import os
import h5py
import numpy as np
from bmtk.utils import io
from bmtk.utils.sonata.utils import add_hdf5_magic, add_hdf5_version
[docs]class CellVarRecorder(object):
"""Used to save cell membrane variables (V, Ca2+, etc) to the described hdf5 format.
For parallel simulations this class will write to a seperate tmp file on each rank, then use the merge method to
combine the results. This is less efficent, but doesn't require the user to install mpi4py and build h5py in
parallel mode. For better performance use the CellVarRecorderParallel class instead.
"""
_io = io
[docs] class DataTable(object):
"""A small struct to keep track of different \*/data (and buffer) tables"""
def __init__(self, var_name):
self.var_name = var_name
# If buffering data, buffer_block will be an in-memory array and will write to data_block during when
# filled. If not buffering buffer_block is an hdf5 dataset and data_block is ignored
self.data_block = None
self.buffer_block = None
def __init__(self, file_name, tmp_dir, variables, buffer_data=True, mpi_rank=0, mpi_size=1, **kwargs):
self._file_name = file_name
self._h5_handle = None
self._tmp_dir = tmp_dir
self._variables = variables if isinstance(variables, list) else [variables]
self._n_vars = len(self._variables) # Used later to keep track if more than one var is saved to the same file.
self._mpi_rank = mpi_rank
self._mpi_size = mpi_size
self._tmp_files = []
self._saved_file = file_name
self._population = kwargs.get('population', None)
self._units = kwargs.get('units', None)
if mpi_size > 1:
self._io.log_warning('Was unable to run h5py in parallel (mpi) mode.' +
' Saving of membrane variable(s) may slow down.')
tmp_fname = os.path.basename(file_name) # make sure file names don't clash if there are multiple reports
self._tmp_files = [os.path.join(tmp_dir, '__bmtk_tmp_cellvars_{}_{}'.format(r, tmp_fname))
for r in range(self._mpi_size)]
self._file_name = self._tmp_files[self._mpi_rank]
self._mapping_gids = [] # list of gids in the order they appear in the data
self._gid_map = {} # table for looking up the gid offsets
self._map_attrs = {} # Used for additonal attributes in /mapping
self._mapping_element_ids = [] # sections
self._mapping_element_pos = [] # segments
self._mapping_index = [0] # index_pointer
self._buffer_data = buffer_data
self._data_blocks = {var_name: self.DataTable(var_name) for var_name in self._variables}
self._last_save_indx = 0 # for buffering, used to keep track of last timestep data was saved to disk
self._buffer_block_size = 0
self._total_steps = 0
# Keep track of gids across the different ranks
self._n_gids_all = 0
self._n_gids_local = 0
self._gids_beg = 0
self._gids_end = 0
# Keep track of segment counts across the different ranks
self._n_segments_all = 0
self._n_segments_local = 0
self._seg_offset_beg = 0
self._seg_offset_end = 0
self._tstart = 0.0
self._tstop = 0.0
self._dt = 0.01
self._is_initialized = False
@property
def tstart(self):
return self._tstart
@tstart.setter
def tstart(self, time_ms):
self._tstart = time_ms
@property
def tstop(self):
return self._tstop
@tstop.setter
def tstop(self, time_ms):
self._tstop = time_ms
@property
def dt(self):
return self._dt
@dt.setter
def dt(self, time_ms):
self._dt = time_ms
@property
def is_initialized(self):
return self._is_initialized
def _calc_offset(self):
self._n_segments_all = self._n_segments_local
self._seg_offset_beg = 0
self._seg_offset_end = self._n_segments_local
self._n_gids_all = self._n_gids_local
self._gids_beg = 0
self._gids_end = self._n_gids_local
def _create_h5_file(self):
self._h5_handle = h5py.File(self._file_name, 'w')
add_hdf5_version(self._h5_handle)
add_hdf5_magic(self._h5_handle)
[docs] def add_cell(self, gid, sec_list, seg_list, **map_attrs):
assert(len(sec_list) == len(seg_list))
# TODO: Check the same gid isn't added twice
n_segs = len(seg_list)
