// Perform one iteration of work
// The first step is to send the local state to the neighbors
void begin_iteration(void) {
itercnt++;
//ckout << "iter=" << itercnt << ":(" << thisIndex.x << "," << thisIndex.y << ") is on " << CkMyPe() << endl;
// Copy left column and right column into temporary arrays
double *left_edge = new double[block_height];
double *right_edge = new double[block_height];
for(int i=0;i<block_height;++i){
left_edge[i] = temperature[i+1][1];
right_edge[i] = temperature[i+1][block_width];
}
// Send my left edge
thisProxy(wrap_x(thisIndex.x-1), thisIndex.y).ghostsFromRight(block_height, left_edge);
// Send my right edge
thisProxy(wrap_x(thisIndex.x+1), thisIndex.y).ghostsFromLeft(block_height, right_edge);
// Send my top edge
thisProxy(thisIndex.x, wrap_y(thisIndex.y-1)).ghostsFromBottom(block_width, &temperature[1][1]);
// Send my bottom edge
thisProxy(thisIndex.x, wrap_y(thisIndex.y+1)).ghostsFromTop(block_width, &temperature[block_height][1]);
hasSent=true;
check_and_compute();
delete [] right_edge;
delete [] left_edge;
}
// Send ghost faces to the six neighbors
void begin_iteration(void) {
if (thisIndex.x == 0 && thisIndex.y == 0 && thisIndex.z == 0) {
// CkPrintf("Start of iteration %d\n", iterations);
if(iterations % PRINT_FREQ == 0) {
average = timing;
timing = CmiWallTimer();
average = (timing - average)/(double)PRINT_FREQ;
CkPrintf("time=%.2f it=%d avg=%.4f\n",timing,iterations,average);
}
}
iterations++;
// Copy different faces into messages
ghostMsg *leftMsg = new (blockDimY*blockDimZ) ghostMsg(RIGHT, blockDimY, blockDimZ);
ghostMsg *rightMsg = new (blockDimY*blockDimZ) ghostMsg(LEFT, blockDimY, blockDimZ);
ghostMsg *topMsg = new (blockDimX*blockDimZ) ghostMsg(BOTTOM, blockDimX, blockDimZ);
ghostMsg *bottomMsg = new (blockDimX*blockDimZ) ghostMsg(TOP, blockDimX, blockDimZ);
ghostMsg *frontMsg = new (blockDimX*blockDimY) ghostMsg(BACK, blockDimX, blockDimY);
ghostMsg *backMsg = new (blockDimX*blockDimY) ghostMsg(FRONT, blockDimX, blockDimY);
CkSetRefNum(leftMsg, iterations);
CkSetRefNum(rightMsg, iterations);
CkSetRefNum(topMsg, iterations);
CkSetRefNum(bottomMsg, iterations);
CkSetRefNum(frontMsg, iterations);
CkSetRefNum(backMsg, iterations);
for(int j=0; j<blockDimY; ++j)
for(int k=0; k<blockDimZ; ++k) {
leftMsg->gh[k*blockDimY+j] = temperature[index(1, j+1, k+1)];
rightMsg->gh[k*blockDimY+j] = temperature[index(blockDimX, j+1, k+1)];
}
for(int i=0; i<blockDimX; ++i)
for(int k=0; k<blockDimZ; ++k) {
topMsg->gh[k*blockDimX+i] = temperature[index(i+1, 1, k+1)];
bottomMsg->gh[k*blockDimX+i] = temperature[index(i+1, blockDimY, k+1)];
}
for(int i=0; i<blockDimX; ++i)
for(int j=0; j<blockDimY; ++j) {
frontMsg->gh[j*blockDimX+i] = temperature[index(i+1, j+1, 1)];
backMsg->gh[j*blockDimX+i] = temperature[index(i+1, j+1, blockDimZ)];
}
// Send my left face
thisProxy(wrap_x(thisIndex.x-1), thisIndex.y, thisIndex.z).receiveGhosts(leftMsg);
// Send my right face
thisProxy(wrap_x(thisIndex.x+1), thisIndex.y, thisIndex.z).receiveGhosts(rightMsg);
// Send my top face
thisProxy(thisIndex.x, wrap_y(thisIndex.y-1), thisIndex.z).receiveGhosts(topMsg);
// Send my bottom face
thisProxy(thisIndex.x, wrap_y(thisIndex.y+1), thisIndex.z).receiveGhosts(bottomMsg);
// Send my front face
