VCluster_base.hpp 31.9 KB
Newer Older
incardon's avatar
incardon committed
1 2
#ifndef VCLUSTER_BASE_HPP_
#define VCLUSTER_BASE_HPP_
incardon's avatar
incardon committed
3

4
#include "config.h"
incardon's avatar
incardon committed
5
#include <mpi.h>
6
#include "MPI_wrapper/MPI_util.hpp"
incardon's avatar
incardon committed
7
#include "Vector/map_vector.hpp"
8 9 10
#include "MPI_wrapper/MPI_IallreduceW.hpp"
#include "MPI_wrapper/MPI_IrecvW.hpp"
#include "MPI_wrapper/MPI_IsendW.hpp"
incardon's avatar
incardon committed
11
#include "MPI_wrapper/MPI_IAllGather.hpp"
incardon's avatar
incardon committed
12
#include "MPI_wrapper/MPI_IBcastW.hpp"
incardon's avatar
incardon committed
13
#include <exception>
14
#include "Vector/map_vector.hpp"
15 16
#ifdef DEBUG
#include "util/check_no_pointers.hpp"
17
#include "util/util_debug.hpp"
18
#endif
incardon's avatar
incardon committed
19
#include "util/Vcluster_log.hpp"
20
#include "memory/BHeapMemory.hpp"
21
#include "Packer_Unpacker/has_max_prop.hpp"
incardon's avatar
incardon committed
22
#include "data_type/aggregate.hpp"
incardon's avatar
incardon committed
23

Pietro Incardona's avatar
Pietro Incardona committed
24 25 26 27
#ifdef HAVE_PETSC
#include <petscvec.h>
#endif

incardon's avatar
incardon committed
28 29
#define MSG_LENGTH 1024
#define MSG_SEND_RECV 1025
30
#define SEND_SPARSE 4096
incardon's avatar
incardon committed
31 32 33
#define NONE 1
#define NEED_ALL_SIZE 2

34 35
#define SERIVCE_MESSAGE_TAG 16384
#define SEND_RECV_BASE 8192
incardon's avatar
incardon committed
36
#define GATHER_BASE 24576
incardon's avatar
incardon committed
37

38 39 40
#define RECEIVE_KNOWN 4
#define KNOWN_ELEMENT_OR_BYTE 8

41
// number of vcluster instances
incardon's avatar
incardon committed
42
extern size_t n_vcluster;
43
// Global MPI initialization
incardon's avatar
incardon committed
44
extern bool global_mpi_init;
Pietro Incardona's avatar
Pietro Incardona committed
45 46
// initialization flag
extern bool ofp_initialized;
incardon's avatar
incardon committed
47 48
extern size_t tot_sent;
extern size_t tot_recv;
incardon's avatar
incardon committed
49

incardon's avatar
incardon committed
50 51 52 53 54 55 56 57
///////////////////// Post functions /////////////

template<typename T> void assign(T * ptr1, T * ptr2)
{
	*ptr1 = *ptr2;
};


incardon's avatar
incardon committed
58
//! temporal buffer for reductions
incardon's avatar
incardon committed
59 60
union red
{
incardon's avatar
incardon committed
61
	//! char
incardon's avatar
incardon committed
62
	char c;
incardon's avatar
incardon committed
63
	//! unsigned char
incardon's avatar
incardon committed
64
	unsigned char uc;
incardon's avatar
incardon committed
65
	//! signed
incardon's avatar
incardon committed
66
	short s;
incardon's avatar
incardon committed
67
	//! unsigned short
incardon's avatar
incardon committed
68
	unsigned short us;
incardon's avatar
incardon committed
69
	//! integer
incardon's avatar
incardon committed
70
	int i;
incardon's avatar
incardon committed
71
	//! unsigned integer
incardon's avatar
incardon committed
72
	unsigned int ui;
incardon's avatar
incardon committed
73
	//! float
incardon's avatar
incardon committed
74
	float f;
incardon's avatar
incardon committed
75
	//! double
incardon's avatar
incardon committed
76 77 78
	double d;
};

79
/*! \brief This class virtualize the cluster of PC as a set of processes that communicate
incardon's avatar
incardon committed
80
 *
81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99
 * At the moment it is an MPI-like interface, with a more type aware, and simple, interface.
 *  It also give some more complex communication functionalities like **Dynamic Sparse Data Exchange**
 *
 * Actually VCluster expose a Computation driven parallelism (MPI-like), with a plan of extending to
 * communication driven parallelism
 *
 * * In computation driven parallelism, the program compute than communicate to the other processors
 *
 * * In a communication driven parallelism, (Charm++ or HPX), the program receive messages, this receiving
 *   messages trigger computation
 *
 * ### An example of sending and receive plain buffers
 * \snippet VCluster_unit_test_util.hpp Send and receive plain buffer data
 * ### An example of sending vectors of primitives with (T=float,double,lont int,...)
 * \snippet VCluster_unit_test_util.hpp Sending and receiving primitives
 * ### An example of sending vectors of complexes object
 * \snippet VCluster_unit_test_util.hpp Send and receive vectors of complex
 * ### An example of gathering numbers from all processors
 * \snippet VCluster_unit_test_util.hpp allGather numbers
incardon's avatar
incardon committed
100 101 102
 *
 */

incardon's avatar
incardon committed
103
class Vcluster_base
incardon's avatar
incardon committed
104
{
incardon's avatar
incardon committed
105
	//! log file
incardon's avatar
incardon committed
106 107
	Vcluster_log log;

