[615] | 1 | from itertools import groupby |
---|
| 2 | import numpy as np |
---|
| 3 | |
---|
| 4 | def inverse_list(lst): return {j:i for i,j in enumerate(lst)} |
---|
| 5 | def ordered_list(lst): |
---|
| 6 | lst = sorted(zip( lst, range(len(lst)) )) |
---|
| 7 | lst, order = zip(*lst) |
---|
| 8 | order = inverse_list(order) |
---|
| 9 | order = [order[i] for i in range(len(lst))] # dict->list |
---|
| 10 | return lst,order |
---|
| 11 | |
---|
| 12 | #----------------------- Parallel access to distributed array ------------------------------# |
---|
| 13 | |
---|
| 14 | # The classes PDim, Get_Indices and PArrayX are used to read NetCDF arrays in parallel |
---|
| 15 | # Each MPI process reads a contiguous chunk of the array once. |
---|
| 16 | # Classes CstPArrayX emulate constant arrays that are not stored in a NetCDF file. |
---|
| 17 | # With class LocArray1D, local data is copied from a user-provided array, rather than from a NetCDF file. |
---|
| 18 | # Using the class Get_Indices, each MPI process can obtain data for an arbitrary list of indices |
---|
| 19 | # Data is not re-read from file but scattered via MPI all2all |
---|
| 20 | |
---|
| 21 | class PDim: |
---|
| 22 | def __init__(self, ncdim, comm): |
---|
| 23 | mpi_rank, mpi_size = comm.Get_rank(), comm.Get_size() |
---|
| 24 | self.ncdim, n = ncdim, len(ncdim) |
---|
| 25 | vtxdist = [(n*i)/mpi_size for i in range(mpi_size+1)] |
---|
| 26 | self.n, self.vtxdist = n, np.asarray(vtxdist, dtype=np.int32) |
---|
| 27 | self.comm, self.start, self.end = comm, vtxdist[mpi_rank], vtxdist[mpi_rank+1] |
---|
| 28 | def getter(self, index_list): return Get_Indices(self, index_list) |
---|
| 29 | def get(self, index_list, data): |
---|
| 30 | getter=self.getter(index_list) |
---|
| 31 | return getter(data) |
---|
| 32 | class Get_Indices: |
---|
| 33 | def __init__(self, dim, index_list): # list MUST be a list of unique indices |
---|
| 34 | index_list, self.order = ordered_list(index_list) |
---|
| 35 | comm, vtxdist = dim.comm, dim.vtxdist |
---|
| 36 | list_local = [ [ j-vtxdist[i] for j in index_list if j>=vtxdist[i] and j<vtxdist[i+1]] |
---|
| 37 | for i in range(len(vtxdist)-1) ] |
---|
| 38 | list_send = comm.alltoall(list_local) |
---|
| 39 | self.comm, self.n, self.list_n = comm, len(index_list), [len(i) for i in list_local] |
---|
| 40 | self.list_local, self.list_send, self.dict = list_local, list_send, inverse_list(index_list) |
---|
| 41 | self.vtxdist=vtxdist |
---|
| 42 | def get(self, get_data, put_data): |
---|
| 43 | data_send = [get_data(indices) for indices in self.list_send] |
---|
| 44 | data_recv = self.comm.alltoall(data_send) |
---|
| 45 | start=0 |
---|
| 46 | for data,n in zip(data_recv,self.list_n): |
---|
| 47 | if n>0: |
---|
| 48 | end = start + n |
---|
| 49 | put_data(start,end,data) |
---|
| 50 | start=end |
---|
| 51 | def __call__(indices, self): # self is a PArrayND |
---|
| 52 | shape = list(self.data.shape) |
---|
| 53 | shape[0] = indices.n |
---|
| 54 | self.data_out = np.zeros(shape, dtype=self.data.dtype) |
---|
| 55 | indices.get(self.get_data, self.put_data ) |
---|
| 56 | # return self.data_out |
---|
| 57 | return self.reorder(indices.order) |
---|
| 58 | |
---|
| 59 | class PArrayND: |
---|
| 60 | def init_data(self, dim, data): |
---|
| 61 | self.dim, self.data = dim, np.asarray(data) # local chunk of data |
