| 1 |
/* |
| 2 |
* (C) Copyright 2013 ECMWF. |
| 3 |
* |
| 4 |
* This software is licensed under the terms of the Apache Licence Version 2.0 |
| 5 |
* which can be obtained at http://www.apache.org/licenses/LICENSE-2.0. |
| 6 |
* In applying this licence, ECMWF does not waive the privileges and immunities |
| 7 |
* granted to it by virtue of its status as an intergovernmental organisation |
| 8 |
* nor does it submit to any jurisdiction. |
| 9 |
*/ |
| 10 |
|
| 11 |
/// @file HaloExchange.cc |
| 12 |
/// @author Willem Deconinck |
| 13 |
/// @date Nov 2013 |
| 14 |
|
| 15 |
#include <memory> |
| 16 |
#include <numeric> |
| 17 |
#include <sstream> |
| 18 |
#include <stdexcept> |
| 19 |
|
| 20 |
#include "pluto/pluto.h" |
| 21 |
|
| 22 |
#include "atlas/array/Array.h" |
| 23 |
#include "atlas/parallel/HaloExchange.h" |
| 24 |
#include "atlas/parallel/mpi/Statistics.h" |
| 25 |
#include "atlas/util/vector.h" |
| 26 |
|
| 27 |
namespace atlas { |
| 28 |
namespace parallel { |
| 29 |
|
| 30 |
namespace { |
| 31 |
struct IsGhostPoint { |
| 32 |
IsGhostPoint(const int mypart, const int part[], const idx_t ridx[], const idx_t base, const int /*N*/) { |
| 33 |
part_ = part; |
| 34 |
ridx_ = ridx; |
| 35 |
base_ = base; |
| 36 |
mypart_ = mypart; |
| 37 |
} |
| 38 |
|
| 39 |
bool operator()(idx_t idx) { |
| 40 |
if (part_[idx] != mypart_) { |
| 41 |
return true; |
| 42 |
} |
| 43 |
if (ridx_[idx] != base_ + idx) { |
| 44 |
return true; |
| 45 |
} |
| 46 |
return false; |
| 47 |
} |
| 48 |
int mypart_; |
| 49 |
const int* part_; |
| 50 |
const idx_t* ridx_; |
| 51 |
idx_t base_; |
| 52 |
}; |
| 53 |
} // namespace |
| 54 |
|
| 55 |
HaloExchange::HaloExchange() : |
| 56 |
HaloExchange("") { |
| 57 |
} |
| 58 |
|
| 59 |
HaloExchange::HaloExchange(const std::string& name): |
| 60 |
name_(name), |
| 61 |
is_setup_(false) { |
| 62 |
} |
| 63 |
|
| 64 |
HaloExchange::~HaloExchange() = default; |
| 65 |
|
| 66 |
void HaloExchange::setup(const int part[], const idx_t remote_idx[], const int base, const idx_t size) { |
| 67 |
setup(mpi::comm().name(), part, remote_idx, base, size); |
| 68 |
} |
| 69 |
|
| 70 |
void HaloExchange::setup(const int part[], const idx_t remote_idx[], const int base, idx_t parsize, idx_t halo_begin) { |
| 71 |
setup(mpi::comm().name(), part, remote_idx, base, parsize, halo_begin); |
| 72 |
} |
| 73 |
|
| 74 |
void HaloExchange::setup(const std::string& mpi_comm, const int part[], const idx_t remote_idx[], const int base, const idx_t size) { |
| 75 |
setup(mpi_comm, part, remote_idx, base, size, 0); |
| 76 |
} |
| 77 |
|
| 78 |
void HaloExchange::setup(const std::string& mpi_comm, const int part[], const idx_t remote_idx[], const int base, idx_t parsize, idx_t halo_begin) { |
| 79 |
ATLAS_TRACE("HaloExchange::setup"); |
| 80 |
comm_ = &mpi::comm(mpi_comm); |
| 81 |
myproc = comm().rank(); |
| 82 |
nproc = comm().size(); |
| 83 |
|
| 84 |
parsize_ = parsize; |
| 85 |
sendcounts_.resize(nproc); |
| 86 |
sendcounts_.assign(nproc, 0); |
| 87 |
recvcounts_.resize(nproc); |
| 88 |