108 109 110
	//! NBX has a potential pitfall that must be addressed,
	//! NBX Send all the messages and probe for incoming messages,
	//! if there is an incoming message it receive it producing
111
	//! an acknowledge notification on the sending processor.
112 113 114 115 116 117
	//! When all the sends has been acknowledged, the processor call the MPI_Ibarrier
	//! when all the processors call MPI_Ibarrier all send has been received.
	//! While the processors are waiting for the MPI_Ibarrier to complete, all processors
	//! are still probing for incoming message, Unfortunately some processor
	//! can quit the MPI_Ibarrier before others and this mean that some
	//! processor can exit the probing status before others, these processors can in theory
118 119 120 121
	//! start new communications while the other processor are still in probing status producing
	//! a wrong send/recv association to
	//! resolve this problem an incremental NBX_cnt is used as message TAG to distinguish that the
	//! messages come from other send or subsequent NBX procedures
incardon's avatar
incardon committed
122
	size_t NBX_cnt;
123

incardon's avatar
incardon committed
124 125
	//! temporal vector used for meta-communication
	//! ( or meta-data before the real communication )
incardon's avatar
incardon committed
126 127
	openfpm::vector<size_t> proc_com;

incardon's avatar
incardon committed
128
	//! vector that contain the scatter map (it is basically an array of one)
incardon's avatar
incardon committed
129 130
	openfpm::vector<int> map_scatter;

incardon's avatar
incardon committed
131
	//! vector of MPI requests
incardon's avatar
incardon committed
132 133
	openfpm::vector<MPI_Request> req;

incardon's avatar
incardon committed
134
	//! vector of MPI status
incardon's avatar
incardon committed
135 136
	openfpm::vector<MPI_Status> stat;

incardon's avatar
incardon committed
137
	//! vector of functions to execute after all the request has been performed
incardon's avatar
incardon committed
138 139 140 141 142 143 144
	std::vector<int> post_exe;

	// Object array


	// Single objects

incardon's avatar
incardon committed
145
	//! number of processes
incardon's avatar
incardon committed
146
	int m_size;
incardon's avatar
incardon committed
147
	//! actual rank
incardon's avatar
incardon committed
148
	int m_rank;
incardon's avatar
incardon committed
149

incardon's avatar
incardon committed
150
	//! number of processing unit per process
incardon's avatar
incardon committed
151 152
	int numPE = 1;

incardon's avatar
incardon committed
153
	/*! This buffer is a temporal buffer for reductions
incardon's avatar
incardon committed
154 155 156 157 158 159 160 161
	 *
	 * MPI_Iallreduce does not accept recv and send buffer to be the same
	 * r is used to overcome this problem (is given as second parameter)
	 * after the execution the data is copied back
	 *
	 */
	std::vector<red> r;

incardon's avatar
incardon committed
162
	//! vector of pointers of send buffers
163 164
	openfpm::vector<void *> ptr_send;

incardon's avatar
incardon committed
165
	//! vector of the size of send buffers
166 167
	openfpm::vector<size_t> sz_send;

incardon's avatar
incardon committed
168
	//! barrier request
169
	MPI_Request bar_req;
incardon's avatar
incardon committed
170 171

	//! barrier status
172 173
	MPI_Status bar_stat;

incardon's avatar
incardon committed
174
	//! disable operator=
incardon's avatar
incardon committed
175
	Vcluster_base & operator=(const Vcluster_base &)	{return *this;};
Pietro Incardona's avatar
Pietro Incardona committed
176

incardon's avatar
incardon committed
177
	//! disable copy constructor
incardon's avatar
incardon committed
178
	Vcluster_base(const Vcluster_base &)
incardon's avatar
incardon committed
179 180
	:NBX_cnt(0)
	{};
Pietro Incardona's avatar
Pietro Incardona committed
181

incardon's avatar
incardon committed
182 183 184 185 186
protected:

	//! Receive buffers
	openfpm::vector<BHeapMemory> recv_buf;

incardon's avatar
incardon committed
187 188 189
	//! tags receiving
	openfpm::vector<size_t> tags;

incardon's avatar
incardon committed
190 191 192
public:

	// Finalize the MPI program
incardon's avatar
incardon committed
193
	~Vcluster_base()
incardon's avatar
incardon committed
194
	{
195 196 197
#ifdef SE_CLASS2
		check_delete(this);
#endif
incardon's avatar
incardon committed
198 199 200 201
		n_vcluster--;

		// if there are no other vcluster instances finalize
		if (n_vcluster == 0)
202
		{
203 204 205 206
			int already_finalised;

			MPI_Finalized(&already_finalised);
			if (!already_finalised)
207
			{
208 209 210 211
				if (MPI_Finalize() != 0)
				{
					std::cerr << __FILE__ << ":" << __LINE__  << " MPI_Finalize FAILED \n";
				}
212
			}
213
		}
incardon's avatar
incardon committed
214 215
	}

216 217 218 219 220 221
	/*! \brief Virtual cluster constructor
	 *
	 * \param argc pointer to arguments counts passed to the program
	 * \param argv pointer to arguments vector passed to the program
	 *
	 */
incardon's avatar
incardon committed
222
	Vcluster_base(int *argc, char ***argv)
incardon's avatar
incardon committed
223
	:NBX_cnt(0)
incardon's avatar
incardon committed
224
	{
225 226 227 228
#ifdef SE_CLASS2
		check_new(this,8,VCLUSTER_EVENT,PRJ_VCLUSTER);
#endif

incardon's avatar
incardon committed
229 230
		n_vcluster++;

231 232 233
		int already_initialised;
		MPI_Initialized(&already_initialised);

incardon's avatar
incardon committed
234
		// Check if MPI is already initialized
235
		if (!already_initialised)
incardon's avatar
incardon committed
236
		{
incardon's avatar
incardon committed
237

incardon's avatar
incardon committed
238 239
			MPI_Init(argc,argv);
		}
incardon's avatar
incardon committed
240

