Skip to content

GitLab

Commit 5252630c authored by incardon's avatar incardon
Browse files

Refactoring HDF5 load

parent 0382f18f
Hide whitespace changes
Inline Side-by-side
src/Grid/grid_dist_id.hpp
View file @ 5252630c
......@@ -1663,42 +1663,11 @@ public:
void map()
{
getGlobalGridsInfo(gdb_ext_global);
/*
std::cout << "Global size: " << gdb_ext_global.size() << std::endl;
for (size_t i = 0; i < gdb_ext_global.size(); i++)
{
std::cout << "(" << gdb_ext_global.get(i).Dbox.getLow(0) << "; " << gdb_ext_global.get(i).Dbox.getLow(1) << "); (" << gdb_ext_global.get(i).Dbox.getHigh(0) << "; " << gdb_ext_global.get(i).Dbox.getHigh(1) << ")" << std::endl;
std::cout << "I = " << i << ", Origin is (" << gdb_ext_global.get(i).origin.get(0) << "; " << gdb_ext_global.get(i).origin.get(1) << ")" << std::endl;
}
if (v_cl.getProcessUnitID() == 0)
{
for (size_t i = 0; i < gdb_ext.size(); i++)
{
Box<dim,long int> box = gdb_ext.get(i).Dbox;
box += gdb_ext.get(i).origin;
std::cout << "(" << box.getLow(0) << "; " << box.getLow(1) << "); (" << box.getHigh(0) << "; " << box.getHigh(1) << ")" << std::endl;
}
}
if (v_cl.getProcessUnitID() == 0)
{
for (size_t i = 0; i < loc_grid_old.size(); i++)
{
Point<dim,St> p1;
Point<dim,St> p2;
for (size_t n = 0; n < dim; n++)
{
p1.get(n) = loc_grid_old.get(i).getGrid().getBox().getLow(n);
p2.get(n) = loc_grid_old.get(i).getGrid().getBox().getHigh(n);
}
std::cout << "Loc_grid_old: (" << p1.get(0) << "; " << p1.get(1) << "); (" << p2.get(0) << "; " << p2.get(1) << "); " << "Gdb_ext_old: (" << gdb_ext_old.get(i).Dbox.getLow(0) << "; " << gdb_ext_old.get(i).Dbox.getLow(1) << "); (" << gdb_ext_old.get(i).Dbox.getHigh(0) << "; " << gdb_ext_old.get(i).Dbox.getHigh(1) << ")" << std::endl;
}
}
*/
this->template map_(dec,cd_sm,loc_grid,loc_grid_old,gdb_ext,gdb_ext_old,gdb_ext_global);
loc_grid_old.clear();
gdb_ext_old.clear();
}
inline void save(const std::string & filename) const
......@@ -1708,368 +1677,14 @@ public:
h5s.save(filename,loc_grid,gdb_ext);
}
void load_block(long int bid,
hssize_t mpi_size_old,
int * metadata_out,
openfpm::vector<size_t> & metadata_accum,
hid_t plist_id,
hid_t dataset_2)
{
/* if (mpi_size >= mpi_size_old)
{
if (mpi_rank >= mpi_size_old)
block[0] = 0;
else
block[0] = {(size_t)metadata_out[mpi_rank]};
}
else
{
int x = mpi_size_old/mpi_size;
int shift = mpi_rank*x;
for (int i = 0; i < x; i++)
{
//block0.get(mpi_rank).add(metadata_out[shift]);
block[0] += metadata_out[shift];
shift++;
}
int y = mpi_size_old%mpi_size;
if (mpi_rank < y)
{
block_add[0] += metadata_out[mpi_size*x+mpi_rank];
//block_add0.get(mpi_rank).add(metadata_out[mpi_size*x+mpi_rank]);
}
}*/
// std::cout << "BID: " << bid << std::endl;
hsize_t offset[1];
hsize_t block[1];
if (bid < mpi_size_old && bid != -1)
{
offset[0] = metadata_accum.get(bid);
block[0] = metadata_out[bid];
}
else
{
offset[0] = 0;
block[0] = 0;
}
// std::cout << "Offset: " << offset[0] << "; Block: " << block[0]<< std::endl;
// hsize_t offset_add[1] = {0};
/* if (mpi_size >= mpi_size_old)
{