recvcounts_.assign(nproc, 0); |
| 89 |
senddispls_.resize(nproc); |
| 90 |
senddispls_.assign(nproc, 0); |
| 91 |
recvdispls_.resize(nproc); |
| 92 |
recvdispls_.assign(nproc, 0); |
| 93 |
|
| 94 |
/* |
| 95 |
Find the amount of nodes this proc has to receive from each other proc |
| 96 |
*/ |
| 97 |
|
| 98 |
IsGhostPoint is_ghost(myproc, part, remote_idx, base, parsize_); |
| 99 |
atlas::vector<idx_t> ghost_points(parsize_); |
| 100 |
idx_t nghost = 0; |
| 101 |
|
| 102 |
atlas_omp_parallel_for(int jj = halo_begin; jj < parsize_; ++jj) { |
| 103 |
if (is_ghost(jj)) { |
| 104 |
int p = part[jj]; |
| 105 |
atlas_omp_critical { |
| 106 |
++recvcounts_[p]; |
| 107 |
ghost_points[nghost] = jj; |
| 108 |
nghost++; |
| 109 |
} |
| 110 |
} |
| 111 |
} |
| 112 |
|
| 113 |
recvcnt_ = std::accumulate(recvcounts_.begin(), recvcounts_.end(), 0); |
| 114 |
|
| 115 |
/* |
| 116 |
Find the amount of nodes this proc has to send to each other proc |
| 117 |
*/ |
| 118 |
ATLAS_TRACE_MPI(ALLTOALL) { comm().allToAll(recvcounts_, sendcounts_); } |
| 119 |
|
| 120 |
sendcnt_ = std::accumulate(sendcounts_.begin(), sendcounts_.end(), 0); |
| 121 |
|
| 122 |
recvdispls_[0] = 0; |
| 123 |
senddispls_[0] = 0; |
| 124 |
for (int jproc = 1; jproc < nproc; ++jproc) // start at 1 |
| 125 |
{ |
| 126 |
recvdispls_[jproc] = recvcounts_[jproc - 1] + recvdispls_[jproc - 1]; |
| 127 |
senddispls_[jproc] = sendcounts_[jproc - 1] + senddispls_[jproc - 1]; |
| 128 |
} |
| 129 |
/* |
| 130 |
Fill vector "send_requests" with remote index of nodes needed, but are on |
| 131 |
other procs |
| 132 |
We can also fill in the vector "recvmap_" which holds local indices of |
| 133 |
requested nodes |
| 134 |
*/ |
| 135 |
std::vector<int> send_requests(recvcnt_); |
| 136 |
std::vector<int> recv_requests(sendcnt_); |
| 137 |
std::vector<int> cnt(nproc, 0); |
| 138 |
recvmap_.resize(recvcnt_); |
| 139 |
#ifdef __PGI |
| 140 |
// No idea why PGI compiler (20.7) in Release build ( -fast -O3 ) decides to vectorize following loop |
| 141 |
#pragma loop novector |
| 142 |
#endif |
| 143 |
for (idx_t jghost = 0; jghost < nghost; ++jghost) { |
| 144 |
const idx_t jj = ghost_points[jghost]; |
| 145 |
const int p = part[jj]; |
| 146 |
const int req_idx = recvdispls_[p] + cnt[p]; |
| 147 |
send_requests[req_idx] = remote_idx[jj] - base; |
| 148 |
recvmap_[req_idx] = jj; |
| 149 |
cnt[p]++; |
| 150 |
} |
| 151 |
|
| 152 |
/* |
| 153 |
Fill vector "recv_requests" with what is needed by other procs |
| 154 |
*/ |
| 155 |
|
| 156 |
ATLAS_TRACE_MPI(ALLTOALL) { |
| 157 |
comm().allToAllv(send_requests.data(), recvcounts_.data(), recvdispls_.data(), recv_requests.data(), |
| 158 |
sendcounts_.data(), senddispls_.data()); |
| 159 |
} |
| 160 |
|
| 161 |
/* |
| 162 |
What needs to be sent to other procs is asked by remote_idx, which is local |
| 163 |
here |
| 164 |
*/ |
| 165 |
sendmap_.resize(sendcnt_); |
| 166 |
for (int jj = 0; jj < sendcnt_; ++jj) { |
| 167 |
sendmap_[jj] = recv_requests[jj]; |
| 168 |
} |
| 169 |
|
| 170 |
is_setup_ = true; |
| 171 |