241 242
		// Get the total number of process
		// and the rank of this process
incardon's avatar
incardon committed
243

incardon's avatar
incardon committed
244 245
		MPI_Comm_size(MPI_COMM_WORLD, &m_size);
		MPI_Comm_rank(MPI_COMM_WORLD, &m_rank);
incardon's avatar
incardon committed
246

247
#ifdef SE_CLASS2
incardon's avatar
incardon committed
248
			process_v_cl = m_rank;
incardon's avatar
incardon committed
249 250
#endif

251
		// create and fill map scatter with one
incardon's avatar
incardon committed
252
		map_scatter.resize(m_size);
incardon's avatar
incardon committed
253 254 255 256 257

		for (size_t i = 0 ; i < map_scatter.size() ; i++)
		{
			map_scatter.get(i) = 1;
		}
incardon's avatar
incardon committed
258 259

		// open the log file
incardon's avatar
incardon committed
260
		log.openLog(m_rank);
261 262 263 264

		// Initialize bar_req
		bar_req = MPI_Request();
		bar_stat = MPI_Status();
incardon's avatar
incardon committed
265 266
	}

incardon's avatar
incardon committed
267
#ifdef SE_CLASS1
268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302

	/*! \brief Check for wrong types
	 *
	 * In general we do not know if a type T make sense to be sent or not, but if it has pointer
	 * inside it does not. This function check if the basic type T has a method called noPointers,
	 * This function in general notify if T has internally pointers. If T has pointer an error
	 * is printed, is T does not have the method a WARNING is printed
	 *
	 * \tparam T type to check
	 *
	 */
	template<typename T> void checkType()
	{
		// if T is a primitive like int, long int, float, double, ... make sense
		// (pointers, l-references and r-references are not fundamentals)
		if (std::is_fundamental<T>::value == true)
			return;

		// if it is a pointer make no sense
		if (std::is_pointer<T>::value == true)
			std::cerr << "Error: " << __FILE__ << ":" << __LINE__ << " the type " << demangle(typeid(T).name()) << " is a pointer, sending pointers values has no sense\n";

		// if it is an l-value reference make no send
		if (std::is_lvalue_reference<T>::value == true)
			std::cerr << "Error: " << __FILE__ << ":" << __LINE__ << " the type " << demangle(typeid(T).name()) << " is a pointer, sending pointers values has no sense\n";

		// if it is an r-value reference make no send
		if (std::is_rvalue_reference<T>::value == true)
			std::cerr << "Error: " << __FILE__ << ":" << __LINE__ << " the type " << demangle(typeid(T).name()) << " is a pointer, sending pointers values has no sense\n";

		// ... if not, check that T has a method called noPointers
		switch (check_no_pointers<T>::value())
		{
			case PNP::UNKNOWN:
			{
incardon's avatar
incardon committed
303
				std::cerr << "Warning: " << __FILE__ << ":" << __LINE__ << " impossible to check the type " << demangle(typeid(T).name()) << " please consider to add a static method \"static bool noPointers()\" \n" ;
304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319
				break;
			}
			case PNP::POINTERS:
			{
				std::cerr << "Error: " << __FILE__ << ":" << __LINE__ << " the type " << demangle(typeid(T).name()) << " has pointers inside, sending pointers values has no sense\n";
				break;
			}
			default:
			{

			}
		}
	}

#endif

Pietro Incardona's avatar
Pietro Incardona committed
320 321 322 323 324 325 326 327 328 329
	/*! \brief Get the MPI_Communicator (or processor group) this VCluster is using
	 *
	 * \return MPI comunicator
	 *
	 */
	MPI_Comm getMPIComm()
	{
		return MPI_COMM_WORLD;
	}

incardon's avatar
incardon committed
330 331 332 333 334
	/*! \brief Get the total number of processors
	 *
	 * \return the total number of processors
	 *
	 */
incardon's avatar
incardon committed
335 336
	size_t getProcessingUnits()
	{
incardon's avatar
incardon committed
337 338 339 340 341 342 343 344 345 346 347 348 349 350 351
		return m_size*numPE;
	}

	/*! \brief Get the total number of processors
	 *
	 * It is the same as getProcessingUnits()
	 *
	 * \see getProcessingUnits()
	 *
	 * \return the total number of processors
	 *
	 */
	size_t size()
	{
		return this->m_size*numPE;
incardon's avatar
incardon committed
352 353
	}

incardon's avatar
incardon committed
354 355 356 357 358
	/*! \brief Get the process unit id
	 *
	 * \return the process ID
	 *
	 */
incardon's avatar
incardon committed
359 360
	size_t getProcessUnitID()
	{
incardon's avatar
incardon committed
361 362 363 364 365 366 367 368 369 370 371 372 373 374 375
		return m_rank;
	}

	/*! \brief Get the process unit id
	 *
	 * It is the same as getProcessUnitID()
	 *
	 * \see getProcessUnitID()
	 *
	 * \return the process ID
	 *
	 */
	size_t rank()
	{
		return m_rank;
incardon's avatar
incardon committed
376 377
	}

incardon's avatar
incardon committed
378

379
	/*! \brief Sum the numbers across all processors and get the result
incardon's avatar
incardon committed
380 381 382 383 384
	 *
	 * \param num to reduce, input and output
	 *
	 */

385
	template<typename T> void sum(T & num)
incardon's avatar
incardon committed
386
	{
incardon's avatar
incardon committed
387
#ifdef SE_CLASS1
388 389 390
		checkType<T>();
#endif

incardon's avatar
incardon committed
391 392 393 394 395 396 397 398 399
		// reduce over MPI

		// Create one request
		req.add();

