libgpuarray/src/gpuarray_array.c at master · jrao1/libgpuarray · GitHub

History

1151 lines (1014 loc) · 30.1 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

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

624

625

626

627

628

629

630

631

632

633

634

635

636

637

638

639

640

641

642

643

644

645

646

647

648

649

650

651

652

653

654

655

656

657

658

659

660

661

662

663

664

665

666

667

668

669

670

671

672

673

674

675

676

677

678

679

680

681

682

683

684

685

686

687

688

689

690

691

692

693

694

695

696

697

698

699

700

701

702

703

704

705

706

707

708

709

710

711

712

713

714

715

716

717

718

719

720

721

722

723

724

725

726

727

728

729

730

731

732

733

734

735

736

737

738

739

740

741

742

743

744

745

746

747

748

749

750

751

752

753

754

755

756

757

758

759

760

761

762

763

764

765

766

767

768

769

770

771

772

773

774

775

776

777

778

779

780

781

782

783

784

785

786

787

788

789

790

791

792

793

794

795

796

797

798

799

800

801

802

803

804

805

806

807

808

809

810

811

812

813

814

815

816

817

818

819

820

821

822

823

824

825

826

827

828

829

830

831

832

833

834

835

836

837

838

839

840

841

842

843

844

845

846

847

848

849

850

851

852

853

854

855

856

857

858

859

860

861

862

863

864

865

866

867

868

869

870

871

872

873

874

875

876

877

878

879

880

881

882

883

884

885

886

887

888

889

890

891

892

893

894

895

896

897

898

899

900

901

902

903

904

905

906

907

908

909

910

911

912

913

914

915

916

917

918

919

920

921

922

923

924

925

926

927

928

929

930

931

932

933

934

935

936

937

938

939

940

941

942

943

944

945

946

947

948

949

950

951

952

953

954

955

956

957

958

959

960

961

962

963

964

965

966

967

968

969

970

971

972

973

974

975

976

977

978

979

980

981

982

983

984

985

986

987

988

989

990

991

992

993

994

995

996

997

998

999

1000

#ifdef _MSC_VER

#define _CRT_SECURE_NO_WARNINGS

#endif

#include <assert.h>

#include <stdarg.h>

#include <stddef.h>

#include <stdlib.h>

#include <string.h>

#include <errno.h>

#include "private.h"

#include "gpuarray/config.h"

#include "gpuarray/array.h"

#include "gpuarray/error.h"

#include "gpuarray/kernel.h"

#include "gpuarray/util.h"

#include "util/strb.h"

#include "util/xxhash.h"

struct extcopy_args {

int itype;

int otype;

};

static int extcopy_eq(cache_key_t _k1, cache_key_t _k2) {

struct extcopy_args *k1 = _k1;

struct extcopy_args *k2 = _k2;

return k1->itype == k2->itype && k1->otype == k2->otype;

}

static void extcopy_free(cache_key_t k) {

free(k);

}

static uint32_t extcopy_hash(cache_key_t k) {

return XXH32(k, sizeof(struct extcopy_args), 42);

}

static int ga_extcopy(GpuArray *dst, const GpuArray *src) {

struct extcopy_args a, *aa;

gpucontext *ctx = gpudata_context(dst->data);

GpuElemwise *k = NULL;

void *args[2];

if (ctx != gpudata_context(src->data))

return GA_INVALID_ERROR;

a.itype = src->typecode;

a.otype = dst->typecode;

if (ctx->extcopy_cache != NULL)

k = cache_get(ctx->extcopy_cache, &a);

if (k == NULL) {

gpuelemwise_arg gargs[2];

gargs[0].name = "src";

gargs[0].typecode = src->typecode;

gargs[0].flags = GE_READ;

gargs[1].name = "dst";

gargs[1].typecode = dst->typecode;

gargs[1].flags = GE_WRITE;

k = GpuElemwise_new(ctx, "", "dst = src", 2, gargs, 0, 0);

if (k == NULL)

return GA_MISC_ERROR;

aa = memdup(&a, sizeof(a));

if (aa == NULL) {

GpuElemwise_free(k);

return GA_MEMORY_ERROR;

}

if (ctx->extcopy_cache == NULL)

ctx->extcopy_cache = cache_twoq(4, 8, 8, 2, extcopy_eq, extcopy_hash,

extcopy_free,

(cache_freev_fn)GpuElemwise_free);

if (ctx->extcopy_cache == NULL)

return GA_MISC_ERROR;

if (cache_add(ctx->extcopy_cache, aa, k) != 0)

return GA_MISC_ERROR;