backdoor.parsize = parsize_; |
| 172 |
} |
| 173 |
|
| 174 |
void HaloExchange::wait_for_send(std::vector<int>& send_counts_init, std::vector<eckit::mpi::Request>& send_req) const { |
| 175 |
ATLAS_TRACE_MPI(WAIT, "mpi-wait send") { |
| 176 |
for (int jproc = 0; jproc < nproc; ++jproc) { |
| 177 |
if (send_counts_init[jproc] > 0) { |
| 178 |
comm().wait(send_req[jproc]); |
| 179 |
} |
| 180 |
} |
| 181 |
} |
| 182 |
} |
| 183 |
|
| 184 |
int HaloExchange::devices() const { |
| 185 |
return pluto::devices(); |
| 186 |
} |
| 187 |
|
| 188 |
bool HaloExchange::is_device_accessible(const void* ptr) const { |
| 189 |
return pluto::is_device_accessible(ptr); |
| 190 |
} |
| 191 |
|
| 192 |
namespace { |
| 193 |
|
| 194 |
template <typename Value> |
| 195 |
void execute_halo_exchange(HaloExchange* This, Value field[], int var_strides[], int var_extents[], int var_rank) { |
| 196 |
// WARNING: Only works if there is only one parallel dimension AND being |
| 197 |
// slowest moving |
| 198 |
|
| 199 |
array::ArrayShape shape{This->backdoor.parsize}; |
| 200 |
for (int j = 0; j < var_rank; ++j) { |
| 201 |
shape.push_back(var_extents[j]); |
| 202 |
} |
| 203 |
|
| 204 |
array::ArrayStrides strides{var_extents[0] * var_strides[0]}; |
| 205 |
for (int j = 0; j < var_rank; ++j) { |
| 206 |
strides.push_back(var_strides[j]); |
| 207 |
} |
| 208 |
|
| 209 |
std::unique_ptr<array::Array> arr(array::Array::wrap(field, array::ArraySpec{shape, strides})); |
| 210 |
|
| 211 |
switch (arr->rank()) { |
| 212 |
case 1: { |
| 213 |
This->execute<Value, 1>(*arr); |
| 214 |
break; |
| 215 |
} |
| 216 |
case 2: { |
| 217 |
This->execute<Value, 2>(*arr); |
| 218 |
break; |
| 219 |
} |
| 220 |
case 3: { |
| 221 |
This->execute<Value, 3>(*arr); |
| 222 |
break; |
| 223 |
} |
| 224 |
case 4: { |
| 225 |
This->execute<Value, 4>(*arr); |
| 226 |
break; |
| 227 |
} |
| 228 |
default: |
| 229 |
throw_NotImplemented("Rank not supported in halo exchange", Here()); |
| 230 |
} |
| 231 |
} |
| 232 |
|
| 233 |
template <typename Value> |
| 234 |
void execute_adjoint_halo_exchange(HaloExchange* This, Value field[], int var_strides[], int var_extents[], |
| 235 |
int var_rank) { |
| 236 |
// WARNING: Only works if there is only one parallel dimension AND being |
| 237 |
// slowest moving |
| 238 |
|
| 239 |
array::ArrayShape shape{This->backdoor.parsize}; |
| 240 |
for (int j = 0; j < var_rank; ++j) { |
| 241 |
shape.push_back(var_extents[j]); |
| 242 |
} |
| 243 |
|
| 244 |
array::ArrayStrides strides{var_extents[0] * var_strides[0]}; |
| 245 |
for (int j = 0; j < var_rank; ++j) { |
| 246 |
strides.push_back(var_strides[j]); |
| 247 |
} |
| 248 |
|
| 249 |
std::unique_ptr<array::Array> arr(array::Array::wrap(field, array::ArraySpec{shape, strides})); |
| 250 |
|
| 251 |
switch (arr->rank()) { |
| 252 |
case 1: { |
| 253 |
This->execute_adjoint<Value, 1>(*arr); |
| 254 |
break; |
| 255 |
} |
| 256 |
case 2: { |
| 257 |
This->execute_adjoint<Value, 2>(*arr); |
| 258 |
break; |
| 259 |
} |
| 260 |
case 3: { |
| 261 |
