1 | /* |
---|
2 | * Copyright (C) 2015, UChicago Argonne, LLC |
---|
3 | * All Rights Reserved |
---|
4 | * |
---|
5 | * Generic IO (ANL-15-066) |
---|
6 | * Hal Finkel, Argonne National Laboratory |
---|
7 | * |
---|
8 | * OPEN SOURCE LICENSE |
---|
9 | * |
---|
10 | * Under the terms of Contract No. DE-AC02-06CH11357 with UChicago Argonne, |
---|
11 | * LLC, the U.S. Government retains certain rights in this software. |
---|
12 | * |
---|
13 | * Redistribution and use in source and binary forms, with or without |
---|
14 | * modification, are permitted provided that the following conditions are met: |
---|
15 | * |
---|
16 | * 1. Redistributions of source code must retain the above copyright notice, |
---|
17 | * this list of conditions and the following disclaimer. |
---|
18 | * |
---|
19 | * 2. Redistributions in binary form must reproduce the above copyright |
---|
20 | * notice, this list of conditions and the following disclaimer in the |
---|
21 | * documentation and/or other materials provided with the distribution. |
---|
22 | * |
---|
23 | * 3. Neither the names of UChicago Argonne, LLC or the Department of Energy |
---|
24 | * nor the names of its contributors may be used to endorse or promote |
---|
25 | * products derived from this software without specific prior written |
---|
26 | * permission. |
---|
27 | * |
---|
28 | * ***************************************************************************** |
---|
29 | * |
---|
30 | * DISCLAIMER |
---|
31 | * THE SOFTWARE IS SUPPLIED “AS IS” WITHOUT WARRANTY OF ANY KIND. NEITHER THE |
---|
32 | * UNTED STATES GOVERNMENT, NOR THE UNITED STATES DEPARTMENT OF ENERGY, NOR |
---|
33 | * UCHICAGO ARGONNE, LLC, NOR ANY OF THEIR EMPLOYEES, MAKES ANY WARRANTY, |
---|
34 | * EXPRESS OR IMPLIED, OR ASSUMES ANY LEGAL LIABILITY OR RESPONSIBILITY FOR THE |
---|
35 | * ACCURACY, COMPLETENESS, OR USEFULNESS OF ANY INFORMATION, DATA, APPARATUS, |
---|
36 | * PRODUCT, OR PROCESS DISCLOSED, OR REPRESENTS THAT ITS USE WOULD NOT INFRINGE |
---|
37 | * PRIVATELY OWNED RIGHTS. |
---|
38 | * |
---|
39 | * ***************************************************************************** |
---|
40 | */ |
---|
41 | |
---|
42 | #define _XOPEN_SOURCE 600 |
---|
43 | #include "CRC64.h" |
---|
44 | #include "GenericIO.h" |
---|
45 | |
---|
46 | extern "C" { |
---|
47 | #include "blosc.h" |
---|
48 | } |
---|
49 | #include "sz.h" |
---|
50 | |
---|
51 | #include <sstream> |
---|
52 | #include <fstream> |
---|
53 | #include <stdexcept> |
---|
54 | #include <iterator> |
---|
55 | #include <algorithm> |
---|
56 | #include <cassert> |
---|
57 | #include <cstddef> |
---|
58 | #include <cstring> |
---|
59 | |
---|
60 | #ifndef GENERICIO_NO_MPI |
---|
61 | #include <ctime> |
---|
62 | #endif |
---|
63 | |
---|
64 | #include <sys/types.h> |
---|
65 | #include <sys/stat.h> |
---|
66 | #include <fcntl.h> |
---|
67 | #include <errno.h> |
---|
68 | |
---|
69 | #ifdef __bgq__ |
---|
70 | #include <mpix.h> |
---|
71 | #endif |
---|
72 | |
---|
73 | #ifndef MPI_UINT64_T |
---|
74 | #define MPI_UINT64_T (sizeof(long) == 8 ? MPI_LONG : MPI_LONG_LONG) |
---|
75 | #endif |
---|
76 | |
---|
77 | using namespace std; |
---|
78 | |
---|
79 | namespace gio { |
---|
80 | |
---|
81 | |
---|
82 | #ifndef GENERICIO_NO_MPI |
---|
83 | GenericFileIO_MPI::~GenericFileIO_MPI() { |
---|
84 | (void) MPI_File_close(&FH); |
---|
85 | } |
---|
86 | |
---|
87 | void GenericFileIO_MPI::open(const std::string &FN, bool ForReading) { |
---|
88 | FileName = FN; |
---|
89 | |
---|
90 | int amode = ForReading ? MPI_MODE_RDONLY : (MPI_MODE_WRONLY | MPI_MODE_CREATE); |
---|
91 | if (MPI_File_open(Comm, const_cast<char *>(FileName.c_str()), amode, |
---|
92 | MPI_INFO_NULL, &FH) != MPI_SUCCESS) |
---|
93 | throw runtime_error((!ForReading ? "Unable to create the file: " : |
---|
94 | "Unable to open the file: ") + |
---|
95 | FileName); |
---|
96 | } |
---|
97 | |
---|
98 | void GenericFileIO_MPI::setSize(size_t sz) { |
---|
99 | if (MPI_File_set_size(FH, sz) != MPI_SUCCESS) |
---|
100 | throw runtime_error("Unable to set size for file: " + FileName); |
---|
101 | } |
---|
102 | |
---|
103 | void GenericFileIO_MPI::read(void *buf, size_t count, off_t offset, |
---|
104 | const std::string &D) { |
---|
105 | while (count > 0) { |
---|
106 | MPI_Status status; |
---|
107 | if (MPI_File_read_at(FH, offset, buf, count, MPI_BYTE, &status) != MPI_SUCCESS) |
---|
108 | throw runtime_error("Unable to read " + D + " from file: " + FileName); |
---|
109 | |
---|
110 | int scount; |
---|
111 | (void) MPI_Get_count(&status, MPI_BYTE, &scount); |
---|
112 | |
---|
113 | count -= scount; |
---|
114 | buf = ((char *) buf) + scount; |
---|
115 | offset += scount; |
---|
116 | } |
---|
117 | } |
---|
118 | |
---|
119 | void GenericFileIO_MPI::write(const void *buf, size_t count, off_t offset, |
---|
120 | const std::string &D) { |
---|
121 | while (count > 0) { |
---|
122 | MPI_Status status; |
---|
123 | if (MPI_File_write_at(FH, offset, (void *) buf, count, MPI_BYTE, &status) != MPI_SUCCESS) |
---|
124 | throw runtime_error("Unable to write " + D + " to file: " + FileName); |
---|
125 | |
---|
126 | int scount = 0; |
---|
127 | // On some systems, MPI_Get_count will not return zero even when count is zero. |
---|
128 | if (count > 0) |
---|
129 | (void) MPI_Get_count(&status, MPI_BYTE, &scount); |
---|
130 | |
---|
131 | count -= scount; |
---|
132 | buf = ((char *) buf) + scount; |
---|
133 | offset += scount; |
---|
134 | } |
---|
135 | } |
---|
136 | |
---|
137 | void GenericFileIO_MPICollective::read(void *buf, size_t count, off_t offset, |
---|
138 | const std::string &D) { |
---|
139 | int Continue = 0; |
---|
140 | |
---|
141 | do { |
---|
142 | MPI_Status status; |
---|
143 | if (MPI_File_read_at_all(FH, offset, buf, count, MPI_BYTE, &status) != MPI_SUCCESS) |
---|
144 | throw runtime_error("Unable to read " + D + " from file: " + FileName); |
---|
145 | |
---|
146 | int scount = 0; |
---|
147 | // On some systems, MPI_Get_count will not return zero even when count is zero. |
---|
148 | if (count > 0) |
---|
149 | (void) MPI_Get_count(&status, MPI_BYTE, &scount); |
---|
150 | |
---|
151 | count -= scount; |
---|
152 | buf = ((char *) buf) + scount; |
---|
153 | offset += scount; |
---|
154 | |
---|
155 | int NeedContinue = (count > 0); |
---|
156 | MPI_Allreduce(&NeedContinue, &Continue, 1, MPI_INT, MPI_SUM, Comm); |
---|
157 | } while (Continue); |
---|
158 | } |
---|
159 | |
---|
160 | void GenericFileIO_MPICollective::write(const void *buf, size_t count, off_t offset, |
---|
161 | const std::string &D) { |
---|
162 | int Continue = 0; |
---|
163 | |
---|
164 | do { |
---|
165 | MPI_Status status; |
---|
166 | if (MPI_File_write_at_all(FH, offset, (void *) buf, count, MPI_BYTE, &status) != MPI_SUCCESS) |
---|
167 | throw runtime_error("Unable to write " + D + " to file: " + FileName); |
---|
168 | |
---|
169 | int scount; |
---|
170 | (void) MPI_Get_count(&status, MPI_BYTE, &scount); |
---|
171 | |
---|
172 | count -= scount; |
---|
173 | buf = ((char *) buf) + scount; |
---|
174 | offset += scount; |
---|
175 | |
---|
176 | int NeedContinue = (count > 0); |
---|
177 | MPI_Allreduce(&NeedContinue, &Continue, 1, MPI_INT, MPI_SUM, Comm); |
---|
178 | } while (Continue); |
---|
179 | } |
---|
180 | #endif |
---|
181 | |
---|
182 | GenericFileIO_POSIX::~GenericFileIO_POSIX() { |
---|
183 | if (FH != -1) close(FH); |
---|
184 | } |
---|
185 | |
---|
186 | void GenericFileIO_POSIX::open(const std::string &FN, bool ForReading) { |
---|
187 | FileName = FN; |
---|
188 | |
---|
189 | int flags = ForReading ? O_RDONLY : (O_WRONLY | O_CREAT); |
---|
190 | int mode = S_IRUSR | S_IWUSR | S_IRGRP; |
---|
191 | errno = 0; |
---|
192 | if ((FH = ::open(FileName.c_str(), flags, mode)) == -1) |
---|
193 | throw runtime_error((!ForReading ? "Unable to create the file: " : |
---|
194 | "Unable to open the file: ") + |
---|
195 | FileName + ": " + strerror(errno)); |
---|
196 | } |
---|
197 | |
---|
198 | void GenericFileIO_POSIX::setSize(size_t sz) { |
---|
199 | if (ftruncate(FH, sz) == -1) |
---|
200 | throw runtime_error("Unable to set size for file: " + FileName); |
---|
201 | } |
---|
202 | |
---|
203 | void GenericFileIO_POSIX::read(void *buf, size_t count, off_t offset, |
---|
204 | const std::string &D) { |
---|
205 | while (count > 0) { |
---|
206 | ssize_t scount; |
---|
207 | errno = 0; |
---|
208 | if ((scount = pread(FH, buf, count, offset)) == -1) { |
---|
209 | if (errno == EINTR) |
---|
210 | continue; |
---|
211 | |
---|
212 | throw runtime_error("Unable to read " + D + " from file: " + FileName + |
---|
213 | ": " + strerror(errno)); |
---|
214 | } |
---|
215 | |
---|
216 | count -= scount; |
---|
217 | buf = ((char *) buf) + scount; |
---|
218 | offset += scount; |
---|
219 | } |
---|
220 | } |
---|
221 | |
---|
222 | void GenericFileIO_POSIX::write(const void *buf, size_t count, off_t offset, |
---|
223 | const std::string &D) { |
---|
224 | while (count > 0) { |
---|
225 | ssize_t scount; |
---|
226 | errno = 0; |
---|
227 | if ((scount = pwrite(FH, buf, count, offset)) == -1) { |
---|
228 | if (errno == EINTR) |
---|
229 | continue; |
