| 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.h |
| 12 |
/// @author Willem Deconinck |
| 13 |
/// @date Nov 2013 |
| 14 |
|
| 15 |
#pragma once |
| 16 |
|
| 17 |
#include <stdexcept> |
| 18 |
#include <string> |
| 19 |
#include <vector> |
| 20 |
|
| 21 |
#include "atlas/parallel/detail/Packer.h" |
| 22 |
#include "atlas/parallel/mpi/Statistics.h" |
| 23 |
#include "atlas/parallel/mpi/mpi.h" |
| 24 |
|
| 25 |
#include "atlas/array/ArrayView.h" |
| 26 |
#include "atlas/array/ArrayViewDefs.h" |
| 27 |
#include "atlas/array/ArrayViewUtil.h" |
| 28 |
#include "atlas/array/SVector.h" |
| 29 |
#include "atlas/array_fwd.h" |
| 30 |
#include "atlas/library/config.h" |
| 31 |
#include "atlas/library/defines.h" |
| 32 |
#include "atlas/runtime/Exception.h" |
| 33 |
#include "atlas/runtime/Log.h" |
| 34 |
#include "atlas/util/Object.h" |
| 35 |
|
| 36 |
namespace atlas { |
| 37 |
namespace parallel { |
| 38 |
|
| 39 |
class HaloExchange : public util::Object { |
| 40 |
public: |
| 41 |
HaloExchange(); |
| 42 |
HaloExchange(const std::string& name); |
| 43 |
|
| 44 |
virtual ~HaloExchange(); |
| 45 |
|
| 46 |
public: // methods |
| 47 |
const std::string& name() const { return name_; } |
| 48 |
|
| 49 |
void setup(const int part[], const idx_t remote_idx[], const int base, idx_t size); |
| 50 |
void setup(const std::string& mpi_comm, const int part[], const idx_t remote_idx[], const int base, idx_t size); |
| 51 |
|
| 52 |
void setup(const int part[], const idx_t remote_idx[], const int base, idx_t size, idx_t halo_begin); |
| 53 |
void setup(const std::string& mpi_comm, const int part[], const idx_t remote_idx[], const int base, idx_t size, idx_t halo_begin); |
| 54 |
|
| 55 |
template <typename DATA_TYPE, int RANK, typename ParallelDim = array::FirstDim> |
| 56 |
void execute(array::Array& field, bool on_device = false) const; |
| 57 |
|
| 58 |
template <typename DATA_TYPE, int RANK, typename ParallelDim = array::FirstDim> |
| 59 |
void execute_adjoint(array::Array& field, bool on_device = false) const; |
| 60 |
|
| 61 |
private: // methods |
| 62 |
idx_t index(idx_t i, idx_t j, idx_t k, idx_t ni, idx_t nj, idx_t /*nk*/) const { return (i + ni * (j + nj * k)); } |
| 63 |
|
| 64 |
idx_t index(idx_t i, idx_t j, idx_t ni, idx_t /*nj*/) const { return (i + ni * j); } |
| 65 |
|
| 66 |
template <typename DATA_TYPE> |
| 67 |
void counts_displs_setup(const idx_t var_size, std::vector<int>& send_counts_init, |
| 68 |
std::vector<int>& recv_counts_init, std::vector<int>& send_counts, |
| 69 |
std::vector<int>& recv_counts, std::vector<int>& send_displs, |
| 70 |
std::vector<int>& recv_displs) const; |
| 71 |
|
| 72 |
|
| 73 |
template <typename DATA_TYPE> |
| 74 |
void ireceive(int tag, std::vector<int>& recv_displs, std::vector<int>& recv_counts, |
| 75 |
std::vector<eckit::mpi::Request>& recv_req, DATA_TYPE* recv_buffer) const; |
| 76 |
|
| 77 |
template <typename DATA_TYPE> |
| 78 |
void isend_and_wait_for_receive(int tag, std::vector<int>& recv_counts_init, |
