VCluster_base.hpp 32.6 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"
23 24 25
#if defined(CUDA_GPU) && defined(__NVCC__)
#include "util/cuda/moderngpu/launch_box.hxx"
#endif
incardon's avatar
incardon committed
26

Pietro Incardona's avatar
Pietro Incardona committed
27 28 29 30
#ifdef HAVE_PETSC
#include <petscvec.h>
#endif

incardon's avatar
incardon committed
31 32
#define MSG_LENGTH 1024
#define MSG_SEND_RECV 1025
33
#define SEND_SPARSE 4096
incardon's avatar
incardon committed
34 35 36
#define NONE 1
#define NEED_ALL_SIZE 2

37 38
#define SERIVCE_MESSAGE_TAG 16384
#define SEND_RECV_BASE 8192
incardon's avatar
incardon committed
39
#define GATHER_BASE 24576
incardon's avatar
incardon committed
40

41 42
#define RECEIVE_KNOWN 4
#define KNOWN_ELEMENT_OR_BYTE 8
43
#define MPI_GPU_DIRECT 16
44

45
// number of vcluster instances
incardon's avatar
incardon committed
46
extern size_t n_vcluster;
47
// Global MPI initialization
incardon's avatar
incardon committed
48
extern bool global_mpi_init;
Pietro Incardona's avatar
Pietro Incardona committed
49 50
// initialization flag
extern bool ofp_initialized;
incardon's avatar
incardon committed
51 52
extern size_t tot_sent;
extern size_t tot_recv;
incardon's avatar
incardon committed
53

incardon's avatar
incardon committed
54 55 56 57 58 59 60 61
///////////////////// Post functions /////////////

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


incardon's avatar
incardon committed
62
//! temporal buffer for reductions
incardon's avatar
incardon committed
63 64
union red
{
incardon's avatar
incardon committed
65
	//! char
incardon's avatar
incardon committed
66
	char c;
incardon's avatar
incardon committed
67
	//! unsigned char
incardon's avatar
incardon committed
68
	unsigned char uc;
incardon's avatar
incardon committed
69
	//! signed
incardon's avatar
incardon committed
70
	short s;
incardon's avatar
incardon committed
71
	//! unsigned short
incardon's avatar
incardon committed
72
	unsigned short us;
incardon's avatar
incardon committed
73
	//! integer
incardon's avatar
incardon committed
74
	int i;
incardon's avatar
incardon committed
75
	//! unsigned integer
incardon's avatar
incardon committed
76
	unsigned int ui;
incardon's avatar
incardon committed
77
	//! float
incardon's avatar
incardon committed
78
	float f;
incardon's avatar
incardon committed
79
	//! double
incardon's avatar
incardon committed
80 81 82
	double d;
};

83
/*! \brief This class virtualize the cluster of PC as a set of processes that communicate
incardon's avatar
incardon committed
84
 *
85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103
 * 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
104 105 106
 *
 */

incardon's avatar
incardon committed
107
class Vcluster_base
incardon's avatar
incardon committed
108
{
incardon's avatar
incardon committed
109
	//! log file
incardon's avatar
incardon committed
110 111
	Vcluster_log log;

112 113 114
	//! 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
115
	//! an acknowledge notification on the sending processor.
116 117 118 119 120 121
	//! 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
122 123 124 125
	//! 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
126
	size_t NBX_cnt;
127

incardon's avatar
incardon committed
128 129
	//! temporal vector used for meta-communication
	//! ( or meta-data before the real communication )
incardon's avatar
incardon committed
130 131
	openfpm::vector<size_t> proc_com;

incardon's avatar
incardon committed
132
	//! vector that contain the scatter map (it is basically an array of one)
incardon's avatar
incardon committed
133 134
	openfpm::vector<int> map_scatter;

incardon's avatar
incardon committed
135
	//! vector of MPI requests
incardon's avatar
incardon committed
136 137
	openfpm::vector<MPI_Request> req;

incardon's avatar
incardon committed
138
	//! vector of MPI status
incardon's avatar
incardon committed
139 140
	openfpm::vector<MPI_Status> stat;

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

144 145 146 147 148 149 150 151 152 153 154
#if defined(CUDA_GPU) && defined(__NVCC__)

	//! standard context for mgpu
	mgpu::standard_context_t * context;

#else

	void * context = NULL;

#endif

incardon's avatar
incardon committed
155 156 157 158 159
	// Object array


	// Single objects

incardon's avatar
incardon committed
160
	//! number of processes
incardon's avatar
incardon committed
161
	int m_size;
incardon's avatar
incardon committed
162
	//! actual rank
incardon's avatar
incardon committed
163
	int m_rank;
incardon's avatar
incardon committed
164

incardon's avatar
incardon committed
165
	//! number of processing unit per process
incardon's avatar
incardon committed
166 167
	int numPE = 1;

incardon's avatar
incardon committed
168
	/*! This buffer is a temporal buffer for reductions
incardon's avatar
incardon committed
169 170 171 172 173 174 175 176
	 *
	 * 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
177
	//! vector of pointers of send buffers
178 179
	openfpm::vector<void *> ptr_send;

incardon's avatar
incardon committed
180
	//! vector of the size of send buffers
181 182
	openfpm::vector<size_t> sz_send;

incardon's avatar
incardon committed
183
	//! barrier request
184
	MPI_Request bar_req;
incardon's avatar
incardon committed
185 186

	//! barrier status
187 188
	MPI_Status bar_stat;

incardon's avatar
incardon committed
189
	//! disable operator=
incardon's avatar
incardon committed
190
	Vcluster_base & operator=(const Vcluster_base &)	{return *this;};
Pietro Incardona's avatar
Pietro Incardona committed
191

incardon's avatar
incardon committed
192
	//! disable copy constructor
incardon's avatar
incardon committed
193
	Vcluster_base(const Vcluster_base &)
incardon's avatar
incardon committed
194 195
	:NBX_cnt(0)
	{};
Pietro Incardona's avatar
Pietro Incardona committed
196

incardon's avatar
incardon committed
197 198 199 200 201
protected:

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

incardon's avatar
incardon committed
202 203 204
	//! tags receiving
	openfpm::vector<size_t> tags;

incardon's avatar
incardon committed
205 206 207
public:

	// Finalize the MPI program
incardon's avatar
incardon committed
208
	~Vcluster_base()
incardon's avatar
incardon committed
209
	{
210 211 212
#ifdef SE_CLASS2
		check_delete(this);
#endif
incardon's avatar
incardon committed
213 214 215 216
		n_vcluster--;

		// if there are no other vcluster instances finalize
		if (n_vcluster == 0)
217
		{
218 219 220 221
			int already_finalised;

			MPI_Finalized(&already_finalised);
			if (!already_finalised)
222
			{
223 224 225 226
				if (MPI_Finalize() != 0)
				{
					std::cerr << __FILE__ << ":" << __LINE__  << " MPI_Finalize FAILED \n";
				}
227
			}
228
		}
229 230 231 232 233 234

#if defined(CUDA_GPU) && defined(__NVCC__)

		delete context;

#endif
incardon's avatar
incardon committed
235 236
	}

237 238 239 240 241 242
	/*! \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
243
	Vcluster_base(int *argc, char ***argv)
incardon's avatar
incardon committed
244
	:NBX_cnt(0)
incardon's avatar
incardon committed
245
	{
246 247 248 249
#ifdef SE_CLASS2
		check_new(this,8,VCLUSTER_EVENT,PRJ_VCLUSTER);
#endif

incardon's avatar
incardon committed
250 251
		n_vcluster++;

252 253 254
		int already_initialised;
		MPI_Initialized(&already_initialised);

incardon's avatar
incardon committed
255
		// Check if MPI is already initialized
256
		if (!already_initialised)
incardon's avatar
incardon committed
257
		{
incardon's avatar
incardon committed
258

incardon's avatar
incardon committed
259 260
			MPI_Init(argc,argv);
		}
incardon's avatar
incardon committed
261