---|
230 | |
---|
231 | throw runtime_error("Unable to write " + D + " to file: " + FileName + |
---|
232 | ": " + strerror(errno)); |
---|
233 | } |
---|
234 | |
---|
235 | count -= scount; |
---|
236 | buf = ((char *) buf) + scount; |
---|
237 | offset += scount; |
---|
238 | } |
---|
239 | } |
---|
240 | |
---|
241 | static bool isBigEndian() { |
---|
242 | const uint32_t one = 1; |
---|
243 | return !(*((char *)(&one))); |
---|
244 | } |
---|
245 | |
---|
246 | static void bswap(void *v, size_t s) { |
---|
247 | char *p = (char *) v; |
---|
248 | for (size_t i = 0; i < s/2; ++i) |
---|
249 | std::swap(p[i], p[s - (i+1)]); |
---|
250 | } |
---|
251 | |
---|
252 | // Using #pragma pack here, instead of __attribute__((packed)) because xlc, at |
---|
253 | // least as of v12.1, won't take __attribute__((packed)) on non-POD and/or |
---|
254 | // templated types. |
---|
255 | #pragma pack(1) |
---|
256 | |
---|
257 | template <typename T, bool IsBigEndian> |
---|
258 | struct endian_specific_value { |
---|
259 | operator T() const { |
---|
260 | T rvalue = value; |
---|
261 | if (IsBigEndian != isBigEndian()) |
---|
262 | bswap(&rvalue, sizeof(T)); |
---|
263 | |
---|
264 | return rvalue; |
---|
265 | }; |
---|
266 | |
---|
267 | endian_specific_value &operator = (T nvalue) { |
---|
268 | if (IsBigEndian != isBigEndian()) |
---|
269 | bswap(&nvalue, sizeof(T)); |
---|
270 | |
---|
271 | value = nvalue; |
---|
272 | return *this; |
---|
273 | } |
---|
274 | |
---|
275 | endian_specific_value &operator += (T nvalue) { |
---|
276 | *this = *this + nvalue; |
---|
277 | return *this; |
---|
278 | } |
---|
279 | |
---|
280 | endian_specific_value &operator -= (T nvalue) { |
---|
281 | *this = *this - nvalue; |
---|
282 | return *this; |
---|
283 | } |
---|
284 | |
---|
285 | private: |
---|
286 | T value; |
---|
287 | }; |
---|
288 | |
---|
289 | static const size_t CRCSize = 8; |
---|
290 | |
---|
291 | static const size_t MagicSize = 8; |
---|
292 | static const char *MagicBE = "HACC01B"; |
---|
293 | static const char *MagicLE = "HACC01L"; |
---|
294 | |
---|
295 | template <bool IsBigEndian> |
---|
296 | struct GlobalHeader { |
---|
297 | char Magic[MagicSize]; |
---|
298 | endian_specific_value<uint64_t, IsBigEndian> HeaderSize; |
---|
299 | endian_specific_value<uint64_t, IsBigEndian> NElems; // The global total |
---|
300 | endian_specific_value<uint64_t, IsBigEndian> Dims[3]; |
---|
301 | endian_specific_value<uint64_t, IsBigEndian> NVars; |
---|
302 | endian_specific_value<uint64_t, IsBigEndian> VarsSize; |
---|
303 | endian_specific_value<uint64_t, IsBigEndian> VarsStart; |
---|
304 | endian_specific_value<uint64_t, IsBigEndian> NRanks; |
---|
305 | endian_specific_value<uint64_t, IsBigEndian> RanksSize; |
---|
306 | endian_specific_value<uint64_t, IsBigEndian> RanksStart; |
---|
307 | endian_specific_value<uint64_t, IsBigEndian> GlobalHeaderSize; |
---|
308 | endian_specific_value<double, IsBigEndian> PhysOrigin[3]; |
---|
309 | endian_specific_value<double, IsBigEndian> PhysScale[3]; |
---|
310 | endian_specific_value<uint64_t, IsBigEndian> BlocksSize; |
---|
311 | endian_specific_value<uint64_t, IsBigEndian> BlocksStart; |
---|
312 | }; |
---|
313 | |
---|
314 | enum { |
---|
315 | FloatValue = (1 << 0), |
---|
316 | SignedValue = (1 << 1), |
---|
317 | ValueIsPhysCoordX = (1 << 2), |
---|
318 | ValueIsPhysCoordY = (1 << 3), |
---|
319 | ValueIsPhysCoordZ = (1 << 4), |
---|
320 | ValueMaybePhysGhost = (1 << 5) |
---|
321 | }; |
---|
322 | |
---|
323 | static const size_t NameSize = 256; |
---|
324 | template <bool IsBigEndian> |
---|
325 | struct VariableHeader { |
---|
326 | char Name[NameSize]; |
---|
327 | endian_specific_value<uint64_t, IsBigEndian> Flags; |
---|
328 | endian_specific_value<uint64_t, IsBigEndian> Size; |
---|
329 | endian_specific_value<uint64_t, IsBigEndian> ElementSize; |
---|
330 | }; |
---|
331 | |
---|
332 | template <bool IsBigEndian> |
---|
333 | struct RankHeader { |
---|
334 | endian_specific_value<uint64_t, IsBigEndian> Coords[3]; |
---|
335 | endian_specific_value<uint64_t, IsBigEndian> NElems; |
---|
336 | endian_specific_value<uint64_t, IsBigEndian> Start; |
---|
337 | endian_specific_value<uint64_t, IsBigEndian> GlobalRank; |
---|
338 | }; |
---|
339 | |
---|
340 | static const size_t FilterNameSize = 8; |
---|
341 | static const size_t MaxFilters = 4; |
---|
342 | template <bool IsBigEndian> |
---|
343 | struct BlockHeader { |
---|
344 | char Filters[MaxFilters][FilterNameSize]; |
---|
345 | endian_specific_value<uint64_t, IsBigEndian> Start; |
---|
346 | endian_specific_value<uint64_t, IsBigEndian> Size; |
---|
347 | }; |
---|
348 | |
---|
349 | template <bool IsBigEndian> |
---|
350 | struct CompressHeader { |
---|
351 | endian_specific_value<uint64_t, IsBigEndian> OrigCRC; |
---|
352 | }; |
---|
353 | const char *CompressName = "BLOSC"; |
---|
354 | |
---|
355 | const char *LossyCompressName = "SZ"; |
---|
356 | |
---|
357 | #pragma pack() |
---|
358 | |
---|
359 | unsigned GenericIO::DefaultFileIOType = FileIOPOSIX; |
---|
360 | int GenericIO::DefaultPartition = 0; |
---|
361 | bool GenericIO::DefaultShouldCompress = false; |
---|
362 | |
---|
363 | #ifndef GENERICIO_NO_MPI |
---|
364 | std::size_t GenericIO::CollectiveMPIIOThreshold = 0; |
---|
365 | #endif |
---|
366 | |
---|
367 | static bool blosc_initialized = false; |
---|
368 | static bool sz_initialized = false; |
---|
369 | |
---|
370 | static int GetSZDT(GenericIO::Variable &Var) { |
---|
371 | if (Var.hasElementType<float>()) |
---|
372 | return SZ_FLOAT; |
---|
373 | else if (Var.hasElementType<double>()) |
---|
374 | return SZ_DOUBLE; |
---|
375 | else if (Var.hasElementType<uint8_t>()) |
---|
376 | return SZ_UINT8; |
---|
377 | else if (Var.hasElementType<int8_t>()) |
---|
378 | return SZ_INT8; |
---|
379 | else if (Var.hasElementType<uint16_t>()) |
---|
380 | return SZ_UINT16; |
---|
381 | else if (Var.hasElementType<int16_t>()) |
---|
382 | return SZ_INT16; |
---|
383 | else if (Var.hasElementType<uint32_t>()) |
---|
384 | return SZ_UINT32; |
---|
385 | else if (Var.hasElementType<int32_t>()) |
---|
386 | return SZ_INT32; |
---|
387 | else if (Var.hasElementType<uint64_t>()) |
---|
388 | return SZ_UINT64; |
---|
389 | else if (Var.hasElementType<int64_t>()) |
---|
390 | return SZ_INT64; |
---|
391 | else |
---|
392 | return -1; |
---|
393 | } |
---|
394 | |
---|
395 | #ifndef GENERICIO_NO_MPI |
---|
396 | void GenericIO::write() { |
---|
397 | if (isBigEndian()) |
---|
398 | write<true>(); |
---|
399 | else |
---|
400 | write<false>(); |
---|
401 | } |
---|
402 | |
---|
403 | // Note: writing errors are not currently recoverable (one rank may fail |
---|
404 | // while the others don't). |
---|
405 | template <bool IsBigEndian> |
---|
406 | void GenericIO::write() { |
---|
407 | const char *Magic = IsBigEndian ? MagicBE : MagicLE; |
---|
408 | |
---|
409 | uint64_t FileSize = 0; |
---|
410 | |
---|
411 | int NRanks, Rank; |
---|
412 | MPI_Comm_rank(Comm, &Rank); |
---|
413 | MPI_Comm_size(Comm, &NRanks); |
---|
414 | |
---|
415 | #ifdef __bgq__ |
---|
416 | MPI_Barrier(Comm); |
---|
417 | #endif |
---|
418 | MPI_Comm_split(Comm, Partition, Rank, &SplitComm); |
---|
419 | |
---|
420 | int SplitNRanks, SplitRank; |
---|
421 | MPI_Comm_rank(SplitComm, &SplitRank); |
---|
422 | MPI_Comm_size(SplitComm, &SplitNRanks); |
---|
423 | |
---|
424 | string LocalFileName; |
---|
425 | if (SplitNRanks != NRanks) { |
---|
426 | if (Rank == 0) { |
---|
427 | // In split mode, the specified file becomes the rank map, and the real |
---|
428 | // data is partitioned. |
---|
429 | |
---|
430 | vector<int> MapRank, MapPartition; |
---|
431 | MapRank.resize(NRanks); |
---|
432 | for (int i = 0; i < NRanks; ++i) MapRank[i] = i; |
---|
433 | |
---|
434 | MapPartition.resize(NRanks); |
---|
435 | MPI_Gather(&Partition, 1, MPI_INT, &MapPartition[0], 1, MPI_INT, 0, Comm); |
---|
436 | |
---|
437 | GenericIO GIO(MPI_COMM_SELF, FileName, FileIOType); |
---|
438 | GIO.setNumElems(NRanks); |
---|
439 | GIO.addVariable("$rank", MapRank); /* this is for use by humans; the reading |
---|
440 | code assumes that the partitions are in |
---|
441 | rank order */ |
---|
442 | GIO.addVariable("$partition", MapPartition); |
---|
443 | |
---|
444 | vector<int> CX, CY, CZ; |
---|
445 | int TopoStatus; |
---|
446 | MPI_Topo_test(Comm, &TopoStatus); |
---|
447 | if (TopoStatus == MPI_CART) { |
---|
448 | CX.resize(NRanks); |
---|
449 | CY.resize(NRanks); |
---|
450 | CZ.resize(NRanks); |
---|
451 | |
---|
452 | for (int i = 0; i < NRanks; ++i) { |
---|
453 | int C[3]; |
---|
454 | MPI_Cart_coords(Comm, i, 3, C); |
---|
455 | |
---|
456 | CX[i] = C[0]; |
---|
457 | CY[i] = C[1]; |
---|
458 | CZ[i] = C[2]; |
---|
459 | } |
---|
460 | |
---|
461 | GIO.addVariable("$x", CX); |
---|
462 | GIO.addVariable("$y", CY); |
---|
463 | GIO.addVariable("$z", CZ); |
---|
464 | } |
---|
465 | |
---|
466 | GIO.write(); |
---|
467 | } else { |
---|
468 | MPI_Gather(&Partition, 1, MPI_INT, 0, 0, MPI_INT, 0, Comm); |
---|
469 | } |
---|
470 | |
---|
471 | stringstream ss; |
---|
472 | ss << FileName << "#" << Partition; |
---|
473 | LocalFileName = ss.str(); |