}

args[0] = (void *)src;

args[1] = (void *)dst;

return GpuElemwise_call(k, args, GE_BROADCAST);

}

/* Value below which a size_t multiplication will never overflow. */

#define MUL_NO_OVERFLOW (1UL << (sizeof(size_t) * 4))

void GpuArray_fix_flags(GpuArray *a) {

/* Only keep the writable flag */

a->flags &= GA_WRITEABLE;

/* Set the other flags if applicable */

if (GpuArray_is_c_contiguous(a)) a->flags |= GA_C_CONTIGUOUS;

if (GpuArray_is_f_contiguous(a)) a->flags |= GA_F_CONTIGUOUS;

if (GpuArray_is_aligned(a)) a->flags |= GA_ALIGNED;

}

int GpuArray_empty(GpuArray *a, gpucontext *ctx, int typecode,

unsigned int nd, const size_t *dims, ga_order ord) {

size_t size = gpuarray_get_elsize(typecode);

unsigned int i;

int res = GA_NO_ERROR;

if (ord == GA_ANY_ORDER)

ord = GA_C_ORDER;

if (ord != GA_C_ORDER && ord != GA_F_ORDER)

return GA_VALUE_ERROR;

for (i = 0; i < nd; i++) {

size_t d = dims[i];

/* Check for overflow */

if ((d >= MUL_NO_OVERFLOW || size >= MUL_NO_OVERFLOW) &&

d > 0 && SIZE_MAX / d < size)

return GA_VALUE_ERROR;

size *= d;

}

a->data = gpudata_alloc(ctx, size, NULL, 0, &res);

if (a->data == NULL) return res;

a->nd = nd;

a->offset = 0;

a->typecode = typecode;

a->dimensions = calloc(nd, sizeof(size_t));

a->strides = calloc(nd, sizeof(ssize_t));

/* F/C distinction comes later */

a->flags = GA_BEHAVED;

if (a->dimensions == NULL || a->strides == NULL) {

GpuArray_clear(a);

return GA_MEMORY_ERROR;

}

/* Mult will not overflow since calloc succeded */

memcpy(a->dimensions, dims, sizeof(size_t)*nd);

size = gpuarray_get_elsize(typecode);

/* mults will not overflow, checked on entry */

switch (ord) {

case GA_C_ORDER:

for (i = nd; i > 0; i--) {

a->strides[i-1] = size;

size *= a->dimensions[i-1];

}

a->flags |= GA_C_CONTIGUOUS;

break;

case GA_F_ORDER:

for (i = 0; i < nd; i++) {

a->strides[i] = size;

size *= a->dimensions[i];

}

a->flags |= GA_F_CONTIGUOUS;

break;

default:

assert(0); /* cannot be reached */

}

if (a->nd <= 1)

a->flags |= GA_F_CONTIGUOUS|GA_C_CONTIGUOUS;

return GA_NO_ERROR;

}

int GpuArray_zeros(GpuArray *a, gpucontext *ctx,

int typecode, unsigned int nd, const size_t *dims,

ga_order ord) {

int err;

err = GpuArray_empty(a, ctx, typecode, nd, dims, ord);

if (err != GA_NO_ERROR)

return err;

err = gpudata_memset(a->data, a->offset, 0);

if (err != GA_NO_ERROR) {

GpuArray_clear(a);

}

return err;

}

int GpuArray_fromdata(GpuArray *a, gpudata *data, size_t offset, int typecode,

unsigned int nd, const size_t *dims,

const ssize_t *strides, int writeable) {

if (gpuarray_get_type(typecode)->typecode != typecode)

return GA_VALUE_ERROR;

assert(data != NULL);

a->data = data;

gpudata_retain(a->data);

a->nd = nd;

a->offset = offset;

a->typecode = typecode;

a->dimensions = calloc(nd, sizeof(size_t));

a->strides = calloc(nd, sizeof(ssize_t));

a->flags = (writeable ? GA_WRITEABLE : 0);

if (a->dimensions == NULL || a->strides == NULL) {

GpuArray_clear(a);

return GA_MEMORY_ERROR;

}

memcpy(a->dimensions, dims, nd*sizeof(size_t));

memcpy(a->strides, strides, nd*sizeof(ssize_t));