		// reduce
		MPI_IallreduceW<T>::reduce(num,MPI_SUM,req.last());
	}

400
	/*! \brief Get the maximum number across all processors (or reduction with infinity norm)
incardon's avatar
incardon committed
401 402 403 404 405 406
	 *
	 * \param num to reduce
	 *
	 */
	template<typename T> void max(T & num)
	{
incardon's avatar
incardon committed
407
#ifdef SE_CLASS1
408 409
		checkType<T>();
#endif
incardon's avatar
incardon committed
410 411 412 413 414 415 416 417 418
		// reduce over MPI

		// Create one request
		req.add();

		// reduce
		MPI_IallreduceW<T>::reduce(num,MPI_MAX,req.last());
	}

tonynsyde's avatar
tonynsyde committed
419 420 421 422 423 424 425 426
	/*! \brief Get the minimum number across all processors (or reduction with insinity norm)
	 *
	 * \param num to reduce
	 *
	 */

	template<typename T> void min(T & num)
	{
incardon's avatar
incardon committed
427
#ifdef SE_CLASS1
tonynsyde's avatar
tonynsyde committed
428 429 430 431 432 433 434 435 436 437 438
		checkType<T>();
#endif
		// reduce over MPI

		// Create one request
		req.add();

		// reduce
		MPI_IallreduceW<T>::reduce(num,MPI_MIN,req.last());
	}

439
	/*! \brief Send and receive multiple messages
440
	 *
441 442
	 * It send multiple messages to a set of processors the and receive
	 * multiple messages from another set of processors, all the processor must call this
443 444 445
	 * function. In this particular case the receiver know from which processor is going
	 * to receive.
	 *
446 447
	 *
	 * suppose the following situation the calling processor want to communicate
448 449 450 451
	 * * 2 messages of size 100 byte to processor 1
	 * * 1 message of size 50 byte to processor 6
	 * * 1 message of size 48 byte to processor 7
	 * * 1 message of size 70 byte to processor 8
452 453
	 *
	 *
454 455
	 * \param prc list of processor with which it should communicate
	 *        [1,1,6,7,8]
456
	 *
incardon's avatar
incardon committed
457 458 459 460
	 * \param data data to send for each processors in contain a pointer to some type T
	 *        this type T must have a method size() that return the size of the data-structure
	 *
	 * \param prc_recv processor that receive data
461
	 *
incardon's avatar
incardon committed
462
	 * \param recv_sz for each processor indicate the size of the data received
463 464 465
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
	 *        for the incoming message and give back a valid pointer, supposing that this call-back has been triggered by
466 467
	 *        the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have
	 *        the following 6 parameters
468 469
	 *        in the call-back are in order:
	 *        * message size required to receive the message [100]
470 471
	 *        * total message size to receive from all the processors (NBX does not provide this information)
	 *        * the total number of processor want to communicate with you (NBX does not provide this information)
472
	 *        * processor id [5]
473
	 *        * ri request id (it is an id that goes from 0 to total_p, and is incremented
474
	 *           every time message_alloc is called)
475
	 *        * void pointer, parameter for additional data to pass to the call-back
476
	 *
incardon's avatar
incardon committed
477 478
	 * \param ptr_arg data passed to the call-back function specified
	 *
479
	 * \param opt options, NONE (ignored in this moment)
480 481
	 *
	 */
incardon's avatar
incardon committed
482 483 484 485
	template<typename T> void sendrecvMultipleMessagesNBX(openfpm::vector< size_t > & prc,
			                                              openfpm::vector< T > & data,
														  openfpm::vector< size_t > prc_recv,
														  openfpm::vector< size_t > & recv_sz ,
incardon's avatar
incardon committed
486
														  void * (* msg_alloc)(size_t,size_t,size_t,size_t,size_t,size_t,void *),
incardon's avatar
incardon committed
487 488
														  void * ptr_arg,
														  long int opt=NONE)
489
	{
490
		// Allocate the buffers
491 492

		for (size_t i = 0 ; i < prc.size() ; i++)
incardon's avatar
incardon committed
493
		{send(prc.get(i),SEND_SPARSE + NBX_cnt,data.get(i).getPointer(),data.get(i).size());}
494 495

		for (size_t i = 0 ; i < prc_recv.size() ; i++)
496
		{
incardon's avatar
incardon committed
497
			void * ptr_recv = msg_alloc(recv_sz.get(i),0,0,prc_recv.get(i),i,SEND_SPARSE + NBX_cnt,ptr_arg);
498 499

			recv(prc_recv.get(i),SEND_SPARSE + NBX_cnt,ptr_recv,recv_sz.get(i));
500 501
		}

502 503 504 505 506 507
		execute();

		// Circular counter
		NBX_cnt = (NBX_cnt + 1) % 1024;
	}

incardon's avatar
incardon committed
508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565

	/*! \brief Send and receive multiple messages
	 *
	 * It send multiple messages to a set of processors the and receive
	 * multiple messages from another set of processors, all the processor must call this
	 * function
	 *
	 * suppose the following situation the calling processor want to communicate
	 * * 2 vector of 100 integers to processor 1
	 * * 1 vector of 50 integers to processor 6
	 * * 1 vector of 48 integers to processor 7
	 * * 1 vector of 70 integers to processor 8
	 *
	 * \param prc list of processors you should communicate with [1,1,6,7,8]
	 *
	 * \param data vector containing the data to send [v=vector<vector<int>>, v.size()=4, T=vector<int>], T at the moment
	 *          is only tested for vectors of 0 or more generic elements (without pointers)
	 *
	 * \param msg_alloc This is a call-back with the purpose to allocate space
	 *        for the incoming messages and give back a valid pointer, supposing that this call-back has been triggered by
	 *        the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have
	 *        the following 6 parameters
	 *        in the call-back in order:
	 *        * message size required to receive the message (100)
	 *        * total message size to receive from all the processors (NBX does not provide this information)
	 *        * the total number of processor want to communicate with you (NBX does not provide this information)
	 *        * processor id (5)
	 *        * ri request id (it is an id that goes from 0 to total_p, and is incremented
	 *           every time message_alloc is called)
	 *        * void pointer, parameter for additional data to pass to the call-back
	 *
	 * \param ptr_arg data passed to the call-back function specified
	 *
	 * \param opt options, only NONE supported
	 *
	 */
	template<typename T>
	void sendrecvMultipleMessagesNBX(openfpm::vector< size_t > & prc,
									 openfpm::vector< T > & data,
									 void * (* msg_alloc)(size_t,size_t,size_t,size_t,size_t,void *),
									 void * ptr_arg, long int opt=NONE)
	{
#ifdef SE_CLASS1
		checkType<typename T::value_type>();
#endif
		// resize the pointer list
		ptr_send.resize(prc.size());
		sz_send.resize(prc.size());