thisProxy(thisIndex.x, thisIndex.y, wrap_z(thisIndex.z-1)).receiveGhosts(frontMsg);
// Send my back face
thisProxy(thisIndex.x, thisIndex.y, wrap_z(thisIndex.z+1)).receiveGhosts(backMsg);
}
int main(int argc, char **argv) {
int myRank, numPes;
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &numPes);
MPI_Comm_rank(MPI_COMM_WORLD, &myRank);
MPI_Request req[4];
MPI_Status status[4];
int blockDimX, blockDimY, arrayDimX, arrayDimY;
int noBarrier = 0;
if (argc != 4 && argc != 6) {
printf("%s [array_size] [block_size] +[no]barrier\n", argv[0]);
printf("%s [array_size_X] [array_size_Y] [block_size_X] [block_size_Y] +[no]barrier\n", argv[0]);
MPI_Abort(MPI_COMM_WORLD, -1);
}
if(argc == 4) {
arrayDimY = arrayDimX = atoi(argv[1]);
blockDimY = blockDimX = atoi(argv[2]);
if(strcasecmp(argv[3], "+nobarrier") == 0)
noBarrier = 1;
else
noBarrier = 0;
if(noBarrier && myRank==0) printf("\nSTENCIL COMPUTATION WITH NO BARRIERS\n");
}
else {
arrayDimX = atoi(argv[1]);
arrayDimY = atoi(argv[2]);
blockDimX = atoi(argv[3]);
blockDimY = atoi(argv[4]);
if(strcasecmp(argv[5], "+nobarrier") == 0)
noBarrier = 1;
else
noBarrier = 0;
if(noBarrier && myRank==0) printf("\nSTENCIL COMPUTATION WITH NO BARRIERS\n");
}
if (arrayDimX < blockDimX || arrayDimX % blockDimX != 0) {
printf("array_size_X mod block_size_X != 0!\n");
MPI_Abort(MPI_COMM_WORLD, -1);
}
if (arrayDimY < blockDimY || arrayDimY % blockDimY != 0) {
printf("array_size_Y mod block_size_Y != 0!\n");
MPI_Abort(MPI_COMM_WORLD, -1);
}
int num_blocks_x = arrayDimX / blockDimX;
int num_blocks_y = arrayDimY / blockDimY;
int iterations = 0, i, j;
double error = 1.0, max_error = 0.0;
if(myRank == 0) {
printf("Running Jacobi on %d processors with (%d, %d) elements\n", numPes, num_blocks_x, num_blocks_y);
printf("Array Dimensions: %d %d\n", arrayDimX, arrayDimY);
printf("Block Dimensions: %d %d\n", blockDimX, blockDimY);
}
double *dataX = new double[blockDimX];
double *dataY = new double[blockDimY];
int myRow = myRank / num_blocks_y;
int myCol = myRank % num_blocks_y;
AMPI_Set_startevent(MPI_COMM_WORLD);
for(; iterations < 5; iterations++) {
#if DO_COMM
/* Receive my right, left, bottom and top edge */
MPI_Irecv(dataX, blockDimX, MPI_DOUBLE, calc_pe(myRow, wrap_y(myCol+1)), RIGHT, MPI_COMM_WORLD, &req[RIGHT-1]);
MPI_Irecv(dataX, blockDimX, MPI_DOUBLE, calc_pe(myRow, wrap_y(myCol-1)), LEFT, MPI_COMM_WORLD, &req[LEFT-1]);
MPI_Irecv(dataY, blockDimY, MPI_DOUBLE, calc_pe(wrap_x(myRow+1), myCol), BOTTOM, MPI_COMM_WORLD, &req[BOTTOM-1]);
MPI_Irecv(dataY, blockDimY, MPI_DOUBLE, calc_pe(wrap_x(myRow-1), myCol), TOP, MPI_COMM_WORLD, &req[TOP-1]);
/* Send my left, right, top and bottom edge */
MPI_Send(dataX, blockDimX, MPI_DOUBLE, calc_pe(myRow, wrap_y(myCol-1)), RIGHT, MPI_COMM_WORLD);
MPI_Send(dataX, blockDimX, MPI_DOUBLE, calc_pe(myRow, wrap_y(myCol+1)), LEFT, MPI_COMM_WORLD);
MPI_Send(dataY, blockDimY, MPI_DOUBLE, calc_pe(wrap_x(myRow-1), myCol), BOTTOM, MPI_COMM_WORLD);
MPI_Send(dataY, blockDimY, MPI_DOUBLE, calc_pe(wrap_x(myRow+1), myCol), TOP, MPI_COMM_WORLD);
MPI_Waitall(4, req, status);
#endif
#if !DO_COMM
printf("%d %d %zd\n", myRank, calc_pe(myRow, wrap_y(myCol-1)), sizeof(double)*blockDimX);
printf("%d %d %zd\n", myRank, calc_pe(myRow, wrap_y(myCol+1)), sizeof(double)*blockDimX);
