cpp-taskflow/taskflow/algorithm/sort.hpp at master · neinnil/cpp-taskflow · GitHub

History

474 lines (379 loc) · 13.9 KB

Raw

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

#pragma once

#include "../core/executor.hpp"

namespace tf {

// threshold whether or not to perform parallel sort

template <typename I>

constexpr size_t parallel_sort_cutoff() {

//using value_type = std::decay_t<decltype(*std::declval<I>())>;

using value_type = typename std::iterator_traits<I>::value_type;

constexpr size_t object_size = sizeof(value_type);

if constexpr(std::is_same_v<value_type, std::string>) {

return 128;

}

else {

if constexpr(object_size < 16) return 4096;

else if constexpr(object_size < 32) return 2048;

else if constexpr(object_size < 64) return 1024;

else if constexpr(object_size < 128) return 768;

else if constexpr(object_size < 256) return 512;

else if constexpr(object_size < 512) return 256;

else return 128;

}

// ----------------------------------------------------------------------------

// pattern-defeating quick sort (pdqsort)

// ----------------------------------------------------------------------------

// Sorts [begin, end) using insertion sort with the given comparison function.

template<typename RandItr, typename Compare>

void insertion_sort(RandItr begin, RandItr end, Compare comp) {

using T = typename std::iterator_traits<RandItr>::value_type;

if (begin == end) {

return;

}

for (RandItr cur = begin + 1; cur != end; ++cur) {

RandItr shift = cur;

RandItr shift_1 = cur - 1;

// Compare first to avoid 2 moves for an element

// already positioned correctly.

if (comp(*shift, *shift_1)) {

T tmp = std::move(*shift);

do {

*shift-- = std::move(*shift_1);

}while (shift != begin && comp(tmp, *--shift_1));

*shift = std::move(tmp);

}

// Sorts [begin, end) using insertion sort with the given comparison function.

// Assumes *(begin - 1) is an element smaller than or equal to any element

// in [begin, end).

template<typename RandItr, typename Compare>

void unguarded_insertion_sort(RandItr begin, RandItr end, Compare comp) {

using T = typename std::iterator_traits<RandItr>::value_type;

if (begin == end) {

return;

}

for (RandItr cur = begin + 1; cur != end; ++cur) {

RandItr shift = cur;

RandItr shift_1 = cur - 1;

// Compare first so we can avoid 2 moves

// for an element already positioned correctly.

if (comp(*shift, *shift_1)) {

T tmp = std::move(*shift);

do {

*shift-- = std::move(*shift_1);

}while (comp(tmp, *--shift_1));

*shift = std::move(tmp);

}

// Attempts to use insertion sort on [begin, end).

// Will return false if more than

// partial_insertion_sort_limit elements were moved,

// and abort sorting. Otherwise it will successfully sort and return true.

template<typename RandItr, typename Compare>

bool partial_insertion_sort(RandItr begin, RandItr end, Compare comp) {

using T = typename std::iterator_traits<RandItr>::value_type;

using D = typename std::iterator_traits<RandItr>::difference_type;

// When we detect an already sorted partition, attempt an insertion sort

// that allows this amount of element moves before giving up.

constexpr auto partial_insertion_sort_limit = D{8};

if (begin == end) return true;

auto limit = D{0};

for (RandItr cur = begin + 1; cur != end; ++cur) {

if (limit > partial_insertion_sort_limit) {

return false;

}

RandItr shift = cur;

RandItr shift_1 = cur - 1;

// Compare first so we can avoid 2 moves

// for an element already positioned correctly.

if (comp(*shift, *shift_1)) {

T tmp = std::move(*shift);

do {

*shift-- = std::move(*shift_1);

}while (shift != begin && comp(tmp, *--shift_1));

*shift = std::move(tmp);

limit += cur - shift;

}

return true;

}

// Partitions [begin, end) around pivot *begin using comparison function comp.

// Elements equal to the pivot are put in the right-hand partition.

// Returns the position of the pivot after partitioning and whether the passed

// sequence already was correctly partitioned.

// Assumes the pivot is a median of at least 3 elements and that [begin, end)

// is at least insertion_sort_threshold long.

template<typename Iter, typename Compare>

std::pair<Iter, bool> partition_right(Iter begin, Iter end, Compare comp) {

using T = typename std::iterator_traits<Iter>::value_type;

// Move pivot into local for speed.

T pivot(std::move(*begin));

Iter first = begin;

Iter last = end;

// Find the first element greater than or equal than the pivot

// (the median of 3 guarantees/ this exists).

while (comp(*++first, pivot));

// Find the first element strictly smaller than the pivot.

