CartDecomposition.hpp 27.3 KB
Newer Older
incardon's avatar
incardon committed
1 2 3
/*
 * CartDecomposition.hpp
 *
4 5
 *  Created on: Oct 07, 2015
 *      Author: Antonio Leo
incardon's avatar
incardon committed
6 7 8 9 10 11
 */

#ifndef CARTDECOMPOSITION_HPP
#define CARTDECOMPOSITION_HPP

#include "config.h"
12
#include <cmath>
13 14
#include "VCluster.hpp"
#include "Graph/CartesianGraphFactory.hpp"
tonynsyde's avatar
tonynsyde committed
15
#include "Graph/DistCartesianGraphFactory.hpp"
incardon's avatar
incardon committed
16
#include "Decomposition.hpp"
incardon's avatar
incardon committed
17
#include "Vector/map_vector.hpp"
incardon's avatar
incardon committed
18 19 20 21 22
#include <vector>
#include <initializer_list>
#include "SubdomainGraphNodes.hpp"
#include "dec_optimizer.hpp"
#include "Space/Shape/Box.hpp"
incardon's avatar
incardon committed
23
#include "Space/Shape/Point.hpp"
incardon's avatar
incardon committed
24
#include "NN/CellList/CellDecomposer.hpp"
incardon's avatar
incardon committed
25 26
#include <unordered_map>
#include "NN/CellList/CellList.hpp"
incardon's avatar
incardon committed
27
#include "Space/Ghost.hpp"
incardon's avatar
incardon committed
28 29
#include "common.hpp"
#include "ie_loc_ghost.hpp"
30 31
#include "ie_ghost.hpp"
#include "nn_processor.hpp"
32 33
#include "GraphMLWriter.hpp"
#include "ParMetisDistribution.hpp"
tonynsyde's avatar
tonynsyde committed
34
#include "DistParMetisDistribution.hpp"
35 36
#include "MetisDistribution.hpp"
#include "DLB.hpp"
incardon's avatar
incardon committed
37

38 39 40 41 42 43 44 45 46 47 48
#define CARTDEC_ERROR 2000lu

// Macro that decide what to do in case of error
#ifdef STOP_ON_ERROR
#define ACTION_ON_ERROR() exit(1);
#elif defined(THROW_ON_ERROR)
#define ACTION_ON_ERROR() throw CARTDEC_ERROR;
#else
#define ACTION_ON_ERROR()
#endif

incardon's avatar
incardon committed
49 50 51 52 53 54 55
/**
 * \brief This class decompose a space into subspaces
 *
 * \tparam dim is the dimensionality of the physical domain we are going to decompose.
 * \tparam T type of the space we decompose, Real, Integer, Complex ...
 * \tparam Memory Memory factory used to allocate memory
 * \tparam Domain Structure that contain the information of your physical domain
tonynsyde's avatar
tonynsyde committed
56
 * \tparam Distribution type of distribution, can be ParMetisDistribution or MetisDistribution
incardon's avatar
incardon committed
57
 *
58 59
 * Given an N-dimensional space, this class decompose the space into a Cartesian grid of small
 * sub-sub-domain. At each sub-sub-domain is assigned  an id that identify which processor is
incardon's avatar
incardon committed
60
 * going to take care of that part of space (in general the space assigned to a processor is
61 62 63 64 65 66
 * simply connected), a second step merge several sub-sub-domain with same id into bigger region
 *  sub-domain with the id. Each sub-domain has an extended space called ghost part
 *
 * Assuming that VCluster.getProcessUnitID(), equivalent to the MPI processor rank, return the processor local
 * processor id, we define
 *
67 68
 * * local processor: processor rank
 * * local sub-domain: sub-domain given to the local processor
69 70 71 72 73 74 75 76 77 78 79 80
 * * external ghost box: (or ghost box) are the boxes that compose the ghost space of the processor, or the
 *   boxes produced expanding every local sub-domain by the ghost extension and intersecting with the sub-domain
 *   of the other processors
 * * Near processors are the processors adjacent to the local processor, where with adjacent we mean all the processor
 *   that has a non-zero intersection with the ghost part of the local processor, or all the processors that
 *   produce non-zero external boxes with the local processor, or all the processor that should communicate
 *   in case of ghost data synchronization
 * * internal ghost box: is the part of ghost of the near processor that intersect the space of the
 *       processor, or the boxes produced expanding the sub-domain of the near processors with the local sub-domain
 * * Near processor sub-domain: is a sub-domain that live in the a near (or contiguous) processor
 * * Near processor list: the list of all the near processor of the local processor (each processor has a list
 *                        of the near processor)
incardon's avatar
incardon committed
81
 * * Local ghosts interal or external are all the ghosts that does not involve inter-processor communications
82 83
 *
 * \see calculateGhostBoxes() for a visualization of internal and external ghost boxes
incardon's avatar
incardon committed
84
 *
85 86 87
 * ### Create a Cartesian decomposition object on a Box space, distribute, calculate internal and external ghost boxes
 * \snippet CartDecomposition_unit_test.hpp Create CartDecomposition
 *
incardon's avatar
incardon committed
88 89
 */

tonynsyde's avatar
tonynsyde committed
90
template<unsigned int dim, typename T, typename Memory = HeapMemory, template<unsigned int, typename > class Domain = Box, typename Distribution = DistParMetisDistribution<dim, T>>
91
class CartDecomposition: public ie_loc_ghost<dim, T>, public nn_prcs<dim, T>, public ie_ghost<dim, T>
incardon's avatar
incardon committed
92
{
incardon's avatar
incardon committed
93

incardon's avatar
incardon committed
94
public:
incardon's avatar
incardon committed
95

incardon's avatar
incardon committed
96 97 98 99
	//! Type of the domain we are going to decompose
	typedef T domain_type;