		for (size_t i = 0 ; i < prc.size() ; i++)
		{
			ptr_send.get(i) = data.get(i).getPointer();
			sz_send.get(i) = data.get(i).size() * sizeof(typename T::value_type);
		}

		sendrecvMultipleMessagesNBX(prc.size(),(size_t *)sz_send.getPointer(),(size_t *)prc.getPointer(),(void **)ptr_send.getPointer(),msg_alloc,ptr_arg,opt);
	}

566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607
	/*! \brief Send and receive multiple messages
	 *
	 * It send multiple messages to a set of processors the and receive
	 * multiple messages from another set of processors, all the processor must call this
	 * function. In this particular case the receiver know from which processor is going
	 * to receive.
	 *
	 *
	 * suppose the following situation the calling processor want to communicate
	 * * 2 messages of size 100 byte to processor 1
	 * * 1 message of size 50 byte to processor 6
	 * * 1 message of size 48 byte to processor 7
	 * * 1 message of size 70 byte to processor 8
	 *
	 * \param n_send number of send for this processor [4]
	 *
	 * \param prc list of processor with which it should communicate
	 *        [1,1,6,7,8]
	 *
	 * \param sz the array contain the size of the message for each processor
	 *        (zeros must not be presents) [100,100,50,48,70]
	 *
	 * \param ptr array that contain the pointers to the message to send
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
	 *        for the incoming message and give back a valid pointer, supposing that this call-back has been triggered by
	 *        the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have
	 *        the following 6 parameters
	 *        in the call-back are in order:
	 *        * message size required to receive the message [100]
	 *        * total message size to receive from all the processors (NBX does not provide this information)
	 *        * the total number of processor want to communicate with you (NBX does not provide this information)
	 *        * processor id [5]
	 *        * ri request id (it is an id that goes from 0 to total_p, and is incremented
	 *           every time message_alloc is called)
	 *        * void pointer, parameter for additional data to pass to the call-back
	 *
	 * \param ptr_arg data passed to the call-back function specified
	 *
	 * \param opt options, NONE (ignored in this moment)
	 *
	 */
incardon's avatar
incardon committed
608 609 610 611 612
	void sendrecvMultipleMessagesNBX(size_t n_send , size_t sz[],
			                         size_t prc[] , void * ptr[],
			                         size_t n_recv, size_t prc_recv[] ,
			                         size_t sz_recv[] ,void * (* msg_alloc)(size_t,size_t,size_t,size_t,size_t, size_t,void *),
			                         void * ptr_arg, long int opt=NONE)
613 614 615 616 617 618 619 620
	{
		// Allocate the buffers

		for (size_t i = 0 ; i < n_send ; i++)
			send(prc[i],SEND_SPARSE + NBX_cnt,ptr[i],sz[i]);

		for (size_t i = 0 ; i < n_recv ; i++)
		{
incardon's avatar
incardon committed
621
			void * ptr_recv = msg_alloc(sz_recv[i],0,0,prc_recv[i],i,SEND_SPARSE + NBX_cnt,ptr_arg);
622 623 624 625 626 627 628 629

			recv(prc_recv[i],SEND_SPARSE + NBX_cnt,ptr_recv,sz_recv[i]);
		}

		execute();

		// Circular counter
		NBX_cnt = (NBX_cnt + 1) % 1024;
630 631
	}

incardon's avatar
incardon committed
632 633
	openfpm::vector<size_t> sz_recv_tmp;

incardon's avatar
incardon committed
634 635
	/*! \brief Send and receive multiple messages
	 *
636
	 * It send multiple messages to a set of processors the and receive
637
	 * multiple messages from another set of processors, all the processor must call this
incardon's avatar
incardon committed
638 639 640
	 * function. In this particular case the receiver know from which processor is going
	 * to receive, but does not know the size.
	 *
641 642
	 *
	 * suppose the following situation the calling processor want to communicate
incardon's avatar
incardon committed
643 644 645 646
	 * * 2 messages of size 100 byte to processor 1
	 * * 1 message of size 50 byte to processor 6
	 * * 1 message of size 48 byte to processor 7
	 * * 1 message of size 70 byte to processor 8
647
	 *
incardon's avatar
incardon committed
648
	 * \param n_send number of send for this processor [4]
649
	 *
incardon's avatar
incardon committed
650 651
	 * \param prc list of processor with which it should communicate
	 *        [1,1,6,7,8]
652
	 *
incardon's avatar
incardon committed
653 654 655 656 657 658 659
	 * \param sz the array contain the size of the message for each processor
	 *        (zeros must not be presents) [100,100,50,48,70]
	 *
	 * \param ptr array that contain the pointers to the message to send
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
	 *        for the incoming message and give back a valid pointer, supposing that this call-back has been triggered by
660 661
	 *        the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have
	 *        the following 6 parameters
incardon's avatar
incardon committed
662 663
	 *        in the call-back are in order:
	 *        * message size required to receive the message [100]
664 665
	 *        * total message size to receive from all the processors (NBX does not provide this information)
	 *        * the total number of processor want to communicate with you (NBX does not provide this information)
incardon's avatar
incardon committed
666
	 *        * processor id [5]
667
	 *        * ri request id (it is an id that goes from 0 to total_p, and is incremented
668
	 *           every time message_alloc is called)
669
	 *        * void pointer, parameter for additional data to pass to the call-back
incardon's avatar
incardon committed
670
	 *
incardon's avatar
incardon committed
671 672
	 * \param ptr_arg data passed to the call-back function specified
	 *
incardon's avatar
incardon committed
673
	 * \param opt options, NONE (ignored in this moment)
incardon's avatar
incardon committed
674 675
	 *
	 */
incardon's avatar
incardon committed
676 677 678 679
	template<typename T>
	void sendrecvMultipleMessagesNBX(openfpm::vector< size_t > & prc, openfpm::vector< T > & data,
			                         void * (* msg_alloc)(size_t,size_t,size_t,size_t,size_t,size_t,void *),
			                         void * ptr_arg, long int opt=NONE)
incardon's avatar
incardon committed
680
	{
incardon's avatar
incardon committed
681
		sz_recv_tmp.resize(n_recv);
incardon's avatar
incardon committed
682