// We have to guard this search if there was no element before *first.

if (first - 1 == begin) while (first < last && !comp(*--last, pivot));

else while (!comp(*--last, pivot));

// If the first pair of elements that should be swapped to partition

// are the same element, the passed in sequence already was correctly

// partitioned.

bool already_partitioned = first >= last;

// Keep swapping pairs of elements that are on the wrong side of the pivot.

// Previously swapped pairs guard the searches,

// which is why the first iteration is special-cased above.

while (first < last) {

std::iter_swap(first, last);

while (comp(*++first, pivot));

while (!comp(*--last, pivot));

}

// Put the pivot in the right place.

Iter pivot_pos = first - 1;

*begin = std::move(*pivot_pos);

*pivot_pos = std::move(pivot);

return std::make_pair(pivot_pos, already_partitioned);

}

// Similar function to the one above, except elements equal to the pivot

// are put to the left of the pivot and it doesn't check or return

// if the passed sequence already was partitioned.

// Since this is rarely used (the many equal case),

// and in that case pdqsort already has O(n) performance,

// no block quicksort is applied here for simplicity.

template<typename RandItr, typename Compare>

RandItr partition_left(RandItr begin, RandItr end, Compare comp) {

using T = typename std::iterator_traits<RandItr>::value_type;

T pivot(std::move(*begin));

RandItr first = begin;

RandItr last = end;

while (comp(pivot, *--last));

if (last + 1 == end) {

while (first < last && !comp(pivot, *++first));

}

else {

while (!comp(pivot, *++first));

}

while (first < last) {

std::iter_swap(first, last);

while (comp(pivot, *--last));

while (!comp(pivot, *++first));

}

RandItr pivot_pos = last;

*begin = std::move(*pivot_pos);

*pivot_pos = std::move(pivot);

return pivot_pos;

}

template<typename Iter, typename Compare>

void parallel_pdqsort(

tf::Subflow& sf,

Iter begin, Iter end, Compare comp,

int bad_allowed, bool leftmost = true

) {

// Partitions below this size are sorted sequentially

constexpr auto cutoff = parallel_sort_cutoff<Iter>();

// Partitions below this size are sorted using insertion sort

constexpr auto insertion_sort_threshold = 24;

// Partitions above this size use Tukey's ninther to select the pivot.

constexpr auto ninther_threshold = 128;

//using diff_t = typename std::iterator_traits<Iter>::difference_type;

// Use a while loop for tail recursion elimination.

while (true) {

//diff_t size = end - begin;

size_t size = end - begin;

// Insertion sort is faster for small arrays.

if (size < insertion_sort_threshold) {

if (leftmost) {

insertion_sort(begin, end, comp);

}

else {

unguarded_insertion_sort(begin, end, comp);

}

return;

}

if(size <= cutoff) {

std::sort(begin, end, comp);

return;

}

// Choose pivot as median of 3 or pseudomedian of 9.

//diff_t s2 = size / 2;

size_t s2 = size >> 1;

if (size > ninther_threshold) {

sort3(begin, begin + s2, end - 1, comp);

sort3(begin + 1, begin + (s2 - 1), end - 2, comp);

sort3(begin + 2, begin + (s2 + 1), end - 3, comp);

sort3(begin + (s2 - 1), begin + s2, begin + (s2 + 1), comp);

std::iter_swap(begin, begin + s2);

}

else {

sort3(begin + s2, begin, end - 1, comp);

}

// If *(begin - 1) is the end of the right partition

// of a previous partition operation, there is no element in [begin, end)

// that is smaller than *(begin - 1).

// Then if our pivot compares equal to *(begin - 1) we change strategy,

// putting equal elements in the left partition,

// greater elements in the right partition.

// We do not have to recurse on the left partition,

// since it's sorted (all equal).

if (!leftmost && !comp(*(begin - 1), *begin)) {

begin = partition_left(begin, end, comp) + 1;

continue;

}

// Partition and get results.

auto pair = partition_right(begin, end, comp);

auto pivot_pos = pair.first;

auto already_partitioned = pair.second;

// Check for a highly unbalanced partition.

//diff_t l_size = pivot_pos - begin;

//diff_t r_size = end - (pivot_pos + 1);

size_t l_size = pivot_pos - begin;

size_t r_size = end - (pivot_pos + 1);

bool highly_unbalanced = l_size < size / 8 || r_size < size / 8;

// If we got a highly unbalanced partition we shuffle elements

// to break many patterns.

if (highly_unbalanced) {

// If we had too many bad partitions, switch to heapsort

// to guarantee O(n log n).

if (--bad_allowed == 0) {

std::make_heap(begin, end, comp);

std::sort_heap(begin, end, comp);

return;

