int main(int argc, char **argv) {
int i, j, rank, nprocs, ret;
int ncfile, ndims, nvars, ngatts, unlimited;
int var_ndims, var_natts;;
MPI_Offset *dim_sizes, var_size;
MPI_Offset *start, *count;
char varname[NC_MAX_NAME+1];
int dimids[NC_MAX_VAR_DIMS];
nc_type type;
int *data=NULL;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
if (argc != 2) {
if (rank == 0) printf("Usage: %s filename\n", argv[0]);
MPI_Finalize();
exit(-1);
}
ret = ncmpi_open(MPI_COMM_WORLD, argv[1], NC_NOWRITE, MPI_INFO_NULL,
&ncfile);
if (ret != NC_NOERR) handle_error(ret, __LINE__);
/* reader knows nothing about dataset, but we can interrogate with query
* routines: ncmpi_inq tells us how many of each kind of "thing"
* (dimension, variable, attribute) we will find in the file */
/* no communication needed after ncmpi_open: all processors have a cached
* view of the metadata once ncmpi_open returns */
ret = ncmpi_inq(ncfile, &ndims, &nvars, &ngatts, &unlimited);
if (ret != NC_NOERR) handle_error(ret, __LINE__);
/* we do not really need the name of the dimension or the variable for
* reading in this example. we could, in a different example, take the
* name of a variable on the command line and read just that one */
dim_sizes = (MPI_Offset*) calloc(ndims, sizeof(MPI_Offset));
/* netcdf dimension identifiers are allocated sequentially starting
* at zero; same for variable identifiers */
for(i=0; i<ndims; i++) {
ret = ncmpi_inq_dimlen(ncfile, i, &(dim_sizes[i]) );
if (ret != NC_NOERR) handle_error(ret, __LINE__);
}
for(i=0; i<nvars; i++) {
/* much less coordination in this case compared to rank 0 doing all
* the i/o: everyone already has the necessary information */
ret = ncmpi_inq_var(ncfile, i, varname, &type, &var_ndims, dimids,
&var_natts);
if (ret != NC_NOERR) handle_error(ret, __LINE__);
start = (MPI_Offset*) calloc(var_ndims, sizeof(MPI_Offset));
count = (MPI_Offset*) calloc(var_ndims, sizeof(MPI_Offset));
/* we will simply decompose along one dimension. Generally the
* application has some algorithm for domain decomposition. Note
* that data decomposition can have an impact on i/o performance.
* Often it's best just to do what is natural for the application,
* but something to consider if performance is not what was
* expected/desired */
start[0] = (dim_sizes[dimids[0]]/nprocs)*rank;
count[0] = (dim_sizes[dimids[0]]/nprocs);
var_size = count[0];
for (j=1; j<var_ndims; j++) {
start[j] = 0;
count[j] = dim_sizes[dimids[j]];
var_size *= count[j];
}
switch(type) {
case NC_INT:
data = (int*) calloc(var_size, sizeof(int));
ret = ncmpi_get_vara_all(ncfile, i, start, count, data,
var_size, MPI_INT);
if (ret != NC_NOERR) handle_error(ret, __LINE__);
break;
default:
/* we can do this for all the known netcdf types but this
* example is already getting too long */
fprintf(stderr, "unsupported NetCDF type \n");
}
free(start);
free(count);
if (data != NULL) free(data);
}
ret = ncmpi_close(ncfile);
if (ret != NC_NOERR) handle_error(ret, __LINE__);
MPI_Finalize();
return 0;
}
IOR_offset_t
IOR_Xfer_NCMPI(int access,
void * fd,
IOR_size_t * buffer,
IOR_offset_t length,
IOR_param_t * param)
{
char * bufferPtr = (char *)buffer;
static int firstReadCheck = FALSE,
startDataSet;
int var_id,
dim_id[NUM_DIMS];
MPI_Offset bufSize[NUM_DIMS],
offset[NUM_DIMS];
IOR_offset_t segmentPosition;
int segmentNum,
transferNum;
/* Wei-keng Liao: In IOR.c line 1979 says "block size must be a multiple
of transfer size." Hence, length should always == param->transferSize
below. I leave it here to double check.