| 79 |
std::vector<eckit::mpi::Request>& recv_req, std::vector<int>& send_displs, |
| 80 |
std::vector<int>& send_counts, std::vector<eckit::mpi::Request>& send_req, |
| 81 |
DATA_TYPE* send_buffer) const; |
| 82 |
|
| 83 |
void wait_for_send(std::vector<int>& send_counts, std::vector<eckit::mpi::Request>& send_req) const; |
| 84 |
|
| 85 |
template <typename DATA_TYPE> |
| 86 |
DATA_TYPE* allocate_buffer(const int buffer_size, const bool on_device) const; |
| 87 |
|
| 88 |
template <typename DATA_TYPE> |
| 89 |
void deallocate_buffer(DATA_TYPE* buffer, const int buffer_size, const bool on_device) const; |
| 90 |
|
| 91 |
template <int ParallelDim, typename DATA_TYPE, int RANK> |
| 92 |
void pack_send_buffer(const array::ArrayView<DATA_TYPE, RANK>& array, |
| 93 |
DATA_TYPE* send_buffer, int send_buffer_size, |
| 94 |
const bool on_device) const; |
| 95 |
|
| 96 |
template <int ParallelDim, typename DATA_TYPE, int RANK> |
| 97 |
void unpack_recv_buffer(const DATA_TYPE* recv_buffer, int recv_buffer_size, |
| 98 |
array::ArrayView<DATA_TYPE, RANK>& array, |
| 99 |
const bool on_device) const; |
| 100 |
|
| 101 |
template <int ParallelDim, typename DATA_TYPE, int RANK> |
| 102 |
void pack_recv_adjoint_buffer(const array::ArrayView<DATA_TYPE, RANK>& array, |
| 103 |
DATA_TYPE* recv_buffer, int recv_buffer_size, |
| 104 |
const bool on_device) const; |
| 105 |
|
| 106 |
template <int ParallelDim, typename DATA_TYPE, int RANK> |
| 107 |
void unpack_send_adjoint_buffer(const DATA_TYPE* send_buffer, int send_buffer_size, |
| 108 |
array::ArrayView<DATA_TYPE, RANK>& array, |
| 109 |
const bool on_device) const; |
| 110 |
|
| 111 |
template <int ParallelDim, typename DATA_TYPE, int RANK> |
| 112 |
void zero_halos(array::ArrayView<DATA_TYPE, RANK>& array, |
| 113 |
DATA_TYPE* recv_buffer, int recv_buffer_size, |
| 114 |
const bool on_device) const; |
| 115 |
|
| 116 |
template <typename DATA_TYPE, int RANK> |
| 117 |
void var_info(const array::ArrayView<DATA_TYPE, RANK>& arr, std::vector<idx_t>& varstrides, |
| 118 |
std::vector<idx_t>& varshape) const; |
| 119 |
|
| 120 |
const mpi::Comm& comm() const { |
| 121 |
return *comm_; |
| 122 |
} |
| 123 |
|
| 124 |
int devices() const; |
| 125 |
bool is_device_accessible(const void* ptr) const; |
| 126 |
|
| 127 |
private: // data |
| 128 |
std::string name_; |
| 129 |
bool is_setup_; |
| 130 |
|
| 131 |
int sendcnt_; |
| 132 |
int recvcnt_; |
| 133 |
std::vector<int> sendcounts_; |
| 134 |
std::vector<int> senddispls_; |
| 135 |
std::vector<int> recvcounts_; |
| 136 |
std::vector<int> recvdispls_; |
| 137 |
array::SVector<int> sendmap_; |
| 138 |
array::SVector<int> recvmap_; |
| 139 |
int parsize_; |
| 140 |
|
| 141 |
int nproc; |
| 142 |
int myproc; |
| 143 |
const mpi::Comm* comm_; |
| 144 |
|
| 145 |
public: |
| 146 |
struct Backdoor { |
| 147 |
int parsize; |
| 148 |
} backdoor; |
| 149 |
}; |
| 150 |
|
| 151 |
template <typename DATA_TYPE, int RANK, typename ParallelDim> |
| 152 |