---|
| 62 | def max(self): |
---|
| 63 | max_loc = self.data.max() |
---|
| 64 | self.dim.comm.allreduce |
---|
| 65 | |
---|
| 66 | class PArray1D(PArrayND): |
---|
| 67 | def __init__(self, dim, data): self.init_data(dim, data[dim.start:dim.end]) |
---|
| 68 | def get_data(self, indices): return np.array(self.data[indices]) |
---|
| 69 | def put_data(self, start, end, data): self.data_out[start:end]=data |
---|
| 70 | def reorder(self, order): return self.data_out[order] |
---|
| 71 | class LocPArray1D(PArray1D): |
---|
| 72 | def __init__(self, dim, data): self.init_data(dim,data) |
---|
| 73 | class CstPArray1D(PArray1D): |
---|
| 74 | def __init__(self, dim, dtype, val): |
---|
| 75 | self.init_data(dim,np.full( (dim.end-dim.start,), val, dtype=dtype)) |
---|
| 76 | |
---|
| 77 | class PArray2D(PArrayND): |
---|
| 78 | def __init__(self, dim, data): self.init_data(dim, data[dim.start:dim.end,:]) |
---|
| 79 | def get_data(self, indices): return self.data[indices,:] |
---|
| 80 | def put_data(self, start, end, data): self.data_out[start:end]=data |
---|
| 81 | def reorder(self, order): return self.data_out[order,:] |
---|
| 82 | |
---|
| 83 | def PArray(dim, data): return {1:PArray1D, 2:PArray2D}[len(data.shape)](dim,data) |
---|
| 84 | |
---|
| 85 | #-------------------------------------- Halo management -----------------------------------# |
---|
| 86 | |
---|
| 87 | # Classes LocalDim, LocalArrayX and Halo_Xchange are used to |
---|
| 88 | # store local arrays with halos and update halos, respectively |
---|
| 89 | # A LocalDim instance is created using a list of (global indices of) cells. The instance contains |
---|
| 90 | # a lookup table associating a local index to a global index. |
---|
| 91 | # A Halo_Xchange instance is created first based on a LocalDim and a partitioning table giving MPI process owning each cell. |
---|
| 92 | # This instance can the be used to create instances of Halo_Xchange. |
---|
| 93 | # LocalArray data can be initialized by reading from a PArray or a NumPy array containing local values. |
---|
| 94 | # The update() methods updates the halo using point-to-point MPI communications (send/recv) |
---|
| 95 | |
---|
| 96 | class LocalDim: |
---|
| 97 | def __init__(self, dim, cells): |
---|
| 98 | # dim : a PDim instance ; cells : a list of (global indices of) cells, in any order |
---|
| 99 | self.dim, self.loc2glob, self.glob2loc = dim, cells, inverse_list(list) |
---|
| 100 | |
---|
| 101 | def dict2list(size, thedict, default): |
---|
| 102 | lst=[default]*size |
---|
| 103 | for rank,val in thedict.items() : lst[rank]=val |
---|
| 104 | return lst |
---|
| 105 | |
---|
| 106 | class Halo_Xchange: |
---|
| 107 | def __init__(self, tag, dim, cells, part, reorder_cells=False): |
---|
| 108 | # cells = global indices, part = rank owning each cell |
---|
| 109 | # if reorder_cells is True, cells will be reordered so that each halo is contiguous |
---|
| 110 | # reordering is not recommended if cells is a telescopic sum of increasing sets C0<C1<C2 ... |
---|
| 111 | comm = dim.comm |
---|
| 112 | mpi_size, mpi_rank = comm.Get_size(), comm.Get_rank() |
---|
| 113 | self.comm, self.tag, self.reorder_cell = comm, tag, reorder_cells |
---|
| 114 | # sort cells by rank, keep track of their global and local index |
---|