*/
if (length != param->transferSize) {
char errMsg[256];
sprintf(errMsg,"length(%lld) != param->transferSize(%lld)\n",
length, param->transferSize);
NCMPI_CHECK(-1, errMsg);
}
/* determine by offset if need to start data set */
if (param->filePerProc == TRUE) {
segmentPosition = (IOR_offset_t)0;
} else {
segmentPosition = (IOR_offset_t)((rank + rankOffset) % param->numTasks)
* param->blockSize;
}
if ((int)(param->offset - segmentPosition) == 0) {
startDataSet = TRUE;
/*
* this toggle is for the read check operation, which passes through
* this function twice; note that this function will open a data set
* only on the first read check and close only on the second
*/
if (access == READCHECK) {
if (firstReadCheck == TRUE) {
firstReadCheck = FALSE;
} else {
firstReadCheck = TRUE;
}
}
}
if (startDataSet == TRUE &&
(access != READCHECK || firstReadCheck == TRUE)) {
if (access == WRITE) {
int numTransfers = param->blockSize / param->transferSize;
/* Wei-keng Liao: change 1D array to 3D array of dimensions:
[segmentCount*numTasksWorld][numTransfers][transferSize]
Requirement: none of these dimensions should be > 4G,
*/
NCMPI_CHECK(ncmpi_def_dim(*(int *)fd, "segments_times_np",
NC_UNLIMITED, &dim_id[0]),
"cannot define data set dimensions");
NCMPI_CHECK(ncmpi_def_dim(*(int *)fd, "number_of_transfers",
numTransfers, &dim_id[1]),
"cannot define data set dimensions");
NCMPI_CHECK(ncmpi_def_dim(*(int *)fd, "transfer_size",
param->transferSize, &dim_id[2]),
"cannot define data set dimensions");
NCMPI_CHECK(ncmpi_def_var(*(int *)fd, "data_var", NC_BYTE,
NUM_DIMS, dim_id, &var_id),
"cannot define data set variables");
NCMPI_CHECK(ncmpi_enddef(*(int *)fd),
"cannot close data set define mode");
} else {
NCMPI_CHECK(ncmpi_inq_varid(*(int *)fd, "data_var", &var_id),
"cannot retrieve data set variable");
}
if (param->collective == FALSE) {
NCMPI_CHECK(ncmpi_begin_indep_data(*(int *)fd),
"cannot enable independent data mode");
}
param->var_id = var_id;
startDataSet = FALSE;
}
var_id = param->var_id;
/* Wei-keng Liao: calculate the segment number */
segmentNum = param->offset / (param->numTasks * param->blockSize);
/* Wei-keng Liao: calculate the transfer number in each block */
transferNum = param->offset % param->blockSize / param->transferSize;
/* Wei-keng Liao: read/write the 3rd dim of the dataset, each is of
amount param->transferSize */
bufSize[0] = 1;
bufSize[1] = 1;
bufSize[2] = param->transferSize;
offset[0] = segmentNum * numTasksWorld + rank;
offset[1] = transferNum;
offset[2] = 0;
/* access the file */
if (access == WRITE) { /* WRITE */
if (param->collective) {
NCMPI_CHECK(ncmpi_put_vara_all(*(int *)fd, var_id, offset, bufSize,
bufferPtr, length, MPI_BYTE),
"cannot write to data set");
} else {
NCMPI_CHECK(ncmpi_put_vara(*(int *)fd, var_id, offset, bufSize,
bufferPtr, length, MPI_BYTE),
"cannot write to data set");
}
} else { /* READ or CHECK */
if (param->collective == TRUE) {
NCMPI_CHECK(ncmpi_get_vara_all(*(int *)fd, var_id, offset, bufSize,
bufferPtr, length, MPI_BYTE),
"cannot read from data set");
} else {
NCMPI_CHECK(ncmpi_get_vara(*(int *)fd, var_id, offset, bufSize,
bufferPtr, length, MPI_BYTE),
"cannot read from data set");
}
}
return(length);
} /* IOR_Xfer_NCMPI() */
void read_pnetcdf(const std::string &filename,
int iVar,
ConfigMap &configMap,
HostArray<double> &localData)
{
int myRank;
MPI_Comm_rank(MPI_COMM_WORLD, &myRank);
int nbMpiProc;
MPI_Comm_size(MPI_COMM_WORLD, &nbMpiProc);
// netcdf file id
int ncFileId;
int err;
// ghostWidth
int ghostWidth = configMap.getInteger("mesh","ghostWidth",3);
// file creation mode
int ncOpenMode = NC_NOWRITE;