---|
474 | } else { |
---|
475 | LocalFileName = FileName; |
---|
476 | } |
---|
477 | |
---|
478 | RankHeader<IsBigEndian> RHLocal; |
---|
479 | int Dims[3], Periods[3], Coords[3]; |
---|
480 | |
---|
481 | int TopoStatus; |
---|
482 | MPI_Topo_test(Comm, &TopoStatus); |
---|
483 | if (TopoStatus == MPI_CART) { |
---|
484 | MPI_Cart_get(Comm, 3, Dims, Periods, Coords); |
---|
485 | } else { |
---|
486 | Dims[0] = NRanks; |
---|
487 | std::fill(Dims + 1, Dims + 3, 1); |
---|
488 | std::fill(Periods, Periods + 3, 0); |
---|
489 | Coords[0] = Rank; |
---|
490 | std::fill(Coords + 1, Coords + 3, 0); |
---|
491 | } |
---|
492 | |
---|
493 | std::copy(Coords, Coords + 3, RHLocal.Coords); |
---|
494 | RHLocal.NElems = NElems; |
---|
495 | RHLocal.Start = 0; |
---|
496 | RHLocal.GlobalRank = Rank; |
---|
497 | |
---|
498 | bool ShouldCompress = DefaultShouldCompress; |
---|
499 | const char *EnvStr = getenv("GENERICIO_COMPRESS"); |
---|
500 | if (EnvStr) { |
---|
501 | int Mod = atoi(EnvStr); |
---|
502 | ShouldCompress = (Mod > 0); |
---|
503 | } |
---|
504 | |
---|
505 | bool NeedsBlockHeaders = ShouldCompress; |
---|
506 | EnvStr = getenv("GENERICIO_FORCE_BLOCKS"); |
---|
507 | if (!NeedsBlockHeaders && EnvStr) { |
---|
508 | int Mod = atoi(EnvStr); |
---|
509 | NeedsBlockHeaders = (Mod > 0); |
---|
510 | } |
---|
511 | |
---|
512 | vector<BlockHeader<IsBigEndian> > LocalBlockHeaders; |
---|
513 | vector<void *> LocalData; |
---|
514 | vector<bool> LocalHasExtraSpace; |
---|
515 | vector<vector<unsigned char> > LocalCData; |
---|
516 | if (NeedsBlockHeaders) { |
---|
517 | LocalBlockHeaders.resize(Vars.size()); |
---|
518 | LocalData.resize(Vars.size()); |
---|
519 | LocalHasExtraSpace.resize(Vars.size()); |
---|
520 | if (ShouldCompress) |
---|
521 | LocalCData.resize(Vars.size()); |
---|
522 | |
---|
523 | for (size_t i = 0; i < Vars.size(); ++i) { |
---|
524 | // Filters null by default, leave null starting address (needs to be |
---|
525 | // calculated by the header-writing rank). |
---|
526 | memset(&LocalBlockHeaders[i], 0, sizeof(BlockHeader<IsBigEndian>)); |
---|
527 | if (ShouldCompress) { |
---|
528 | void *OrigData = Vars[i].Data; |
---|
529 | bool FreeOrigData = false; |
---|
530 | size_t OrigUnitSize = Vars[i].Size; |
---|
531 | size_t OrigDataSize = NElems*Vars[i].Size; |
---|
532 | |
---|
533 | int FilterIdx = 0; |
---|
534 | if (Vars[i].LCI.Mode != LossyCompressionInfo::LCModeNone) { |
---|
535 | #ifdef _OPENMP |
---|
536 | #pragma omp master |
---|
537 | { |
---|
538 | #endif |
---|
539 | if (!sz_initialized) { |
---|
540 | SZ_Init(NULL); |
---|
541 | sz_initialized = true; |
---|
542 | } |
---|
543 | |
---|
544 | #ifdef _OPENMP |
---|
545 | } |
---|
546 | #endif |
---|
547 | int SZDT = GetSZDT(Vars[i]); |
---|
548 | if (SZDT == -1) |
---|
549 | goto nosz; |
---|
550 | |
---|
551 | int EBM; |
---|
552 | switch (Vars[i].LCI.Mode) { |
---|
553 | case LossyCompressionInfo::LCModeAbs: |
---|
554 | EBM = ABS; |
---|
555 | break; |
---|
556 | case LossyCompressionInfo::LCModeRel: |
---|
557 | EBM = REL; |
---|
558 | break; |
---|
559 | case LossyCompressionInfo::LCModeAbsAndRel: |
---|
560 | EBM = ABS_AND_REL; |
---|
561 | break; |
---|
562 | case LossyCompressionInfo::LCModeAbsOrRel: |
---|
563 | EBM = ABS_OR_REL; |
---|
564 | break; |
---|
565 | case LossyCompressionInfo::LCModePSNR: |
---|
566 | EBM = PSNR; |
---|
567 | break; |
---|
568 | } |
---|
569 | |
---|
570 | size_t LOutSize; |
---|
571 | unsigned char *LCompressedData = SZ_compress_args(SZDT, Vars[i].Data, &LOutSize, EBM, |
---|
572 | Vars[i].LCI.AbsErrThreshold, Vars[i].LCI.RelErrThreshold, |
---|
573 | Vars[i].LCI.PSNRThreshold, 0, 0, 0, 0, NElems); |
---|
574 | if (!LCompressedData) |
---|
575 | goto nosz; |
---|
576 | if (LOutSize >= NElems*Vars[i].Size) { |
---|
577 | free(LCompressedData); |
---|
578 | goto nosz; |
---|
579 | } |
---|
580 | |
---|
581 | OrigData = LCompressedData; |
---|
582 | FreeOrigData = true; |
---|
583 | OrigUnitSize = 1; |
---|
584 | OrigDataSize = LOutSize; |
---|
585 | |
---|
586 | strncpy(LocalBlockHeaders[i].Filters[FilterIdx++], LossyCompressName, FilterNameSize); |
---|
587 | } |
---|
588 | nosz: |
---|
589 | |
---|
590 | LocalCData[i].resize(sizeof(CompressHeader<IsBigEndian>)); |
---|
591 | |
---|
592 | CompressHeader<IsBigEndian> *CH = (CompressHeader<IsBigEndian>*) &LocalCData[i][0]; |
---|
593 | CH->OrigCRC = crc64_omp(OrigData, OrigDataSize); |
---|
594 | |
---|
595 | #ifdef _OPENMP |
---|
596 | #pragma omp master |
---|
597 | { |
---|
598 | #endif |
---|
599 | |
---|
600 | if (!blosc_initialized) { |
---|
601 | blosc_init(); |
---|
602 | blosc_initialized = true; |
---|
603 | } |
---|
604 | |
---|
605 | #ifdef _OPENMP |
---|
606 | blosc_set_nthreads(omp_get_max_threads()); |
---|
607 | } |
---|
608 | #endif |
---|
609 | |
---|
610 | LocalCData[i].resize(LocalCData[i].size() + OrigDataSize); |
---|
611 | if (blosc_compress(9, 1, OrigUnitSize, OrigDataSize, OrigData, |
---|
612 | &LocalCData[i][0] + sizeof(CompressHeader<IsBigEndian>), |
---|
613 | OrigDataSize) <= 0) { |
---|
614 | if (FreeOrigData) |
---|
615 | free(OrigData); |
---|
616 | |
---|
617 | goto nocomp; |
---|
618 | } |
---|
619 | |
---|
620 | if (FreeOrigData) |
---|
621 | free(OrigData); |
---|
622 | |
---|
623 | strncpy(LocalBlockHeaders[i].Filters[FilterIdx++], CompressName, FilterNameSize); |
---|
624 | size_t CNBytes, CCBytes, CBlockSize; |
---|
625 | blosc_cbuffer_sizes(&LocalCData[i][0] + sizeof(CompressHeader<IsBigEndian>), |
---|
626 | &CNBytes, &CCBytes, &CBlockSize); |
---|
627 | LocalCData[i].resize(CCBytes + sizeof(CompressHeader<IsBigEndian>)); |
---|
628 | |
---|
629 | LocalBlockHeaders[i].Size = LocalCData[i].size(); |
---|
630 | LocalCData[i].resize(LocalCData[i].size() + CRCSize); |
---|
631 | LocalData[i] = &LocalCData[i][0]; |
---|
632 | LocalHasExtraSpace[i] = true; |
---|
633 | } else { |
---|
634 | nocomp: |
---|
635 | LocalBlockHeaders[i].Size = NElems*Vars[i].Size; |
---|
636 | LocalData[i] = Vars[i].Data; |
---|
637 | LocalHasExtraSpace[i] = Vars[i].HasExtraSpace; |
---|
638 | } |
---|
639 | } |
---|
640 | } |
---|
641 | |
---|
642 | double StartTime = MPI_Wtime(); |
---|
643 | |
---|
644 | if (SplitRank == 0) { |
---|
645 | uint64_t HeaderSize = sizeof(GlobalHeader<IsBigEndian>) + Vars.size()*sizeof(VariableHeader<IsBigEndian>) + |
---|
646 | SplitNRanks*sizeof(RankHeader<IsBigEndian>) + CRCSize; |
---|
647 | if (NeedsBlockHeaders) |
---|
648 | HeaderSize += SplitNRanks*Vars.size()*sizeof(BlockHeader<IsBigEndian>); |
---|
649 | |
---|
650 | vector<char> Header(HeaderSize, 0); |
---|
651 | GlobalHeader<IsBigEndian> *GH = (GlobalHeader<IsBigEndian> *) &Header[0]; |
---|
652 | std::copy(Magic, Magic + MagicSize, GH->Magic); |
---|
653 | GH->HeaderSize = HeaderSize - CRCSize; |
---|
654 | GH->NElems = NElems; // This will be updated later |
---|
655 | std::copy(Dims, Dims + 3, GH->Dims); |
---|
656 | GH->NVars = Vars.size(); |
---|
657 | GH->VarsSize = sizeof(VariableHeader<IsBigEndian>); |
---|
658 | GH->VarsStart = sizeof(GlobalHeader<IsBigEndian>); |
---|
659 | GH->NRanks = SplitNRanks; |
---|
660 | GH->RanksSize = sizeof(RankHeader<IsBigEndian>); |
---|
661 | GH->RanksStart = GH->VarsStart + Vars.size()*sizeof(VariableHeader<IsBigEndian>); |
---|
662 | GH->GlobalHeaderSize = sizeof(GlobalHeader<IsBigEndian>); |
---|
663 | std::copy(PhysOrigin, PhysOrigin + 3, GH->PhysOrigin); |
---|
664 | std::copy(PhysScale, PhysScale + 3, GH->PhysScale); |
---|
665 | if (!NeedsBlockHeaders) { |
---|
666 | GH->BlocksSize = GH->BlocksStart = 0; |
---|
667 | } else { |
---|
668 | GH->BlocksSize = sizeof(BlockHeader<IsBigEndian>); |
---|
669 | GH->BlocksStart = GH->RanksStart + SplitNRanks*sizeof(RankHeader<IsBigEndian>); |
---|
670 | } |
---|
671 | |
---|
672 | uint64_t RecordSize = 0; |
---|
673 | VariableHeader<IsBigEndian> *VH = (VariableHeader<IsBigEndian> *) &Header[GH->VarsStart]; |
---|
674 | for (size_t i = 0; i < Vars.size(); ++i, ++VH) { |
---|
675 | string VName(Vars[i].Name); |
---|
676 | VName.resize(NameSize); |
---|
677 | |
---|
678 | std::copy(VName.begin(), VName.end(), VH->Name); |
---|
679 | uint64_t VFlags = 0; |
---|
680 | if (Vars[i].IsFloat) VFlags |= FloatValue; |
---|
681 | if (Vars[i].IsSigned) VFlags |= SignedValue; |
---|
682 | if (Vars[i].IsPhysCoordX) VFlags |= ValueIsPhysCoordX; |
---|
683 | if (Vars[i].IsPhysCoordY) VFlags |= ValueIsPhysCoordY; |
---|
684 | if (Vars[i].IsPhysCoordZ) VFlags |= ValueIsPhysCoordZ; |
---|
685 | if (Vars[i].MaybePhysGhost) VFlags |= ValueMaybePhysGhost; |
---|
686 | VH->Flags = VFlags; |
---|
687 | RecordSize += VH->Size = Vars[i].Size; |
---|
688 | VH->ElementSize = Vars[i].ElementSize; |
---|
689 | } |
---|
690 | |
---|
691 | MPI_Gather(&RHLocal, sizeof(RHLocal), MPI_BYTE, |
---|
692 | &Header[GH->RanksStart], sizeof(RHLocal), |
---|
693 | MPI_BYTE, 0, SplitComm); |
---|
694 | |
---|
695 | if (NeedsBlockHeaders) { |
---|
696 | MPI_Gather(&LocalBlockHeaders[0], |
---|
697 | Vars.size()*sizeof(BlockHeader<IsBigEndian>), MPI_BYTE, |
---|
698 | &Header[GH->BlocksStart], |
---|
699 | Vars.size()*sizeof(BlockHeader<IsBigEndian>), MPI_BYTE, |
---|