if (mpi_rank >= mpi_size_old)
offset[0] = 0;
else
{
for (int i = 0; i < mpi_rank; i++)
offset[0] += metadata_out[i];
}
}
else
{
int x = mpi_size_old/mpi_size;
int shift = mpi_rank*x;
for (int i = 0; i < shift; i++)
{
offset[0] += metadata_out[i];
//offset0.get(mpi_rank).add(metadata_out[i]);
}
int y = mpi_size_old%mpi_size;
if (mpi_rank < y)
{
for (int i = 0; i < mpi_size*x + mpi_rank; i++)
{
offset_add[0] += metadata_out[i];
//offset_add0.get(mpi_rank).add(metadata_out[i]);
}
}
}*/
//hsize_t stride[1] = {1};
hsize_t count[1] = {1};
//std::cout << "LOAD: MPI rank: " << mpi_rank << ", MPI size: " << mpi_size << ", Offset: " << offset[0] << ", Offset_add: " << offset_add[0] << ", Block: " << block[0] << ", Block_add: " << block_add[0] << std::endl;
/*
std::cout << "LOAD: MPI rank: " << mpi_rank << ", MPI size: " << mpi_size << std::endl;
for (size_t i = 0; i < offset0.get(mpi_rank).size(); i++)
std::cout << ", Offset: " << offset0.get(mpi_rank).get(i) << std::endl;
for (size_t i = 0; i < offset_add0.get(mpi_rank).size(); i++)
std::cout << ", Offset_add: " << offset_add0.get(mpi_rank).get(i) << std::endl;
for (size_t i = 0; i < block0.get(mpi_rank).size(); i++)
std::cout << ", Block: " << block0.get(mpi_rank).get(i) << std::endl;
for (size_t i = 0; i < block_add0.get(mpi_rank).size(); i++)
std::cout << ", Block_add: " << block_add0.get(mpi_rank).get(i) << std::endl;
*/
//Select file dataspace
hid_t file_dataspace_id_2 = H5Dget_space(dataset_2);
H5Sselect_hyperslab(file_dataspace_id_2, H5S_SELECT_SET, offset, NULL, count, block);
//Select file dataspace
/* hid_t file_dataspace_id_3 = H5Dget_space(dataset_2);
H5Sselect_hyperslab(file_dataspace_id_3, H5S_SELECT_SET, offset_add, NULL, count, block_add);*/
hsize_t mdim_2[1] = {block[0]};
// hsize_t mdim_3[1] = {block_add[0]};
//Size for data space in memory
/*if (mpi_rank >= mpi_size_old)
mdim_2[0] = 0;
else
mdim_2[0] = metadata_out[mpi_rank];*/
//Create data space in memory
hid_t mem_dataspace_id_2 = H5Screate_simple(1, mdim_2, NULL);
// hid_t mem_dataspace_id_3 = H5Screate_simple(1, mdim_3, NULL);
//if (mpi_rank == 0)
/* {
hssize_t size2;
size2 = H5Sget_select_npoints (mem_dataspace_id_2);
printf ("\nLOAD: memspace_id_2 size: %llu\n", size2);
size2 = H5Sget_select_npoints (file_dataspace_id_2);
printf ("LOAD: dataspace_id_2 size: %llu\n", size2);
}*/
/*
if (mpi_rank == 0)
{
hssize_t size2;
size2 = H5Sget_select_npoints (mem_dataspace_id_3);
printf ("\nLOAD: memspace_id_3 size: %llu\n", size2);
size2 = H5Sget_select_npoints (file_dataspace_id_3);
printf ("LOAD: dataspace_id_3 size: %llu\n", size2);
}
*/
/* size_t sum = 0;
for (int i = 0; i < mpi_size_old; i++)
{
sum += metadata_out[i];
}*/
// std::cout << "LOAD: sum: " << sum << std::endl;
// allocate the memory
HeapMemory pmem;
// HeapMemory pmem2;
//pmem.allocate(req);
ExtPreAlloc<HeapMemory> & mem = *(new ExtPreAlloc<HeapMemory>(block[0],pmem));
mem.incRef();
// ExtPreAlloc<HeapMemory> & mem2 = *(new ExtPreAlloc<HeapMemory>(block_add[0],pmem2));
// mem2.incRef();
// Read the dataset.