int varIds[8];
MPI_Offset starts[3], counts[3]; // read_size, sum_read_size;
MPI_Info mpi_info_used;
// domain local size
int nx,ny,nz;
// sizes to read
//int nx_r, ny_r, nz_r; // logical sizes / per sub-domain
//int nx_g, ny_g, nz_g; // sizes with ghost zones included / per sub-domain
/* read domain sizes */
nx=configMap.getInteger("mesh","nx",32);
ny=configMap.getInteger("mesh","ny",32);
nz=configMap.getInteger("mesh","nz",32);
// nx_g = nx+2*ghostWidth;
// ny_g = ny+2*ghostWidth;
// nz_g = nz+2*ghostWidth;
// get input filename from configMap
//std::string filename = configMap.getString("input", "filename", "");
/*
* Open NetCDF file
*/
err = ncmpi_open(MPI_COMM_WORLD, filename.c_str(),
ncOpenMode,
MPI_INFO_NULL, &ncFileId);
if (err != NC_NOERR) {
printf("Error: ncmpi_open() file %s (%s)\n",filename.c_str(),ncmpi_strerror(err));
MPI_Abort(MPI_COMM_WORLD, -1);
exit(1);
}
/*
* Query NetCDF mode
*/
int NC_mode;
err = ncmpi_inq_version(ncFileId, &NC_mode);
if (myRank==0) {
if (NC_mode == NC_64BIT_DATA)
std::cout << "Pnetcdf Input mode : NC_64BIT_DATA (CDF-5)\n";
else if (NC_mode == NC_64BIT_OFFSET)
std::cout << "Pnetcdf Input mode : NC_64BIT_OFFSET (CDF-2)\n";
else
std::cout << "Pnetcdf Input mode : unknown\n";
}
/*
* Query information about variables
*/
{
int ndims, nvars, ngatts, unlimited;
err = ncmpi_inq(ncFileId, &ndims, &nvars, &ngatts, &unlimited);
PNETCDF_HANDLE_ERROR;
err = ncmpi_inq_varid(ncFileId, "rho", &varIds[ID]);
PNETCDF_HANDLE_ERROR;
err = ncmpi_inq_varid(ncFileId, "E", &varIds[IP]);
PNETCDF_HANDLE_ERROR;
err = ncmpi_inq_varid(ncFileId, "rho_vx", &varIds[IU]);
PNETCDF_HANDLE_ERROR;
err = ncmpi_inq_varid(ncFileId, "rho_vy", &varIds[IV]);
PNETCDF_HANDLE_ERROR;
err = ncmpi_inq_varid(ncFileId, "rho_vz", &varIds[IW]);
PNETCDF_HANDLE_ERROR;
err = ncmpi_inq_varid(ncFileId, "Bx", &varIds[IA]);
PNETCDF_HANDLE_ERROR;
err = ncmpi_inq_varid(ncFileId, "By", &varIds[IB]);
PNETCDF_HANDLE_ERROR;
err = ncmpi_inq_varid(ncFileId, "Bz", &varIds[IC]);
PNETCDF_HANDLE_ERROR;
} // end query information
/*
* Define expected data types (no conversion done here)
*/
//nc_type ncDataType;
MPI_Datatype mpiDataType;
//ncDataType = NC_DOUBLE;
mpiDataType = MPI_DOUBLE;
/*
* Get all the MPI_IO hints used (just in case, we want to print it after
* reading data...
*/
err = ncmpi_get_file_info(ncFileId, &mpi_info_used);
PNETCDF_HANDLE_ERROR;
/*
* Read heavy data (take care of row-major / column major format !)
*/
// use overlapping domains
// counts[IZ] = nx_rg;
// counts[IY] = ny_rg;
// counts[IX] = nz_rg;
// starts[IZ] = 0;
// starts[IY] = 0;
// starts[IX] = myRank*nz_r;
counts[IZ] = nx;
counts[IY] = ny;
counts[IX] = nz;
starts[IZ] = ghostWidth;
starts[IY] = ghostWidth;
starts[IX] = ghostWidth+myRank*nz;
int nItems = counts[IX]*counts[IY]*counts[IZ];
/*
* Actual reading
*/
{
double* data;
//data = &(localData(0,0,0,0));
data = localData.data();
err = ncmpi_get_vara_all(ncFileId, varIds[iVar],
starts, counts, data, nItems, mpiDataType);
PNETCDF_HANDLE_ERROR;
} // end for loop reading heavy data
/*
* close the file
*/
err = ncmpi_close(ncFileId);
PNETCDF_HANDLE_ERROR;
} // read_pnetcdf
/**
* Read a single-precision parallel-nedcdf file.
*
* We assume here that localData is a scalar.
*
* Pnetcdf uses row-major format (same as FFTW).
*
* \param[in] filename : PnetCDF filename
* \param[in] starts : offset to where to start reading data
* \param[in] counts : number of elements read (3D sub-domain inside global)
* \param[in] gsizes : global sizes
* \param[out] localData : actual data buffer (size : nx*ny*nz*sizeof(float))
*
* localData must have been allocated prior to calling this routine.