700 | 0, SplitComm); |
---|
701 | |
---|
702 | BlockHeader<IsBigEndian> *BH = (BlockHeader<IsBigEndian> *) &Header[GH->BlocksStart]; |
---|
703 | for (int i = 0; i < SplitNRanks; ++i) |
---|
704 | for (size_t j = 0; j < Vars.size(); ++j, ++BH) { |
---|
705 | if (i == 0 && j == 0) |
---|
706 | BH->Start = HeaderSize; |
---|
707 | else |
---|
708 | BH->Start = BH[-1].Start + BH[-1].Size + CRCSize; |
---|
709 | } |
---|
710 | |
---|
711 | RankHeader<IsBigEndian> *RH = (RankHeader<IsBigEndian> *) &Header[GH->RanksStart]; |
---|
712 | RH->Start = HeaderSize; ++RH; |
---|
713 | for (int i = 1; i < SplitNRanks; ++i, ++RH) { |
---|
714 | RH->Start = |
---|
715 | ((BlockHeader<IsBigEndian> *) &Header[GH->BlocksStart])[i*Vars.size()].Start; |
---|
716 | GH->NElems += RH->NElems; |
---|
717 | } |
---|
718 | |
---|
719 | // Compute the total file size. |
---|
720 | uint64_t LastData = BH[-1].Size + CRCSize; |
---|
721 | FileSize = BH[-1].Start + LastData; |
---|
722 | } else { |
---|
723 | RankHeader<IsBigEndian> *RH = (RankHeader<IsBigEndian> *) &Header[GH->RanksStart]; |
---|
724 | RH->Start = HeaderSize; ++RH; |
---|
725 | for (int i = 1; i < SplitNRanks; ++i, ++RH) { |
---|
726 | uint64_t PrevNElems = RH[-1].NElems; |
---|
727 | uint64_t PrevData = PrevNElems*RecordSize + CRCSize*Vars.size(); |
---|
728 | RH->Start = RH[-1].Start + PrevData; |
---|
729 | GH->NElems += RH->NElems; |
---|
730 | } |
---|
731 | |
---|
732 | // Compute the total file size. |
---|
733 | uint64_t LastNElems = RH[-1].NElems; |
---|
734 | uint64_t LastData = LastNElems*RecordSize + CRCSize*Vars.size(); |
---|
735 | FileSize = RH[-1].Start + LastData; |
---|
736 | } |
---|
737 | |
---|
738 | // Now that the starting offset has been computed, send it back to each rank. |
---|
739 | MPI_Scatter(&Header[GH->RanksStart], sizeof(RHLocal), |
---|
740 | MPI_BYTE, &RHLocal, sizeof(RHLocal), |
---|
741 | MPI_BYTE, 0, SplitComm); |
---|
742 | |
---|
743 | if (NeedsBlockHeaders) |
---|
744 | MPI_Scatter(&Header[GH->BlocksStart], |
---|
745 | sizeof(BlockHeader<IsBigEndian>)*Vars.size(), MPI_BYTE, |
---|
746 | &LocalBlockHeaders[0], |
---|
747 | sizeof(BlockHeader<IsBigEndian>)*Vars.size(), MPI_BYTE, |
---|
748 | 0, SplitComm); |
---|
749 | |
---|
750 | uint64_t HeaderCRC = crc64_omp(&Header[0], HeaderSize - CRCSize); |
---|
751 | crc64_invert(HeaderCRC, &Header[HeaderSize - CRCSize]); |
---|
752 | |
---|
753 | if (FileIOType == FileIOMPI) |
---|
754 | FH.get() = new GenericFileIO_MPI(MPI_COMM_SELF); |
---|
755 | else if (FileIOType == FileIOMPICollective) |
---|
756 | FH.get() = new GenericFileIO_MPICollective(MPI_COMM_SELF); |
---|
757 | else |
---|
758 | FH.get() = new GenericFileIO_POSIX(); |
---|
759 | |
---|
760 | FH.get()->open(LocalFileName); |
---|
761 | FH.get()->setSize(FileSize); |
---|
762 | FH.get()->write(&Header[0], HeaderSize, 0, "header"); |
---|
763 | |
---|
764 | close(); |
---|
765 | } else { |
---|
766 | MPI_Gather(&RHLocal, sizeof(RHLocal), MPI_BYTE, 0, 0, MPI_BYTE, 0, SplitComm); |
---|
767 | if (NeedsBlockHeaders) |
---|
768 | MPI_Gather(&LocalBlockHeaders[0], Vars.size()*sizeof(BlockHeader<IsBigEndian>), |
---|
769 | MPI_BYTE, 0, 0, MPI_BYTE, 0, SplitComm); |
---|
770 | MPI_Scatter(0, 0, MPI_BYTE, &RHLocal, sizeof(RHLocal), MPI_BYTE, 0, SplitComm); |
---|
771 | if (NeedsBlockHeaders) |
---|
772 | MPI_Scatter(0, 0, MPI_BYTE, &LocalBlockHeaders[0], sizeof(BlockHeader<IsBigEndian>)*Vars.size(), |
---|
773 | MPI_BYTE, 0, SplitComm); |
---|
774 | } |
---|
775 | |
---|
776 | MPI_Barrier(SplitComm); |
---|
777 | |
---|
778 | if (FileIOType == FileIOMPI) |
---|
779 | FH.get() = new GenericFileIO_MPI(SplitComm); |
---|
780 | else if (FileIOType == FileIOMPICollective) |
---|
781 | FH.get() = new GenericFileIO_MPICollective(SplitComm); |
---|
782 | else |
---|
783 | FH.get() = new GenericFileIO_POSIX(); |
---|
784 | |
---|
785 | FH.get()->open(LocalFileName); |
---|
786 | |
---|
787 | uint64_t Offset = RHLocal.Start; |
---|
788 | for (size_t i = 0; i < Vars.size(); ++i) { |
---|
789 | uint64_t WriteSize = NeedsBlockHeaders ? |
---|
790 | LocalBlockHeaders[i].Size : NElems*Vars[i].Size; |
---|
791 | void *Data = NeedsBlockHeaders ? LocalData[i] : Vars[i].Data; |
---|
792 | uint64_t CRC = crc64_omp(Data, WriteSize); |
---|
793 | bool HasExtraSpace = NeedsBlockHeaders ? |
---|
794 | LocalHasExtraSpace[i] : Vars[i].HasExtraSpace; |
---|
795 | char *CRCLoc = HasExtraSpace ? ((char *) Data) + WriteSize : (char *) &CRC; |
---|
796 | |
---|
797 | if (NeedsBlockHeaders) |
---|
798 | Offset = LocalBlockHeaders[i].Start; |
---|
799 | |
---|
800 | // When using extra space for the CRC write, preserve the original contents. |
---|
801 | char CRCSave[CRCSize]; |
---|
802 | if (HasExtraSpace) |
---|
803 | std::copy(CRCLoc, CRCLoc + CRCSize, CRCSave); |
---|
804 | |
---|
805 | crc64_invert(CRC, CRCLoc); |
---|
806 | |
---|
807 | if (HasExtraSpace) { |
---|
808 | FH.get()->write(Data, WriteSize + CRCSize, Offset, Vars[i].Name + " with CRC"); |
---|
809 | } else { |
---|
810 | FH.get()->write(Data, WriteSize, Offset, Vars[i].Name); |
---|
811 | FH.get()->write(CRCLoc, CRCSize, Offset + WriteSize, Vars[i].Name + " CRC"); |
---|
812 | } |
---|
813 | |
---|
814 | if (HasExtraSpace) |
---|
815 | std::copy(CRCSave, CRCSave + CRCSize, CRCLoc); |
---|
816 | |
---|
817 | Offset += WriteSize + CRCSize; |
---|
818 | } |
---|
819 | |
---|
820 | close(); |
---|
821 | MPI_Barrier(Comm); |
---|
822 | |
---|
823 | double EndTime = MPI_Wtime(); |
---|
824 | double TotalTime = EndTime - StartTime; |
---|
825 | double MaxTotalTime; |
---|
826 | MPI_Reduce(&TotalTime, &MaxTotalTime, 1, MPI_DOUBLE, MPI_MAX, 0, Comm); |
---|
827 | |
---|
828 | if (SplitNRanks != NRanks) { |
---|
829 | uint64_t ContribFileSize = (SplitRank == 0) ? FileSize : 0; |
---|
830 | MPI_Reduce(&ContribFileSize, &FileSize, 1, MPI_UINT64_T, MPI_SUM, 0, Comm); |
---|
831 | } |
---|
832 | |
---|
833 | if (Rank == 0) { |
---|
834 | double Rate = ((double) FileSize) / MaxTotalTime / (1024.*1024.); |
---|
835 | std::cout << "Wrote " << Vars.size() << " variables to " << FileName << |
---|
836 | " (" << FileSize << " bytes) in " << MaxTotalTime << "s: " << |
---|
837 | Rate << " MB/s" << std::endl; |
---|
838 | } |
---|
839 | |
---|
840 | MPI_Comm_free(&SplitComm); |
---|
841 | SplitComm = MPI_COMM_NULL; |
---|
842 | } |
---|
843 | #endif // GENERICIO_NO_MPI |
---|
844 | |
---|
845 | template <bool IsBigEndian> |
---|
846 | void GenericIO::readHeaderLeader(void *GHPtr, MismatchBehavior MB, int NRanks, |
---|
847 | int Rank, int SplitNRanks, |
---|
848 | string &LocalFileName, uint64_t &HeaderSize, |
---|
849 | vector<char> &Header) { |
---|
850 | GlobalHeader<IsBigEndian> &GH = *(GlobalHeader<IsBigEndian> *) GHPtr; |
---|
851 | |
---|
852 | if (MB == MismatchDisallowed) { |
---|
853 | if (SplitNRanks != (int) GH.NRanks) { |
---|
854 | stringstream ss; |
---|
855 | ss << "Won't read " << LocalFileName << ": communicator-size mismatch: " << |
---|
856 | "current: " << SplitNRanks << ", file: " << GH.NRanks; |
---|
857 | throw runtime_error(ss.str()); |
---|
858 | } |
---|
859 | |
---|
860 | #ifndef GENERICIO_NO_MPI |
---|
861 | int TopoStatus; |
---|
862 | MPI_Topo_test(Comm, &TopoStatus); |
---|
863 | if (TopoStatus == MPI_CART) { |
---|
864 | int Dims[3], Periods[3], Coords[3]; |
---|
865 | MPI_Cart_get(Comm, 3, Dims, Periods, Coords); |
---|
866 | |
---|
867 | bool DimsMatch = true; |
---|
868 | for (int i = 0; i < 3; ++i) { |
---|
869 | if ((uint64_t) Dims[i] != GH.Dims[i]) { |
---|
870 | DimsMatch = false; |
---|
871 | break; |
---|
872 | } |
---|
873 | } |
---|
874 | |
---|
875 | if (!DimsMatch) { |
---|
876 | stringstream ss; |
---|
877 | ss << "Won't read " << LocalFileName << |
---|
878 | ": communicator-decomposition mismatch: " << |
---|
879 | "current: " << Dims[0] << "x" << Dims[1] << "x" << Dims[2] << |
---|
880 | ", file: " << GH.Dims[0] << "x" << GH.Dims[1] << "x" << |
---|
881 | GH.Dims[2]; |
---|
882 | throw runtime_error(ss.str()); |
---|
883 | } |
---|
884 | } |
---|
885 | #endif |
---|
886 | } else if (MB == MismatchRedistribute && !Redistributing) { |
---|
887 | Redistributing = true; |
---|
888 | |
---|
889 | int NFileRanks = RankMap.empty() ? (int) GH.NRanks : (int) RankMap.size(); |
---|
890 | int NFileRanksPerRank = NFileRanks/NRanks; |
---|
891 | int NRemFileRank = NFileRanks % NRanks; |
---|
892 | |
---|
893 | if (!NFileRanksPerRank) { |
---|
894 | // We have only the remainder, so the last NRemFileRank ranks get one |
---|
895 | // file rank, and the others don't. |
---|
896 | if (NRemFileRank && NRanks - Rank <= NRemFileRank) |
---|
897 | SourceRanks.push_back(NRanks - (Rank + 1)); |
---|
898 | } else { |
---|
899 | // Since NRemFileRank < NRanks, and we don't want to put any extra memory |
---|
900 | // load on rank 0 (because rank 0's memory load is normally higher than |
---|
901 | // the other ranks anyway), the last NRemFileRank will each take |
---|
902 | // (NFileRanksPerRank+1) file ranks. |
---|
903 | |
---|
904 | int FirstFileRank = 0, LastFileRank = NFileRanksPerRank - 1; |
---|
905 | for (int i = 1; i <= Rank; ++i) { |
---|
906 | FirstFileRank = LastFileRank + 1; |
---|