}

if (l_size >= insertion_sort_threshold) {

std::iter_swap(begin, begin + l_size / 4);

std::iter_swap(pivot_pos - 1, pivot_pos - l_size / 4);

if (l_size > ninther_threshold) {

std::iter_swap(begin + 1, begin + (l_size / 4 + 1));

std::iter_swap(begin + 2, begin + (l_size / 4 + 2));

std::iter_swap(pivot_pos - 2, pivot_pos - (l_size / 4 + 1));

std::iter_swap(pivot_pos - 3, pivot_pos - (l_size / 4 + 2));

}

if (r_size >= insertion_sort_threshold) {

std::iter_swap(pivot_pos + 1, pivot_pos + (1 + r_size / 4));

std::iter_swap(end - 1, end - r_size / 4);

if (r_size > ninther_threshold) {

std::iter_swap(pivot_pos + 2, pivot_pos + (2 + r_size / 4));

std::iter_swap(pivot_pos + 3, pivot_pos + (3 + r_size / 4));

std::iter_swap(end - 2, end - (1 + r_size / 4));

std::iter_swap(end - 3, end - (2 + r_size / 4));

}

// decently balanced

else {

// sequence try to use insertion sort.

if (already_partitioned &&

partial_insertion_sort(begin, pivot_pos, comp) &&

partial_insertion_sort(pivot_pos + 1, end, comp)

) {

return;

}

// Sort the left partition first using recursion and

// do tail recursion elimination for the right-hand partition.

sf.silent_async(

[&sf, begin, pivot_pos, comp, bad_allowed, leftmost] () mutable {

parallel_pdqsort(sf, begin, pivot_pos, comp, bad_allowed, leftmost);

}

);

begin = pivot_pos + 1;

leftmost = false;

}

// ----------------------------------------------------------------------------

// 3-way quick sort

// ----------------------------------------------------------------------------

// 3-way quick sort

template <typename RandItr, typename C>

void parallel_3wqsort(tf::Subflow& sf, RandItr first, RandItr last, C compare) {

using namespace std::string_literals;

constexpr auto cutoff = parallel_sort_cutoff<RandItr>();

sort_partition:

if(static_cast<size_t>(last - first) < cutoff) {

std::sort(first, last+1, compare);

return;

}

auto m = pseudo_median_of_nine(first, last, compare);

if(m != first) {

std::iter_swap(first, m);

}

auto l = first;

auto r = last;

auto f = std::next(first, 1);

bool is_swapped_l = false;

bool is_swapped_r = false;

while(f <= r) {

if(compare(*f, *l)) {

is_swapped_l = true;

std::iter_swap(l, f);

l++;

f++;

}

else if(compare(*l, *f)) {

is_swapped_r = true;

std::iter_swap(r, f);

r--;

}

else {

f++;

}

if(l - first > 1 && is_swapped_l) {

//sf.emplace([&](tf::Subflow& sfl) mutable {

// parallel_3wqsort(sfl, first, l-1, compare);

//});

sf.silent_async([&sf, first, l, &compare] () mutable {

parallel_3wqsort(sf, first, l-1, compare);

});

}

if(last - r > 1 && is_swapped_r) {

//sf.emplace([&](tf::Subflow& sfr) mutable {

// parallel_3wqsort(sfr, r+1, last, compare);

//});

//sf.silent_async([&sf, r, last, &compare] () mutable {

// parallel_3wqsort(sf, r+1, last, compare);

//});

first = r+1;

goto sort_partition;

}

//sf.join();

}

// ----------------------------------------------------------------------------

// tf::Taskflow::sort

// ----------------------------------------------------------------------------

// Function: sort

template <typename B, typename E, typename C>

Task FlowBuilder::sort(B beg, E end, C cmp) {

using B_t = std::decay_t<unwrap_ref_decay_t<B>>;

using E_t = std::decay_t<unwrap_ref_decay_t<E>>;

Task task = emplace([b=beg, e=end, cmp] (Subflow& sf) mutable {

// fetch the iterator values

B_t beg = b;

E_t end = e;

if(beg == end) {

return;

}

size_t W = sf._executor.num_workers();

size_t N = std::distance(beg, end);

// only myself - no need to spawn another graph

if(W <= 1 || N <= parallel_sort_cutoff<B_t>()) {

std::sort(beg, end, cmp);

return;

}

//parallel_3wqsort(sf, beg, end-1, cmp);

parallel_pdqsort(sf, beg, end, cmp, log2(end - beg));

sf.join();

});

return task;

}

// Function: sort

template <typename B, typename E>

Task FlowBuilder::sort(B beg, E end) {

using value_type = std::decay_t<decltype(*std::declval<B>())>;

return sort(beg, end, std::less<value_type>{});

}

} // namespace tf ------------------------------------------------------------

cpp-taskflow/taskflow/algorithm/sort.hpp at master · neinnil/cpp-taskflow · GitHub

Expand file tree

Breadcrumbs

sort.hpp

Latest commit

History

Breadcrumbs

sort.hpp

File metadata and controls