*/
void read_pnetcdf(const std::string &filename,
MPI_Offset starts[3],
MPI_Offset counts[3],
int gsizes[3],
float *localData)
{
int myRank;
MPI_Comm_rank(MPI_COMM_WORLD, &myRank);
// netcdf file id
int ncFileId;
int err;
// file opening mode
int ncOpenMode = NC_NOWRITE;
int nbVar=1;
int varIds[nbVar];
MPI_Info mpi_info_used;
/*
* Open NetCDF file
*/
err = ncmpi_open(MPI_COMM_WORLD, filename.c_str(),
ncOpenMode,
MPI_INFO_NULL, &ncFileId);
if (err != NC_NOERR) {
printf("Error: ncmpi_open() file %s (%s)\n",filename.c_str(),ncmpi_strerror(err));
MPI_Abort(MPI_COMM_WORLD, -1);
exit(1);
}
/*
* Query NetCDF mode
*/
int NC_mode;
err = ncmpi_inq_version(ncFileId, &NC_mode);
if (myRank==0) {
if (NC_mode == NC_64BIT_DATA)
std::cout << "Pnetcdf Input mode : NC_64BIT_DATA (CDF-5)\n";
else if (NC_mode == NC_64BIT_OFFSET)
std::cout << "Pnetcdf Input mode : NC_64BIT_OFFSET (CDF-2)\n";
else
std::cout << "Pnetcdf Input mode : unknown\n";
}
/*
* Query information about variable named "data"
*/
{
int ndims, nvars, ngatts, unlimited;
err = ncmpi_inq(ncFileId, &ndims, &nvars, &ngatts, &unlimited);
PNETCDF_HANDLE_ERROR;
err = ncmpi_inq_varid(ncFileId, "data", &varIds[0]);
PNETCDF_HANDLE_ERROR;
}
/*
* Define expected data types (no conversion done here)
*/
MPI_Datatype mpiDataType = MPI_FLOAT;
/*
* Get all the MPI_IO hints used (just in case, we want to print it after
* reading data...
*/
err = ncmpi_get_file_info(ncFileId, &mpi_info_used);
PNETCDF_HANDLE_ERROR;
/*
* Read heavy data (take care of row-major / column major format !)
*/
int nItems = counts[IX]*counts[IY]*counts[IZ];
{
err = ncmpi_get_vara_all(ncFileId,
varIds[0],
starts,
counts,
localData,
nItems,
mpiDataType);
PNETCDF_HANDLE_ERROR;
} // end reading heavy data
/*
* close the file
*/
err = ncmpi_close(ncFileId);
PNETCDF_HANDLE_ERROR;
} // read_pnetcdf
static
int get_var_and_verify(int ncid,
int varid,
MPI_Offset *start,
MPI_Offset *count,
int **buf,
MPI_Datatype buftype,
MPI_Datatype ghost_buftype,
MPI_Datatype filetype)
{
int i, j, rank, err, *ncbuf, nerrs=0;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
ncbuf = (int *) malloc((count[0]+4)*(count[1]+4)*sizeof(int));
/* clear the contents of the read buffer */
for (j=0; j<count[0]; j++) for (i=0; i<count[1]; i++) buf[j][i] = -1;
/* read back using regular vara API */
err = ncmpi_get_vara_int_all(ncid, varid, start, count, buf[0]); ERR
/* check if the contents of buf are expected */
CHECK_VALUE_PERMUTED
/* clear the contents of the read buffer */
for (j=0; j<count[0]; j++) for (i=0; i<count[1]; i++) buf[j][i] = -1;
/* read back using flexible vara API */
err = ncmpi_get_vara_all(ncid, varid, start, count, buf[1], 1, buftype); ERR
/* check if the contents of buf are expected */
CHECK_VALUE
/* clear the contents of the read buffer */
for (j=0; j<count[0]; j++) for (i=0; i<count[1]; i++) buf[j][i] = -1;
/* read back using vard API and permuted buftype */
err = ncmpi_get_vard_all(ncid, varid, filetype, buf[1], 1, buftype); ERR
/* check if the contents of buf are expected */
CHECK_VALUE
/* clear the contents of the read buffer */
for (j=0; j<count[0]; j++) for (i=0; i<count[1]; i++) buf[j][i] = -1;
/* read back using vard API and no buftype */
err = ncmpi_get_vard_all(ncid, varid, filetype, buf[0], 0, MPI_DATATYPE_NULL); ERR
/* check if the contents of buf are expected */
CHECK_VALUE_PERMUTED
/* clear the contents of the read buffer */
for (i=0; i<(count[0]+4)*(count[1]+4); i++) ncbuf[i] = -1;
/* read back using ghost buftype */
err = ncmpi_get_vard_all(ncid, varid, filetype, ncbuf, 1, ghost_buftype); ERR
for (j=0; j<count[0]; j++) {
for (i=0; i<count[1]; i++)
if (buf[j][i] != ncbuf[(j+2)*(count[1]+4)+(i+2)]) {
printf("Error at line %d: expecting ncbuf[%d][%d]=%d but got %d\n",
__LINE__,j,i,buf[j][i],ncbuf[(j+2)*(count[1]+4)+(i+2)]);
nerrs++;
}
}
free(ncbuf);
return nerrs;
}