H5Dread(dataset_2, H5T_NATIVE_CHAR, mem_dataspace_id_2, file_dataspace_id_2, plist_id, (char *)mem.getPointer());
// Read the dataset.
// H5Dread(dataset_2, H5T_NATIVE_CHAR, mem_dataspace_id_3, file_dataspace_id_3, plist_id, (char *)mem2.getPointer());
mem.allocate(pmem.size());
// mem2.allocate(pmem2.size());
// std::cout << "Mem.size(): " << mem.size() << " = " << block[0] << std::endl;
Unpack_stat ps;
openfpm::vector<device_grid> loc_grid_old_unp;
openfpm::vector<GBoxes<device_grid::dims>> gdb_ext_old_unp;
Unpacker<decltype(loc_grid_old),HeapMemory>::unpack(mem,loc_grid_old_unp,ps,1);
Unpacker<decltype(gdb_ext_old),HeapMemory>::unpack(mem,gdb_ext_old_unp,ps,1);
/*
std::cout << "Loc_grid_old.size() before merge: " << loc_grid_old.size() << std::endl;
std::cout << "Gdb_ext_old.size() before merge: " << gdb_ext_old.size() << std::endl;
std::cout << "Loc_grid_old_unp.size() before merge: " << loc_grid_old_unp.size() << std::endl;
std::cout << "Gdb_ext_old_unp.size() before merge: " << gdb_ext_old_unp.size() << std::endl;
*/
for (size_t i = 0; i < loc_grid_old_unp.size(); i++)
loc_grid_old.add(loc_grid_old_unp.get(i));
for (size_t i = 0; i < gdb_ext_old_unp.size(); i++)
gdb_ext_old.add(gdb_ext_old_unp.get(i));
// std::cout << "Loc_grid_old.size() after merge: " << loc_grid_old.size() << std::endl;
// std::cout << "Gdb_ext_old.size() after merge: " << gdb_ext_old.size() << std::endl;
// std::cout << "*********************************" << std::endl;
mem.decRef();
delete &mem;
}
inline void load(const std::string & filename)
{
MPI_Comm comm = v_cl.getMPIComm();
MPI_Info info = MPI_INFO_NULL;
int mpi_rank = v_cl.getProcessUnitID();
//int mpi_size = v_cl.getProcessingUnits();
// Set up file access property list with parallel I/O access
hid_t plist_id = H5Pcreate(H5P_FILE_ACCESS);
H5Pset_fapl_mpio(plist_id, comm, info);
//Open a file
hid_t file = H5Fopen (filename.c_str(), H5F_ACC_RDONLY, plist_id);
H5Pclose(plist_id);
//Open dataset
hid_t dataset = H5Dopen (file, "metadata", H5P_DEFAULT);
//Create property list for collective dataset read
plist_id = H5Pcreate(H5P_DATASET_XFER);
H5Pset_dxpl_mpio(plist_id, H5FD_MPIO_COLLECTIVE);
//Select file dataspace
hid_t file_dataspace_id = H5Dget_space(dataset);
hssize_t mpi_size_old = H5Sget_select_npoints (file_dataspace_id);
//if (mpi_rank == 0)
//printf ("\nOld MPI size: %llu\n", mpi_size_old);
//Where to read metadata
int metadata_out[mpi_size_old];
for (int i = 0; i < mpi_size_old; i++)
{
metadata_out[i] = 0;
}
//Size for data space in memory
hsize_t mdim[1] = {(size_t)mpi_size_old};
//Create data space in memory
hid_t mem_dataspace_id = H5Screate_simple(1, mdim, NULL);
/*
if (mpi_rank == 0)
{
hssize_t size;
size = H5Sget_select_npoints (mem_dataspace_id);
printf ("\nmemspace_id size: %llu\n", size);
size = H5Sget_select_npoints (file_dataspace_id);
printf ("dataspace_id size: %llu\n", size);
}
*/
// Read the dataset.
