1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
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
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
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
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
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
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
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
|
// license:GPL-2.0+
// copyright-holders:Couriersud
/*
* nld_matrix_solver.cpp
*
*/
#include "nld_matrix_solver.h"
#include "plib/putil.h"
#include <cmath> // <<= needed by windows build
namespace netlist
{
namespace solver
{
terms_for_net_t::terms_for_net_t(analog_net_t * net)
: m_net(net)
, m_railstart(0)
{
}
void terms_for_net_t::add_terminal(terminal_t *term, int net_other, bool sorted)
{
if (sorted)
for (std::size_t i=0; i < m_connected_net_idx.size(); i++)
{
if (m_connected_net_idx[i] > net_other)
{
plib::container::insert_at(m_terms, i, term);
plib::container::insert_at(m_connected_net_idx, i, net_other);
return;
}
}
m_terms.push_back(term);
m_connected_net_idx.push_back(net_other);
}
// ----------------------------------------------------------------------------------------
// matrix_solver
// ----------------------------------------------------------------------------------------
matrix_solver_t::matrix_solver_t(netlist_state_t &anetlist, const pstring &name,
const analog_net_t::list_t &nets,
const solver_parameters_t *params)
: device_t(anetlist, name)
, m_params(*params)
, m_stat_calculations(*this, "m_stat_calculations", 0)
, m_stat_newton_raphson(*this, "m_stat_newton_raphson", 0)
, m_stat_vsolver_calls(*this, "m_stat_vsolver_calls", 0)
, m_iterative_fail(*this, "m_iterative_fail", 0)
, m_iterative_total(*this, "m_iterative_total", 0)
, m_last_step(*this, "m_last_step", netlist_time::zero())
, m_fb_sync(*this, "FB_sync")
, m_Q_sync(*this, "Q_sync")
, m_ops(0)
{
connect_post_start(m_fb_sync, m_Q_sync);
setup_base(nets);
/* now setup the matrix */
setup_matrix();
}
void matrix_solver_t::setup_base(const analog_net_t::list_t &nets)
{
log().debug("New solver setup\n");
m_terms.clear();
for (auto & net : nets)
{
m_terms.emplace_back(net);
m_rails_temp.emplace_back();
}
for (std::size_t k = 0; k < nets.size(); k++)
{
analog_net_t *net = nets[k];
log().debug("setting up net\n");
net->set_solver(this);
for (auto &p : net->core_terms())
{
log().debug("{1} {2} {3}\n", p->name(), net->name(), net->isRailNet());
switch (p->type())
{
case detail::terminal_type::TERMINAL:
if (p->device().is_timestep())
if (!plib::container::contains(m_step_devices, &p->device()))
m_step_devices.push_back(&p->device());
if (p->device().is_dynamic())
if (!plib::container::contains(m_dynamic_devices, &p->device()))
m_dynamic_devices.push_back(&p->device());
{
auto *pterm = dynamic_cast<terminal_t *>(p);
add_term(k, pterm);
}
log().debug("Added terminal {1}\n", p->name());
break;
case detail::terminal_type::INPUT:
{
proxied_analog_output_t *net_proxy_output = nullptr;
for (auto & input : m_inps)
if (input->proxied_net() == &p->net())
{
net_proxy_output = input.get();
break;
}
if (net_proxy_output == nullptr)
{
pstring nname(this->name() + "." + pstring(plib::pfmt("m{1}")(m_inps.size())));
nl_assert(p->net().is_analog());
auto net_proxy_output_u = pool().make_unique<proxied_analog_output_t>(*this, nname, static_cast<analog_net_t *>(&p->net()));
net_proxy_output = net_proxy_output_u.get();
m_inps.emplace_back(std::move(net_proxy_output_u));
}
net_proxy_output->net().add_terminal(*p);
// FIXME: repeated calling - kind of brute force
net_proxy_output->net().rebuild_list();
log().debug("Added input {1}", net_proxy_output->name());
}
break;
case detail::terminal_type::OUTPUT:
log().fatal(MF_UNHANDLED_ELEMENT_1_FOUND(p->name()));
break;
}
}
log().debug("added net with {1} populated connections\n", net->core_terms().size());
}
}
void matrix_solver_t::sort_terms(matrix_sort_type_e sort)
{
/* Sort in descending order by number of connected matrix voltages.
* The idea is, that for Gauss-Seidel algo the first voltage computed
* depends on the greatest number of previous voltages thus taking into
* account the maximum amout of information.
*
* This actually improves performance on popeye slightly. Average
* GS computations reduce from 2.509 to 2.370
*
* Smallest to largest : 2.613
* Unsorted : 2.509
* Largest to smallest : 2.370
*
* Sorting as a general matrix pre-conditioning is mentioned in
* literature but I have found no articles about Gauss Seidel.
*
* For Gaussian Elimination however increasing order is better suited.
* NOTE: Even better would be to sort on elements right of the matrix diagonal.
*
*/
const std::size_t iN = m_terms.size();
switch (sort)
{
case matrix_sort_type_e::PREFER_BAND_MATRIX:
{
for (std::size_t k = 0; k < iN - 1; k++)
{
auto pk = get_weight_around_diag(k,k);
for (std::size_t i = k+1; i < iN; i++)
{
auto pi = get_weight_around_diag(i,k);
if (pi < pk)
{
std::swap(m_terms[i], m_terms[k]);
pk = get_weight_around_diag(k,k);
}
}
}
}
break;
case matrix_sort_type_e::PREFER_IDENTITY_TOP_LEFT:
{
for (std::size_t k = 0; k < iN - 1; k++)
{
auto pk = get_left_right_of_diag(k,k);
for (std::size_t i = k+1; i < iN; i++)
{
auto pi = get_left_right_of_diag(i,k);
if (pi.first <= pk.first && pi.second >= pk.second)
{
std::swap(m_terms[i], m_terms[k]);
pk = get_left_right_of_diag(k,k);
}
}
}
}
break;
case matrix_sort_type_e::ASCENDING:
case matrix_sort_type_e::DESCENDING:
{
int sort_order = (sort == matrix_sort_type_e::DESCENDING ? 1 : -1);
for (std::size_t k = 0; k < iN - 1; k++)
for (std::size_t i = k+1; i < iN; i++)
{
if ((static_cast<int>(m_terms[k].railstart()) - static_cast<int>(m_terms[i].railstart())) * sort_order < 0)
{
std::swap(m_terms[i], m_terms[k]);
}
}
}
break;
case matrix_sort_type_e::NOSORT:
break;
}
/* rebuild */
for (auto &term : m_terms)
{
int *other = term.m_connected_net_idx.data();
for (std::size_t i = 0; i < term.count(); i++)
//FIXME: this is weird
if (other[i] != -1)
other[i] = get_net_idx(&term.terms()[i]->connected_terminal()->net());
}
}
void matrix_solver_t::setup_matrix()
{
const std::size_t iN = m_terms.size();
for (std::size_t k = 0; k < iN; k++)
{
m_terms[k].set_railstart(m_terms[k].count());
for (std::size_t i = 0; i < m_rails_temp[k].count(); i++)
this->m_terms[k].add_terminal(m_rails_temp[k].terms()[i], m_rails_temp[k].m_connected_net_idx.data()[i], false);
}
// free all - no longer needed
m_rails_temp.clear();
sort_terms(m_params.m_sort_type);
this->set_pointers();
/* create a list of non zero elements. */
for (unsigned k = 0; k < iN; k++)
{
terms_for_net_t & t = m_terms[k];
/* pretty brutal */
int *other = t.m_connected_net_idx.data();
t.m_nz.clear();
for (std::size_t i = 0; i < t.railstart(); i++)
if (!plib::container::contains(t.m_nz, static_cast<unsigned>(other[i])))
t.m_nz.push_back(static_cast<unsigned>(other[i]));
t.m_nz.push_back(k); // add diagonal
/* and sort */
std::sort(t.m_nz.begin(), t.m_nz.end());
}
/* create a list of non zero elements right of the diagonal
* These list anticipate the population of array elements by
* Gaussian elimination.
*/
for (std::size_t k = 0; k < iN; k++)
{
terms_for_net_t & t = m_terms[k];
/* pretty brutal */
int *other = t.m_connected_net_idx.data();
if (k==0)
t.m_nzrd.clear();
else
{
t.m_nzrd = m_terms[k-1].m_nzrd;
for (auto j = t.m_nzrd.begin(); j != t.m_nzrd.end(); )
{
if (*j < k + 1)
j = t.m_nzrd.erase(j);
else
++j;
}
}
for (std::size_t i = 0; i < t.railstart(); i++)
if (!plib::container::contains(t.m_nzrd, static_cast<unsigned>(other[i])) && other[i] >= static_cast<int>(k + 1))
t.m_nzrd.push_back(static_cast<unsigned>(other[i]));
/* and sort */
std::sort(t.m_nzrd.begin(), t.m_nzrd.end());
}
/* create a list of non zero elements below diagonal k
* This should reduce cache misses ...
*/
std::vector<std::vector<bool>> touched(iN, std::vector<bool>(iN));
for (std::size_t k = 0; k < iN; k++)
{
for (std::size_t j = 0; j < iN; j++)
touched[k][j] = false;
for (std::size_t j = 0; j < m_terms[k].m_nz.size(); j++)
touched[k][m_terms[k].m_nz[j]] = true;
}
m_ops = 0;
for (unsigned k = 0; k < iN; k++)
{
m_ops++; // 1/A(k,k)
for (unsigned row = k + 1; row < iN; row++)
{
if (touched[row][k])
{
m_ops++;
if (!plib::container::contains(m_terms[k].m_nzbd, row))
m_terms[k].m_nzbd.push_back(row);
for (std::size_t col = k + 1; col < iN; col++)
if (touched[k][col])
{
touched[row][col] = true;
m_ops += 2;
}
}
}
}
log().verbose("Number of mults/adds for {1}: {2}", name(), m_ops);
if ((false))
for (std::size_t k = 0; k < iN; k++)
{
pstring line = plib::pfmt("{1:3}")(k);
for (const auto & nzrd : m_terms[k].m_nzrd)
line += plib::pfmt(" {1:3}")(nzrd);
log().verbose("{1}", line);
}
/*
* save states
*/
m_last_V.resize(iN, plib::constants<nl_fptype>::zero());
m_DD_n_m_1.resize(iN, plib::constants<nl_fptype>::zero());
m_h_n_m_1.resize(iN, plib::constants<nl_fptype>::zero());
state().save(*this, m_last_V.as_base(), this->name(), "m_last_V");
state().save(*this, m_DD_n_m_1.as_base(), this->name(), "m_DD_n_m_1");
state().save(*this, m_h_n_m_1.as_base(), this->name(), "m_h_n_m_1");
for (std::size_t k = 0; k < iN; k++)
{
pstring num = plib::pfmt("{1}")(k);
// FIXME: This shouldn't be necessary, recalculate on each entry ...
state().save(*this, m_gonn[k],"GO" + num, this->name(), m_terms[k].count());
state().save(*this, m_gtn[k],"GT" + num, this->name(), m_terms[k].count());
state().save(*this, m_Idrn[k],"IDR" + num, this->name(), m_terms[k].count());
}
}
void matrix_solver_t::update_inputs()
{
// avoid recursive calls. Inputs are updated outside this call
for (auto &inp : m_inps)
inp->push(inp->proxied_net()->Q_Analog());
}
void matrix_solver_t::update_dynamic()
{
/* update all non-linear devices */
for (auto &dyn : m_dynamic_devices)
dyn->update_terminals();
}
void matrix_solver_t::reset()
{
m_last_step = netlist_time::zero();
}
void matrix_solver_t::update() NL_NOEXCEPT
{
const netlist_time new_timestep = solve(exec().time());
update_inputs();
if (m_params.m_dynamic_ts && has_timestep_devices() && new_timestep > netlist_time::zero())
{
m_Q_sync.net().toggle_and_push_to_queue(new_timestep);
}
}
/* update_forced is called from within param_update
*
* this should only occur outside of execution and thus
* using time should be safe.
*
*/
void matrix_solver_t::update_forced()
{
const netlist_time new_timestep = solve(exec().time());
plib::unused_var(new_timestep);
update_inputs();
if (m_params.m_dynamic_ts && has_timestep_devices())
{
m_Q_sync.net().toggle_and_push_to_queue(netlist_time::from_double(m_params.m_min_timestep));
}
}
void matrix_solver_t::step(const netlist_time &delta)
{
const nl_fptype dd = delta.as_double();
for (auto &d : m_step_devices)
d->timestep(dd);
}
const netlist_time matrix_solver_t::solve(netlist_time now)
{
const netlist_time delta = now - m_last_step;
// We are already up to date. Avoid oscillations.
// FIXME: Make this a parameter!
if (delta < netlist_time::quantum())
return netlist_time::zero();
/* update all terminals for new time step */
m_last_step = now;
step(delta);
++m_stat_vsolver_calls;
if (has_dynamic_devices())
{
std::size_t this_resched(0);
std::size_t newton_loops = 0;
do
{
update_dynamic();
// Gauss-Seidel will revert to Gaussian elemination if steps exceeded.
this_resched = this->vsolve_non_dynamic(true);
newton_loops++;
} while (this_resched > 1 && newton_loops < m_params.m_nr_loops);
m_stat_newton_raphson += newton_loops;
// reschedule ....
if (this_resched > 1 && !m_Q_sync.net().is_queued())
{
log().warning(MW_NEWTON_LOOPS_EXCEEDED_ON_NET_1(this->name()));
m_Q_sync.net().toggle_and_push_to_queue(netlist_time::from_double(m_params.m_nr_recalc_delay));
}
}
else
{
this->vsolve_non_dynamic(false);
}
const netlist_time next_time_step = compute_next_timestep(delta.as_double());
return next_time_step;
}
int matrix_solver_t::get_net_idx(const analog_net_t *net) const noexcept
{
for (std::size_t k = 0; k < m_terms.size(); k++)
if (m_terms[k].isNet(net))
return static_cast<int>(k);
return -1;
}
std::pair<int, int> matrix_solver_t::get_left_right_of_diag(std::size_t irow, std::size_t idiag)
{
/*
* return the maximum column left of the diagonal (-1 if no cols found)
* return the minimum column right of the diagonal (999999 if no cols found)
*/
const auto row = static_cast<int>(irow);
const auto diag = static_cast<int>(idiag);
int colmax = -1;
int colmin = 999999;
auto &term = m_terms[irow];
for (std::size_t i = 0; i < term.count(); i++)
{
auto col = get_net_idx(&term.terms()[i]->connected_terminal()->net());
if (col != -1)
{
if (col==row) col = diag;
else if (col==diag) col = row;
if (col > diag && col < colmin)
colmin = col;
else if (col < diag && col > colmax)
colmax = col;
}
}
return {colmax, colmin};
}
nl_fptype matrix_solver_t::get_weight_around_diag(std::size_t row, std::size_t diag)
{
{
/*
* return average absolute distance
*/
std::vector<bool> touched(1024, false); // FIXME!
nl_fptype weight = 0.0;
auto &term = m_terms[row];
for (std::size_t i = 0; i < term.count(); i++)
{
auto col = get_net_idx(&term.terms()[i]->connected_terminal()->net());
if (col >= 0)
{
auto colu = static_cast<std::size_t>(col);
if (!touched[colu])
{
if (colu==row) colu = static_cast<unsigned>(diag);
else if (colu==diag) colu = static_cast<unsigned>(row);
weight = weight + std::abs(static_cast<nl_fptype>(colu) - static_cast<nl_fptype>(diag));
touched[colu] = true;
}
}
}
return weight; // / static_cast<nl_fptype>(term.railstart());
}
}
void matrix_solver_t::add_term(std::size_t k, terminal_t *term)
{
if (term->connected_terminal()->net().isRailNet())
{
m_rails_temp[k].add_terminal(term, -1, false);
}
else
{
int ot = get_net_idx(&term->connected_terminal()->net());
if (ot>=0)
{
m_terms[k].add_terminal(term, ot, true);
}
/* Should this be allowed ? */
else // if (ot<0)
{
m_rails_temp[k].add_terminal(term, ot, true);
log().fatal(MF_FOUND_TERM_WITH_MISSING_OTHERNET(term->name()));
}
}
}
netlist_time matrix_solver_t::compute_next_timestep(const nl_fptype cur_ts)
{
nl_fptype new_solver_timestep = m_params.m_max_timestep;
if (m_params.m_dynamic_ts)
{
for (std::size_t k = 0; k < m_terms.size(); k++)
{
auto &t = m_terms[k];
//const nl_fptype DD_n = (n->Q_Analog() - t->m_last_V);
// avoid floating point exceptions
const nl_fptype DD_n = std::max(-fp_constants<nl_fptype>::TIMESTEP_MAXDIFF, std::min(+fp_constants<nl_fptype>::TIMESTEP_MAXDIFF,(t.getV() - m_last_V[k])));
const nl_fptype hn = cur_ts;
//printf("%g %g %g %g\n", DD_n, hn, t.m_DD_n_m_1, t.m_h_n_m_1);
nl_fptype DD2 = (DD_n / hn - m_DD_n_m_1[k] / m_h_n_m_1[k]) / (hn + m_h_n_m_1[k]);
nl_fptype new_net_timestep(0);
m_h_n_m_1[k] = hn;
m_DD_n_m_1[k] = DD_n;
if (std::fabs(DD2) > fp_constants<nl_fptype>::TIMESTEP_MINDIV) // avoid div-by-zero
new_net_timestep = std::sqrt(m_params.m_dynamic_lte / std::fabs(plib::constants<nl_fptype>::cast(0.5)*DD2));
else
new_net_timestep = m_params.m_max_timestep;
if (new_net_timestep < new_solver_timestep)
new_solver_timestep = new_net_timestep;
m_last_V[k] = t.getV();
}
if (new_solver_timestep < m_params.m_min_timestep)
{
new_solver_timestep = m_params.m_min_timestep;
}
}
//if (new_solver_timestep > 10.0 * hn)
// new_solver_timestep = 10.0 * hn;
/*
* FIXME: Factor 2 below is important. Without, we get timing issues. This must be a bug elsewhere.
*/
return std::max(netlist_time::from_double(new_solver_timestep), netlist_time::quantum() * 2);
}
void matrix_solver_t::log_stats()
{
if (this->m_stat_calculations != 0 && this->m_stat_vsolver_calls && log().verbose.is_enabled())
{
log().verbose("==============================================");
log().verbose("Solver {1}", this->name());
log().verbose(" ==> {1} nets", this->m_terms.size()); //, (*(*groups[i].first())->m_core_terms.first())->name());
log().verbose(" has {1} elements", this->has_dynamic_devices() ? "dynamic" : "no dynamic");
log().verbose(" has {1} elements", this->has_timestep_devices() ? "timestep" : "no timestep");
log().verbose(" {1:6.3} average newton raphson loops",
static_cast<nl_fptype>(this->m_stat_newton_raphson) / static_cast<nl_fptype>(this->m_stat_vsolver_calls));
log().verbose(" {1:10} invocations ({2:6.0} Hz) {3:10} gs fails ({4:6.2} %) {5:6.3} average",
this->m_stat_calculations,
static_cast<nl_fptype>(this->m_stat_calculations) / this->exec().time().as_double(),
this->m_iterative_fail,
100.0 * static_cast<nl_fptype>(this->m_iterative_fail)
/ static_cast<nl_fptype>(this->m_stat_calculations),
static_cast<nl_fptype>(this->m_iterative_total) / static_cast<nl_fptype>(this->m_stat_calculations));
}
}
} // namespace solver
} // namespace netlist
|