	//! It simplify to access the SpaceBox element
100
	typedef SpaceBox<dim, T> Box;
incardon's avatar
incardon committed
101 102 103

private:

incardon's avatar
incardon committed
104
	//! This is the key type to access  data_s, for example in the case of vector
incardon's avatar
incardon committed
105
	//! acc_key is size_t
tonynsyde's avatar
tonynsyde committed
106
	typedef typename openfpm::vector<SpaceBox<dim, T>, Memory, openfpm::vector_grow_policy_default, openfpm::vect_isel<SpaceBox<dim, T>>::value>::access_key acc_key;
incardon's avatar
incardon committed
107 108

	//! the set of all local sub-domain as vector
109
	openfpm::vector<SpaceBox<dim, T>> sub_domains;
incardon's avatar
incardon committed
110

incardon's avatar
incardon committed
111
	//! for each sub-domain, contain the list of the neighborhood processors
incardon's avatar
incardon committed
112 113
	openfpm::vector<openfpm::vector<long unsigned int> > box_nn_processor;

incardon's avatar
incardon committed
114
	//! Structure that contain for each sub-sub-domain box the processor id
incardon's avatar
Add ORB  
incardon committed
115
	//! exist for efficient global communication
incardon's avatar
incardon committed
116 117
	openfpm::vector<size_t> fine_s;

incardon's avatar
incardon committed
118
	//! Structure that store the cartesian grid information
119
	grid_sm<dim, void> gr;
incardon's avatar
incardon committed
120

incardon's avatar
incardon committed
121 122
	//! Structure that decompose your structure into cell without creating them
	//! useful to convert positions to CellId or sub-domain id in this case
123
	CellDecomposer_sm<dim, T> cd;
incardon's avatar
incardon committed
124 125

	//! rectangular domain to decompose
126
	Domain<dim, T> domain;
incardon's avatar
incardon committed
127 128 129 130 131 132 133

	//! Box Spacing
	T spacing[dim];

	//! Runtime virtual cluster machine
	Vcluster & v_cl;

134 135 136
	//! Create ditribution
	Distribution dist;

incardon's avatar
incardon committed
137
	//! Cell-list that store the geometrical information of the local internal ghost boxes
138
	CellList<dim, T, FAST> lgeo_cell;
incardon's avatar
incardon committed
139

140
	/*! \brief Constructor, it decompose and distribute the sub-domains across the processors
incardon's avatar
incardon committed
141
	 *
142 143
	 * \param v_cl Virtual cluster, used internally for communications
	 *
incardon's avatar
incardon committed
144 145 146
	 */
	void CreateDecomposition(Vcluster & v_cl)
	{
147 148 149 150 151 152 153
#ifdef SE_CLASS1
		if (&v_cl == NULL)
		{
			std::cerr << __FILE__ << ":" << __LINE__ << " error VCluster instance is null, check that you ever initialized it \n";
			ACTION_ON_ERROR()
		}
#endif
154 155 156

		int p_id = v_cl.getProcessUnitID();

incardon's avatar
incardon committed
157 158 159
		// Calculate the total number of box and and the spacing
		// on each direction
		// Get the box containing the domain
160
		SpaceBox<dim, T> bs = domain.getBox();
incardon's avatar
incardon committed
161

162
		for (unsigned int i = 0; i < dim; i++)
incardon's avatar
incardon committed
163 164
		{
			// Calculate the spacing
incardon's avatar
incardon committed
165
			spacing[i] = (bs.getHigh(i) - bs.getLow(i)) / gr.size(i);
incardon's avatar
incardon committed
166 167
		}

incardon's avatar
Add ORB  
incardon committed
168
		// fill the structure that store the processor id for each sub-domain
incardon's avatar
incardon committed
169
		fine_s.resize(gr.size());
incardon's avatar
Add ORB  
incardon committed
170

incardon's avatar
incardon committed
171 172
		// Optimize the decomposition creating bigger spaces
		// And reducing Ghost over-stress
tonynsyde's avatar
tonynsyde committed
173
		dec_optimizer<dim, DistGraph_CSR<nm_v, nm_e>> d_o(dist.getGraph(), gr.getSize());
incardon's avatar
incardon committed
174 175