self._gid_map[gid] = (self._n_segments_local, self._n_segments_local + n_segs)
self._mapping_gids.append(gid)
self._mapping_element_ids.extend(sec_list)
self._mapping_element_pos.extend(seg_list)
self._mapping_index.append(self._mapping_index[-1] + n_segs)
self._n_segments_local += n_segs
self._n_gids_local += 1
for k, v in map_attrs.items():
if k not in self._map_attrs:
self._map_attrs[k] = v
else:
self._map_attrs[k].extend(v)
[docs] def initialize(self, n_steps, buffer_size=0):
self._calc_offset()
self._create_h5_file()
#root_name = '/report/{}'.format(self._population) if self._population is not None else '/'
var_grp = self._h5_handle.create_group('/mapping')
# TODO: gids --> node_ids
var_grp.create_dataset('gids', shape=(self._n_gids_all,), dtype=np.uint)
# TODO: element_id --> element_ids
var_grp.create_dataset('element_id', shape=(self._n_segments_all,), dtype=np.uint)
var_grp.create_dataset('element_pos', shape=(self._n_segments_all,), dtype=np.float)
var_grp.create_dataset('index_pointer', shape=(self._n_gids_all+1,), dtype=np.uint64)
var_grp.create_dataset('time', data=[self.tstart, self.tstop, self.dt])
# TODO: Let user determine value
var_grp['time'].attrs['units'] = 'ms'
for k, v in self._map_attrs.items():
var_grp.create_dataset(k, shape=(self._n_segments_all,), dtype=type(v[0]))
var_grp['gids'][self._gids_beg:self._gids_end] = self._mapping_gids
var_grp['element_id'][self._seg_offset_beg:self._seg_offset_end] = self._mapping_element_ids
var_grp['element_pos'][self._seg_offset_beg:self._seg_offset_end] = self._mapping_element_pos
var_grp['index_pointer'][self._gids_beg:(self._gids_end+1)] = self._mapping_index
for k, v in self._map_attrs.items():
var_grp[k][self._seg_offset_beg:self._seg_offset_end] = v
self._total_steps = n_steps
self._buffer_block_size = buffer_size
if not self._buffer_data:
# If data is not being buffered and instead written to the main block, we have to add a rank offset
# to the gid offset
for gid, gid_offset in self._gid_map.items():
self._gid_map[gid] = (gid_offset[0] + self._seg_offset_beg, gid_offset[1] + self._seg_offset_beg)
for var_name, data_tables in self._data_blocks.items():
# If users are trying to save multiple variables in the same file put data table in its own /{var} group
# (not sonata compliant). Otherwise the data table is located at the root
data_grp = self._h5_handle if self._n_vars == 1 else self._h5_handle.create_group('/{}'.format(var_name))
if self._buffer_data:
# Set up in-memory block to buffer recorded variables before writing to the dataset
data_tables.buffer_block = np.zeros((buffer_size, self._n_segments_local), dtype=np.float)
data_tables.data_block = data_grp.create_dataset('data', shape=(n_steps, self._n_segments_all),
dtype=np.float, chunks=True)
# TODO: Remove Variable name
data_tables.data_block.attrs['variable_name'] = var_name
if self._units is not None:
data_tables.data_block.attrs['units'] = self._units
else:
# Since we are not buffering data, we just write directly to the on-disk dataset
data_tables.buffer_block = data_grp.create_dataset('data', shape=(n_steps, self._n_segments_all),
dtype=np.float, chunks=True)
data_tables.buffer_block.attrs['variable_name'] = var_name
if self._units is not None:
data_tables.buffer_block.attrs['units'] = self._units
self._is_initialized = True
[docs] def record_cell(self, gid, var_name, seg_vals, tstep):
"""Record cell parameters.
:param gid: gid of cell.
:param var_name: name of variable being recorded.
:param seg_vals: list of all segment values
:param tstep: time step
"""
gid_beg, gid_end = self._gid_map[gid]
buffer_block = self._data_blocks[var_name].buffer_block
update_index = (tstep - self._last_save_indx)
buffer_block[update_index, gid_beg:gid_end] = seg_vals
[docs] def record_cell_block(self, gid, var_name, seg_vals):
"""Save cell parameters one block at a time
:param gid: gid of cell.
:param var_name: name of variable being recorded.