GpuArray_fix_flags(a);

return GA_NO_ERROR;

}

int GpuArray_copy_from_host(GpuArray *a, gpucontext *ctx, void *buf,

int typecode, unsigned int nd, const size_t *dims,

const ssize_t *strides) {

char *base = (char *)buf;

size_t offset = 0;

size_t size = gpuarray_get_elsize(typecode);

gpudata *b;

int err;

unsigned int i;

for (i = 0; i < nd; i++) {

if (dims[i] == 0) {

size = 0;

base = (char *)buf;

break;

}

if (strides[i] < 0)

base += (dims[i]-1) * strides[i];

else

size += (dims[i]-1) * strides[i];

}

offset = (char *)buf - base;

size += offset;

b = gpudata_alloc(ctx, size, base, GA_BUFFER_INIT, &err);

if (b == NULL) return err;

err = GpuArray_fromdata(a, b, offset, typecode, nd, dims, strides, 1);

gpudata_release(b);

return err;

}

int GpuArray_view(GpuArray *v, const GpuArray *a) {

v->data = a->data;

gpudata_retain(a->data);

v->nd = a->nd;

v->offset = a->offset;

v->typecode = a->typecode;

v->flags = a->flags;

v->dimensions = calloc(v->nd, sizeof(size_t));

v->strides = calloc(v->nd, sizeof(ssize_t));

if (v->dimensions == NULL || v->strides == NULL) {

GpuArray_clear(v);

return GA_MEMORY_ERROR;

}

memcpy(v->dimensions, a->dimensions, v->nd*sizeof(size_t));

memcpy(v->strides, a->strides, v->nd*sizeof(ssize_t));

return GA_NO_ERROR;

}

int GpuArray_sync(GpuArray *a) {

return gpudata_sync(a->data);

}

int GpuArray_index_inplace(GpuArray *a, const ssize_t *starts,

const ssize_t *stops, const ssize_t *steps) {

unsigned int i, new_i;

unsigned int new_nd = a->nd;

size_t *newdims;

ssize_t *newstrs;

size_t new_offset = a->offset;

if ((starts == NULL) || (stops == NULL) || (steps == NULL))

return GA_VALUE_ERROR;

for (i = 0; i < a->nd; i++) {

if (steps[i] == 0) new_nd -= 1;

}

newdims = calloc(new_nd, sizeof(size_t));

newstrs = calloc(new_nd, sizeof(ssize_t));

if (newdims == NULL || newstrs == NULL) {

free(newdims);

free(newstrs);

return GA_MEMORY_ERROR;

}

new_i = 0;

for (i = 0; i < a->nd; i++) {

if (starts[i] < -1 || (starts[i] > 0 &&

(size_t)starts[i] > a->dimensions[i])) {

free(newdims);

free(newstrs);

return GA_VALUE_ERROR;

}

if (steps[i] == 0 &&

(starts[i] == -1 || (size_t)starts[i] >= a->dimensions[i])) {

free(newdims);

free(newstrs);

return GA_VALUE_ERROR;

}

new_offset += starts[i] * a->strides[i];

if (steps[i] != 0) {

if ((stops[i] < -1 || (stops[i] > 0 &&

(size_t)stops[i] > a->dimensions[i])) ||

(stops[i]-starts[i])/steps[i] < 0) {

free(newdims);

free(newstrs);

return GA_VALUE_ERROR;

}

newstrs[new_i] = steps[i] * a->strides[i];

newdims[new_i] = (stops[i]-starts[i]+steps[i]-

(steps[i] < 0? -1 : 1))/steps[i];

new_i++;

}

a->nd = new_nd;

a->offset = new_offset;

free(a->dimensions);

a->dimensions = newdims;

free(a->strides);

a->strides = newstrs;

GpuArray_fix_flags(a);

return GA_NO_ERROR;

}

int GpuArray_index(GpuArray *r, const GpuArray *a, const ssize_t *starts,

const ssize_t *stops, const ssize_t *steps) {

int err;

err = GpuArray_view(r, a);

if (err != GA_NO_ERROR) return err;

err = GpuArray_index_inplace(r, starts, stops, steps);

if (err != GA_NO_ERROR) GpuArray_clear(r);

return err;