907 | LastFileRank = FirstFileRank + NFileRanksPerRank - 1; |
---|
908 | |
---|
909 | if (NRemFileRank && NRanks - i <= NRemFileRank) |
---|
910 | ++LastFileRank; |
---|
911 | } |
---|
912 | |
---|
913 | for (int i = FirstFileRank; i <= LastFileRank; ++i) |
---|
914 | SourceRanks.push_back(i); |
---|
915 | } |
---|
916 | } |
---|
917 | |
---|
918 | HeaderSize = GH.HeaderSize; |
---|
919 | Header.resize(HeaderSize + CRCSize, 0xFE /* poison */); |
---|
920 | FH.get()->read(&Header[0], HeaderSize + CRCSize, 0, "header"); |
---|
921 | |
---|
922 | uint64_t CRC = crc64_omp(&Header[0], HeaderSize + CRCSize); |
---|
923 | if (CRC != (uint64_t) -1) { |
---|
924 | throw runtime_error("Header CRC check failed: " + LocalFileName); |
---|
925 | } |
---|
926 | } |
---|
927 | |
---|
928 | // Note: Errors from this function should be recoverable. This means that if |
---|
929 | // one rank throws an exception, then all ranks should. |
---|
930 | void GenericIO::openAndReadHeader(MismatchBehavior MB, int EffRank, bool CheckPartMap) { |
---|
931 | int NRanks, Rank; |
---|
932 | #ifndef GENERICIO_NO_MPI |
---|
933 | MPI_Comm_rank(Comm, &Rank); |
---|
934 | MPI_Comm_size(Comm, &NRanks); |
---|
935 | #else |
---|
936 | Rank = 0; |
---|
937 | NRanks = 1; |
---|
938 | #endif |
---|
939 | |
---|
940 | if (EffRank == -1) |
---|
941 | EffRank = MB == MismatchRedistribute ? 0 : Rank; |
---|
942 | |
---|
943 | if (RankMap.empty() && CheckPartMap) { |
---|
944 | // First, check to see if the file is a rank map. |
---|
945 | unsigned long RanksInMap = 0; |
---|
946 | if (Rank == 0) { |
---|
947 | try { |
---|
948 | #ifndef GENERICIO_NO_MPI |
---|
949 | GenericIO GIO(MPI_COMM_SELF, FileName, FileIOType); |
---|
950 | #else |
---|
951 | GenericIO GIO(FileName, FileIOType); |
---|
952 | #endif |
---|
953 | GIO.openAndReadHeader(MismatchDisallowed, 0, false); |
---|
954 | RanksInMap = GIO.readNumElems(); |
---|
955 | |
---|
956 | RankMap.resize(RanksInMap + GIO.requestedExtraSpace()/sizeof(int)); |
---|
957 | GIO.addVariable("$partition", RankMap, true); |
---|
958 | |
---|
959 | GIO.readData(0, false); |
---|
960 | RankMap.resize(RanksInMap); |
---|
961 | } catch (...) { |
---|
962 | RankMap.clear(); |
---|
963 | RanksInMap = 0; |
---|
964 | } |
---|
965 | } |
---|
966 | |
---|
967 | #ifndef GENERICIO_NO_MPI |
---|
968 | MPI_Bcast(&RanksInMap, 1, MPI_UNSIGNED_LONG, 0, Comm); |
---|
969 | if (RanksInMap > 0) { |
---|
970 | RankMap.resize(RanksInMap); |
---|
971 | MPI_Bcast(&RankMap[0], RanksInMap, MPI_INT, 0, Comm); |
---|
972 | } |
---|
973 | #endif |
---|
974 | } |
---|
975 | |
---|
976 | #ifndef GENERICIO_NO_MPI |
---|
977 | if (SplitComm != MPI_COMM_NULL) |
---|
978 | MPI_Comm_free(&SplitComm); |
---|
979 | #endif |
---|
980 | |
---|
981 | string LocalFileName; |
---|
982 | if (RankMap.empty()) { |
---|
983 | LocalFileName = FileName; |
---|
984 | #ifndef GENERICIO_NO_MPI |
---|
985 | MPI_Comm_dup(MB == MismatchRedistribute ? MPI_COMM_SELF : Comm, &SplitComm); |
---|
986 | #endif |
---|
987 | } else { |
---|
988 | stringstream ss; |
---|
989 | ss << FileName << "#" << RankMap[EffRank]; |
---|
990 | LocalFileName = ss.str(); |
---|
991 | #ifndef GENERICIO_NO_MPI |
---|
992 | if (MB == MismatchRedistribute) { |
---|
993 | MPI_Comm_dup(MPI_COMM_SELF, &SplitComm); |
---|
994 | } else { |
---|
995 | #ifdef __bgq__ |
---|
996 | MPI_Barrier(Comm); |
---|
997 | #endif |
---|
998 | MPI_Comm_split(Comm, RankMap[EffRank], Rank, &SplitComm); |
---|
999 | } |
---|
1000 | #endif |
---|
1001 | } |
---|
1002 | |
---|
1003 | if (LocalFileName == OpenFileName) |
---|
1004 | return; |
---|
1005 | FH.close(); |
---|
1006 | |
---|
1007 | int SplitNRanks, SplitRank; |
---|
1008 | #ifndef GENERICIO_NO_MPI |
---|
1009 | MPI_Comm_rank(SplitComm, &SplitRank); |
---|
1010 | MPI_Comm_size(SplitComm, &SplitNRanks); |
---|
1011 | #else |
---|
1012 | SplitRank = 0; |
---|
1013 | SplitNRanks = 1; |
---|
1014 | #endif |
---|
1015 | |
---|
1016 | uint64_t HeaderSize; |
---|
1017 | vector<char> Header; |
---|
1018 | |
---|
1019 | if (SplitRank == 0) { |
---|
1020 | #ifndef GENERICIO_NO_MPI |
---|
1021 | if (FileIOType == FileIOMPI) |
---|
1022 | FH.get() = new GenericFileIO_MPI(MPI_COMM_SELF); |
---|
1023 | else if (FileIOType == FileIOMPICollective) |
---|
1024 | FH.get() = new GenericFileIO_MPICollective(MPI_COMM_SELF); |
---|
1025 | else |
---|
1026 | #endif |
---|
1027 | FH.get() = new GenericFileIO_POSIX(); |
---|
1028 | |
---|
1029 | #ifndef GENERICIO_NO_MPI |
---|
1030 | char True = 1, False = 0; |
---|
1031 | #endif |
---|
1032 | |
---|
1033 | try { |
---|
1034 | FH.get()->open(LocalFileName, true); |
---|
1035 | |
---|
1036 | GlobalHeader<false> GH; // endianness does not matter yet... |
---|
1037 | FH.get()->read(&GH, sizeof(GlobalHeader<false>), 0, "global header"); |
---|
1038 | |
---|
1039 | if (string(GH.Magic, GH.Magic + MagicSize - 1) == MagicLE) { |
---|
1040 | readHeaderLeader<false>(&GH, MB, NRanks, Rank, SplitNRanks, LocalFileName, |
---|
1041 | HeaderSize, Header); |
---|
1042 | } else if (string(GH.Magic, GH.Magic + MagicSize - 1) == MagicBE) { |
---|
1043 | readHeaderLeader<true>(&GH, MB, NRanks, Rank, SplitNRanks, LocalFileName, |
---|
1044 | HeaderSize, Header); |
---|
1045 | } else { |
---|
1046 | string Error = "invalid file-type identifier"; |
---|
1047 | throw runtime_error("Won't read " + LocalFileName + ": " + Error); |
---|
1048 | } |
---|
1049 | |
---|
1050 | #ifndef GENERICIO_NO_MPI |
---|
1051 | close(); |
---|
1052 | MPI_Bcast(&True, 1, MPI_BYTE, 0, SplitComm); |
---|
1053 | #endif |
---|
1054 | } catch (...) { |
---|
1055 | #ifndef GENERICIO_NO_MPI |
---|
1056 | MPI_Bcast(&False, 1, MPI_BYTE, 0, SplitComm); |
---|
1057 | #endif |
---|
1058 | close(); |
---|
1059 | throw; |
---|
1060 | } |
---|
1061 | } else { |
---|
1062 | #ifndef GENERICIO_NO_MPI |
---|
1063 | char Okay; |
---|
1064 | MPI_Bcast(&Okay, 1, MPI_BYTE, 0, SplitComm); |
---|
1065 | if (!Okay) |
---|
1066 | throw runtime_error("Failure broadcast from rank 0"); |
---|
1067 | #endif |
---|
1068 | } |
---|
1069 | |
---|
1070 | #ifndef GENERICIO_NO_MPI |
---|
1071 | MPI_Bcast(&HeaderSize, 1, MPI_UINT64_T, 0, SplitComm); |
---|
1072 | #endif |
---|
1073 | |
---|
1074 | Header.resize(HeaderSize, 0xFD /* poison */); |
---|
1075 | #ifndef GENERICIO_NO_MPI |
---|
1076 | MPI_Bcast(&Header[0], HeaderSize, MPI_BYTE, 0, SplitComm); |
---|
1077 | #endif |
---|
1078 | |
---|
1079 | FH.getHeaderCache().clear(); |
---|
1080 | |
---|
1081 | GlobalHeader<false> *GH = (GlobalHeader<false> *) &Header[0]; |
---|
1082 | FH.setIsBigEndian(string(GH->Magic, GH->Magic + MagicSize - 1) == MagicBE); |
---|
1083 | |
---|
1084 | FH.getHeaderCache().swap(Header); |
---|
1085 | OpenFileName = LocalFileName; |
---|
1086 | |
---|
1087 | #ifndef GENERICIO_NO_MPI |
---|
1088 | if (!DisableCollErrChecking) |
---|
1089 | MPI_Barrier(Comm); |
---|
1090 | |
---|
1091 | if (FileIOType == FileIOMPI) |
---|
1092 | FH.get() = new GenericFileIO_MPI(SplitComm); |
---|
1093 | else if (FileIOType == FileIOMPICollective) |
---|
1094 | FH.get() = new GenericFileIO_MPICollective(SplitComm); |
---|
1095 | else |
---|
1096 | FH.get() = new GenericFileIO_POSIX(); |
---|
1097 | |
---|
1098 | int OpenErr = 0, TotOpenErr; |
---|
1099 | try { |
---|
1100 | FH.get()->open(LocalFileName, true); |
---|
1101 | MPI_Allreduce(&OpenErr, &TotOpenErr, 1, MPI_INT, MPI_SUM, |
---|
1102 | DisableCollErrChecking ? MPI_COMM_SELF : Comm); |
---|
1103 | } catch (...) { |
---|
1104 | OpenErr = 1; |
---|
1105 | MPI_Allreduce(&OpenErr, &TotOpenErr, 1, MPI_INT, MPI_SUM, |
---|
1106 | DisableCollErrChecking ? MPI_COMM_SELF : Comm); |
---|
1107 | throw; |
---|
1108 | } |
---|
1109 | |
---|
1110 | if (TotOpenErr > 0) { |
---|
1111 | stringstream ss; |
---|
1112 | ss << TotOpenErr << " ranks failed to open file: " << LocalFileName; |
---|
1113 | throw runtime_error(ss.str()); |
---|
1114 | } |
---|
1115 | #endif |
---|
1116 | } |
---|
1117 | |
---|
1118 | int GenericIO::readNRanks() { |
---|
1119 | if (FH.isBigEndian()) |
---|
1120 | return readNRanks<true>(); |
---|
1121 | return readNRanks<false>(); |
---|
1122 | } |
---|
1123 | |
---|
1124 | template <bool IsBigEndian> |
---|
1125 | int GenericIO::readNRanks() { |
---|
1126 | if (RankMap.size()) |
---|
1127 | return RankMap.size(); |
---|
1128 | |
---|
1129 | assert(FH.getHeaderCache().size() && "HeaderCache must not be empty"); |
---|
1130 | GlobalHeader<IsBigEndian> *GH = (GlobalHeader<IsBigEndian> *) &FH.getHeaderCache()[0]; |
---|
1131 | return (int) GH->NRanks; |
---|
1132 | } |
---|
1133 | |
---|
1134 | void GenericIO::readDims(int Dims[3]) { |
---|
1135 | if (FH.isBigEndian()) |
---|
1136 | readDims<true>(Dims); |
---|
1137 | else |
---|
1138 | readDims<false>(Dims); |
---|
1139 | } |
---|
1140 | |
---|
1141 | template <bool IsBigEndian> |
---|
1142 | void GenericIO::readDims(int Dims[3]) { |
---|
1143 | assert(FH.getHeaderCache().size() && "HeaderCache must not be empty"); |
---|
1144 | GlobalHeader<IsBigEndian> *GH = (GlobalHeader<IsBigEndian> *) &FH.getHeaderCache()[0]; |
---|
1145 | std::copy(GH->Dims, GH->Dims + 3, Dims); |
---|
1146 | } |
---|
1147 | |
---|
1148 | uint64_t GenericIO::readTotalNumElems() { |