void HaloExchange::execute(array::Array& field, bool on_device) const { |
| 153 |
ATLAS_TRACE(std::string("HaloExchange::execute")+(on_device?"[device]":"[host]"), {std::string("halo-exchange"), std::string("halo-exchange")+(on_device?".device":".host")}); |
| 154 |
if (!is_setup_) { |
| 155 |
throw_Exception("HaloExchange was not setup", Here()); |
| 156 |
} |
| 157 |
|
| 158 |
on_device = on_device && devices() > 0; |
| 159 |
|
| 160 |
bool update_device_at_end = false; |
| 161 |
if (on_device) { |
| 162 |
ATLAS_ASSERT(field.deviceNeedsUpdate() == false); |
| 163 |
|
| 164 |
if (not ATLAS_HAVE_GPU_AWARE_MPI) { |
| 165 |
// Without GPU aware MPI, we need to do halo exchange on host and copy back to device |
| 166 |
update_device_at_end = on_device; |
| 167 |
on_device = false; |
| 168 |
field.syncHost(); |
| 169 |
} |
| 170 |
else { |
| 171 |
field.syncDevice(); |
| 172 |
} |
| 173 |
} |
| 174 |
else { |
| 175 |
field.syncHost(); |
| 176 |
} |
| 177 |
|
| 178 |
auto array_view = |
| 179 |
on_device ? array::make_device_view<DATA_TYPE, RANK>(field) : array::make_host_view<DATA_TYPE, RANK>(field); |
| 180 |
|
| 181 |
constexpr int parallelDim = array::get_parallel_dim<ParallelDim>(array_view); |
| 182 |
idx_t var_size = array::get_var_size<parallelDim>(array_view); |
| 183 |
|
| 184 |
int tag(1); |
| 185 |
std::size_t nproc_loc(static_cast<std::size_t>(nproc)); |
| 186 |
std::vector<int> inner_counts(nproc_loc), halo_counts(nproc_loc); |
| 187 |
std::vector<int> inner_counts_init(nproc_loc), halo_counts_init(nproc_loc); |
| 188 |
std::vector<int> inner_displs(nproc_loc), halo_displs(nproc_loc); |
| 189 |
std::vector<eckit::mpi::Request> inner_req(nproc_loc), halo_req(nproc_loc); |
| 190 |
|
| 191 |
int inner_size = sendcnt_ * var_size; |
| 192 |
int halo_size = recvcnt_ * var_size; |
| 193 |
DATA_TYPE* inner_buffer = allocate_buffer<DATA_TYPE>(inner_size, on_device); |
| 194 |
DATA_TYPE* halo_buffer = allocate_buffer<DATA_TYPE>(halo_size, on_device); |
| 195 |
|
| 196 |
if (on_device) { |
| 197 |
ATLAS_ASSERT( is_device_accessible(inner_buffer) ); |
| 198 |
ATLAS_ASSERT( is_device_accessible(halo_buffer) ); |
| 199 |
ATLAS_ASSERT( is_device_accessible(array_view.data()) ); |
| 200 |
} |
| 201 |
|
| 202 |
counts_displs_setup<DATA_TYPE>(var_size, inner_counts_init, halo_counts_init, inner_counts, halo_counts, |
| 203 |
inner_displs, halo_displs); |
| 204 |
|
| 205 |
ireceive<DATA_TYPE>(tag, halo_displs, halo_counts, halo_req, halo_buffer); |
| 206 |
|
| 207 |
/// Pack |
| 208 |
pack_send_buffer<parallelDim>(array_view, inner_buffer, inner_size, on_device); |
| 209 |
|
| 210 |
isend_and_wait_for_receive<DATA_TYPE>(tag, halo_counts_init, halo_req, inner_displs, inner_counts, inner_req, |
| 211 |
inner_buffer); |
| 212 |
|
| 213 |
/// Unpack |
| 214 |
unpack_recv_buffer<parallelDim>(halo_buffer, halo_size, array_view, on_device); |
| 215 |
|
| 216 |
wait_for_send(inner_counts_init, inner_req); |
| 217 |
|
| 218 |
deallocate_buffer<DATA_TYPE>(inner_buffer, inner_size, on_device); |