		// set of Boxes produced by the decomposition optimizer
176
		openfpm::vector<::Box<dim, size_t>> loc_box;
incardon's avatar
incardon committed
177

incardon's avatar
incardon committed
178
		// optimize the decomposition
179
		d_o.template optimize<nm_v::sub_id, nm_v::proc_id>(dist.getGraph(), p_id, loc_box, box_nn_processor);
incardon's avatar
incardon committed
180

incardon's avatar
incardon committed
181 182 183
		// Initialize ss_box and bbox
		if (loc_box.size() >= 0)
		{
184 185
			SpaceBox<dim, size_t> sub_dc = loc_box.get(0);
			SpaceBox<dim, T> sub_d(sub_dc);
incardon's avatar
incardon committed
186 187 188
			sub_d.mul(spacing);
			sub_d.expand(spacing);

189 190 191 192 193
			// Fixing sub-domains to cover all the domain

			// Fixing sub_d
			// if (loc_box) is a the boundary we have to ensure that the box span the full
			// domain (avoiding rounding off error)
194
			for (size_t i = 0; i < dim; i++)
195 196 197
			{
				if (sub_dc.getHigh(i) == cd.getGrid().size(i) - 1)
				{
198
					sub_d.setHigh(i, domain.getHigh(i));
199 200 201
				}
			}

incardon's avatar
incardon committed
202 203 204 205
			// add the sub-domain
			sub_domains.add(sub_d);

			ss_box = sub_d;
incardon's avatar
incardon committed
206
			ss_box -= ss_box.getP1();
incardon's avatar
incardon committed
207 208 209
			bbox = sub_d;
		}

incardon's avatar
incardon committed
210
		// convert into sub-domain
211
		for (size_t s = 1; s < loc_box.size(); s++)
incardon's avatar
incardon committed
212
		{
213 214
			SpaceBox<dim, size_t> sub_dc = loc_box.get(s);
			SpaceBox<dim, T> sub_d(sub_dc);
incardon's avatar
incardon committed
215

incardon's avatar
incardon committed
216 217 218
			// re-scale and add spacing (the end is the starting point of the next domain + spacing)
			sub_d.mul(spacing);
			sub_d.expand(spacing);
incardon's avatar
incardon committed
219

220 221 222 223 224
			// Fixing sub-domains to cover all the domain

			// Fixing sub_d
			// if (loc_box) is a the boundary we have to ensure that the box span the full
			// domain (avoiding rounding off error)
225
			for (size_t i = 0; i < dim; i++)
226 227 228
			{
				if (sub_dc.getHigh(i) == cd.getGrid().size(i) - 1)
				{
229
					sub_d.setHigh(i, domain.getHigh(i));
230 231 232
				}
			}

incardon's avatar
incardon committed
233 234
			// add the sub-domain
			sub_domains.add(sub_d);
incardon's avatar
incardon committed
235 236 237 238 239 240

			// Calculate the bound box
			bbox.enclose(sub_d);

			// Create the smallest box contained in all sub-domain
			ss_box.contained(sub_d);
incardon's avatar
incardon committed
241
		}
incardon's avatar
incardon committed
242

243
		nn_prcs<dim, T>::create(box_nn_processor, sub_domains);
244

incardon's avatar
incardon committed
245
		// fill fine_s structure
incardon's avatar
incardon committed
246 247 248
		// fine_s structure contain the processor id for each sub-sub-domain
		// with sub-sub-domain we mean the sub-domain decomposition before
		// running dec_optimizer (before merging sub-domains)
249
		auto it = dist.getGraph().getVertexIterator();
incardon's avatar
incardon committed
250 251 252 253 254 255

		while (it.isNext())
		{
			size_t key = it.get();

			// fill with the fine decomposition
256
			fine_s.get(key) = dist.getGraph().template vertex_p<nm_v::proc_id>(key);
incardon's avatar
incardon committed
257 258 259

			++it;
		}
260 261

		// Get the smallest sub-division on each direction
262
		::Box<dim, T> unit = getSmallestSubdivision();
263
		// Get the processor bounding Box
264
		::Box<dim, T> bound = getProcessorBounds();
265

incardon's avatar
incardon committed
266
		// calculate the sub-divisions
267
		size_t div[dim];
268 269
		for (size_t i = 0; i < dim; i++)
			div[i] = (size_t) ((bound.getHigh(i) - bound.getLow(i)) / unit.getHigh(i));
270 271

		// Create shift
272
		Point<dim, T> orig;
273 274

		// p1 point of the Processor bound box is the shift
275
		for (size_t i = 0; i < dim; i++)
276 277 278
			orig.get(i) = bound.getLow(i);

		// Initialize the geo_cell structure
279 280 281
		ie_ghost<dim, T>::Initialize_geo_cell(domain, div, orig);
		lgeo_cell.Initialize(domain, div, orig);

incardon's avatar
incardon committed
282 283
	}

284 285 286 287 288 289
	/*! \brief Calculate communication and migration costs
	 *
	 * \param ts how many timesteps have passed since last calculation, used to approximate the cost
	 */
	void computeCommunicationAndMigrationCosts(size_t ts)
	{
tonynsyde's avatar
tonynsyde committed
290
		float migration = 0;
291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314

		SpaceBox<dim, T> cellBox = cd.getCellBox();
		float b_s = cellBox.getHigh(0);
		float gh_s = ghost.getHigh(0);

		// compute the gh_area for 2 dim case
		float gh_v = (gh_s * b_s);