---|
1149 | if (FH.isBigEndian()) |
---|
1150 | return readTotalNumElems<true>(); |
---|
1151 | return readTotalNumElems<false>(); |
---|
1152 | } |
---|
1153 | |
---|
1154 | template <bool IsBigEndian> |
---|
1155 | uint64_t GenericIO::readTotalNumElems() { |
---|
1156 | if (RankMap.size()) |
---|
1157 | return (uint64_t) -1; |
---|
1158 | |
---|
1159 | assert(FH.getHeaderCache().size() && "HeaderCache must not be empty"); |
---|
1160 | GlobalHeader<IsBigEndian> *GH = (GlobalHeader<IsBigEndian> *) &FH.getHeaderCache()[0]; |
---|
1161 | return GH->NElems; |
---|
1162 | } |
---|
1163 | |
---|
1164 | void GenericIO::readPhysOrigin(double Origin[3]) { |
---|
1165 | if (FH.isBigEndian()) |
---|
1166 | readPhysOrigin<true>(Origin); |
---|
1167 | else |
---|
1168 | readPhysOrigin<false>(Origin); |
---|
1169 | } |
---|
1170 | |
---|
1171 | // Define a "safe" version of offsetof (offsetof itself might not work for |
---|
1172 | // non-POD types, and at least xlC v12.1 will complain about this if you try). |
---|
1173 | #define offsetof_safe(S, F) (size_t(&(S)->F) - size_t(S)) |
---|
1174 | |
---|
1175 | template <bool IsBigEndian> |
---|
1176 | void GenericIO::readPhysOrigin(double Origin[3]) { |
---|
1177 | assert(FH.getHeaderCache().size() && "HeaderCache must not be empty"); |
---|
1178 | GlobalHeader<IsBigEndian> *GH = (GlobalHeader<IsBigEndian> *) &FH.getHeaderCache()[0]; |
---|
1179 | if (offsetof_safe(GH, PhysOrigin) >= GH->GlobalHeaderSize) { |
---|
1180 | std::fill(Origin, Origin + 3, 0.0); |
---|
1181 | return; |
---|
1182 | } |
---|
1183 | |
---|
1184 | std::copy(GH->PhysOrigin, GH->PhysOrigin + 3, Origin); |
---|
1185 | } |
---|
1186 | |
---|
1187 | void GenericIO::readPhysScale(double Scale[3]) { |
---|
1188 | if (FH.isBigEndian()) |
---|
1189 | readPhysScale<true>(Scale); |
---|
1190 | else |
---|
1191 | readPhysScale<false>(Scale); |
---|
1192 | } |
---|
1193 | |
---|
1194 | template <bool IsBigEndian> |
---|
1195 | void GenericIO::readPhysScale(double Scale[3]) { |
---|
1196 | assert(FH.getHeaderCache().size() && "HeaderCache must not be empty"); |
---|
1197 | GlobalHeader<IsBigEndian> *GH = (GlobalHeader<IsBigEndian> *) &FH.getHeaderCache()[0]; |
---|
1198 | if (offsetof_safe(GH, PhysScale) >= GH->GlobalHeaderSize) { |
---|
1199 | std::fill(Scale, Scale + 3, 0.0); |
---|
1200 | return; |
---|
1201 | } |
---|
1202 | |
---|
1203 | std::copy(GH->PhysScale, GH->PhysScale + 3, Scale); |
---|
1204 | } |
---|
1205 | |
---|
1206 | template <bool IsBigEndian> |
---|
1207 | static size_t getRankIndex(int EffRank, GlobalHeader<IsBigEndian> *GH, |
---|
1208 | vector<int> &RankMap, vector<char> &HeaderCache) { |
---|
1209 | if (RankMap.empty()) |
---|
1210 | return EffRank; |
---|
1211 | |
---|
1212 | for (size_t i = 0; i < GH->NRanks; ++i) { |
---|
1213 | RankHeader<IsBigEndian> *RH = (RankHeader<IsBigEndian> *) &HeaderCache[GH->RanksStart + |
---|
1214 | i*GH->RanksSize]; |
---|
1215 | if (offsetof_safe(RH, GlobalRank) >= GH->RanksSize) |
---|
1216 | return EffRank; |
---|
1217 | |
---|
1218 | if ((int) RH->GlobalRank == EffRank) |
---|
1219 | return i; |
---|
1220 | } |
---|
1221 | |
---|
1222 | assert(false && "Index requested of an invalid rank"); |
---|
1223 | return (size_t) -1; |
---|
1224 | } |
---|
1225 | |
---|
1226 | int GenericIO::readGlobalRankNumber(int EffRank) { |
---|
1227 | if (FH.isBigEndian()) |
---|
1228 | return readGlobalRankNumber<true>(EffRank); |
---|
1229 | return readGlobalRankNumber<false>(EffRank); |
---|
1230 | } |
---|
1231 | |
---|
1232 | template <bool IsBigEndian> |
---|
1233 | int GenericIO::readGlobalRankNumber(int EffRank) { |
---|
1234 | if (EffRank == -1) { |
---|
1235 | #ifndef GENERICIO_NO_MPI |
---|
1236 | MPI_Comm_rank(Comm, &EffRank); |
---|
1237 | #else |
---|
1238 | EffRank = 0; |
---|
1239 | #endif |
---|
1240 | } |
---|
1241 | |
---|
1242 | openAndReadHeader(MismatchAllowed, EffRank, false); |
---|
1243 | |
---|
1244 | assert(FH.getHeaderCache().size() && "HeaderCache must not be empty"); |
---|
1245 | |
---|
1246 | GlobalHeader<IsBigEndian> *GH = (GlobalHeader<IsBigEndian> *) &FH.getHeaderCache()[0]; |
---|
1247 | size_t RankIndex = getRankIndex<IsBigEndian>(EffRank, GH, RankMap, FH.getHeaderCache()); |
---|
1248 | |
---|
1249 | assert(RankIndex < GH->NRanks && "Invalid rank specified"); |
---|
1250 | |
---|
1251 | RankHeader<IsBigEndian> *RH = (RankHeader<IsBigEndian> *) &FH.getHeaderCache()[GH->RanksStart + |
---|
1252 | RankIndex*GH->RanksSize]; |
---|
1253 | |
---|
1254 | if (offsetof_safe(RH, GlobalRank) >= GH->RanksSize) |
---|
1255 | return EffRank; |
---|
1256 | |
---|
1257 | return (int) RH->GlobalRank; |
---|
1258 | } |
---|
1259 | |
---|
1260 | void GenericIO::getSourceRanks(vector<int> &SR) { |
---|
1261 | SR.clear(); |
---|
1262 | |
---|
1263 | if (Redistributing) { |
---|
1264 | std::copy(SourceRanks.begin(), SourceRanks.end(), std::back_inserter(SR)); |
---|
1265 | return; |
---|
1266 | } |
---|
1267 | |
---|
1268 | int Rank; |
---|
1269 | #ifndef GENERICIO_NO_MPI |
---|
1270 | MPI_Comm_rank(Comm, &Rank); |
---|
1271 | #else |
---|
1272 | Rank = 0; |
---|
1273 | #endif |
---|
1274 | |
---|
1275 | SR.push_back(Rank); |
---|
1276 | } |
---|
1277 | |
---|
1278 | size_t GenericIO::readNumElems(int EffRank) { |
---|
1279 | if (EffRank == -1 && Redistributing) { |
---|
1280 | DisableCollErrChecking = true; |
---|
1281 | |
---|
1282 | size_t TotalSize = 0; |
---|
1283 | for (int i = 0, ie = SourceRanks.size(); i != ie; ++i) |
---|
1284 | TotalSize += readNumElems(SourceRanks[i]); |
---|
1285 | |
---|
1286 | DisableCollErrChecking = false; |
---|
1287 | return TotalSize; |
---|
1288 | } |
---|
1289 | |
---|
1290 | if (FH.isBigEndian()) |
---|
1291 | return readNumElems<true>(EffRank); |
---|
1292 | return readNumElems<false>(EffRank); |
---|
1293 | } |
---|
1294 | |
---|
1295 | template <bool IsBigEndian> |
---|
1296 | size_t GenericIO::readNumElems(int EffRank) { |
---|
1297 | if (EffRank == -1) { |
---|
1298 | #ifndef GENERICIO_NO_MPI |
---|
1299 | MPI_Comm_rank(Comm, &EffRank); |
---|
1300 | #else |
---|
1301 | EffRank = 0; |
---|
1302 | #endif |
---|
1303 | } |
---|
1304 | |
---|
1305 | openAndReadHeader(Redistributing ? MismatchRedistribute : MismatchAllowed, |
---|
1306 | EffRank, false); |
---|
1307 | |
---|
1308 | assert(FH.getHeaderCache().size() && "HeaderCache must not be empty"); |
---|
1309 | |
---|
1310 | GlobalHeader<IsBigEndian> *GH = (GlobalHeader<IsBigEndian> *) &FH.getHeaderCache()[0]; |
---|
1311 | size_t RankIndex = getRankIndex<IsBigEndian>(EffRank, GH, RankMap, FH.getHeaderCache()); |
---|
1312 | |
---|
1313 | assert(RankIndex < GH->NRanks && "Invalid rank specified"); |
---|
1314 | |
---|
1315 | RankHeader<IsBigEndian> *RH = (RankHeader<IsBigEndian> *) &FH.getHeaderCache()[GH->RanksStart + |
---|
1316 | RankIndex*GH->RanksSize]; |
---|
1317 | return (size_t) RH->NElems; |
---|
1318 | } |
---|
1319 | |
---|
1320 | void GenericIO::readCoords(int Coords[3], int EffRank) { |
---|
1321 | if (EffRank == -1 && Redistributing) { |
---|
1322 | std::fill(Coords, Coords + 3, 0); |
---|
1323 | return; |
---|
1324 | } |
---|
1325 | |
---|
1326 | if (FH.isBigEndian()) |
---|
1327 | readCoords<true>(Coords, EffRank); |
---|
1328 | else |
---|
1329 | readCoords<false>(Coords, EffRank); |
---|
1330 | } |
---|
1331 | |
---|
1332 | template <bool IsBigEndian> |
---|
1333 | void GenericIO::readCoords(int Coords[3], int EffRank) { |
---|
1334 | if (EffRank == -1) { |
---|
1335 | #ifndef GENERICIO_NO_MPI |
---|
1336 | MPI_Comm_rank(Comm, &EffRank); |
---|
1337 | #else |
---|
1338 | EffRank = 0; |
---|
1339 | #endif |
---|
1340 | } |
---|
1341 | |
---|
1342 | openAndReadHeader(MismatchAllowed, EffRank, false); |
---|
1343 | |
---|
1344 | assert(FH.getHeaderCache().size() && "HeaderCache must not be empty"); |
---|
1345 | |
---|
1346 | GlobalHeader<IsBigEndian> *GH = (GlobalHeader<IsBigEndian> *) &FH.getHeaderCache()[0]; |
---|
1347 | size_t RankIndex = getRankIndex<IsBigEndian>(EffRank, GH, RankMap, FH.getHeaderCache()); |
---|
1348 | |
---|
1349 | assert(RankIndex < GH->NRanks && "Invalid rank specified"); |
---|
1350 | |
---|
1351 | RankHeader<IsBigEndian> *RH = (RankHeader<IsBigEndian> *) &FH.getHeaderCache()[GH->RanksStart + |
---|
1352 | RankIndex*GH->RanksSize]; |
---|
1353 | |
---|
1354 | std::copy(RH->Coords, RH->Coords + 3, Coords); |
---|
1355 | } |
---|
1356 | |
---|
1357 | void GenericIO::readData(int EffRank, bool PrintStats, bool CollStats) { |
---|
1358 | int Rank; |
---|
1359 | #ifndef GENERICIO_NO_MPI |
---|
1360 | MPI_Comm_rank(Comm, &Rank); |
---|
1361 | #else |
---|
1362 | Rank = 0; |
---|
1363 | #endif |
---|
1364 | |
---|
1365 | uint64_t TotalReadSize = 0; |
---|
1366 | #ifndef GENERICIO_NO_MPI |
---|
1367 | double StartTime = MPI_Wtime(); |
---|
1368 | #else |
---|
1369 | double StartTime = double(clock())/CLOCKS_PER_SEC; |
---|
1370 | #endif |
---|
1371 | |
---|
1372 | int NErrs[3] = { 0, 0, 0 }; |
---|
1373 | |
---|
1374 | if (EffRank == -1 && Redistributing) { |
---|
1375 | DisableCollErrChecking = true; |
---|
1376 | |
---|
1377 | size_t RowOffset = 0; |
---|
1378 | for (int i = 0, ie = SourceRanks.size(); i != ie; ++i) { |