| 219 |
deallocate_buffer<DATA_TYPE>(halo_buffer, halo_size, on_device); |
| 220 |
|
| 221 |
on_device ? field.setHostNeedsUpdate(true) : field.setDeviceNeedsUpdate(true); |
| 222 |
if (update_device_at_end) { |
| 223 |
field.syncDevice(); |
| 224 |
} |
| 225 |
} |
| 226 |
|
| 227 |
template <typename DATA_TYPE, int RANK, typename ParallelDim> |
| 228 |
void HaloExchange::execute_adjoint(array::Array& field, bool on_device) const { |
| 229 |
if (!is_setup_) { |
| 230 |
throw_Exception("HaloExchange was not setup", Here()); |
| 231 |
} |
| 232 |
ATLAS_TRACE(std::string("HaloExchange::execute_adjoint")+(on_device?"[device]":"[host]"), {std::string("halo-exchange-adjoint"),std::string("halo-exchange-adjoint")+(on_device?".device":".host")}); |
| 233 |
|
| 234 |
on_device = on_device && devices() > 0; |
| 235 |
|
| 236 |
// adjoint on device is not yet implemented! |
| 237 |
// So copy to host and do halo-exchange there, and copy back to device |
| 238 |
bool update_device_at_end = on_device; |
| 239 |
if (on_device) { |
| 240 |
ATLAS_ASSERT(field.deviceNeedsUpdate() == false); |
| 241 |
field.syncHost(); |
| 242 |
on_device = false; |
| 243 |
} |
| 244 |
|
| 245 |
auto array_view = |
| 246 |
on_device ? array::make_device_view<DATA_TYPE, RANK>(field) : array::make_host_view<DATA_TYPE, RANK>(field); |
| 247 |
|
| 248 |
constexpr int parallelDim = array::get_parallel_dim<ParallelDim>(array_view); |
| 249 |
idx_t var_size = array::get_var_size<parallelDim>(array_view); |
| 250 |
|
| 251 |
int tag(1); |
| 252 |
std::size_t nproc_loc(static_cast<std::size_t>(nproc)); |
| 253 |
std::vector<int> halo_counts(nproc_loc), inner_counts(nproc_loc); |
| 254 |
std::vector<int> halo_counts_init(nproc_loc), inner_counts_init(nproc_loc); |
| 255 |
std::vector<int> halo_displs(nproc_loc), inner_displs(nproc_loc); |
| 256 |
std::vector<eckit::mpi::Request> halo_req(nproc_loc), inner_req(nproc_loc); |
| 257 |
|
| 258 |
int halo_size = sendcnt_ * var_size; |
| 259 |
int inner_size = recvcnt_ * var_size; |
| 260 |
DATA_TYPE* halo_buffer = allocate_buffer<DATA_TYPE>(halo_size, on_device); |
| 261 |
DATA_TYPE* inner_buffer = allocate_buffer<DATA_TYPE>(inner_size, on_device); |
| 262 |
|
| 263 |
counts_displs_setup<DATA_TYPE>(var_size, halo_counts_init, inner_counts_init, halo_counts, inner_counts, |
| 264 |
halo_displs, inner_displs); |
| 265 |
|
| 266 |
ireceive<DATA_TYPE>(tag, halo_displs, halo_counts, halo_req, halo_buffer); |
| 267 |
|
| 268 |
/// Pack |
| 269 |
pack_recv_adjoint_buffer<parallelDim>(array_view, inner_buffer, inner_size, on_device); |
| 270 |
|
| 271 |
/// Send |
| 272 |
isend_and_wait_for_receive<DATA_TYPE>(tag, halo_counts_init, halo_req, inner_displs, inner_counts, inner_req, |
| 273 |
inner_buffer); |
| 274 |
|
| 275 |
/// Unpack |
| 276 |
unpack_send_adjoint_buffer<parallelDim>(halo_buffer, halo_size, array_view, on_device); |
| 277 |
|
| 278 |
/// Wait for sending to finish |
| 279 |
wait_for_send(inner_counts_init, inner_req); |