incardon's avatar
incardon committed
683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701
		// First we understand the receive size for each processor

		for (size_t i = 0 ; i < n_send ; i++)
		{send(prc[i],SEND_SPARSE + NBX_cnt,&sz[i],sizeof(size_t));}

		for (size_t i = 0 ; i < n_recv ; i++)
		{recv(prc_recv[i],SEND_SPARSE + NBX_cnt,&sz_recv_tmp.get(i),sizeof(size_t));}

		execute();

		// Circular counter
		NBX_cnt = (NBX_cnt + 1) % 1024;

		// Allocate the buffers

		for (size_t i = 0 ; i < n_send ; i++)
		{send(prc[i],SEND_SPARSE + NBX_cnt,ptr[i],sz[i]);}

		for (size_t i = 0 ; i < n_recv ; i++)
incardon's avatar
incardon committed
702
		{
incardon's avatar
incardon committed
703
			void * ptr_recv = msg_alloc(sz_recv_tmp.get(i),0,0,prc_recv[i],i,ptr_arg);
incardon's avatar
incardon committed
704

incardon's avatar
incardon committed
705 706
			recv(prc_recv[i],SEND_SPARSE + NBX_cnt,ptr_recv,sz_recv_tmp.get(i));
		}
707

incardon's avatar
incardon committed
708
		execute();
709

incardon's avatar
incardon committed
710 711 712
		// Circular counter
		NBX_cnt = (NBX_cnt + 1) % 1024;
	}
713

714
	/*! \brief Send and receive multiple messages
715
	 *
716 717
	 * It send multiple messages to a set of processors the and receive
	 * multiple messages from another set of processors, all the processor must call this
718
	 * function
719
	 *
720 721 722 723 724
	 * suppose the following situation the calling processor want to communicate
	 * * 2 messages of size 100 byte to processor 1
	 * * 1 message of size 50 byte to processor 6
	 * * 1 message of size 48 byte to processor 7
	 * * 1 message of size 70 byte to processor 8
725
	 *
726
	 * \param n_send number of send for this processor [4]
727 728
	 *
	 * \param prc list of processor with which it should communicate
729
	 *        [1,1,6,7,8]
730
	 *
731 732 733 734
	 * \param sz the array contain the size of the message for each processor
	 *        (zeros must not be presents) [100,100,50,48,70]
	 *
	 * \param ptr array that contain the pointers to the message to send
735 736
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
737 738 739
	 *        for the incoming message and give back a valid pointer, supposing that this call-back has been triggered by
	 *        the processor of id 5 that want to communicate with me a message of size 100 byte the call-back will have
	 *        the following 6 parameters
740
	 *        in the call-back are in order:
741 742 743 744 745
	 *        * message size required to receive the message [100]
	 *        * total message size to receive from all the processors (NBX does not provide this information)
	 *        * the total number of processor want to communicate with you (NBX does not provide this information)
	 *        * processor id [5]
	 *        * ri request id (it is an id that goes from 0 to total_p, and is incremented
746
	 *           every time message_alloc is called)
747
	 *        * void pointer, parameter for additional data to pass to the call-back
748
	 *
incardon's avatar
incardon committed
749 750
	 * \param ptr_arg data passed to the call-back function specified
	 *
751 752 753
	 * \param opt options, NONE (ignored in this moment)
	 *
	 */
incardon's avatar
incardon committed
754 755 756 757
	void sendrecvMultipleMessagesNBX(size_t n_send , size_t sz[],
			                         size_t prc[] , void * ptr[],
			                         void * (* msg_alloc)(size_t,size_t,size_t,size_t,size_t,size_t,void *),
			                         void * ptr_arg, long int opt = NONE)
758
	{
759
		if (stat.size() != 0 || req.size() != 0)
incardon's avatar
incardon committed
760
			std::cerr << "Error: " << __FILE__ << ":" << __LINE__ << " this function must be called when no other requests are in progress. Please remember that if you use function like max(),sum(),send(),recv() check that you did not miss to call the function execute() \n";
761 762 763


		stat.clear();
764 765 766 767 768
		req.clear();
		// Do MPI_Issend

		for (size_t i = 0 ; i < n_send ; i++)
		{
incardon's avatar
incardon committed
769 770 771
			if (sz[i] != 0)
			{
				req.add();
772 773 774 775 776

