; j < contents[i].Length; j++)
{
if (contents[i][j] == what) // An occurrence has been found
{
if (verbose)
{
Console.WriteLine("Process {0} has found the character at {1}:{2}", comm.Rank, i + 1, j + 1);
}
totalOccurrences += 1;
}
totalCharacters += 1;
}
}
// All processes must reach here before we stop the timer
comm.Barrier();
double end = MPI.Environment.Time; // The end of time measurement
allOccurrences = comm.Reduce(totalOccurrences, Operation <int> .Add, ROOT_PROCESS_ID); // Accumulate all occurrencess
allCharacters = comm.Reduce(totalCharacters, Operation <int> .Add, ROOT_PROCESS_ID); // Accumulate all characters
totalTime = comm.Reduce(end - start, Operation <double> .Max, ROOT_PROCESS_ID); // The overall time is the longest time of all processes
Console.WriteLine("Total occurrences found by process {0} is {1}", comm.Rank, totalOccurrences);
Console.WriteLine("Total characters processed by {0} is {1}", comm.Rank, totalCharacters);
// Finally, if we're here, print all information
if (comm.Rank == ROOT_PROCESS_ID)
{
Console.WriteLine("Overall occurrences = {0}", allOccurrences);
Console.WriteLine("Overall characters = {0}", allCharacters);
Console.WriteLine("Overall time = {0}s", totalTime);
}
}
return(0);
}
public static Matrix Inverse(Matrix matrix, Intracommunicator comm, ref double timeS, ref double timeP, ref double timeC)
{
if (comm.Rank == 0 && !matrix.isSquare)
{
Exception e = new Exception("Matrix must be square!");
throw e;
}
benchmark bm = new benchmark(), bm2 = new benchmark();
bm.start();
int n = 0;
int[] perm = new int[10]; int toggle = 0; Matrix lum = null;
if (comm.Rank == 0)
{
n = matrix.dim1;
lum = LUPDecompose(matrix, out perm, out toggle);
}
bm.pause();
timeS += bm.getElapsedSeconds();
bm.start();
comm.Broadcast(ref n, 0);
comm.Broadcast(ref lum, 0);
if (comm.Rank != 0)
{
perm = new int[n];
}
comm.Broadcast(ref perm, 0);
comm.Broadcast(ref toggle, 0);
comm.Barrier();
bm.pause();
timeC += bm.getElapsedSeconds();
if (lum == null)
{
return(zeroLike(matrix));
}
bm.start();
Double det = 0;
if (comm.Rank == 0)
{
det = Determinant(lum, perm, toggle);
}
bm.pause();
timeS += bm.getElapsedSeconds();
bm.start();
comm.Broadcast(ref det, 0);
comm.Barrier();
bm.pause();
timeC += bm.getElapsedSeconds();
if (det == 0) // not invertible
{
// still return for the sake of simplicity
// Zero matrix * any matrix = zero matrix
// so it's never a valid answer
return(zeroLike(matrix));
}
bm.pause();
int slaves = comm.Size;
Matrix jobDistro = Utils.splitJob(n, slaves);
int startCol = 0, endCol = 0, size = (int)jobDistro[0, comm.Rank];
for (int p = 0; p < slaves; p++)
{
if (p != comm.Rank)
{
startCol += (int)jobDistro[0, p];
}
else
{
endCol = startCol + (int)jobDistro[0, p] - 1;
break;
}
}
bm.pause();
timeP += bm.getElapsedSeconds();
bm.start();
Matrix result = new Matrix(n, size);
for (int i = startCol; i < startCol + size; ++i)
{
double[] b = new double[n];
for (int j = 0; j < n; ++j)
{
if (i == perm[j])
{
b[j] = 1.0;
}
else
{
b[j] = 0.0;
}
}
double[] x = HelperSolve(lum, b);
for (int j = 0; j < n; ++j)
{
result[j, i - startCol] = x[j];
}
}
bm.pause();
timeP += bm.getElapsedSeconds();
bm.start();
// collect result
result = comm.Reduce(result, ConcatenateColumn, 0);
bm.pause();
timeP += bm.getElapsedSeconds();
return(result);
}