---|
1379 | readData(SourceRanks[i], RowOffset, Rank, TotalReadSize, NErrs); |
---|
1380 | RowOffset += readNumElems(SourceRanks[i]); |
---|
1381 | } |
---|
1382 | |
---|
1383 | DisableCollErrChecking = false; |
---|
1384 | } else { |
---|
1385 | readData(EffRank, 0, Rank, TotalReadSize, NErrs); |
---|
1386 | } |
---|
1387 | |
---|
1388 | int AllNErrs[3]; |
---|
1389 | #ifndef GENERICIO_NO_MPI |
---|
1390 | MPI_Allreduce(NErrs, AllNErrs, 3, MPI_INT, MPI_SUM, Comm); |
---|
1391 | #else |
---|
1392 | AllNErrs[0] = NErrs[0]; AllNErrs[1] = NErrs[1]; AllNErrs[2] = NErrs[2]; |
---|
1393 | #endif |
---|
1394 | |
---|
1395 | if (AllNErrs[0] > 0 || AllNErrs[1] > 0 || AllNErrs[2] > 0) { |
---|
1396 | stringstream ss; |
---|
1397 | ss << "Experienced " << AllNErrs[0] << " I/O error(s), " << |
---|
1398 | AllNErrs[1] << " CRC error(s) and " << AllNErrs[2] << |
---|
1399 | " decompression CRC error(s) reading: " << OpenFileName; |
---|
1400 | throw runtime_error(ss.str()); |
---|
1401 | } |
---|
1402 | |
---|
1403 | #ifndef GENERICIO_NO_MPI |
---|
1404 | MPI_Barrier(Comm); |
---|
1405 | #endif |
---|
1406 | |
---|
1407 | #ifndef GENERICIO_NO_MPI |
---|
1408 | double EndTime = MPI_Wtime(); |
---|
1409 | #else |
---|
1410 | double EndTime = double(clock())/CLOCKS_PER_SEC; |
---|
1411 | #endif |
---|
1412 | |
---|
1413 | double TotalTime = EndTime - StartTime; |
---|
1414 | double MaxTotalTime; |
---|
1415 | #ifndef GENERICIO_NO_MPI |
---|
1416 | if (CollStats) |
---|
1417 | MPI_Reduce(&TotalTime, &MaxTotalTime, 1, MPI_DOUBLE, MPI_MAX, 0, Comm); |
---|
1418 | else |
---|
1419 | #endif |
---|
1420 | MaxTotalTime = TotalTime; |
---|
1421 | |
---|
1422 | uint64_t AllTotalReadSize; |
---|
1423 | #ifndef GENERICIO_NO_MPI |
---|
1424 | if (CollStats) |
---|
1425 | MPI_Reduce(&TotalReadSize, &AllTotalReadSize, 1, MPI_UINT64_T, MPI_SUM, 0, Comm); |
---|
1426 | else |
---|
1427 | #endif |
---|
1428 | AllTotalReadSize = TotalReadSize; |
---|
1429 | |
---|
1430 | if (Rank == 0 && PrintStats) { |
---|
1431 | double Rate = ((double) AllTotalReadSize) / MaxTotalTime / (1024.*1024.); |
---|
1432 | std::cout << "Read " << Vars.size() << " variables from " << FileName << |
---|
1433 | " (" << AllTotalReadSize << " bytes) in " << MaxTotalTime << "s: " << |
---|
1434 | Rate << " MB/s [excluding header read]" << std::endl; |
---|
1435 | } |
---|
1436 | } |
---|
1437 | |
---|
1438 | void GenericIO::readData(int EffRank, size_t RowOffset, int Rank, |
---|
1439 | uint64_t &TotalReadSize, int NErrs[3]) { |
---|
1440 | if (FH.isBigEndian()) |
---|
1441 | readData<true>(EffRank, RowOffset, Rank, TotalReadSize, NErrs); |
---|
1442 | else |
---|
1443 | readData<false>(EffRank, RowOffset, Rank, TotalReadSize, NErrs); |
---|
1444 | } |
---|
1445 | |
---|
1446 | // Note: Errors from this function should be recoverable. This means that if |
---|
1447 | // one rank throws an exception, then all ranks should. |
---|
1448 | template <bool IsBigEndian> |
---|
1449 | void GenericIO::readData(int EffRank, size_t RowOffset, int Rank, |
---|
1450 | uint64_t &TotalReadSize, int NErrs[3]) { |
---|
1451 | openAndReadHeader(Redistributing ? MismatchRedistribute : MismatchAllowed, |
---|
1452 | EffRank, false); |
---|
1453 | |
---|
1454 | assert(FH.getHeaderCache().size() && "HeaderCache must not be empty"); |
---|
1455 | |
---|
1456 | if (EffRank == -1) |
---|
1457 | EffRank = Rank; |
---|
1458 | |
---|
1459 | GlobalHeader<IsBigEndian> *GH = (GlobalHeader<IsBigEndian> *) &FH.getHeaderCache()[0]; |
---|
1460 | size_t RankIndex = getRankIndex<IsBigEndian>(EffRank, GH, RankMap, FH.getHeaderCache()); |
---|
1461 | |
---|
1462 | assert(RankIndex < GH->NRanks && "Invalid rank specified"); |
---|
1463 | |
---|
1464 | RankHeader<IsBigEndian> *RH = (RankHeader<IsBigEndian> *) &FH.getHeaderCache()[GH->RanksStart + |
---|
1465 | RankIndex*GH->RanksSize]; |
---|
1466 | |
---|
1467 | for (size_t i = 0; i < Vars.size(); ++i) { |
---|
1468 | uint64_t Offset = RH->Start; |
---|
1469 | bool VarFound = false; |
---|
1470 | for (uint64_t j = 0; j < GH->NVars; ++j) { |
---|
1471 | VariableHeader<IsBigEndian> *VH = (VariableHeader<IsBigEndian> *) &FH.getHeaderCache()[GH->VarsStart + |
---|
1472 | j*GH->VarsSize]; |
---|
1473 | |
---|
1474 | string VName(VH->Name, VH->Name + NameSize); |
---|
1475 | size_t VNameNull = VName.find('\0'); |
---|
1476 | if (VNameNull < NameSize) |
---|
1477 | VName.resize(VNameNull); |
---|
1478 | |
---|
1479 | uint64_t ReadSize = RH->NElems*VH->Size + CRCSize; |
---|
1480 | if (VName != Vars[i].Name) { |
---|
1481 | Offset += ReadSize; |
---|
1482 | continue; |
---|
1483 | } |
---|
1484 | |
---|
1485 | size_t ElementSize = VH->Size; |
---|
1486 | if (offsetof_safe(VH, ElementSize) < GH->VarsSize) |
---|
1487 | ElementSize = VH->ElementSize; |
---|
1488 | |
---|
1489 | VarFound = true; |
---|
1490 | bool IsFloat = (bool) (VH->Flags & FloatValue), |
---|
1491 | IsSigned = (bool) (VH->Flags & SignedValue); |
---|
1492 | if (VH->Size != Vars[i].Size) { |
---|
1493 | stringstream ss; |
---|
1494 | ss << "Size mismatch for variable " << Vars[i].Name << |
---|
1495 | " in: " << OpenFileName << ": current: " << Vars[i].Size << |
---|
1496 | ", file: " << VH->Size; |
---|
1497 | throw runtime_error(ss.str()); |
---|
1498 | } else if (ElementSize != Vars[i].ElementSize) { |
---|
1499 | stringstream ss; |
---|
1500 | ss << "Element size mismatch for variable " << Vars[i].Name << |
---|
1501 | " in: " << OpenFileName << ": current: " << Vars[i].ElementSize << |
---|
1502 | ", file: " << ElementSize; |
---|
1503 | throw runtime_error(ss.str()); |
---|
1504 | } else if (IsFloat != Vars[i].IsFloat) { |
---|
1505 | string Float("float"), Int("integer"); |
---|
1506 | stringstream ss; |
---|
1507 | ss << "Type mismatch for variable " << Vars[i].Name << |
---|
1508 | " in: " << OpenFileName << ": current: " << |
---|
1509 | (Vars[i].IsFloat ? Float : Int) << |
---|
1510 | ", file: " << (IsFloat ? Float : Int); |
---|
1511 | throw runtime_error(ss.str()); |
---|
1512 | } else if (IsSigned != Vars[i].IsSigned) { |
---|
1513 | string Signed("signed"), Uns("unsigned"); |
---|
1514 | stringstream ss; |
---|
1515 | ss << "Type mismatch for variable " << Vars[i].Name << |
---|
1516 | " in: " << OpenFileName << ": current: " << |
---|
1517 | (Vars[i].IsSigned ? Signed : Uns) << |
---|
1518 | ", file: " << (IsSigned ? Signed : Uns); |
---|
1519 | throw runtime_error(ss.str()); |
---|
1520 | } |
---|
1521 | |
---|
1522 | size_t VarOffset = RowOffset*Vars[i].Size; |
---|
1523 | void *VarData = ((char *) Vars[i].Data) + VarOffset; |
---|
1524 | |
---|
1525 | vector<unsigned char> LData; |
---|
1526 | bool HasSZ = false; |
---|
1527 | void *Data = VarData; |
---|
1528 | bool HasExtraSpace = Vars[i].HasExtraSpace; |
---|
1529 | if (offsetof_safe(GH, BlocksStart) < GH->GlobalHeaderSize && |
---|
1530 | GH->BlocksSize > 0) { |
---|
1531 | BlockHeader<IsBigEndian> *BH = (BlockHeader<IsBigEndian> *) |
---|
1532 | &FH.getHeaderCache()[GH->BlocksStart + |
---|
1533 | (RankIndex*GH->NVars + j)*GH->BlocksSize]; |
---|
1534 | ReadSize = BH->Size + CRCSize; |
---|
1535 | Offset = BH->Start; |
---|
1536 | |
---|
1537 | int FilterIdx = 0; |
---|
1538 | |
---|
1539 | if (strncmp(BH->Filters[FilterIdx], LossyCompressName, FilterNameSize) == 0) { |
---|
1540 | ++FilterIdx; |
---|
1541 | HasSZ = true; |
---|
1542 | } |
---|
1543 | |
---|
1544 | if (strncmp(BH->Filters[FilterIdx], CompressName, FilterNameSize) == 0) { |
---|
1545 | LData.resize(ReadSize); |
---|
1546 | Data = &LData[0]; |
---|
1547 | HasExtraSpace = true; |
---|
1548 | } else if (BH->Filters[FilterIdx][0] != '\0') { |
---|
1549 | stringstream ss; |
---|
1550 | ss << "Unknown filter \"" << BH->Filters[0] << "\" on variable " << Vars[i].Name; |
---|
1551 | throw runtime_error(ss.str()); |
---|
1552 | } |
---|
1553 | } |
---|
1554 | |
---|
1555 | assert(HasExtraSpace && "Extra space required for reading"); |
---|
1556 | |
---|
1557 | char CRCSave[CRCSize]; |
---|
1558 | char *CRCLoc = ((char *) Data) + ReadSize - CRCSize; |
---|
1559 | if (HasExtraSpace) |
---|
1560 | std::copy(CRCLoc, CRCLoc + CRCSize, CRCSave); |
---|
1561 | |
---|
1562 | int Retry = 0; |
---|
1563 | { |
---|
1564 | int RetryCount = 300; |
---|
1565 | const char *EnvStr = getenv("GENERICIO_RETRY_COUNT"); |
---|
1566 | if (EnvStr) |
---|
1567 | RetryCount = atoi(EnvStr); |
---|
1568 | |
---|
1569 | int RetrySleep = 100; // ms |
---|
1570 | EnvStr = getenv("GENERICIO_RETRY_SLEEP"); |
---|
1571 | if (EnvStr) |
---|
1572 | RetrySleep = atoi(EnvStr); |
---|
1573 | |
---|
1574 | for (; Retry < RetryCount; ++Retry) { |
---|
1575 | try { |
---|
1576 | FH.get()->read(Data, ReadSize, Offset, Vars[i].Name); |
---|
1577 | break; |
---|
1578 | } catch (...) { } |
---|
1579 | |
---|
1580 | usleep(1000*RetrySleep); |
---|
1581 | } |
---|
1582 | |
---|
1583 | if (Retry == RetryCount) { |
---|
1584 | ++NErrs[0]; |
---|
1585 | break; |
---|
1586 | } else if (Retry > 0) { |
---|
1587 | EnvStr = getenv("GENERICIO_VERBOSE"); |
---|
1588 | if (EnvStr) { |
---|
1589 | int Mod = atoi(EnvStr); |
---|
1590 | if (Mod > 0) { |
---|
1591 | int Rank; |
---|