| 280 |
|
| 281 |
zero_halos<parallelDim>(array_view, halo_buffer, halo_size, on_device); |
| 282 |
|
| 283 |
deallocate_buffer<DATA_TYPE>(halo_buffer, halo_size, on_device); |
| 284 |
deallocate_buffer<DATA_TYPE>(inner_buffer, inner_size, on_device); |
| 285 |
|
| 286 |
on_device ? field.setHostNeedsUpdate(true) : field.setDeviceNeedsUpdate(true); |
| 287 |
if (update_device_at_end) { |
| 288 |
field.syncDevice(); |
| 289 |
} |
| 290 |
} |
| 291 |
|
| 292 |
template <typename DATA_TYPE> |
| 293 |
DATA_TYPE* HaloExchange::allocate_buffer(const int buffer_size, const bool on_device) const { |
| 294 |
DATA_TYPE* buffer{nullptr}; |
| 295 |
|
| 296 |
if (on_device) { |
| 297 |
util::allocate_devicemem(buffer, buffer_size); |
| 298 |
} |
| 299 |
else { |
| 300 |
util::allocate_hostmem(buffer, buffer_size); |
| 301 |
} |
| 302 |
|
| 303 |
return buffer; |
| 304 |
} |
| 305 |
|
| 306 |
|
| 307 |
template <typename DATA_TYPE> |
| 308 |
void HaloExchange::deallocate_buffer(DATA_TYPE* buffer, const int buffer_size, const bool on_device) const { |
| 309 |
if (on_device) { |
| 310 |
util::delete_devicemem(buffer, buffer_size); |
| 311 |
} |
| 312 |
else { |
| 313 |
util::delete_hostmem(buffer, buffer_size); |
| 314 |
} |
| 315 |
} |
| 316 |
|
| 317 |
|
| 318 |
template <typename DATA_TYPE> |
| 319 |
void HaloExchange::counts_displs_setup(const idx_t var_size, std::vector<int>& send_counts_init, |
| 320 |
std::vector<int>& recv_counts_init, std::vector<int>& send_counts, |
| 321 |
std::vector<int>& recv_counts, std::vector<int>& send_displs, |
| 322 |
std::vector<int>& recv_displs) const { |
| 323 |
for (size_t jproc = 0; jproc < static_cast<size_t>(nproc); ++jproc) { |
| 324 |
send_counts_init[jproc] = sendcounts_[jproc]; |
| 325 |
recv_counts_init[jproc] = recvcounts_[jproc]; |
| 326 |
send_counts[jproc] = sendcounts_[jproc] * var_size; |
| 327 |
recv_counts[jproc] = recvcounts_[jproc] * var_size; |
| 328 |
send_displs[jproc] = senddispls_[jproc] * var_size; |
| 329 |
recv_displs[jproc] = recvdispls_[jproc] * var_size; |
| 330 |
} |
| 331 |
} |
| 332 |
|
| 333 |
template <typename DATA_TYPE> |
| 334 |
void HaloExchange::ireceive(int tag, std::vector<int>& recv_displs, std::vector<int>& recv_counts, |
| 335 |
std::vector<eckit::mpi::Request>& recv_req, DATA_TYPE* recv_buffer) const { |
| 336 |
ATLAS_TRACE_MPI(IRECEIVE) { |
| 337 |
/// Let MPI know what we like to receive |
| 338 |
for (size_t jproc = 0; jproc < static_cast<size_t>(nproc); ++jproc) { |
| 339 |
if (recv_counts[jproc] > 0) { |
| 340 |
recv_req[jproc] = |
| 341 |
comm().iReceive(recv_buffer+recv_displs[jproc], recv_counts[jproc], jproc, tag); |
| 342 |
} |
| 343 |
} |
| 344 |
} |
| 345 |
} |
| 346 |
|
| 347 |
template <typename DATA_TYPE> |
| 348 |
void HaloExchange::isend_and_wait_for_receive(int tag, std::vector<int>& recv_counts_init, |
| 349 |
std::vector<eckit::mpi::Request>& recv_req, std::vector<int>& send_displs, |
| 350 |
std::vector<int>& send_counts, std::vector<eckit::mpi::Request>& send_req, |