printf("%d %d %zd\n", myRank, calc_pe(wrap_x(myRow-1), myCol), sizeof(double)*blockDimY);
printf("%d %d %zd\n", myRank, calc_pe(wrap_x(myRow+1), myCol), sizeof(double)*blockDimY);
#endif
} /* end of while loop */
if(myRank == 0) {
printf("Completed %d iterations\n", iterations);
}
MPI_Finalize();
return 0;
} /* end function main */
Main(CkArgMsg* m) {
if ( (m->argc != 3) && (m->argc != 7) ) {
CkPrintf("%s [array_size] [block_size]\n", m->argv[0]);
CkPrintf("OR %s [array_size_X] [array_size_Y] [array_size_Z] [block_size_X] [block_size_Y] [block_size_Z]\n", m->argv[0]);
CkAbort("Abort");
}
// set iteration counter to zero
iterations = 0;
// store the main proxy
mainProxy = thisProxy;
if(m->argc == 3) {
arrayDimX = arrayDimY = arrayDimZ = atoi(m->argv[1]);
blockDimX = blockDimY = blockDimZ = atoi(m->argv[2]);
}
else if (m->argc == 7) {
arrayDimX = atoi(m->argv[1]);
arrayDimY = atoi(m->argv[2]);
arrayDimZ = atoi(m->argv[3]);
blockDimX = atoi(m->argv[4]);
blockDimY = atoi(m->argv[5]);
blockDimZ = atoi(m->argv[6]);
}
if (arrayDimX < blockDimX || arrayDimX % blockDimX != 0)
CkAbort("array_size_X % block_size_X != 0!");
if (arrayDimY < blockDimY || arrayDimY % blockDimY != 0)
CkAbort("array_size_Y % block_size_Y != 0!");
if (arrayDimZ < blockDimZ || arrayDimZ % blockDimZ != 0)
CkAbort("array_size_Z % block_size_Z != 0!");
num_chare_x = arrayDimX / blockDimX;
num_chare_y = arrayDimY / blockDimY;
num_chare_z = arrayDimZ / blockDimZ;
// print info
CkPrintf("\nSTENCIL COMPUTATION WITH NO BARRIERS\n");
CkPrintf("Running Jacobi on %d processors with (%d, %d, %d) chares\n", CkNumPes(), num_chare_x, num_chare_y, num_chare_z);
CkPrintf("Array Dimensions: %d %d %d\n", arrayDimX, arrayDimY, arrayDimZ);
CkPrintf("Block Dimensions: %d %d %d\n", blockDimX, blockDimY, blockDimZ);
// Create new array of worker chares
#if USE_TOPOMAP
CProxy_JacobiMap map = CProxy_JacobiMap::ckNew(num_chare_x, num_chare_y, num_chare_z);
CkPrintf("Topology Mapping is being done ... \n");
CkArrayOptions opts(num_chare_x, num_chare_y, num_chare_z);
opts.setMap(map);
array = CProxy_Jacobi::ckNew(opts);
#else
array = CProxy_Jacobi::ckNew(num_chare_x, num_chare_y, num_chare_z);
#endif
TopoManager tmgr;
CkArray *jarr = array.ckLocalBranch();
int jmap[num_chare_x][num_chare_y][num_chare_z];
int hops=0, p;
for(int i=0; i<num_chare_x; i++)
for(int j=0; j<num_chare_y; j++)
for(int k=0; k<num_chare_z; k++) {
jmap[i][j][k] = jarr->procNum(CkArrayIndex3D(i, j, k));
}
for(int i=0; i<num_chare_x; i++)
for(int j=0; j<num_chare_y; j++)
for(int k=0; k<num_chare_z; k++) {
p = jmap[i][j][k];
hops += tmgr.getHopsBetweenRanks(p, jmap[wrap_x(i+1)][j][k]);
hops += tmgr.getHopsBetweenRanks(p, jmap[wrap_x(i-1)][j][k]);
hops += tmgr.getHopsBetweenRanks(p, jmap[i][wrap_y(j+1)][k]);
hops += tmgr.getHopsBetweenRanks(p, jmap[i][wrap_y(j-1)][k]);
hops += tmgr.getHopsBetweenRanks(p, jmap[i][j][wrap_z(k+1)]);
hops += tmgr.getHopsBetweenRanks(p, jmap[i][j][wrap_z(k-1)]);
}
CkPrintf("Total Hops: %d\n", hops);
#ifdef JACOBI_OPENMP
CProxy_OmpInitializer ompInit =
CProxy_OmpInitializer::ckNew(4);
#else
//Start the computation
start();
#endif
}
int main(int argc, char **argv) {
int myRank, numPes;
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &numPes);
MPI_Comm_rank(MPI_COMM_WORLD, &myRank);