}

static int gen_take1_kernel(GpuKernel *k, gpucontext *ctx, char **err_str,

GpuArray *a, const GpuArray *v,

const GpuArray *ind, int addr32) {

strb sb = STRB_STATIC_INIT;

int *atypes;

size_t nargs, apos;

char *sz, *ssz;

unsigned int i, i2;

int flags = GA_USE_CLUDA;

int res;

nargs = 9 + 2 * v->nd;

atypes = calloc(nargs, sizeof(int));

if (atypes == NULL)

return GA_MEMORY_ERROR;

if (addr32) {

sz = "ga_uint";

ssz = "ga_int";

} else {

sz = "ga_size";

ssz = "ga_ssize";

}

apos = 0;

strb_appendf(&sb, "KERNEL void take1(GLOBAL_MEM %s *r, ga_size r_off, "

"GLOBAL_MEM const %s *v, ga_size v_off,",

gpuarray_get_type(a->typecode)->cluda_name,

gpuarray_get_type(v->typecode)->cluda_name);

atypes[apos++] = GA_BUFFER;

atypes[apos++] = GA_SIZE;

atypes[apos++] = GA_BUFFER;

atypes[apos++] = GA_SIZE;

for (i = 0; i < v->nd; i++) {

strb_appendf(&sb, " ga_ssize s%u, ga_size d%u,", i, i);

atypes[apos++] = GA_SSIZE;

atypes[apos++] = GA_SIZE;

}

strb_appendf(&sb, " GLOBAL_MEM const %s *ind, ga_size i_off, "

"ga_size n0, ga_size n1, GLOBAL_MEM int* err) {\n",

gpuarray_get_type(ind->typecode)->cluda_name);

atypes[apos++] = GA_BUFFER;

atypes[apos++] = GA_SIZE;

atypes[apos++] = GA_BUFFER;

assert(apos == nargs);

strb_appendf(&sb, " const %s idx0 = LDIM_0 * GID_0 + LID_0;\n"

" const %s numThreads0 = LDIM_0 * GDIM_0;\n"

" const %s idx1 = LDIM_1 * GID_1 + LID_1;\n"

" const %s numThreads1 = LDIM_1 * GDIM_1;\n"

" %s i0, i1;\n", sz, sz, sz, sz, sz);

strb_appends(&sb, " if (idx0 >= n0 || idx1 >= n1) return;\n");

strb_appendf(&sb, " r = (GLOBAL_MEM %s *)(((char *)r) + r_off);\n"

" ind = (GLOBAL_MEM %s *)(((char *)ind) + i_off);\n",

gpuarray_get_type(a->typecode)->cluda_name,

gpuarray_get_type(ind->typecode)->cluda_name);

strb_appendf(&sb, " for (i0 = idx0; i0 < n0; i0 += numThreads0) {\n"

" %s ii0 = ind[i0];\n"

" %s pos0 = v_off;\n"

" if (ii0 < 0) ii0 += d0;\n"

" if ((ii0 < 0) || (ii0 >= d0)) {\n"

" *err = -1;\n"

" continue;\n"

" }\n"

" pos0 += ii0 * (%s)s0;\n"

" for (i1 = idx1; i1 < n1; i1 += numThreads1) {\n"

" %s p = pos0;\n", ssz, sz, sz, sz);

if (v->nd > 1) {

strb_appendf(&sb, " %s pos, ii = i1;\n", sz);

for (i2 = v->nd; i2 > 1; i2--) {

i = i2 - 1;

if (i > 1)

strb_appendf(&sb, " pos = ii %% (%s)d%u;\n"

" ii /= (%s)d%u;\n", sz, i, sz, i);

else

strb_appends(&sb, " pos = ii;\n");

strb_appendf(&sb, " p += pos * (%s)s%u;\n", ssz, i);

}

strb_appendf(&sb, " r[i0*((%s)n1) + i1] = *((GLOBAL_MEM %s *)(((GLOBAL_MEM char *)v) + p));\n",

sz, gpuarray_get_type(v->typecode)->cluda_name);

strb_appends(&sb, " }\n"

" }\n"

"}\n");

if (strb_error(&sb)) {

res = GA_MEMORY_ERROR;

goto bail;

}

flags |= gpuarray_type_flags(a->typecode, v->typecode, GA_BYTE, -1);

res = GpuKernel_init(k, ctx, 1, (const char **)&sb.s, &sb.l, "take1",

nargs, atypes, flags, err_str);

bail:

free(atypes);

strb_clear(&sb);

return res;