| 351 |
DATA_TYPE* send_buffer) const { |
| 352 |
/// Send |
| 353 |
ATLAS_TRACE_MPI(ISEND) { |
| 354 |
for (size_t jproc = 0; jproc < static_cast<size_t>(nproc); ++jproc) { |
| 355 |
if (send_counts[jproc] > 0) { |
| 356 |
send_req[jproc] = comm().iSend(send_buffer+send_displs[jproc], send_counts[jproc], jproc, tag); |
| 357 |
} |
| 358 |
} |
| 359 |
} |
| 360 |
|
| 361 |
/// Wait for receiving to finish |
| 362 |
ATLAS_TRACE_MPI(WAIT, "mpi-wait receive") { |
| 363 |
for (size_t jproc = 0; jproc < static_cast<size_t>(nproc); ++jproc) { |
| 364 |
if (recv_counts_init[jproc] > 0) { |
| 365 |
comm().wait(recv_req[jproc]); |
| 366 |
} |
| 367 |
} |
| 368 |
} |
| 369 |
} |
| 370 |
|
| 371 |
template <int ParallelDim, typename DATA_TYPE, int RANK> |
| 372 |
void HaloExchange::zero_halos(array::ArrayView<DATA_TYPE, RANK>& array, |
| 373 |
DATA_TYPE* recv_buffer, int recv_size, |
| 374 |
ATLAS_MAYBE_UNUSED const bool on_device) const { |
| 375 |
ATLAS_TRACE(); |
| 376 |
detail::Zeroer<ParallelDim, DATA_TYPE, RANK>::zero(recvcnt_, recvmap_, array, recv_buffer, recv_size, on_device); |
| 377 |
} |
| 378 |
|
| 379 |
template <int ParallelDim, typename DATA_TYPE, int RANK> |
| 380 |
void HaloExchange::pack_send_buffer(const array::ArrayView<DATA_TYPE, RANK>& array, |
| 381 |
DATA_TYPE* send_buffer, int send_size, |
| 382 |
const bool on_device) const { |
| 383 |
ATLAS_TRACE(); |
| 384 |
detail::Packer<ParallelDim, DATA_TYPE, RANK>::pack(sendcnt_, sendmap_, array, send_buffer, |
| 385 |
send_size, on_device); |
| 386 |
} |
| 387 |
|
| 388 |
template <int ParallelDim, typename DATA_TYPE, int RANK> |
| 389 |
void HaloExchange::unpack_recv_buffer(const DATA_TYPE* recv_buffer, int recv_size, |
| 390 |
array::ArrayView<DATA_TYPE, RANK>& array, |
| 391 |
const bool on_device) const { |
| 392 |
ATLAS_TRACE(); |
| 393 |
detail::Packer<ParallelDim, DATA_TYPE, RANK>::unpack(recvcnt_, recvmap_, recv_buffer, recv_size, array, on_device); |
| 394 |
} |
| 395 |
|
| 396 |
template <int ParallelDim, typename DATA_TYPE, int RANK> |
| 397 |
void HaloExchange::pack_recv_adjoint_buffer(const array::ArrayView<DATA_TYPE, RANK>& array, |
| 398 |
DATA_TYPE* recv_buffer, int recv_size, const bool on_device) const { |
| 399 |
ATLAS_TRACE(); |
| 400 |
detail::AdjointPacker<ParallelDim, DATA_TYPE, RANK>::pack(recvcnt_, recvmap_, array, recv_buffer, recv_size, on_device); |
| 401 |
} |
| 402 |
|
| 403 |
template <int ParallelDim, typename DATA_TYPE, int RANK> |
| 404 |
void HaloExchange::unpack_send_adjoint_buffer(const DATA_TYPE* send_buffer, int send_size, |
| 405 |
array::ArrayView<DATA_TYPE, RANK>& array, |
| 406 |
ATLAS_MAYBE_UNUSED const bool on_device) const { |
| 407 |
ATLAS_TRACE(); |
| 408 |
detail::AdjointPacker<ParallelDim, DATA_TYPE, RANK>::unpack(sendcnt_, sendmap_, send_buffer, send_size, array, on_device); |
| 409 |
} |
| 410 |
|
| 411 |
// template<typename DATA_TYPE> |
| 412 |
// void HaloExchange::execute( DATA_TYPE field[], idx_t nb_vars ) const |