#ifdef SE_CLASS2
				check_valid(ptr[i],sz[i]);
#endif

incardon's avatar
incardon committed
777
				tot_sent += sz[i];
incardon's avatar
incardon committed
778
				MPI_SAFE_CALL(MPI_Issend(ptr[i], sz[i], MPI_BYTE, prc[i], SEND_SPARSE + NBX_cnt*131072 + i, MPI_COMM_WORLD,&req.last()));
incardon's avatar
incardon committed
779
				log.logSend(prc[i]);
incardon's avatar
incardon committed
780
			}
781 782 783 784 785 786 787
		}

		size_t rid = 0;
		int flag = false;

		bool reached_bar_req = false;

incardon's avatar
incardon committed
788 789
		log.start(10);

790 791 792
		// Wait that all the send are acknowledge
		do
		{
incardon's avatar
incardon committed
793

794 795 796 797 798
			// flag that notify that this processor reach the barrier
			// Barrier request

			MPI_Status stat_t;
			int stat = false;
incardon's avatar
incardon committed
799
			MPI_SAFE_CALL(MPI_Iprobe(MPI_ANY_SOURCE,MPI_ANY_TAG/*SEND_SPARSE + NBX_cnt*/,MPI_COMM_WORLD,&stat,&stat_t));
800

801
			// If I have an incoming message and is related to this NBX communication
802 803 804
			if (stat == true)
			{
				int msize;
incardon's avatar
incardon committed
805 806

				// Get the message tag and size
807 808
				MPI_SAFE_CALL(MPI_Get_count(&stat_t,MPI_BYTE,&msize));

incardon's avatar
incardon committed
809 810 811 812 813
				// Ok we check if the TAG come from one of our send TAG
				if (stat_t.MPI_TAG >= (int)(SEND_SPARSE + NBX_cnt*131072) && stat_t.MPI_TAG < (int)(SEND_SPARSE + (NBX_cnt + 1)*131072))
				{
					// Get the pointer to receive the message
					void * ptr = msg_alloc(msize,0,0,stat_t.MPI_SOURCE,rid,stat_t.MPI_TAG,ptr_arg);
814

incardon's avatar
incardon committed
815 816
					// Log the receiving request
					log.logRecv(stat_t);
incardon's avatar
incardon committed
817

incardon's avatar
incardon committed
818
					rid++;
819

incardon's avatar
incardon committed
820
					// Check the pointer
821
#ifdef SE_CLASS2
incardon's avatar
incardon committed
822
					check_valid(ptr,msize);
823
#endif
incardon's avatar
incardon committed
824 825
					tot_recv += msize;
					MPI_SAFE_CALL(MPI_Recv(ptr,msize,MPI_BYTE,stat_t.MPI_SOURCE,stat_t.MPI_TAG,MPI_COMM_WORLD,&stat_t));
826 827

#ifdef SE_CLASS2
incardon's avatar
incardon committed
828
					check_valid(ptr,msize);
829
#endif
incardon's avatar
incardon committed
830
				}
831 832 833 834 835 836 837
			}

			// Check the status of all the MPI_issend and call the barrier if finished

			if (reached_bar_req == false)
			{
				int flag = false;
incardon's avatar
incardon committed
838 839 840 841
				if (req.size() != 0)
				{MPI_SAFE_CALL(MPI_Testall(req.size(),&req.get(0),&flag,MPI_STATUSES_IGNORE));}
				else
					flag = true;
842 843 844 845 846 847

				// If all send has been completed
				if (flag == true)
				{MPI_SAFE_CALL(MPI_Ibarrier(MPI_COMM_WORLD,&bar_req));reached_bar_req = true;}
			}

848
			// Check if all processor reached the async barrier
849
			if (reached_bar_req)
850
			{MPI_SAFE_CALL(MPI_Test(&bar_req,&flag,&bar_stat))};
incardon's avatar
incardon committed
851 852

			// produce a report if communication get stuck
853
			log.NBXreport(NBX_cnt,req,reached_bar_req,bar_stat);
incardon's avatar
incardon committed
854

855 856 857 858 859 860
		} while (flag == false);

		// Remove the executed request

		req.clear();
		stat.clear();
incardon's avatar
incardon committed
861
		log.clear();
862 863 864

		// Circular counter
		NBX_cnt = (NBX_cnt + 1) % 1024;
incardon's avatar
incardon committed
865 866
	}

867 868 869 870 871 872
	/*! \brief Send data to a processor
	 *
	 * \warning In order to avoid deadlock every send must be coupled with a recv
	 *          in case you want to send data without knowledge from the other side
	 *          consider to use sendRecvMultipleMessages
	 *
incardon's avatar
incardon committed
873 874 875 876 877 878 879 880 881 882 883 884
	 * \warning operation is asynchronous execute must be called to ensure they are executed
	 *
	 * \see sendRecvMultipleMessages
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param mem buffer with the data to send
	 * \param sz size
	 *
	 * \return true if succeed false otherwise
	 *
	 */
885
	bool send(size_t proc, size_t tag, const void * mem, size_t sz)
incardon's avatar
incardon committed
886 887 888 889 890 891 892 893 894 895 896 897
	{
		// send over MPI

		// Create one request
		req.add();

		// send
		MPI_IsendWB::send(proc,SEND_RECV_BASE + tag,mem,sz,req.last());

		return true;
	}

incardon's avatar
incardon committed
898

incardon's avatar
incardon committed
899 900 901 902 903 904 905
	/*! \brief Send data to a processor
	 *
	 * \warning In order to avoid deadlock every send must be coupled with a recv
	 *          in case you want to send data without knowledge from the other side
	 *          consider to use sendRecvMultipleMessages
	 *
	 * \warning operation is asynchronous execute must be called to ensure they are executed
906 907 908 909 910 911 912
	 *
	 * \see sendRecvMultipleMessages
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param v buffer to send
	 *
913 914
	 * \return true if succeed false otherwise
	 *
915
	 */
916
	template<typename T, typename Mem, typename gr> bool send(size_t proc, size_t tag, openfpm::vector<T,Mem,gr> & v)
917
	{
incardon's avatar
incardon committed
918
#ifdef SE_CLASS1
919 920 921 922 923 924 925 926
		checkType<T>();
#endif