H5Dread(dataset, H5T_NATIVE_INT, mem_dataspace_id, file_dataspace_id, plist_id, metadata_out);
/*
if (mpi_rank == 0)
{
std::cout << "Metadata_out[]: ";
for (int i = 0; i < mpi_size_old; i++)
{
std::cout << metadata_out[i] << " ";
}
std::cout << " " << std::endl;
}
*/
openfpm::vector<size_t> metadata_accum;
metadata_accum.resize(mpi_size_old);
metadata_accum.get(0) = 0;
for (int i = 1 ; i < mpi_size_old ; i++)
metadata_accum.get(i) = metadata_accum.get(i-1) + metadata_out[i-1];
//Open dataset
hid_t dataset_2 = H5Dopen (file, "grid_dist", H5P_DEFAULT);
HDF5_reader<GRID_DIST> h5l;
//Create property list for collective dataset read
plist_id = H5Pcreate(H5P_DATASET_XFER);
H5Pset_dxpl_mpio(plist_id, H5FD_MPIO_COLLECTIVE);
/////////////////////////////////////
openfpm::vector<size_t> n_block;
n_block.resize(v_cl.getProcessingUnits());
for(size_t i = 0 ; i < n_block.size() ; i++)
n_block.get(i) = mpi_size_old / v_cl.getProcessingUnits();
size_t rest_block = mpi_size_old % v_cl.getProcessingUnits();
// std::cout << "MPI size old: " << mpi_size_old << std::endl;
//std::cout << "MPI size: " << v_cl.getProcessingUnits() << std::endl;
// std::cout << "Rest block: " << rest_block << std::endl;
size_t max_block;
if (rest_block != 0)
max_block = n_block.get(0) + 1;
else
max_block = n_block.get(0);
//for(size_t i = 0 ; i < n_block.size() ; i++)
for(size_t i = 0 ; i < rest_block ; i++)
n_block.get(i) += 1;
//for(size_t i = 0 ; i < n_block.size() ; i++)
//std::cout << "n_block.get(i): " << n_block.get(i) << std::endl;
size_t start_block = 0;
size_t stop_block = 0;
if (v_cl.getProcessUnitID() != 0)
{
for(size_t i = 0 ; i < v_cl.getProcessUnitID() ; i++)
start_block += n_block.get(i);
}
stop_block = start_block + n_block.get(v_cl.getProcessUnitID());
// std::cout << "ID: " << v_cl.getProcessUnitID() << "; Start block: " << start_block << "; " << "Stop block: " << stop_block << std::endl;
if (mpi_rank >= mpi_size_old)
load_block(start_block,mpi_size_old,metadata_out,metadata_accum,plist_id,dataset_2);
else
{
size_t n_bl = 0;
size_t lb = start_block;
for ( ; lb < stop_block ; lb++, n_bl++)
load_block(lb,mpi_size_old,metadata_out,metadata_accum,plist_id,dataset_2);
if (n_bl < max_block)
load_block(-1,mpi_size_old,metadata_out,metadata_accum,plist_id,dataset_2);
}
////////////////////////////////////
//std::cout << "LOAD: sum: " << sum << std::endl;
// Close the dataset.
H5Dclose(dataset);
H5Dclose(dataset_2);
// Close the file.
H5Fclose(file);
H5Pclose(plist_id);
h5l.load<device_grid>(filename,loc_grid_old,gdb_ext_old);
// Map the distributed grid
map();
/*
for (size_t i = 0; i < loc_grid.size(); i++)
{
std::cout << "loc_grids loaded: (" << loc_grid.get(i).getGrid().getBox().getLow(0) << "; " << loc_grid.get(i).getGrid().getBox().getLow(1) << "); (" << loc_grid.get(i).getGrid().getBox().getHigh(0) << "; " << loc_grid.get(i).getGrid().getBox().getHigh(1) << ")" << std::endl;
}*/
}
//! Define friend classes
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment