Skip to content

Commit dab7391

Browse files
brunowubrunowu
committed
update the flops counter to support both real and cmplx scalars
1 parent 534bcc2 commit dab7391

2 files changed

Lines changed: 46 additions & 47 deletions

File tree

‎algorithm/performance.hpp‎

Lines changed: 45 additions & 46 deletions
Original file line numberDiff line numberDiff line change
@@ -36,6 +36,7 @@ namespace chase
3636
timings are stored in a vector of objects derived by the class
3737
template `std::chrono::duration`.
3838
*/
39+
template <class T>
3940
class ChasePerfData
4041
{
4142
public:
@@ -114,61 +115,58 @@ class ChasePerfData
114115
std::size_t get_flops(std::size_t N)
115116
{
116117
std::size_t flop_count = 0;
118+
int factor = std::pow(4, int(sizeof(T) / sizeof(Base<T>)) - 1) ;
117119
for (auto block : chase_iter_blocksizes)
118120
{
119121
// QR //
120-
121-
// https://software.intel.com/en-us/mkl-developer-reference-fortran-2018-beta-geqrf
122-
// QR m=N n=block
123-
//(2/3)n^2(3m-n) *4
124-
flop_count += (8. / 3.) *
125-
static_cast<double>(block * block * (3 * N - block));
126-
127-
// https://software.intel.com/en-us/mkl-developer-reference-c-2018-beta-ungqr
128-
// m = N, n=k=block
129-
// (8/3)*n^2*(3m - n)
130-
flop_count += (8. / 3.) * static_cast<double>(block * block *
131-
(3 * N - block) * 4);
132-
133-
// RR //
122+
// assume always to use cholqr-2
123+
// syrk
124+
flop_count += 2. * N * block * block;
125+
//cholesky factorizatin
126+
flop_count += 2. * block * block * block;
127+
// dtrsm
128+
flop_count += 2. * N * block * block;
129+
130+
// RR //
134131

135132
// W = H*V
136133
// https://software.intel.com/en-us/mkl-developer-reference-fortran-2018-beta-gemm
137134
// 8MNK + 18MN
138135
// m = N, k = N, n = block
139-
flop_count += 8 * N * block * N + 18 * N * block;
136+
flop_count += 2 * N * block * N;
140137

141138
// A = W' * W
142139
// 8MNK + 18MN
143140
// m = block, k = N, n = block
144-
flop_count += 8 * block * block * N + 18 * block * block;
141+
flop_count += 2 * block * block * N;
145142

146-
// reduction to tridiagonal
147-
// https://software.intel.com/en-us/mkl-developer-reference-fortran-2018-beta-hetrd
148-
// (16/3)n^3
149-
flop_count += 16. / 3. * static_cast<double>(block * block * block);
143+
// https://en.wikipedia.org/wiki/Divide-and-conquer_eigenvalue_algorithm
144+
// 4M^3
145+
flop_count += 4 * block * block * block;
150146

151147
// W = V*Z
152-
// 8MNK + 18MN
153-
flop_count += 8 * N * block * block + 18 * N * block;
148+
// 2MNK
149+
flop_count += 2 * N * block * block;
154150

155151
// resid //
156152
// W = H*V
157-
flop_count += 8 * N * block * N + 18 * N * block;
153+
flop_count += 2 * N * block * N;
158154

159155
// V[:,i] - lambda W[:,i]
160156
// 3*block
161157
flop_count += 3 * block * N;
162158

163159
// ||V[:,i]||
164-
// 4*N
165-
flop_count += 4 * N * block;
160+
// N
161+
flop_count += N * block;
166162
}
167163

168164
// filter
169165
// 8MNK + 18MN
170166
flop_count +=
171-
8 * N * chase_filtered_vecs * N + 16 * N * chase_filtered_vecs;
167+
2 * N * chase_filtered_vecs * N;
168+
169+
flop_count *= factor;
172170

173171
return flop_count / 1e9;
174172
}
@@ -190,8 +188,9 @@ class ChasePerfData
190188
*/
191189
std::size_t get_filter_flops(std::size_t N)
192190
{
193-
return (8 * N * chase_filtered_vecs * N +
194-
16 * N * chase_filtered_vecs) /
191+
int factor = std::pow(4, int(sizeof(T) / sizeof(Base<T>)) - 1) ;
192+
193+
return 2 * factor * N * chase_filtered_vecs * N /
195194
1e9;
196195
}
197196

@@ -261,8 +260,8 @@ class ChasePerfData
261260
{
262261
std::size_t flops = get_flops(N);
263262
std::size_t filter_flops = get_filter_flops(N);
264-
std::cout << " | " << static_cast<double>(flops);
265-
std::cout << " | " << static_cast<double>(filter_flops);
263+
std::cout << " | " << std::setw(11) << static_cast<double>(flops);
264+
std::cout << " | " << std::setw(15) << static_cast<double>(filter_flops);
266265
}
267266

268267
std::cout << " | " << std::setw(5) << nprocs;
@@ -338,15 +337,15 @@ class PerformanceDecoratorChase : public chase::Chase<T>
338337
void initVecs(bool random)
339338
{
340339
chase_->initVecs(random);
341-
perf_.start_clock(ChasePerfData::TimePtrs::Lanczos);
340+
perf_.start_clock(ChasePerfData<T>::TimePtrs::Lanczos);
342341
}
343342

344343
void Shift(T c, bool isunshift = false)
345344
{
346345
if (isunshift)
347-
perf_.end_clock(ChasePerfData::TimePtrs::Filter);
346+
perf_.end_clock(ChasePerfData<T>::TimePtrs::Filter);
348347
else
349-
perf_.start_clock(ChasePerfData::TimePtrs::Filter);
348+
perf_.start_clock(ChasePerfData<T>::TimePtrs::Filter);
350349

351350
chase_->Shift(c, isunshift);
352351
}
@@ -358,29 +357,29 @@ class PerformanceDecoratorChase : public chase::Chase<T>
358357

359358
void QR(std::size_t fixednev, Base<T> cond)
360359
{
361-
perf_.start_clock(ChasePerfData::TimePtrs::Qr);
360+
perf_.start_clock(ChasePerfData<T>::TimePtrs::Qr);
362361
chase_->QR(fixednev, cond);
363-
perf_.end_clock(ChasePerfData::TimePtrs::Qr);
362+
perf_.end_clock(ChasePerfData<T>::TimePtrs::Qr);
364363
}
365364

366365
void RR(Base<T>* ritzv, std::size_t block)
367366
{
368-
perf_.start_clock(ChasePerfData::TimePtrs::Rr);
367+
perf_.start_clock(ChasePerfData<T>::TimePtrs::Rr);
369368
chase_->RR(ritzv, block);
370369
perf_.add_iter_blocksize(block);
371-
perf_.end_clock(ChasePerfData::TimePtrs::Rr);
370+
perf_.end_clock(ChasePerfData<T>::TimePtrs::Rr);
372371
}
373372
void Resd(Base<T>* ritzv, Base<T>* resd, std::size_t fixednev)
374373
{
375-
perf_.start_clock(ChasePerfData::TimePtrs::Resids_Locking);
374+
perf_.start_clock(ChasePerfData<T>::TimePtrs::Resids_Locking);
376375
chase_->Resd(ritzv, resd, fixednev);
377376
// We end with ->chase_->Lock()
378377
}
379378
void Lanczos(std::size_t m, Base<T>* upperb)
380379
{
381-
perf_.start_clock(ChasePerfData::TimePtrs::Lanczos);
380+
perf_.start_clock(ChasePerfData<T>::TimePtrs::Lanczos);
382381
chase_->Lanczos(m, upperb);
383-
perf_.end_clock(ChasePerfData::TimePtrs::Lanczos);
382+
perf_.end_clock(ChasePerfData<T>::TimePtrs::Lanczos);
384383
}
385384
void Lanczos(std::size_t M, std::size_t idx, Base<T>* upperb,
386385
Base<T>* ritzv, Base<T>* Tau, Base<T>* ritzV)
@@ -390,21 +389,21 @@ class PerformanceDecoratorChase : public chase::Chase<T>
390389
void LanczosDos(std::size_t idx, std::size_t m, T* ritzVc)
391390
{
392391
chase_->LanczosDos(idx, m, ritzVc);
393-
perf_.end_clock(ChasePerfData::TimePtrs::Lanczos);
392+
perf_.end_clock(ChasePerfData<T>::TimePtrs::Lanczos);
394393
}
395394

396395
void Swap(std::size_t i, std::size_t j) { chase_->Swap(i, j); }
397396
void Lock(std::size_t new_converged)
398397
{
399398
chase_->Lock(new_converged);
400-
perf_.end_clock(ChasePerfData::TimePtrs::Resids_Locking);
399+
perf_.end_clock(ChasePerfData<T>::TimePtrs::Resids_Locking);
401400
perf_.add_iter_count(1);
402401
}
403402
void Start()
404403
{
405404
chase_->Start();
406405
perf_.Reset();
407-
perf_.start_clock(ChasePerfData::TimePtrs::All);
406+
perf_.start_clock(ChasePerfData<T>::TimePtrs::All);
408407
perf_.set_nprocs(chase_->get_nprocs());
409408
}
410409

@@ -413,7 +412,7 @@ class PerformanceDecoratorChase : public chase::Chase<T>
413412
void End()
414413
{
415414
chase_->End();
416-
perf_.end_clock(ChasePerfData::TimePtrs::All);
415+
perf_.end_clock(ChasePerfData<T>::TimePtrs::All);
417416
}
418417

419418
std::size_t GetN() const { return chase_->GetN(); }
@@ -422,15 +421,15 @@ class PerformanceDecoratorChase : public chase::Chase<T>
422421
Base<T>* GetRitzv() { return chase_->GetRitzv(); }
423422
Base<T>* GetResid() { return chase_->GetResid(); }
424423
ChaseConfig<T>& GetConfig() { return chase_->GetConfig(); }
425-
ChasePerfData& GetPerfData() { return perf_; }
424+
ChasePerfData<T>& GetPerfData() { return perf_; }
426425

427426
#ifdef CHASE_OUTPUT
428427
void Output(std::string str) { chase_->Output(str); }
429428
#endif
430429

431430
private:
432431
Chase<T>* chase_;
433-
ChasePerfData perf_;
432+
ChasePerfData<T> perf_;
434433
};
435434

436435
} // namespace chase

‎examples/2_input_output/2_input_output.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -621,7 +621,7 @@ int do_chase(ChASE_DriverProblemConfig& conf)
621621
{
622622
std::cout << " ChASE timings: "
623623
<< "\n";
624-
performanceDecorator.GetPerfData().print();
624+
performanceDecorator.GetPerfData().print(N);
625625
#ifdef PRINT_EIGENVALUES
626626
Base<T>* resid = single.GetResid();
627627
std::cout << "Finished Problem \n";

0 commit comments

Comments
 (0)