		// multiply for sub-sub-domain side for each domain
		for (int i = 2; i < dim; i++)
			gh_v *= b_s;

		size_t norm = (size_t) (1.0 / gh_v);

		migration = pow(b_s, dim);

		size_t prev = 0;

		for (size_t i = 0; i < dist.getNSubSubDomains(); i++)
		{
			dist.setMigrationCost(i, norm * migration * dist.getVertexWeight(i));

			for (size_t s = 0; s < dist.getNSubSubDomainNeighbors(i); s++)
			{
tonynsyde's avatar
tonynsyde committed
315
				dist.setCommunicationCost(i, s, 1 * dist.getVertexWeight(i) * ts);
316 317 318 319
			}
			prev += dist.getNSubSubDomainNeighbors(i);
		}
	}
incardon's avatar
incardon committed
320

321 322
	// Save the ghost boundaries
	Ghost<dim, T> ghost;
incardon's avatar
incardon committed
323

incardon's avatar
incardon committed
324 325 326 327 328 329
	/*! \brief Create the subspaces that decompose your domain
	 *
	 */
	void CreateSubspaces()
	{
		// Create a grid where each point is a space
330
		grid_sm<dim, void> g(div);
incardon's avatar
incardon committed
331 332 333 334 335 336 337 338 339 340 341

		// create a grid_key_dx iterator
		grid_key_dx_iterator<dim> gk_it(g);

		// Divide the space into subspaces
		while (gk_it.isNext())
		{
			//! iterate through all subspaces
			grid_key_dx<dim> key = gk_it.get();

			//! Create a new subspace
342
			SpaceBox<dim, T> tmp;
incardon's avatar
incardon committed
343 344

			//! fill with the Margin of the box
345
			for (int i = 0; i < dim; i++)
incardon's avatar
incardon committed
346
			{
347 348
				tmp.setHigh(i, (key.get(i) + 1) * spacing[i]);
				tmp.setLow(i, key.get(i) * spacing[i]);
incardon's avatar
incardon committed
349 350 351 352 353 354 355 356 357 358
			}

			//! add the space box
			sub_domains.add(tmp);

			// add the iterator
			++gk_it;
		}
	}

incardon's avatar
incardon committed
359 360 361 362 363 364 365 366 367
	// Heap memory receiver
	HeapMemory hp_recv;

	// vector v_proc
	openfpm::vector<size_t> v_proc;

	// Receive counter
	size_t recv_cnt;

incardon's avatar
incardon committed
368 369 370
	// reference counter of the object in case is shared between object
	long int ref_cnt;

incardon's avatar
incardon committed
371 372
public:

373
	/*! \brief Cart decomposition constructor
incardon's avatar
incardon committed
374
	 *
375
	 * \param v_cl Virtual cluster, used internally to handle or pipeline communication
incardon's avatar
incardon committed
376 377
	 *
	 */
378 379
	CartDecomposition(Vcluster & v_cl) :
			nn_prcs<dim, T>(v_cl), v_cl(v_cl), dist(v_cl)
incardon's avatar
incardon committed
380 381 382 383
	{
		// Reset the box to zero
		bbox.zero();
	}
incardon's avatar
incardon committed
384

385
	//! Cart decomposition destructor
incardon's avatar
incardon committed
386
	~CartDecomposition()
387 388
	{
	}
incardon's avatar
incardon committed
389

390 391 392 393 394 395 396 397 398 399 400 401 402 403
	/*! \brief class to select the returned id by ghost_processorID
	 *
	 */
	class box_id
	{
	public:
		/*! \brief Return the box id
		 *
		 * \param p structure containing the id informations
		 * \param b_id box_id
		 *
		 * \return box id
		 *
		 */
404
		inline static size_t id(p_box<dim, T> & p, size_t b_id)
405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423
		{
			return b_id;
		}
	};

	/*! \brief class to select the returned id by ghost_processorID
	 *
	 */
	class processor_id
	{
	public:
		/*! \brief Return the processor id
		 *
		 * \param p structure containing the id informations
		 * \param b_id box_id
		 *
		 * \return processor id
		 *
		 */
424
		inline static size_t id(p_box<dim, T> & p, size_t b_id)
425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443
		{
			return p.proc;
		}
	};

	/*! \brief class to select the returned id by ghost_processorID
	 *
	 */
	class lc_processor_id
	{
	public:
		/*! \brief Return the near processor id
		 *
		 * \param p structure containing the id informations
		 * \param b_id box_id
		 *
		 * \return local processor id
		 *
		 */
444
		inline static size_t id(p_box<dim, T> & p, size_t b_id)
445 446 447 448 449
		{
			return p.lc_proc;
		}
	};

incardon's avatar
incardon committed
450
	/*! It calculate the internal ghost boxes
incardon's avatar
incardon committed
451 452 453 454 455
	 *
	 * Example: Processor 10 calculate
	 * B8_0 B9_0 B9_1 and B5_0
	 *
	 *
456 457 458
	 *
	 \verbatim