		// send over MPI

		// Create one request
		req.add();

927
		// send
928
		MPI_IsendW<T,Mem,gr>::send(proc,SEND_RECV_BASE + tag,v,req.last());
929 930

		return true;
931 932 933 934 935 936
	}

	/*! \brief Recv data from a processor
	 *
	 * \warning In order to avoid deadlock every recv must be coupled with a send
	 *          in case you want to send data without knowledge from the other side
937
	 *          consider to use or sendrecvMultipleMessagesNBX
938
	 *
incardon's avatar
incardon committed
939
	 * \warning operation is asynchronous execute must be called to ensure they are executed
940
	 *
941
	 * \see sendrecvMultipleMessagesNBX
942 943 944 945
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param v buffer to send
incardon's avatar
incardon committed
946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963
	 * \param sz size of the buffer
	 *
	 * \return true if succeed false otherwise
	 *
	 */
	bool recv(size_t proc, size_t tag, void * v, size_t sz)
	{
		// recv over MPI

		// Create one request
		req.add();

		// receive
		MPI_IrecvWB::recv(proc,SEND_RECV_BASE + tag,v,sz,req.last());

		return true;
	}

incardon's avatar
incardon committed
964 965 966 967
    /*! \brief Recv data from a processor
     *
     * \warning In order to avoid deadlock every recv must be coupled with a send
     *          in case you want to send data without knowledge from the other side
968
     *          consider to use sendrecvMultipleMessagesNBX
incardon's avatar
incardon committed
969 970 971
     *
     * \warning operation is asynchronous execute must be called to ensure they are executed
     *
972
     * \see sendrecvMultipleMessagesNBX
incardon's avatar
incardon committed
973 974 975 976 977 978 979 980 981 982
     *
     * \param proc processor id
     * \param tag id
     * \param v vector to send
     *
     * \return true if succeed false otherwise
     *
     */
    template<typename T, typename Mem, typename gr> bool recv(size_t proc, size_t tag, openfpm::vector<T,Mem,gr> & v)
    {
incardon's avatar
incardon committed
983
#ifdef SE_CLASS1
incardon's avatar
incardon committed
984 985 986 987 988 989 990 991 992 993 994 995 996 997 998
            checkType<T>();
#endif

            // recv over MPI

            // Create one request
            req.add();

            // receive
            MPI_IrecvW<T>::recv(proc,SEND_RECV_BASE + tag,v,req.last());

            return true;
    }

	/*! \brief Gather the data from all processors
999 1000
	 *
	 * send a primitive data T receive the same primitive T from all the other processors
incardon's avatar
incardon committed
1001 1002 1003
	 *
	 * \warning operation is asynchronous execute must be called to ensure they are executed
	 *
1004
	 * \param v vector to receive (automaticaly resized)
incardon's avatar
incardon committed
1005
	 * \param send data to send
1006
	 *
1007 1008
	 * \return true if succeed false otherwise
	 *
1009
	 */
incardon's avatar
incardon committed
1010
	template<typename T, typename Mem, typename gr> bool allGather(T & send, openfpm::vector<T,Mem,gr> & v)
1011
	{
incardon's avatar
incardon committed
1012
#ifdef SE_CLASS1
1013 1014 1015 1016 1017 1018
		checkType<T>();
#endif

		// Create one request
		req.add();

incardon's avatar
incardon committed
1019 1020 1021
		// Number of processors
		v.resize(getProcessingUnits());

1022
		// gather
incardon's avatar
incardon committed
1023
		MPI_IAllGatherW<T>::gather(&send,1,v.getPointer(),1,req.last());
1024 1025

		return true;
1026 1027
	}

incardon's avatar
incardon committed
1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058
	/*! \brief Broadcast the data to all processors
	 *
	 * broadcast a vector of primitives.
	 *
	 * \warning operation is asynchronous execute must be called to ensure the operation is executed
	 *
	 * \warning the non-root processor must resize the vector to the exact receive size. This mean the
	 *          each processor must known a priory the receiving size
	 *
	 * \param v vector to send in the case of the root processor and vector where to receive in the case of
	 *          non-root
	 * \param root processor (who broadcast)
	 *
	 * \return true if succeed false otherwise
	 *
	 */
	template<typename T, typename Mem, typename gr> bool Bcast(openfpm::vector<T,Mem,gr> & v, size_t root)
	{
#ifdef SE_CLASS1
		checkType<T>();
#endif

		// Create one request
		req.add();

		// gather
		MPI_IBcastW<T>::bcast(root,v,req.last());

		return true;
	}

1059
	/*! \brief Execute all the requests
incardon's avatar
incardon committed
1060 1061 1062 1063
	 *
	 */
	void execute()
	{
1064 1065 1066 1067
		// if req == 0 return
		if (req.size() == 0)
			return;

incardon's avatar
incardon committed
1068
		// Wait for all the requests
1069
		stat.resize(req.size());
1070
		MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));
incardon's avatar
incardon committed
1071

1072
		// Remove executed request and status
incardon's avatar
incardon committed
1073 1074 1075
		req.clear();
		stat.clear();
	}
1076 1077 1078 1079 1080 1081 1082 1083 1084

	/*! \brief Release the buffer used for communication
	 *
	 *
	 */
	void clear()
	{
		recv_buf.clear();
	}
incardon's avatar
incardon committed
1085 1086
};

1087

Pietro Incardona's avatar
Pietro Incardona committed
1088

1089

incardon's avatar
incardon committed
1090 1091
#endif