:param seg_vals: A vector/matrix of values being recorded
"""
gid_beg, gid_end = self._gid_map[gid]
buffer_block = self._data_blocks[var_name].buffer_block
if gid_end - gid_beg == 1:
buffer_block[:, gid_beg] = seg_vals
else:
buffer_block[:, gid_beg:gid_end] = seg_vals
[docs] def flush(self):
"""Move data from memory to dataset"""
if self._buffer_data:
blk_beg = self._last_save_indx
blk_end = blk_beg + self._buffer_block_size
if blk_end > self._total_steps:
# Need to handle the case that simulation doesn't end on a block step
blk_end = blk_beg + self._total_steps - blk_beg
block_size = blk_end - blk_beg
self._last_save_indx += block_size
for _, data_table in self._data_blocks.items():
data_table.data_block[blk_beg:blk_end, :] = data_table.buffer_block[:block_size, :]
[docs] def close(self):
self._h5_handle.close()
[docs] def merge(self):
if self._mpi_size > 1 and self._mpi_rank == 0:
h5final = h5py.File(self._saved_file, 'w')
tmp_h5_handles = [h5py.File(name, 'r') for name in self._tmp_files]
# Find the gid and segment offsets for each temp h5 file
gid_ranges = [] # list of (gid-beg, gid-end)
gid_offset = 0
total_gid_count = 0 # total number of gids across all ranks
seg_ranges = []
seg_offset = 0
total_seg_count = 0 # total number of segments across all ranks
time_ds = None
for h5_tmp in tmp_h5_handles:
seg_count = len(h5_tmp['/mapping/element_pos'])
seg_ranges.append((seg_offset, seg_offset+seg_count))
seg_offset += seg_count
total_seg_count += seg_count
gid_count = len(h5_tmp['mapping/gids'])
gid_ranges.append((gid_offset, gid_offset+gid_count))
gid_offset += gid_count
total_gid_count += gid_count
time_ds = h5_tmp['mapping/time']
mapping_grp = h5final.create_group('mapping')
if time_ds:
mapping_grp.create_dataset('time', data=time_ds)
element_id_ds = mapping_grp.create_dataset('element_id', shape=(total_seg_count,), dtype=np.uint)
el_pos_ds = mapping_grp.create_dataset('element_pos', shape=(total_seg_count,), dtype=np.float)
gids_ds = mapping_grp.create_dataset('gids', shape=(total_gid_count,), dtype=np.uint)
index_pointer_ds = mapping_grp.create_dataset('index_pointer', shape=(total_gid_count+1,), dtype=np.uint)
for k, v in self._map_attrs.items():
mapping_grp.create_dataset(k, shape=(total_seg_count,), dtype=type(v[0]))
# combine the /mapping datasets
for i, h5_tmp in enumerate(tmp_h5_handles):
tmp_mapping_grp = h5_tmp['mapping']
beg, end = seg_ranges[i]
element_id_ds[beg:end] = tmp_mapping_grp['element_id']
el_pos_ds[beg:end] = tmp_mapping_grp['element_pos']
for k, v in self._map_attrs.items():
mapping_grp[k][beg:end] = v
# shift the index pointer values
index_pointer = np.array(tmp_mapping_grp['index_pointer'])
update_index = beg + index_pointer
beg, end = gid_ranges[i]
gids_ds[beg:end] = tmp_mapping_grp['gids']
index_pointer_ds[beg:(end+1)] = update_index
# combine the /var/data datasets
for var_name in self._variables:
data_name = '/data' if self._n_vars == 1 else '/{}/data'.format(var_name)
# data_name = '/{}/data'.format(var_name)
var_data = h5final.create_dataset(data_name, shape=(self._total_steps, total_seg_count), dtype=np.float)
var_data.attrs['variable_name'] = var_name
for i, h5_tmp in enumerate(tmp_h5_handles):
beg, end = seg_ranges[i]
var_data[:, beg:end] = h5_tmp[data_name]
for tmp_file in self._tmp_files:
os.remove(tmp_file)
[docs]class CellVarRecorderParallel(CellVarRecorder):
"""
Unlike the parent, this take advantage of parallel h5py to writting to the results file across different ranks.
"""
def __init__(self, file_name, tmp_dir, variables, buffer_data=True, **kwargs):
super(CellVarRecorder, self).__init__(file_name, tmp_dir, variables, buffer_data=buffer_data, mpi_rank=0,
mpi_size=1, **kwargs)
def _calc_offset(self):
from mpi4py import MPI # Just needed for UNSIGNED_INT
comm = io.bmtk_world_comm.comm
rank = comm.Get_rank()
nhosts = comm.Get_size()
# iterate through the ranks let rank r determine the offset from rank r-1
for r in range(comm.Get_size()):
if rank == r:
if rank < (nhosts - 1):
# pass the num of segments and num of gids to the next rank
offsets = np.array([self._n_segments_local, self._n_gids_local], dtype=np.uint)
comm.Send([offsets, MPI.UNSIGNED_INT], dest=(rank+1))
if rank > 0:
# get num of segments and gids from prev. rank and calculate offsets
offset = np.empty(2, dtype=np.uint)
comm.Recv([offsets, MPI.UNSIGNED_INT], source=(r-1))
self._seg_offset_beg = offsets[0]
self._seg_offset_end = self._seg_offset_beg + self._n_segments_local
self._gids_beg = offset[1]
self._gids_end = self._gids_beg + self._n_gids_local
comm.Barrier()
# broadcast the total num of gids/segments from the final rank to all the others
if rank == (nhosts - 1):
total_counts = np.array([self._seg_offset_end, self._gids_end], dtype=np.uint)
else:
total_counts = np.empty(2, dtype=np.uint)
comm.Bcast(total_counts, root=(nhosts-1))
self._n_segments_all = total_counts[0]
self._n_gids_all = total_counts[1]
def _create_h5_file(self):
self._h5_handle = h5py.File(self._file_name, 'w', driver='mpio', comm=io.bmtk_world_comm.comm)
add_hdf5_version(self._h5_handle)
add_hdf5_magic(self._h5_handle)