| 115 | recv = sorted(zip(part,cells,range(len(cells)))) |
---|
| 116 | # group by MPI rank ; lst is a list of (rank,global,local) tuples so zip(*it) is a tuple of lists (rank,global,local) |
---|
| 117 | recv = { rank : zip(*lst) for rank,lst in groupby(recv,lambda x:x[0]) } |
---|
| 118 | # remove mpi_rank from dict and get list of own cells |
---|
| 119 | junk, own_global, own_local = recv.pop(mpi_rank) |
---|
| 120 | own_len = len(own_local) |
---|
| 121 | if reorder_cells : own_local = range(own_len) |
---|
| 122 | self.own_len, self.own_local, self.get_own = own_len, list(own_local), Get_Indices(dim, own_global) |
---|
| 123 | # figure out data we want to receive from other CPUs |
---|
| 124 | recv_rank = sorted(recv.keys()) |
---|
| 125 | recv_glob = { rank : recv[rank][1] for rank in recv_rank} # global indices of cells to receive |
---|
| 126 | recv_len = { rank : len(recv_glob[rank]) for rank in recv_rank} |
---|
| 127 | if reorder_cells : |
---|
| 128 | recv_loc, start = {}, own_len |
---|
| 129 | for rank in recv_rank: |
---|
| 130 | end = start+recv_len[rank] |
---|
| 131 | recv_loc[rank] = range(start,end) |
---|
| 132 | start=end |
---|
| 133 | cells = sum([recv_glob[rank] for rank in recv_rank], own_global) |
---|
| 134 | else: |
---|
| 135 | recv_loc = { rank : recv[rank][2] for rank in recv_rank} |
---|
| 136 | self.recv_list = [(rank, list(recv_loc[rank])) for rank in recv_rank] |
---|
| 137 | self.cells, self.get_all = cells, Get_Indices(dim, cells) |
---|
| 138 | # now figure out the data we must send |
---|
| 139 | send=comm.alltoall(dict2list(mpi_size, recv_glob, [])) |
---|
| 140 | send_len = [len(lst) for lst in send] |
---|
| 141 | send_rank = [rank for rank in range(mpi_size) if (send_len[rank]>0)] # CPUs asking for data |
---|
| 142 | print 'send_rank %d'%mpi_rank, send_rank |
---|
| 143 | # asssociate local index to global index |
---|
| 144 | own_dict = { glob:loc for glob,loc in zip(own_global, own_local) } |
---|
| 145 | self.send_list = [(rank, [own_dict[i] for i in send[rank]]) for rank in send_rank] |
---|
| 146 | |
---|
| 147 | class LocalArray: # a base class for arrays with halos |
---|
| 148 | def read_own(self, parray): self.put(self.halo.own_local, self.halo.get_own(parray)) |
---|
| 149 | def read_all(self, parray): self.data = self.halo.get_all(parray) |
---|
| 150 | def update(self): |
---|
| 151 | halo=self.halo |
---|
| 152 | comm, tag = halo.comm, halo.tag |
---|
| 153 | for rank,loc in halo.send_list: |
---|
| 154 | comm.send(self.get(loc), dest=rank, tag=tag) |
---|
| 155 | for rank,loc in halo.recv_list: |
---|
| 156 | self.put(loc, comm.recv(source=rank, tag=tag) ) |
---|
| 157 | |
---|
| 158 | class LocalArray1(LocalArray): # a 1D array with halo |
---|
| 159 | def __init__(self, halo): |
---|
| 160 | self.halo, self.data = halo, np.zeros((len(halo.cells),)) |
---|
| 161 | def get(self, cells): return np.asarray([self.data[i] for i in cells]) |
---|
| 162 | def put(self,cells,data): self.data[cells]=data |
---|
| 163 | class LocalArray2(LocalArray): # a 2D array with halo |
---|
| 164 | def __init__(self, halo, llm): |
---|
| 165 | self.halo, self.llm, self.data = halo, llm, np.array((halo.local_len,llm)) |
---|
| 166 | def get(self, cells): return self.data[cells,:] |
---|
| 167 | def put(self,cells,data): self.data[cells,:]=data |
---|