262 263
		// Get the total number of process
		// and the rank of this process
incardon's avatar
incardon committed
264

incardon's avatar
incardon committed
265 266
		MPI_Comm_size(MPI_COMM_WORLD, &m_size);
		MPI_Comm_rank(MPI_COMM_WORLD, &m_rank);
incardon's avatar
incardon committed
267

268
#ifdef SE_CLASS2
incardon's avatar
incardon committed
269
			process_v_cl = m_rank;
incardon's avatar
incardon committed
270 271
#endif

272
		// create and fill map scatter with one
incardon's avatar
incardon committed
273
		map_scatter.resize(m_size);
incardon's avatar
incardon committed
274 275 276 277 278

		for (size_t i = 0 ; i < map_scatter.size() ; i++)
		{
			map_scatter.get(i) = 1;
		}
incardon's avatar
incardon committed
279 280

		// open the log file
incardon's avatar
incardon committed
281
		log.openLog(m_rank);
282 283 284 285

		// Initialize bar_req
		bar_req = MPI_Request();
		bar_stat = MPI_Status();
286 287 288 289 290 291

#if defined(CUDA_GPU) && defined(__NVCC__)

		context = new mgpu::standard_context_t();

#endif
incardon's avatar
incardon committed
292 293
	}

incardon's avatar
incardon committed
294
#ifdef SE_CLASS1
295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329

	/*! \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
330
				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" ;
331 332 333 334 335 336 337 338 339 340 341 342 343 344
				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:
			{

			}
		}
	}

345 346 347 348 349 350 351 352 353 354 355 356 357
#endif

#if defined(CUDA_GPU) && defined(__NVCC__)

	/*! \brief If nvidia cuda is activated return a mgpu context
	 *
	 *
	 */
	mgpu::standard_context_t & getmgpuContext()
	{
		return *context;
	}

358 359
#endif

Pietro Incardona's avatar
Pietro Incardona committed
360 361 362 363 364 365 366 367 368 369
	/*! \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
370 371 372 373 374
	/*! \brief Get the total number of processors
	 *
	 * \return the total number of processors
	 *
	 */
incardon's avatar
incardon committed
375 376
	size_t getProcessingUnits()
	{
incardon's avatar
incardon committed
377 378 379 380 381 382 383 384 385 386 387 388 389 390 391
		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
392 393
	}

incardon's avatar
incardon committed
394 395 396 397 398
	/*! \brief Get the process unit id
	 *
	 * \return the process ID
	 *
	 */
incardon's avatar
incardon committed
399 400
	size_t getProcessUnitID()
	{
incardon's avatar
incardon committed
401 402 403 404 405 406 407 408 409 410 411 412 413 414 415
		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
416 417
	}

incardon's avatar
incardon committed
418

419
	/*! \brief Sum the numbers across all processors and get the result
incardon's avatar
incardon committed
420 421 422 423 424
	 *
	 * \param num to reduce, input and output
	 *
	 */

425
	template<typename T> void sum(T & num)
incardon's avatar
incardon committed
426
	{
incardon's avatar
incardon committed
427
#ifdef SE_CLASS1
428 429 430
		checkType<T>();
#endif

incardon's avatar
incardon committed
431 432 433 434 435 436 437 438 439
		// reduce over MPI

		// Create one request
		req.add();

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

440
	/*! \brief Get the maximum number across all processors (or reduction with infinity norm)
incardon's avatar
incardon committed
441 442 443 444 445 446
	 *
	 * \param num to reduce
	 *
	 */
	template<typename T> void max(T & num)
	{
incardon's avatar
incardon committed
447
#ifdef SE_CLASS1
448 449
		checkType<T>();
#endif
incardon's avatar
incardon committed
450 451 452 453 454 455 456 457 458
		// reduce over MPI

		// Create one request
		req.add();

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

tonynsyde's avatar
tonynsyde committed
459 460 461 462 463 464 465 466
	/*! \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
467
#ifdef SE_CLASS1
tonynsyde's avatar
tonynsyde committed
468 469 470 471 472 473 474 475 476 477 478
		checkType<T>();
#endif
		// reduce over MPI

		// Create one request
		req.add();

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

479
	/*! \brief Send and receive multiple messages
480
	 *
481 482
	 * 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
483 484 485
	 * function. In this particular case the receiver know from which processor is going
	 * to receive.
	 *
486 487
	 *
	 * suppose the following situation the calling processor want to communicate
488 489 490 491
	 * * 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
492 493
	 *
	 *
494 495
	 * \param prc list of processor with which it should communicate
	 *        [1,1,6,7,8]
496
	 *
incardon's avatar
incardon committed
497 498 499 500
	 * \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
501
	 *
incardon's avatar
incardon committed
502
	 * \param recv_sz for each processor indicate the size of the data received
503 504 505
	 *
	 * \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
506 507
	 *        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
508 509
	 *        in the call-back are in order:
	 *        * message size required to receive the message [100]
510 511
	 *        * 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)
512
	 *        * processor id [5]
513
	 *        * ri request id (it is an id that goes from 0 to total_p, and is incremented
514
	 *           every time message_alloc is called)
515
	 *        * void pointer, parameter for additional data to pass to the call-back
516
	 *
incardon's avatar
incardon committed
517 518
	 * \param ptr_arg data passed to the call-back function specified
	 *
519
	 * \param opt options, NONE (ignored in this moment)
520 521
	 *
	 */
incardon's avatar
incardon committed
522 523 524 525
	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
526
														  void * (* msg_alloc)(size_t,size_t,size_t,size_t,size_t,size_t,void *),
incardon's avatar
incardon committed
527 528
														  void * ptr_arg,
														  long int opt=NONE)
529
	{
530
		// Allocate the buffers
531 532

		for (size_t i = 0 ; i < prc.size() ; i++)
incardon's avatar
incardon committed
533
		{send(prc.get(i),SEND_SPARSE + NBX_cnt,data.get(i).getPointer(),data.get(i).size());}
534 535

		for (size_t i = 0 ; i < prc_recv.size() ; i++)
536
		{
incardon's avatar
incardon committed
537
			void * ptr_recv = msg_alloc(recv_sz.get(i),0,0,prc_recv.get(i),i,SEND_SPARSE + NBX_cnt,ptr_arg);
538 539

			recv(prc_recv.get(i),SEND_SPARSE + NBX_cnt,ptr_recv,recv_sz.get(i));
540 541
		}

542 543 544 545 546 547
		execute();

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

incardon's avatar
incardon committed
548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586

	/*! \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,
587
									 void * (* msg_alloc)(size_t,size_t,size_t,size_t,size_t,size_t,void *),
incardon's avatar
incardon committed
588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605
									 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);
	}

606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647
	/*! \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
648 649 650 651 652
	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)
653 654 655 656 657 658 659 660
	{
		// 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
661
			void * ptr_recv = msg_alloc(sz_recv[i],0,0,prc_recv[i],i,SEND_SPARSE + NBX_cnt,ptr_arg);
662 663 664 665 666 667 668 669

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

		execute();

		// Circular counter
		NBX_cnt = (NBX_cnt + 1) % 1024;
670 671
	}

incardon's avatar
incardon committed
672 673
	openfpm::vector<size_t> sz_recv_tmp;

incardon's avatar
incardon committed
674 675
	/*! \brief Send and receive multiple messages
	 *
676
	 * It send multiple messages to a set of processors the and receive
677
	 * multiple messages from another set of processors, all the processor must call this
incardon's avatar
incardon committed
678 679 680
	 * function. In this particular case the receiver know from which processor is going
	 * to receive, but does not know the size.
	 *
681 682
	 *
	 * suppose the following situation the calling processor want to communicate
incardon's avatar
incardon committed
683 684 685 686
	 * * 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
687
	 *
incardon's avatar
incardon committed
688
	 * \param n_send number of send for this processor [4]
689
	 *
incardon's avatar
incardon committed
690 691
	 * \param prc list of processor with which it should communicate
	 *        [1,1,6,7,8]
692
	 *
incardon's avatar
incardon committed
693 694 695 696 697 698 699
	 * \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
700 701
	 *        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
702 703
	 *        in the call-back are in order:
	 *        * message size required to receive the message [100]
704 705
	 *        * 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
706
	 *        * processor id [5]
707
	 *        * ri request id (it is an id that goes from 0 to total_p, and is incremented
708
	 *           every time message_alloc is called)
709
	 *        * void pointer, parameter for additional data to pass to the call-back
incardon's avatar
incardon committed
710
	 *
incardon's avatar
incardon committed
711 712
	 * \param ptr_arg data passed to the call-back function specified
	 *
incardon's avatar
incardon committed
713
	 * \param opt options, NONE (ignored in this moment)
incardon's avatar
incardon committed
714 715
	 *
	 */
716 717 718 719
	void sendrecvMultipleMessagesNBX(size_t n_send , size_t sz[], size_t prc[] ,
									 void * ptr[], size_t n_recv, size_t prc_recv[] ,
									 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
720
	{
incardon's avatar
incardon committed
721
		sz_recv_tmp.resize(n_recv);
incardon's avatar
incardon committed
722

incardon's avatar
incardon committed
723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741
		// 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
742
		{
743
			void * ptr_recv = msg_alloc(sz_recv_tmp.get(i),0,0,prc_recv[i],i,0,ptr_arg);
incardon's avatar
incardon committed
744

incardon's avatar
incardon committed
745 746
			recv(prc_recv[i],SEND_SPARSE + NBX_cnt,ptr_recv,sz_recv_tmp.get(i));
		}
747

incardon's avatar
incardon committed
748
		execute();
749

incardon's avatar
incardon committed
750 751 752
		// Circular counter
		NBX_cnt = (NBX_cnt + 1) % 1024;
	}
753

754
	/*! \brief Send and receive multiple messages
755
	 *
756 757
	 * 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
758
	 * function
759
	 *
760 761 762 763 764
	 * 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
765
	 *
766
	 * \param n_send number of send for this processor [4]
767 768
	 *
	 * \param prc list of processor with which it should communicate
769
	 *        [1,1,6,7,8]
770
	 *
771 772 773 774
	 * \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
775 776
	 *
	 * \param msg_alloc This is a call-back with the purpose of allocate space
777 778 779
	 *        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
780
	 *        in the call-back are in order:
781 782 783 784 785
	 *        * 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
786
	 *           every time message_alloc is called)
787
	 *        * void pointer, parameter for additional data to pass to the call-back
788
	 *
incardon's avatar
incardon committed
789 790
	 * \param ptr_arg data passed to the call-back function specified
	 *
791 792 793
	 * \param opt options, NONE (ignored in this moment)
	 *
	 */
incardon's avatar
incardon committed
794 795 796 797
	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)
798
	{
799
		if (stat.size() != 0 || req.size() != 0)
incardon's avatar
incardon committed
800
			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";
801 802 803


		stat.clear();
804 805 806 807 808
		req.clear();
		// Do MPI_Issend

		for (size_t i = 0 ; i < n_send ; i++)
		{
incardon's avatar
incardon committed
809 810 811
			if (sz[i] != 0)
			{
				req.add();
812 813 814 815 816

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

incardon's avatar
incardon committed
817
				tot_sent += sz[i];
incardon's avatar
incardon committed
818
				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
819
				log.logSend(prc[i]);
incardon's avatar
incardon committed
820
			}
821 822 823 824 825 826 827
		}

		size_t rid = 0;
		int flag = false;

		bool reached_bar_req = false;

incardon's avatar
incardon committed
828 829
		log.start(10);

830 831 832
		// Wait that all the send are acknowledge
		do
		{
incardon's avatar
incardon committed
833

834 835 836 837 838
			// flag that notify that this processor reach the barrier
			// Barrier request

			MPI_Status stat_t;
			int stat = false;
incardon's avatar
incardon committed
839
			MPI_SAFE_CALL(MPI_Iprobe(MPI_ANY_SOURCE,MPI_ANY_TAG/*SEND_SPARSE + NBX_cnt*/,MPI_COMM_WORLD,&stat,&stat_t));
840

841
			// If I have an incoming message and is related to this NBX communication
842 843 844
			if (stat == true)
			{
				int msize;
incardon's avatar
incardon committed
845 846

				// Get the message tag and size
847 848
				MPI_SAFE_CALL(MPI_Get_count(&stat_t,MPI_BYTE,&msize));

incardon's avatar
incardon committed
849 850 851 852 853
				// 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);
854

incardon's avatar
incardon committed
855 856
					// Log the receiving request
					log.logRecv(stat_t);
incardon's avatar
incardon committed
857

incardon's avatar
incardon committed
858
					rid++;
859

incardon's avatar
incardon committed
860
					// Check the pointer
861
#ifdef SE_CLASS2
incardon's avatar
incardon committed
862
					check_valid(ptr,msize);
863
#endif
incardon's avatar
incardon committed
864 865
					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));
866 867

#ifdef SE_CLASS2
incardon's avatar
incardon committed
868
					check_valid(ptr,msize);
869
#endif
incardon's avatar
incardon committed
870
				}
871 872 873 874 875 876 877
			}

			// 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
878 879 880 881
				if (req.size() != 0)
				{MPI_SAFE_CALL(MPI_Testall(req.size(),&req.get(0),&flag,MPI_STATUSES_IGNORE));}
				else
					flag = true;
882 883 884 885 886 887

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

888
			// Check if all processor reached the async barrier
889
			if (reached_bar_req)
890
			{MPI_SAFE_CALL(MPI_Test(&bar_req,&flag,&bar_stat))};
incardon's avatar
incardon committed
891 892

			// produce a report if communication get stuck
893
			log.NBXreport(NBX_cnt,req,reached_bar_req,bar_stat);
incardon's avatar
incardon committed
894

895 896 897 898 899 900
		} while (flag == false);

		// Remove the executed request

		req.clear();
		stat.clear();
incardon's avatar
incardon committed
901
		log.clear();
902 903 904

		// Circular counter
		NBX_cnt = (NBX_cnt + 1) % 1024;
incardon's avatar
incardon committed
905 906
	}

907 908 909 910 911 912
	/*! \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
913 914 915 916 917 918 919 920 921 922 923 924
	 * \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
	 *
	 */
925
	bool send(size_t proc, size_t tag, const void * mem, size_t sz)
incardon's avatar
incardon committed
926 927 928 929 930 931 932 933 934 935 936 937
	{
		// 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
938

incardon's avatar
incardon committed
939 940 941 942 943 944 945
	/*! \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
946 947 948 949 950 951 952
	 *
	 * \see sendRecvMultipleMessages
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param v buffer to send
	 *
953 954
	 * \return true if succeed false otherwise
	 *
955
	 */
956
	template<typename T, typename Mem, typename gr> bool send(size_t proc, size_t tag, openfpm::vector<T,Mem,gr> & v)
957
	{
incardon's avatar
incardon committed
958
#ifdef SE_CLASS1
959 960 961 962 963 964 965 966
		checkType<T>();
#endif

		// send over MPI

		// Create one request
		req.add();

967
		// send
968
		MPI_IsendW<T,Mem,gr>::send(proc,SEND_RECV_BASE + tag,v,req.last());
969 970

		return true;
971 972 973 974 975 976
	}

	/*! \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
977
	 *          consider to use or sendrecvMultipleMessagesNBX
978
	 *
incardon's avatar
incardon committed
979
	 * \warning operation is asynchronous execute must be called to ensure they are executed
980
	 *
981
	 * \see sendrecvMultipleMessagesNBX
982 983 984 985
	 *
	 * \param proc processor id
	 * \param tag id
	 * \param v buffer to send
incardon's avatar
incardon committed
986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003
	 * \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
1004 1005 1006 1007
    /*! \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
1008
     *          consider to use sendrecvMultipleMessagesNBX
incardon's avatar
incardon committed
1009 1010 1011
     *
     * \warning operation is asynchronous execute must be called to ensure they are executed
     *
1012
     * \see sendrecvMultipleMessagesNBX
incardon's avatar
incardon committed
1013 1014 1015 1016 1017 1018 1019 1020 1021 1022
     *
     * \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
1023
#ifdef SE_CLASS1
incardon's avatar
incardon committed
1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
            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
1039 1040
	 *
	 * send a primitive data T receive the same primitive T from all the other processors
incardon's avatar
incardon committed
1041 1042 1043
	 *
	 * \warning operation is asynchronous execute must be called to ensure they are executed
	 *
1044
	 * \param v vector to receive (automaticaly resized)
incardon's avatar
incardon committed
1045
	 * \param send data to send
1046
	 *
1047 1048
	 * \return true if succeed false otherwise
	 *
1049
	 */
incardon's avatar
incardon committed
1050
	template<typename T, typename Mem, typename gr> bool allGather(T & send, openfpm::vector<T,Mem,gr> & v)
1051
	{
incardon's avatar
incardon committed
1052
#ifdef SE_CLASS1
1053 1054 1055 1056 1057 1058
		checkType<T>();
#endif

		// Create one request
		req.add();

incardon's avatar
incardon committed
1059 1060 1061
		// Number of processors
		v.resize(getProcessingUnits());

1062
		// gather
incardon's avatar
incardon committed
1063
		MPI_IAllGatherW<T>::gather(&send,1,v.getPointer(),1,req.last());
1064 1065

		return true;
1066 1067
	}

incardon's avatar
incardon committed
1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083
	/*! \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
	 *
	 */
1084 1085
	template<typename T, typename Mem, typename lt_type, template<typename> class layout_base >
	bool Bcast(openfpm::vector<T,Mem,lt_type,layout_base> & v, size_t root)
incardon's avatar
incardon committed
1086 1087 1088 1089 1090
	{
#ifdef SE_CLASS1
		checkType<T>();
#endif

1091
		b_cast_helper<openfpm::vect_isel<T>::value == STD_VECTOR || is_layout_mlin<layout_base<T>>::value >::bcast_(req,v,root);
incardon's avatar
incardon committed
1092 1093 1094 1095

		return true;
	}

1096
	/*! \brief Execute all the requests
incardon's avatar
incardon committed
1097 1098 1099 1100
	 *
	 */
	void execute()
	{
1101 1102 1103 1104
		// if req == 0 return
		if (req.size() == 0)
			return;

incardon's avatar
incardon committed
1105
		// Wait for all the requests
1106
		stat.resize(req.size());
1107
		MPI_SAFE_CALL(MPI_Waitall(req.size(),&req.get(0),&stat.get(0)));
incardon's avatar
incardon committed
1108

1109
		// Remove executed request and status
incardon's avatar
incardon committed
1110 1111 1112
		req.clear();
		stat.clear();
	}
1113 1114 1115 1116 1117 1118 1119 1120 1121

	/*! \brief Release the buffer used for communication
	 *
	 *
	 */
	void clear()
	{
		recv_buf.clear();
	}
incardon's avatar
incardon committed
1122 1123
};

1124

Pietro Incardona's avatar
Pietro Incardona committed
1125

1126

incardon's avatar
incardon committed
1127 1128
#endif