MPI_Request sreq[2], rreq[2];
int blockDimX, arrayDimX, arrayDimY;
if (argc != 2 && argc != 3) {
printf("%s [array_size] \n", argv[0]);
printf("%s [array_size_X] [array_size_Y] \n", argv[0]);
MPI_Abort(MPI_COMM_WORLD, -1);
}
if(argc == 2) {
arrayDimY = arrayDimX = atoi(argv[1]);
}
else {
arrayDimX = atoi(argv[1]);
arrayDimY = atoi(argv[2]);
}
if (arrayDimX % numPes != 0) {
printf("array_size_X % numPes != 0!\n");
MPI_Abort(MPI_COMM_WORLD, -1);
}
blockDimX = arrayDimX / numPes;
int iterations = 0, i, j;
double error = 1.0, max_error = 0.0;
if(myRank == 0) {
printf("Running Jacobi on %d processors\n", numPes);
printf("Array Dimensions: %d %d\n", arrayDimX, arrayDimY);
printf("Block Dimensions: %d\n", blockDimX);
}
double **temperature;
double **new_temperature;
/* allocate two dimensional arrays */
temperature = new double*[blockDimX+2];
new_temperature = new double*[blockDimX+2];
for (i=0; i<blockDimX+2; i++) {
temperature[i] = new double[arrayDimY];
new_temperature[i] = new double[arrayDimY];
}
for(i=0; i<blockDimX+2; i++) {
for(j=0; j<arrayDimY; j++) {
temperature[i][j] = 0.5;
new_temperature[i][j] = 0.5;
}
}
// boundary conditions
if(myRank < numPes/2) {
for(i=1; i<=blockDimX; i++)
temperature[i][0] = 1.0;
}
if(myRank == numPes-1) {
for(j=arrayDimY/2; j<arrayDimY; j++)
temperature[blockDimX][j] = 0.0;
}
MPI_Barrier(MPI_COMM_WORLD);
MPI_Pcontrol(1);
startTime = MPI_Wtime();
while(/*error > 0.001 &&*/ iterations < MAX_ITER) {
iterations++;
/* Receive my bottom and top edge */
MPI_Irecv(&temperature[blockDimX+1][0], arrayDimY, MPI_DOUBLE, wrap_x(myRank+1), BOTTOM, MPI_COMM_WORLD, &rreq[BOTTOM-1]);
MPI_Irecv(&temperature[0][0], arrayDimY, MPI_DOUBLE, wrap_x(myRank-1), TOP, MPI_COMM_WORLD, &rreq[TOP-1]);
/* Send my top and bottom edge */
MPI_Isend(&temperature[1][0], arrayDimY, MPI_DOUBLE, wrap_x(myRank-1), BOTTOM, MPI_COMM_WORLD, &sreq[BOTTOM-1]);
MPI_Isend(&temperature[blockDimX][0], arrayDimY, MPI_DOUBLE, wrap_x(myRank+1), TOP, MPI_COMM_WORLD, &sreq[TOP-1]);
MPI_Waitall(2, rreq, MPI_STATUSES_IGNORE);
MPI_Waitall(2, sreq, MPI_STATUSES_IGNORE);
for(i=1; i<blockDimX+1; i++) {
for(j=0; j<arrayDimY; j++) {
/* update my value based on the surrounding values */
new_temperature[i][j] = (temperature[i-1][j]+temperature[i+1][j]+temperature[i][wrap_y(j-1)]+temperature[i][wrap_y(j+1)]+temperature[i][j]) * 0.2;
}
}
max_error = error = 0.0;
for(i=1; i<blockDimX+1; i++) {
for(j=0; j<arrayDimY; j++) {
error = fabs(new_temperature[i][j] - temperature[i][j]);
if(error > max_error)
max_error = error;
}
}
double **tmp;
tmp = temperature;
temperature = new_temperature;
new_temperature = tmp;
// boundary conditions
if(myRank < numPes/2) {
for(i=1; i<=blockDimX; i++)
temperature[i][0] = 1.0;
}
if(myRank == numPes-1) {
for(j=arrayDimY/2; j<arrayDimY; j++)
temperature[blockDimX][j] = 0.0;
}
//if(myRank == 0) printf("Iteration %d %f %f %f\n", iterations, max_error, temperature[1][0], temperature[1][1]);
MPI_Allreduce(&max_error, &error, 1, MPI_DOUBLE, MPI_MAX, MPI_COMM_WORLD);
} /* end of while loop */
MPI_Barrier(MPI_COMM_WORLD);
MPI_Pcontrol(0);
if(myRank == 0) {
endTime = MPI_Wtime();
printf("Completed %d iterations\n", iterations);
printf("Time elapsed: %f\n", endTime - startTime);
}
MPI_Finalize();
return 0;
} /* end function main */