459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517
	 +----------------------------------------------------+
	 |                                                    |
	 |                 Processor 8                        |
	 |                 Sub-domain 0                       +-----------------------------------+
	 |                                                    |                                   |
	 |                                                    |                                   |
	 ++--------------+---+---------------------------+----+        Processor 9                |
	 |              |   |     B8_0                  |    |        Subdomain 0                |
	 |              +------------------------------------+                                   |
	 |              |   |                           |    |                                   |
	 |              |   |  XXXXXXXXXXXXX XX         |B9_0|                                   |
	 |              | B |  X Processor 10 X         |    |                                   |
	 | Processor 5  | 5 |  X Sub-domain 0 X         |    |                                   |
	 | Subdomain 0  | _ |  X              X         +----------------------------------------+
	 |              | 0 |  XXXXXXXXXXXXXXXX         |    |                                   |
	 |              |   |                           |    |                                   |
	 |              |   |                           |    |        Processor 9                |
	 |              |   |                           |B9_1|        Subdomain 1                |
	 |              |   |                           |    |                                   |
	 |              |   |                           |    |                                   |
	 |              |   |                           |    |                                   |
	 +--------------+---+---------------------------+----+                                   |
	 |                                   |
	 +-----------------------------------+

	 \endverbatim

	 and also
	 G8_0 G9_0 G9_1 G5_0 (External ghost boxes)

	 \verbatim

	 +----------------------------------------------------+
	 |                                                    |
	 |                 Processor 8                        |
	 |                 Sub-domain 0                       +-----------------------------------+
	 |           +---------------------------------------------+                              |
	 |           |         G8_0                           |    |                              |
	 ++--------------+------------------------------------+    |   Processor 9                |
	 |          |   |                                    |    |   Subdomain 0                |
	 |          |   |                                    |G9_0|                              |
	 |          |   |                                    |    |                              |
	 |          |   |      XXXXXXXXXXXXX XX              |    |                              |
	 |          |   |      X Processor 10 X              |    |                              |
	 | Processor|5  |      X Sub-domain 0 X              |    |                              |
	 | Subdomain|0  |      X              X              +-----------------------------------+
	 |          |   |      XXXXXXXXXXXXXXXX              |    |                              |
	 |          | G |                                    |    |                              |
	 |          | 5 |                                    |    |   Processor 9                |
	 |          | | |                                    |    |   Subdomain 1                |
	 |          | 0 |                                    |G9_1|                              |
	 |          |   |                                    |    |                              |
	 |          |   |                                    |    |                              |
	 +--------------+------------------------------------+    |                              |
	 |                                        |    |                              |
	 +----------------------------------------+----+------------------------------+


	 \endverbatim
518

incardon's avatar
incardon committed
519 520 521 522 523
	 *
	 *
	 *
	 * \param ghost margins for each dimensions (p1 negative part) (p2 positive part)
	 *
524 525
	 *
	 \verbatim
526 527 528 529 530 531 532 533 534 535 536 537
	 ^ p2[1]
	 |
	 |
	 +----+----+
	 |         |
	 |         |
	 p1[0]<-----+         +----> p2[0]
	 |         |
	 |         |
	 +----+----+
	 |
	 v  p1[1]
incardon's avatar
incardon committed
538

539
	 \endverbatim
540

incardon's avatar
incardon committed
541 542 543
	 *
	 *
	 */
incardon's avatar
incardon committed
544
	void calculateGhostBoxes()
incardon's avatar
incardon committed
545 546 547 548 549
	{
#ifdef DEBUG
		// the ghost margins are assumed to be smaller
		// than one sub-domain

550
		for (size_t i = 0; i < dim; i++)
incardon's avatar
incardon committed
551
		{
552 553
			if (ghost.template getLow(i) >= domain.template getHigh(i) / gr.size(i)
					|| ghost.template getHigh(i) >= domain.template getHigh(i) / gr.size(i))
incardon's avatar
incardon committed
554
			{
555
				std::cerr << "Error " << __FILE__ << ":" << __LINE__ << " : Ghost are bigger than one domain" << "\n";
incardon's avatar
incardon committed
556 557 558 559 560 561
			}
		}
#endif

		// Intersect all the local sub-domains with the sub-domains of the contiguous processors

incardon's avatar
incardon committed
562
		// create the internal structures that store ghost information
563 564
		ie_ghost<dim, T>::create_box_nn_processor_ext(v_cl, ghost, sub_domains, box_nn_processor, *this);
		ie_ghost<dim, T>::create_box_nn_processor_int(v_cl, ghost, sub_domains, box_nn_processor, *this);
incardon's avatar
incardon committed
565

incardon's avatar
incardon committed
566
		// ebox must come after ibox (in this case)
incardon's avatar
incardon committed
567

568 569
		ie_loc_ghost<dim, T>::create_loc_ghost_ibox(ghost, sub_domains);
		ie_loc_ghost<dim, T>::create_loc_ghost_ebox(ghost, sub_domains);
incardon's avatar
incardon committed
570 571

		// get the smallest sub-domain dimension on each direction
572
		for (size_t i = 0; i < dim; i++)
incardon's avatar
incardon committed
573
		{
tonynsyde's avatar
tonynsyde committed
574
			if (ghost.template getLow(i) >= ss_box.getHigh(i) || ghost.template getHigh(i) >= domain.template getHigh(i) / gr.size(i))
incardon's avatar
incardon committed
575
			{
576
				std::cerr << "Error " << __FILE__ << ":" << __LINE__ << " : Ghost are bigger than one domain" << "\n";
incardon's avatar
incardon committed
577 578
			}
		}
incardon's avatar
incardon committed
579 580
	}

581 582 583 584 585 586 587 588 589 590 591
	/*! \brief The default grid size
	 *
	 *  The default grid is always an isotropic grid that adapt with the number of processors,
	 *  it define in how many cell it will be divided the space for a particular required minimum
	 *  number of sub-domain
	 *
	 */
	static size_t getDefaultGrid(size_t n_sub)
	{
		// Calculate the number of sub-sub-domain on
		// each dimension
592
		return openfpm::math::round_big_2(pow(n_sub, 1.0 / dim));
593 594
	}

595
	/*! \brief Given a point return in which processor the particle should go
incardon's avatar
incardon committed
596 597 598 599
	 *
	 * \return processorID
	 *
	 */
600
	template<typename Mem> size_t inline processorID(encapc<1, Point<dim, T>, Mem> p)
incardon's avatar
incardon committed
601
	{
incardon's avatar
incardon committed
602
		return fine_s.get(cd.getCell(p));
incardon's avatar
incardon committed
603 604
	}

incardon's avatar
incardon committed
605
	// Smallest subdivision on each direction
606
	::Box<dim, T> ss_box;
incardon's avatar
incardon committed
607 608 609 610 611 612

	/*! \brief Get the smallest subdivision of the domain on each direction
	 *
	 * \return a box p1 is set to zero
	 *
	 */
613
	const ::Box<dim, T> & getSmallestSubdivision()
incardon's avatar
incardon committed
614 615 616 617
	{
		return ss_box;
	}

618
	/*! \brief Given a point return in which processor the particle should go
incardon's avatar
incardon committed
619 620 621 622 623
	 *
	 * \return processorID
	 *
	 */

624
	size_t inline processorID(const T (&p)[dim]) const
incardon's avatar
incardon committed
625
	{
incardon's avatar
incardon committed
626
		return fine_s.get(cd.getCell(p));
incardon's avatar
incardon committed
627 628
	}

incardon's avatar
incardon committed
629 630
	/*! \brief Set the parameter of the decomposition
	 *
631 632
	 * \param div_ storing into how many domain to decompose on each dimension
	 * \param domain_ domain to decompose
incardon's avatar
incardon committed
633 634
	 *
	 */
635
	void setParameters(const size_t (&div_)[dim], Domain<dim, T> domain_, Ghost<dim, T> ghost = Ghost<dim, T>())
incardon's avatar
incardon committed
636
	{
incardon's avatar
incardon committed
637 638
		// set the ghost
		this->ghost = ghost;
incardon's avatar
incardon committed
639

640
		// Set the decomposition parameters
incardon's avatar
incardon committed
641
		gr.setDimensions(div_);
incardon's avatar
incardon committed
642
		domain = domain_;
643
		cd.setDimensions(domain, div_, 0);
incardon's avatar
incardon committed
644

645 646 647 648 649 650 651 652 653 654 655 656 657
		// init distribution
		dist.init(gr, domain);

	}

	/*! \brief Start decomposition
	 *
	 */
	void decompose()
	{
		computeCommunicationAndMigrationCosts(1);

		dist.decompose();
incardon's avatar
incardon committed
658

tonynsyde's avatar
tonynsyde committed
659
		//CreateDecomposition(v_cl);
incardon's avatar
incardon committed
660 661
	}

662 663 664 665 666 667 668 669 670 671 672 673
	/*! \brief Refine the decomposition, available only for ParMetis distribution, for Metis it is a null call
	 *
	 */
	void rebalance()
	{
		computeCommunicationAndMigrationCosts(1);

		dist.refine();
	}

	/*! \brief Refine the decomposition, available only for ParMetis distribution, for Metis it is a null call
	 *
tonynsyde's avatar
tonynsyde committed
674
	 * \return true if the re-balance has been executed, false otherwise
675
	 */
tonynsyde's avatar
tonynsyde committed
676
	bool rebalance(DLB & dlb)
677
	{
tonynsyde's avatar
tonynsyde committed
678 679 680 681 682 683 684 685 686 687 688
		// if the DLB heuristic to use is the "Unbalance Threshold" get unbalance percentage
		if (dlb.getHeurisitc() == DLB::Heuristic::UNBALANCE_THRLD)
		{
			float unbalance = dist.getUnbalance();
			dlb.setUnbalance(unbalance);
			if (v_cl.getProcessUnitID() == 0)
			{
				std::cout << std::setprecision(3) << unbalance << "\n";
			}
		}

689 690 691 692
		if (dlb.rebalanceNeeded())
		{
			computeCommunicationAndMigrationCosts(dlb.getNTimeStepSinceDLB());
			dist.refine();
tonynsyde's avatar
tonynsyde committed
693
			return true;
694
		}
tonynsyde's avatar
tonynsyde committed
695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713
		return false;
	}

	/*! \brief Get the current un-balance value
	 *
	 * \return the un-balance percentage value
	 */
	float getUnbalance()
	{
		return dist.getUnbalance();
	}

	/*! \brief Compute the processor load counting the total weights of its vertices
	 *
	 * \return the current processor load
	 */
	size_t getProcessorLoad()
	{
		return dist.getProcessorLoad();
714 715 716 717 718 719 720 721
	}

	/*! \brief function that return the position of the cell in the space
	 *
	 * \param id vertex id
	 * \param pos vector that will contain x, y, z
	 *
	 */
tonynsyde's avatar
tonynsyde committed
722
	inline void getSubSubDomainPosition(size_t id, T (&pos)[dim])
723 724 725 726
	{
		dist.getVertexPosition(id, pos);
	}

tonynsyde's avatar
tonynsyde committed
727 728 729 730 731 732 733 734 735
	/*! \brief Get the number of sub-sub-domains in this sub-graph
	 *
	 * @return number of sub-sub-domains in this sub-graph
	 */
	size_t getNSubSubDomains()
	{
		return dist.getNSubSubDomains();
	}

736 737 738 739 740 741 742 743 744 745
	/*! \brief function that set the weight of the vertex
	 *
	 * \param id vertex id
	 *
	 */
	inline void setSubSubDomainComputationCost(size_t id, size_t weight)
	{
		dist.setVertexWeight(id, weight);
	}

tonynsyde's avatar
tonynsyde committed
746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764
	/*! \brief function that set the weight of the vertex
	 *
	 * \param id vertex id
	 *
	 */
	inline size_t getSubSubDomainComputationCost(size_t id)
	{
		return dist.getVertexWeight(id);
	}

	/*! \brief Operator to access the size of the sub-graph
	 *
	 * \return the size of the subgraph
	 */
	size_t subSize()
	{
		return dist.subSize();
	}

765
	/*! \brief Get the number of local sub-domains
incardon's avatar
incardon committed
766 767 768 769 770 771 772 773 774
	 *
	 * \return the number of sub-domains
	 *
	 */
	size_t getNLocalHyperCube()
	{
		return sub_domains.size();
	}

incardon's avatar
incardon committed
775
	/*! \brief Get the local sub-domain
incardon's avatar
incardon committed
776
	 *
incardon's avatar
incardon committed
777 778
	 * \param i (each local processor can have more than one sub-domain)
	 * \return the sub-domain
incardon's avatar
incardon committed
779 780
	 *
	 */
781
	SpaceBox<dim, T> getLocalHyperCube(size_t lc)
incardon's avatar
incardon committed
782 783
	{
		// Create a space box
784
		SpaceBox<dim, T> sp;
incardon's avatar
incardon committed
785 786 787

		// fill the space box

788
		for (size_t k = 0; k < dim; k++)
incardon's avatar
incardon committed
789 790
		{
			// create the SpaceBox Low and High
791 792
			sp.setLow(k, sub_domains.template get<Box::p1>(lc)[k]);
			sp.setHigh(k, sub_domains.template get<Box::p2>(lc)[k]);
incardon's avatar
incardon committed
793 794 795 796 797
		}

		return sp;
	}

incardon's avatar
incardon committed
798 799 800 801 802 803
	/*! \brief Get the local sub-domain with ghost extension
	 *
	 * \param i (each local processor can have more than one sub-domain)
	 * \return the sub-domain
	 *
	 */
804
	SpaceBox<dim, T> getSubDomainWithGhost(size_t lc)
incardon's avatar
incardon committed
805 806
	{
		// Create a space box
807
		SpaceBox<dim, T> sp = sub_domains.get(lc);
incardon's avatar
incardon committed
808 809 810 811 812 813 814

		// enlarge with ghost
		sp.enlarge(ghost);

		return sp;
	}

incardon's avatar
incardon committed
815 816 817 818 819
	/*! \brief Return the structure that store the physical domain
	 *
	 * \return The physical domain
	 *
	 */
820
	Domain<dim, T> & getDomain()
incardon's avatar
incardon committed
821 822 823 824
	{
		return domain;
	}

incardon's avatar
incardon committed
825 826 827
	/*! \brief Check if the particle is local
	 *
	 * \param p object position
incardon's avatar
incardon committed
828
	 *
incardon's avatar
incardon committed
829
	 * \return true if it is local
incardon's avatar
incardon committed
830 831
	 *
	 */
832
	template<typename Mem> bool isLocal(const encapc<1, Point<dim, T>, Mem> p) const
incardon's avatar
incardon committed
833
	{
834
		return processorID<Mem>(p) == v_cl.getProcessUnitID();
incardon's avatar
incardon committed
835
	}
incardon's avatar
incardon committed
836

incardon's avatar
incardon committed
837
	/*! \brief Check if the particle is local
incardon's avatar
incardon committed
838
	 *
incardon's avatar
incardon committed
839
	 * \param p object position
incardon's avatar
incardon committed
840
	 *
incardon's avatar
incardon committed
841
	 * \return true if it is local
incardon's avatar
incardon committed
842 843
	 *
	 */
844
	bool isLocal(const T (&pos)[dim]) const
incardon's avatar
incardon committed
845
	{
incardon's avatar
incardon committed
846 847
		return processorID(pos) == v_cl.getProcessUnitID();
	}
incardon's avatar
incardon committed
848

849
	::Box<dim, T> bbox;
incardon's avatar
incardon committed
850

851
	/*! \brief Return the bounding box containing union of all the sub-domains for the local processor
incardon's avatar
incardon committed
852 853 854 855
	 *
	 * \return The bounding box
	 *
	 */
856
	::Box<dim, T> & getProcessorBounds()
incardon's avatar
incardon committed
857 858 859
	{
		return bbox;
	}
incardon's avatar
incardon committed
860

861 862
	////////////// Functions to get decomposition information ///////////////

incardon's avatar
incardon committed
863 864 865 866
	/*! \brief Write the decomposition as VTK file
	 *
	 * The function generate several files
	 *
867 868 869 870 871 872
	 * * subdomains_X.vtk domain for the local processor (X) as union of sub-domain
	 * * subdomains_adjacent_X.vtk sub-domains adjacent to the local processor (X)
	 * * internal_ghost_X.vtk Internal ghost boxes for the local processor (X)
	 * * external_ghost_X.vtk External ghost boxes for the local processor (X)
	 * * local_internal_ghost_X.vtk internal local ghost boxes for the local processor (X)
	 * * local_external_ghost_X.vtk external local ghost boxes for the local processor (X)
incardon's avatar
incardon committed
873
	 *
incardon's avatar
incardon committed
874
	 * where X is the local processor rank
incardon's avatar
incardon committed
875 876 877 878 879 880
	 *
	 * \param output directory where to write the files
	 *
	 */
	bool write(std::string output) const
	{
881
		//! subdomains_X.vtk domain for the local processor (X) as union of sub-domain
882
		VTKWriter<openfpm::vector<::SpaceBox<dim, T>>, VECTOR_BOX> vtk_box1;
incardon's avatar
incardon committed
883
		vtk_box1.add(sub_domains);
tonynsyde's avatar
tonynsyde committed
884
		vtk_box1.write(output + std::string("subdomains_") + std::to_string(v_cl.getProcessUnitID()) + std::string(".vtk"));
incardon's avatar
incardon committed
885

886 887 888
		nn_prcs<dim, T>::write(output);
		ie_ghost<dim, T>::write(output, v_cl.getProcessUnitID());
		ie_loc_ghost<dim, T>::write(output, v_cl.getProcessUnitID());
incardon's avatar
incardon committed
889

890
		return true;
incardon's avatar
incardon committed
891
	}
incardon's avatar
incardon committed
892 893 894 895 896 897 898 899

	/*! \brief function to check the consistency of the information of the decomposition
	 *
	 * \return false if is inconsistent
	 *
	 */
	bool check_consistency()
	{
900
		if (ie_loc_ghost<dim, T>::check_consistency(getNLocalHyperCube()) == false)
incardon's avatar
incardon committed
901
			return false;
incardon's avatar
incardon committed
902 903 904

		return true;
	}
905

906 907 908
	/*! \brief Print subdomains, external and internal ghost boxes
	 *
	 */
909 910
	void debugPrint()
	{
911
		std::cout << "Subdomains\n";
912
		for (size_t p = 0; p < sub_domains.size(); p++)
913
		{
914
			std::cout << ::SpaceBox<dim, T>(sub_domains.get(p)).toString() << "\n";
915 916
		}

917
		std::cout << "External ghost box\n";
918

919
		for (size_t p = 0; p<nn_prcs < dim, T>::getNNProcessors(); p++)
920
		{
921
			for (size_t i = 0; i<ie_ghost < dim, T>::getProcessorNEGhost(p); i++)
922
			{
tonynsyde's avatar
tonynsyde committed
923
				std::cout << ie_ghost<dim, T>::getProcessorEGhostBox(p, i).toString() << "   prc=" << nn_prcs<dim, T>::IDtoProc(p) << "   id=" << ie_ghost<dim, T>::getProcessorEGhostId(p, i) << "\n";
924
			}
925
		}
926

927
		std::cout << "Internal ghost box\n";
928

929
		for (size_t p = 0; p<nn_prcs < dim, T>::getNNProcessors(); p++)
930
		{
931
			for (size_t i = 0; i<ie_ghost < dim, T>::getProcessorNIGhost(p); i++)
932
			{
tonynsyde's avatar
tonynsyde committed
933
				std::cout << ie_ghost<dim, T>::getProcessorIGhostBox(p, i).toString() << "   prc=" << nn_prcs<dim, T>::IDtoProc(p) << "   id=" << ie_ghost<dim, T>::getProcessorIGhostId(p, i) << "\n";
934 935
			}
		}
936
	}
incardon's avatar
incardon committed
937

938 939 940 941 942 943 944
	/*! \brief Print current graph and save it to file with name test_graph_[id]
	 *
	 * \param id to attach to the filename
	 *
	 */
	void printCurrentDecomposition(int id)
	{
tonynsyde's avatar
tonynsyde committed
945
		dist.printCurrentDecomposition(id);
946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965
	}

	//! Increment the reference counter
	void incRef()
	{
		ref_cnt++;
	}

	//! Decrement the reference counter
	void decRef()
	{
		ref_cnt--;
	}

	//! Return the reference counter
	long int ref()
	{
		return ref_cnt;
	}
};
incardon's avatar
incardon committed
966 967

#endif