| 413 |
//{ |
| 414 |
// throw_AssertionFailed("Call not supported"); |
| 415 |
|
| 416 |
// idx_t strides[] = {1}; |
| 417 |
// idx_t shape[] = {nb_vars}; |
| 418 |
// execute( field, strides, shape, 1); |
| 419 |
//} |
| 420 |
|
| 421 |
// template <typename DATA_TYPE, int RANK> |
| 422 |
// void HaloExchange::execute( array::ArrayView<DATA_TYPE,RANK>&& field ) const |
| 423 |
//{ |
| 424 |
// execute(field); |
| 425 |
//} |
| 426 |
|
| 427 |
//---------------------------------------------------------------------------------------------------------------------- |
| 428 |
// C wrapper interfaces to C++ routines |
| 429 |
extern "C" { |
| 430 |
HaloExchange* atlas__HaloExchange__new(); |
| 431 |
void atlas__HaloExchange__delete(HaloExchange* This); |
| 432 |
void atlas__HaloExchange__setup(HaloExchange* This, int part[], idx_t remote_idx[], int base, int size); |
| 433 |
void atlas__HaloExchange__execute_strided_int(HaloExchange* This, int field[], int var_strides[], int var_shape[], |
| 434 |
int var_rank); |
| 435 |
void atlas__HaloExchange__execute_strided_long(HaloExchange* This, long field[], int var_strides[], int var_shape[], |
| 436 |
int var_rank); |
| 437 |
void atlas__HaloExchange__execute_strided_float(HaloExchange* This, float field[], int var_strides[], int var_shape[], |
| 438 |
int var_rank); |
| 439 |
void atlas__HaloExchange__execute_strided_double(HaloExchange* This, double field[], int var_strides[], int var_shape[], |
| 440 |
int var_rank); |
| 441 |
void atlas__HaloExchange__execute_int(HaloExchange* This, int field[], int var_rank); |
| 442 |
void atlas__HaloExchange__execute_float(HaloExchange* This, float field[], int var_rank); |
| 443 |
void atlas__HaloExchange__execute_double(HaloExchange* This, double field[], int var_rank); |
| 444 |
|
| 445 |
void atlas__HaloExchange__execute_adjoint_strided_int(HaloExchange* This, int field[], int var_strides[], |
| 446 |
int var_shape[], int var_rank); |
| 447 |
void atlas__HaloExchange__execute_adjoint_strided_long(HaloExchange* This, long field[], int var_strides[], |
| 448 |
int var_shape[], int var_rank); |
| 449 |
void atlas__HaloExchange__execute_adjoint_strided_float(HaloExchange* This, float field[], int var_strides[], |
| 450 |
int var_shape[], int var_rank); |
| 451 |
void atlas__HaloExchange__execute_adjoint_strided_double(HaloExchange* This, double field[], int var_strides[], |
| 452 |
int var_shape[], int var_rank); |
| 453 |
void atlas__HaloExchange__execute_adjoint_int(HaloExchange* This, int field[], int var_rank); |
| 454 |
void atlas__HaloExchange__execute_adjoint_float(HaloExchange* This, float field[], int var_rank); |
| 455 |
void atlas__HaloExchange__execute_adjoint_double(HaloExchange* This, double field[], int var_rank); |
| 456 |
} |
| 457 |
|
| 458 |
//---------------------------------------------------------------------------------------------------------------------- |
| 459 |
|
| 460 |
} // namespace parallel |
| 461 |
} // namespace atlas |