1592 | #ifndef GENERICIO_NO_MPI |
---|
1593 | MPI_Comm_rank(MPI_COMM_WORLD, &Rank); |
---|
1594 | #else |
---|
1595 | Rank = 0; |
---|
1596 | #endif |
---|
1597 | |
---|
1598 | std::cerr << "Rank " << Rank << ": " << Retry << |
---|
1599 | " I/O retries were necessary for reading " << |
---|
1600 | Vars[i].Name << " from: " << OpenFileName << "\n"; |
---|
1601 | |
---|
1602 | std::cerr.flush(); |
---|
1603 | } |
---|
1604 | } |
---|
1605 | } |
---|
1606 | } |
---|
1607 | |
---|
1608 | TotalReadSize += ReadSize; |
---|
1609 | |
---|
1610 | uint64_t CRC = crc64_omp(Data, ReadSize); |
---|
1611 | if (CRC != (uint64_t) -1) { |
---|
1612 | ++NErrs[1]; |
---|
1613 | |
---|
1614 | int Rank; |
---|
1615 | #ifndef GENERICIO_NO_MPI |
---|
1616 | MPI_Comm_rank(MPI_COMM_WORLD, &Rank); |
---|
1617 | #else |
---|
1618 | Rank = 0; |
---|
1619 | #endif |
---|
1620 | |
---|
1621 | // All ranks will do this and have a good time! |
---|
1622 | string dn = "gio_crc_errors"; |
---|
1623 | mkdir(dn.c_str(), 0777); |
---|
1624 | |
---|
1625 | srand(time(0)); |
---|
1626 | int DumpNum = rand(); |
---|
1627 | stringstream ssd; |
---|
1628 | ssd << dn << "/gio_crc_error_dump." << Rank << "." << DumpNum << ".bin"; |
---|
1629 | |
---|
1630 | stringstream ss; |
---|
1631 | ss << dn << "/gio_crc_error_log." << Rank << ".txt"; |
---|
1632 | |
---|
1633 | ofstream ofs(ss.str().c_str(), ofstream::out | ofstream::app); |
---|
1634 | ofs << "On-Disk CRC Error Report:\n"; |
---|
1635 | ofs << "Variable: " << Vars[i].Name << "\n"; |
---|
1636 | ofs << "File: " << OpenFileName << "\n"; |
---|
1637 | ofs << "I/O Retries: " << Retry << "\n"; |
---|
1638 | ofs << "Size: " << ReadSize << " bytes\n"; |
---|
1639 | ofs << "Offset: " << Offset << " bytes\n"; |
---|
1640 | ofs << "CRC: " << CRC << " (expected is -1)\n"; |
---|
1641 | ofs << "Dump file: " << ssd.str() << "\n"; |
---|
1642 | ofs << "\n"; |
---|
1643 | ofs.close(); |
---|
1644 | |
---|
1645 | ofstream dofs(ssd.str().c_str(), ofstream::out); |
---|
1646 | dofs.write((const char *) Data, ReadSize); |
---|
1647 | dofs.close(); |
---|
1648 | |
---|
1649 | uint64_t RawCRC = crc64_omp(Data, ReadSize - CRCSize); |
---|
1650 | unsigned char *UData = (unsigned char *) Data; |
---|
1651 | crc64_invert(RawCRC, &UData[ReadSize - CRCSize]); |
---|
1652 | uint64_t NewCRC = crc64_omp(Data, ReadSize); |
---|
1653 | std::cerr << "Recalulated CRC: " << NewCRC << ((NewCRC == -1) ? "ok" : "bad") << "\n"; |
---|
1654 | break; |
---|
1655 | } |
---|
1656 | |
---|
1657 | if (HasExtraSpace) |
---|
1658 | std::copy(CRCSave, CRCSave + CRCSize, CRCLoc); |
---|
1659 | |
---|
1660 | if (LData.size()) { |
---|
1661 | CompressHeader<IsBigEndian> *CH = (CompressHeader<IsBigEndian>*) &LData[0]; |
---|
1662 | |
---|
1663 | #ifdef _OPENMP |
---|
1664 | #pragma omp master |
---|
1665 | { |
---|
1666 | #endif |
---|
1667 | |
---|
1668 | if (!blosc_initialized) { |
---|
1669 | blosc_init(); |
---|
1670 | blosc_initialized = true; |
---|
1671 | } |
---|
1672 | |
---|
1673 | if (!sz_initialized) { |
---|
1674 | SZ_Init(NULL); |
---|
1675 | sz_initialized = true; |
---|
1676 | } |
---|
1677 | |
---|
1678 | #ifdef _OPENMP |
---|
1679 | blosc_set_nthreads(omp_get_max_threads()); |
---|
1680 | } |
---|
1681 | #endif |
---|
1682 | |
---|
1683 | void *OrigData = VarData; |
---|
1684 | size_t OrigDataSize = Vars[i].Size*RH->NElems; |
---|
1685 | |
---|
1686 | if (HasSZ) { |
---|
1687 | size_t CNBytes, CCBytes, CBlockSize; |
---|
1688 | blosc_cbuffer_sizes(&LData[0] + sizeof(CompressHeader<IsBigEndian>), |
---|
1689 | &CNBytes, &CCBytes, &CBlockSize); |
---|
1690 | |
---|
1691 | OrigData = malloc(CNBytes); |
---|
1692 | OrigDataSize = CNBytes; |
---|
1693 | } |
---|
1694 | |
---|
1695 | blosc_decompress(&LData[0] + sizeof(CompressHeader<IsBigEndian>), |
---|
1696 | OrigData, OrigDataSize); |
---|
1697 | |
---|
1698 | if (CH->OrigCRC != crc64_omp(OrigData, OrigDataSize)) { |
---|
1699 | ++NErrs[2]; |
---|
1700 | break; |
---|
1701 | } |
---|
1702 | |
---|
1703 | if (HasSZ) { |
---|
1704 | int SZDT = GetSZDT(Vars[i]); |
---|
1705 | size_t LDSz = SZ_decompress_args(SZDT, (unsigned char *)OrigData, OrigDataSize, |
---|
1706 | VarData, 0, 0, 0, 0, RH->NElems); |
---|
1707 | free(OrigData); |
---|
1708 | |
---|
1709 | if (LDSz != Vars[i].Size*RH->NElems) |
---|
1710 | throw runtime_error("Variable " + Vars[i].Name + |
---|
1711 | ": SZ decompression yielded the wrong amount of data"); |
---|
1712 | } |
---|
1713 | } |
---|
1714 | |
---|
1715 | // Byte swap the data if necessary. |
---|
1716 | if (IsBigEndian != isBigEndian() && !HasSZ) |
---|
1717 | for (size_t j = 0; |
---|
1718 | j < RH->NElems*(Vars[i].Size/Vars[i].ElementSize); ++j) { |
---|
1719 | char *Offset = ((char *) VarData) + j*Vars[i].ElementSize; |
---|
1720 | bswap(Offset, Vars[i].ElementSize); |
---|
1721 | } |
---|
1722 | |
---|
1723 | break; |
---|
1724 | } |
---|
1725 | |
---|
1726 | if (!VarFound) |
---|
1727 | throw runtime_error("Variable " + Vars[i].Name + |
---|
1728 | " not found in: " + OpenFileName); |
---|
1729 | |
---|
1730 | // This is for debugging. |
---|
1731 | if (NErrs[0] || NErrs[1] || NErrs[2]) { |
---|
1732 | const char *EnvStr = getenv("GENERICIO_VERBOSE"); |
---|
1733 | if (EnvStr) { |
---|
1734 | int Mod = atoi(EnvStr); |
---|
1735 | if (Mod > 0) { |
---|
1736 | int Rank; |
---|
1737 | #ifndef GENERICIO_NO_MPI |
---|
1738 | MPI_Comm_rank(MPI_COMM_WORLD, &Rank); |
---|
1739 | #else |
---|
1740 | Rank = 0; |
---|
1741 | #endif |
---|
1742 | |
---|
1743 | std::cerr << "Rank " << Rank << ": " << NErrs[0] << " I/O error(s), " << |
---|
1744 | NErrs[1] << " CRC error(s) and " << NErrs[2] << |
---|
1745 | " decompression CRC error(s) reading: " << Vars[i].Name << |
---|
1746 | " from: " << OpenFileName << "\n"; |
---|
1747 | |
---|
1748 | std::cerr.flush(); |
---|
1749 | } |
---|
1750 | } |
---|
1751 | } |
---|
1752 | |
---|
1753 | if (NErrs[0] || NErrs[1] || NErrs[2]) |
---|
1754 | break; |
---|
1755 | } |
---|
1756 | } |
---|
1757 | |
---|
1758 | void GenericIO::getVariableInfo(vector<VariableInfo> &VI) { |
---|
1759 | if (FH.isBigEndian()) |
---|
1760 | getVariableInfo<true>(VI); |
---|
1761 | else |
---|
1762 | getVariableInfo<false>(VI); |
---|
1763 | } |
---|
1764 | |
---|
1765 | template <bool IsBigEndian> |
---|
1766 | void GenericIO::getVariableInfo(vector<VariableInfo> &VI) { |
---|
1767 | assert(FH.getHeaderCache().size() && "HeaderCache must not be empty"); |
---|
1768 | |
---|
1769 | GlobalHeader<IsBigEndian> *GH = (GlobalHeader<IsBigEndian> *) &FH.getHeaderCache()[0]; |
---|
1770 | for (uint64_t j = 0; j < GH->NVars; ++j) { |
---|
1771 | VariableHeader<IsBigEndian> *VH = (VariableHeader<IsBigEndian> *) &FH.getHeaderCache()[GH->VarsStart + |
---|
1772 | j*GH->VarsSize]; |
---|
1773 | |
---|
1774 | string VName(VH->Name, VH->Name + NameSize); |
---|
1775 | size_t VNameNull = VName.find('\0'); |
---|
1776 | if (VNameNull < NameSize) |
---|
1777 | VName.resize(VNameNull); |
---|
1778 | |
---|
1779 | size_t ElementSize = VH->Size; |
---|
1780 | if (offsetof_safe(VH, ElementSize) < GH->VarsSize) |
---|
1781 | ElementSize = VH->ElementSize; |
---|
1782 | |
---|
1783 | bool IsFloat = (bool) (VH->Flags & FloatValue), |
---|
1784 | IsSigned = (bool) (VH->Flags & SignedValue), |
---|
1785 | IsPhysCoordX = (bool) (VH->Flags & ValueIsPhysCoordX), |
---|
1786 | IsPhysCoordY = (bool) (VH->Flags & ValueIsPhysCoordY), |
---|
1787 | IsPhysCoordZ = (bool) (VH->Flags & ValueIsPhysCoordZ), |
---|
1788 | MaybePhysGhost = (bool) (VH->Flags & ValueMaybePhysGhost); |
---|
1789 | VI.push_back(VariableInfo(VName, (size_t) VH->Size, IsFloat, IsSigned, |
---|
1790 | IsPhysCoordX, IsPhysCoordY, IsPhysCoordZ, |
---|
1791 | MaybePhysGhost, ElementSize)); |
---|
1792 | } |
---|
1793 | } |
---|
1794 | |
---|
1795 | void GenericIO::setNaturalDefaultPartition() { |
---|
1796 | #ifdef __bgq__ |
---|
1797 | DefaultPartition = MPIX_IO_link_id(); |
---|
1798 | #else |
---|
1799 | #ifndef GENERICIO_NO_MPI |
---|
1800 | bool UseName = true; |
---|
1801 | const char *EnvStr = getenv("GENERICIO_PARTITIONS_USE_NAME"); |
---|
1802 | if (EnvStr) { |
---|
1803 | int Mod = atoi(EnvStr); |
---|
1804 | UseName = (Mod != 0); |
---|
1805 | } |
---|
1806 | |
---|
1807 | if (UseName) { |
---|
1808 | // This is a heuristic to generate ~256 partitions based on the |
---|
1809 | // names of the nodes. |
---|
1810 | char Name[MPI_MAX_PROCESSOR_NAME]; |
---|
1811 | int Len = 0; |
---|
1812 | |
---|
1813 | MPI_Get_processor_name(Name, &Len); |
---|
1814 | unsigned char color = 0; |
---|
1815 | for (int i = 0; i < Len; ++i) |
---|
1816 | color += (unsigned char) Name[i]; |
---|
1817 | |
---|
1818 | DefaultPartition = color; |
---|
1819 | } |
---|
1820 | |
---|
1821 | // This is for debugging. |
---|
1822 | EnvStr = getenv("GENERICIO_RANK_PARTITIONS"); |
---|
1823 | if (EnvStr) { |
---|
1824 | int Mod = atoi(EnvStr); |
---|
1825 | if (Mod > 0) { |
---|
1826 | int Rank; |
---|
1827 | MPI_Comm_rank(MPI_COMM_WORLD, &Rank); |
---|
1828 | DefaultPartition += Rank % Mod; |
---|
1829 | } |
---|
1830 | } |
---|
1831 | #endif |
---|
1832 | #endif |
---|
1833 | } |
---|
1834 | |
---|
1835 | } /* END namespace cosmotk */ |
---|