This->execute_adjoint<Value, 3>(*arr); |
| 262 |
break; |
| 263 |
} |
| 264 |
case 4: { |
| 265 |
This->execute_adjoint<Value, 4>(*arr); |
| 266 |
break; |
| 267 |
} |
| 268 |
default: |
| 269 |
throw_NotImplemented("Rank not supported in halo exchange", Here()); |
| 270 |
} |
| 271 |
} |
| 272 |
} // namespace |
| 273 |
|
| 274 |
extern "C" { |
| 275 |
|
| 276 |
HaloExchange* atlas__HaloExchange__new() { |
| 277 |
return new HaloExchange(); |
| 278 |
} |
| 279 |
|
| 280 |
void atlas__HaloExchange__delete(HaloExchange* This) { |
| 281 |
delete This; |
| 282 |
} |
| 283 |
|
| 284 |
void atlas__HaloExchange__setup(HaloExchange* This, int part[], idx_t remote_idx[], int base, int size) { |
| 285 |
This->setup(part, remote_idx, base, size); |
| 286 |
} |
| 287 |
|
| 288 |
|
| 289 |
void atlas__HaloExchange__execute_adjoint_strided_int(HaloExchange* This, int field[], int var_strides[], |
| 290 |
int var_extents[], int var_rank) { |
| 291 |
execute_adjoint_halo_exchange(This, field, var_strides, var_extents, var_rank); |
| 292 |
} |
| 293 |
|
| 294 |
void atlas__HaloExchange__execute_adjoint_strided_long(HaloExchange* This, long field[], int var_strides[], |
| 295 |
int var_extents[], int var_rank) { |
| 296 |
execute_adjoint_halo_exchange(This, field, var_strides, var_extents, var_rank); |
| 297 |
} |
| 298 |
|
| 299 |
void atlas__HaloExchange__execute_adjoint_strided_float(HaloExchange* This, float field[], int var_strides[], |
| 300 |
int var_extents[], int var_rank) { |
| 301 |
execute_adjoint_halo_exchange(This, field, var_strides, var_extents, var_rank); |
| 302 |
} |
| 303 |
|
| 304 |
void atlas__HaloExchange__execute_adjoint_strided_double(HaloExchange* This, double field[], int var_strides[], |
| 305 |
int var_extents[], int var_rank) { |
| 306 |
execute_adjoint_halo_exchange(This, field, var_strides, var_extents, var_rank); |
| 307 |
} |
| 308 |
|
| 309 |
void atlas__HaloExchange__execute_strided_int(HaloExchange* This, int field[], int var_strides[], int var_extents[], |
| 310 |
int var_rank) { |
| 311 |
execute_halo_exchange(This, field, var_strides, var_extents, var_rank); |
| 312 |
} |
| 313 |
|
| 314 |
void atlas__HaloExchange__execute_strided_long(HaloExchange* This, long field[], int var_strides[], int var_extents[], |
| 315 |
int var_rank) { |
| 316 |
execute_halo_exchange(This, field, var_strides, var_extents, var_rank); |
| 317 |
} |
| 318 |
|
| 319 |
void atlas__HaloExchange__execute_strided_float(HaloExchange* This, float field[], int var_strides[], int var_extents[], |
| 320 |
int var_rank) { |
| 321 |
execute_halo_exchange(This, field, var_strides, var_extents, var_rank); |
| 322 |
} |
| 323 |
|
| 324 |
void atlas__HaloExchange__execute_strided_double(HaloExchange* This, double field[], int var_strides[], |
| 325 |
int var_extents[], int var_rank) { |
| 326 |
execute_halo_exchange(This, field, var_strides, var_extents, var_rank); |
| 327 |
} |
| 328 |
} |
| 329 |
|
| 330 |
///////////////////// |
| 331 |
|
| 332 |
} // namespace parallel |
| 333 |
} // namespace atlas |