void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
if((nrhs < 1) || (nrhs > 2) || (nlhs != 1)) mexErrMsgTxt("Call is min = mpi_min(array) or min = mpi_min(array, communicator)");
MPI_Comm commune = MPI_COMM_WORLD;
/*if(nrhs == 2) commune = (MPI_Comm)*mxGetPr(prhs[1]);*/
/* Grab some basic meta info about the input arrays */
/* God help you if they aren't all the same size. */
long numel = mxGetNumberOfElements(prhs[0]);
mwSize ndims = mxGetNumberOfDimensions(prhs[0]);
const mwSize *arrDims = mxGetDimensions(prhs[0]);
mxClassID arraytype = mxGetClassID(prhs[0]);
/* Create our output array containing the reduced values, on every host */
plhs[0] = mxCreateNumericArray(ndims, arrDims, arraytype, mxREAL);
MPI_Datatype mtype = typeid_ml2mpi(arraytype);
void *src = mxGetData(prhs[0]);
void *dst = mxGetData(plhs[0]);
MPI_Allreduce(src, dst, numel, mtype, MPIOPERATION, commune);
}
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
if((nrhs < 1) || (nrhs > 2) || (nlhs != 1)) mexErrMsgTxt("Call is TF = mpi_all(array)");
MPI_Comm commune = MPI_COMM_WORLD;
/*if(nrhs == 2) commune = (MPI_Comm)*mxGetPr(prhs[1]);*/
/* Grab some basic meta info about the input arrays */
/* God help you if they aren't all the same size. */
long numel = mxGetNumberOfElements(prhs[0]);
mwSize ndims = mxGetNumberOfDimensions(prhs[0]);
const mwSize *arrDims = mxGetDimensions(prhs[0]);
mxClassID arraytype = mxGetClassID(prhs[0]);
/* Create our output array containing the reduced values, on every host */
plhs[0] = mxCreateNumericArray(ndims, arrDims, arraytype, mxREAL);
MPI_Datatype mtype = typeid_ml2mpi(arraytype);
void *src; void *dst;
src = mxGetData(prhs[0]);
dst = mxGetData(plhs[0]);
/* Safe to do on the Matlab data */
if(mpidatatypeIsInteger(mtype)) {
MPI_Allreduce(src, dst, numel, mtype, MPI_BOR, commune);
} else {
/* Whelp, this just got slower */
uint8_t *logicSrc = (uint8_t *)malloc(numel*2);
uint8_t *logicDst = logicSrc + numel;
long k;
if(arraytype == mxDOUBLE_CLASS) {
double *d = (double *)src;
for(k = 0; k < numel; k++)
logicSrc[k] = (d[k] != 0);
MPI_Allreduce(logicSrc, logicDst, numel, MPI_BYTE, MPI_BOR, commune);
d = (double *)dst;
for(k = 0; k < numel; k++)
d[k] = (double)logicDst[k];
}
if(arraytype == mxSINGLE_CLASS) {
float *d = (float *)src;
for(k = 0; k < numel; k++)
logicSrc[k] = (d[k] != 0);
MPI_Allreduce(logicSrc, logicDst, numel, MPI_BYTE, MPI_BOR, commune);
d = (float *)dst;
for(k = 0; k < numel; k++)
d[k] = (float)logicDst[k];
}
free(logicSrc);
}
}
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
if((nrhs < 1) || (nrhs > 2) || (nlhs != 1)) mexErrMsgTxt("Call is min = mpi_min(array) or min = mpi_min(array, communicator)");
MPI_Comm commune = MPI_COMM_WORLD;
/*if(nrhs == 2) commune = (MPI_Comm)*mxGetPr(prhs[1]);*/
/* Grab some basic meta info about the input arrays */
/* God help you if they aren't all the same size. */
// FIXME: Accept a 'paranoid' flag to check if they are & avoid crashing the MPI runtime?
long numel = mxGetNumberOfElements(prhs[0]);
mwSize ndims = mxGetNumberOfDimensions(prhs[0]);
const mwSize *arrDims = mxGetDimensions(prhs[0]);
mxClassID arraytype = mxGetClassID(prhs[0]);
/* Create our output array containing the reduced values, on every host */
plhs[0] = mxCreateNumericArray(ndims, arrDims, arraytype, mxREAL);
MPI_Datatype mtype = typeid_ml2mpi(arraytype);
void *src = mxGetData(prhs[0]);
void *dst = mxGetData(plhs[0]);
if(mtype == MPI_BYTE) { // reduce hates mpi_byte for me...
int *srcI = (int *)malloc(2*numel*sizeof(int));
int *dstI = srcI + numel;
int j;
for(j = 0; j < numel; j++) {
srcI[j] = (int)(( (uint8_t *)src)[j]);
}
MPI_Allreduce((void *)srcI, (void *)dstI, numel, MPI_INT, MPIOPERATION, commune);
for(j = 0; j < numel; j++) {
(( (uint8_t *)dst)[j]) = (uint8_t)dstI[j];
}
free(srcI);
} else {
MPI_Allreduce(src, dst, numel, mtype, MPIOPERATION, commune);
}
}