}

int GpuArray_take1(GpuArray *a, const GpuArray *v, const GpuArray *i,

int check_error) {

size_t n[2], ls[2] = {0, 0}, gs[2] = {0, 0};

size_t pl;

gpudata *errbuf;

#if DEBUG

char *errstr = NULL;

#endif

size_t argp;

GpuKernel k;

unsigned int j;

int err, kerr = 0;

int addr32 = 0;

if (!GpuArray_ISWRITEABLE(a))

return GA_INVALID_ERROR;

if (!GpuArray_ISALIGNED(a) || !GpuArray_ISALIGNED(v) ||

!GpuArray_ISALIGNED(i))

return GA_UNALIGNED_ERROR;

/* a and i have to be C contiguous */

if (!GpuArray_IS_C_CONTIGUOUS(a) || !GpuArray_IS_C_CONTIGUOUS(i))

return GA_INVALID_ERROR;

/* Check that the dimensions match namely a[0] == i[0] and a[>0] == v[>0] */

if (v->nd == 0 || a->nd == 0 || i->nd != 1 || a->nd != v->nd ||

a->dimensions[0] != i->dimensions[0])

return GA_INVALID_ERROR;

n[0] = i->dimensions[0];

n[1] = 1;

for (j = 1; j < v->nd; j++) {

if (a->dimensions[j] != v->dimensions[j])

return GA_INVALID_ERROR;

n[1] *= v->dimensions[j];

}

if (n[0] * n[1] < SADDR32_MAX) {

addr32 = 1;

}

err = gpudata_property(v->data, GA_CTX_PROP_ERRBUF, &errbuf);

if (err != GA_NO_ERROR)

return err;

err = gen_take1_kernel(&k, GpuArray_context(a),

#if DEBUG

&errstr,

#else

NULL,

#endif

a, v, i, addr32);

#if DEBUG

if (errstr != NULL) {

fprintf(stderr, "%s\n", errstr);

free(errstr);

}

#endif

if (err != GA_NO_ERROR)

return err;

err = GpuKernel_sched(&k, n[0]*n[1], &ls[1], &gs[1]);

if (err != GA_NO_ERROR)

goto out;

/* This may not be the best scheduling, but it's good enough */

err = gpukernel_property(k.k, GA_KERNEL_PROP_PREFLSIZE, &pl);

ls[0] = ls[1] / pl;

ls[1] = pl;

if (n[1] > n[0]) {

pl = ls[0];

ls[0] = ls[1];

ls[1] = pl;

gs[0] = 1;

} else {

gs[0] = gs[1];

gs[1] = 1;

}

argp = 0;

GpuKernel_setarg(&k, argp++, a->data);

GpuKernel_setarg(&k, argp++, (void *)&a->offset);

GpuKernel_setarg(&k, argp++, v->data);

/* The cast is to avoid a warning about const */

GpuKernel_setarg(&k, argp++, (void *)&v->offset);

for (j = 0; j < v->nd; j++) {

GpuKernel_setarg(&k, argp++, &v->strides[j]);

GpuKernel_setarg(&k, argp++, &v->dimensions[j]);

}

GpuKernel_setarg(&k, argp++, i->data);

GpuKernel_setarg(&k, argp++, (void *)&i->offset);

GpuKernel_setarg(&k, argp++, &n[0]);

GpuKernel_setarg(&k, argp++, &n[1]);

GpuKernel_setarg(&k, argp++, errbuf);

err = GpuKernel_call(&k, 2, ls, gs, 0, NULL);

if (check_error && err == GA_NO_ERROR) {

err = gpudata_read(&kerr, errbuf, 0, sizeof(int));

if (err == GA_NO_ERROR && kerr != 0) {

err = GA_VALUE_ERROR;

kerr = 0;

/* We suppose this will not fail */

gpudata_write(errbuf, 0, &kerr, sizeof(int));

}

out:

GpuKernel_clear(&k);

return err;

}

int GpuArray_setarray(GpuArray *a, const GpuArray *v) {

GpuArray tv;

size_t sz;

ssize_t *strs;

unsigned int i, off;

int err = GA_NO_ERROR;

int simple_move = 1;

if (a->nd < v->nd)

return GA_VALUE_ERROR;

if (!GpuArray_ISWRITEABLE(a))

return GA_VALUE_ERROR;

if (!GpuArray_ISALIGNED(v) || !GpuArray_ISALIGNED(a))

return GA_UNALIGNED_ERROR;

off = a->nd - v->nd;

for (i = 0; i < v->nd; i++) {

if (v->dimensions[i] != a->dimensions[i+off]) {

if (v->dimensions[i] != 1)

return GA_VALUE_ERROR;

else

simple_move = 0;

}

if (simple_move && GpuArray_ISONESEGMENT(a) && GpuArray_ISONESEGMENT(v) &&

GpuArray_ISFORTRAN(a) == GpuArray_ISFORTRAN(v) &&

a->typecode == v->typecode &&

a->nd == v->nd) {

sz = gpuarray_get_elsize(a->typecode);

for (i = 0; i < a->nd; i++) sz *= a->dimensions[i];

return gpudata_move(a->data, a->offset, v->data, v->offset, sz);

}

strs = calloc(a->nd, sizeof(ssize_t));

if (strs == NULL)

return GA_MEMORY_ERROR;

for (i = off; i < a->nd; i++) {

if (v->dimensions[i-off] == a->dimensions[i]) {

strs[i] = v->strides[i-off];

}

memcpy(&tv, v, sizeof(GpuArray));

tv.nd = a->nd;

tv.dimensions = a->dimensions;

tv.strides = strs;

if (tv.nd != 0)

GpuArray_fix_flags(&tv);

err = ga_extcopy(a, &tv);

free(strs);

return err;

}

int GpuArray_reshape(GpuArray *res, const GpuArray *a, unsigned int nd,

const size_t *newdims, ga_order ord, int nocopy) {

int err;

err = GpuArray_view(res, a);

if (err != GA_NO_ERROR) return err;

err = GpuArray_reshape_inplace(res, nd, newdims, ord);

if (err == GA_COPY_ERROR && !nocopy) {

GpuArray_clear(res);

GpuArray_copy(res, a, ord);

err = GpuArray_reshape_inplace(res, nd, newdims, ord);

}

if (err != GA_NO_ERROR) GpuArray_clear(res);

return err;

}

int GpuArray_reshape_inplace(GpuArray *a, unsigned int nd,

const size_t *newdims, ga_order ord) {

ssize_t *newstrides;

size_t *tmpdims;

size_t np;

size_t op;

size_t newsize = 1;

size_t oldsize = 1;

unsigned int ni = 0;

unsigned int oi = 0;

unsigned int nj = 1;

unsigned int oj = 1;

unsigned int nk;

unsigned int ok;

unsigned int i;

if (ord == GA_ANY_ORDER && GpuArray_ISFORTRAN(a) && a->nd > 1)

ord = GA_F_ORDER;

for (i = 0; i < a->nd; i++) {

oldsize *= a->dimensions[i];

}

for (i = 0; i < nd; i++) {

size_t d = newdims[i];

/* Check for overflow */

if ((d >= MUL_NO_OVERFLOW || newsize >= MUL_NO_OVERFLOW) &&

d > 0 && SIZE_MAX / d < newsize)

return GA_INVALID_ERROR;

newsize *= d;

}

if (newsize != oldsize) return GA_INVALID_ERROR;

/* If the source and desired layouts are the same, then just copy

strides and dimensions */

if ((ord == GA_C_ORDER && GpuArray_CHKFLAGS(a, GA_C_CONTIGUOUS)) ||

(ord == GA_F_ORDER && GpuArray_CHKFLAGS(a, GA_F_CONTIGUOUS))) {

goto do_final_copy;

}

newstrides = calloc(nd, sizeof(ssize_t));

if (newstrides == NULL)

return GA_MEMORY_ERROR;

while (ni < nd && oi < a->nd) {

np = newdims[ni];

op = a->dimensions[oi];

while (np != op) {

if (np < op) {

np *= newdims[nj++];

} else {

op *= a->dimensions[oj++];

}

for (ok = oi; ok < oj - 1; ok++) {

if (ord == GA_F_ORDER) {

if (a->strides[ok+1] != (ssize_t)a->dimensions[ok]*a->strides[ok])

goto need_copy;

} else {

if (a->strides[ok] != (ssize_t)a->dimensions[ok+1]*a->strides[ok+1])

goto need_copy;

}

if (ord == GA_F_ORDER) {

newstrides[ni] = a->strides[oi];

for (nk = ni + 1; nk < nj; nk++) {

newstrides[nk] = newstrides[nk - 1]*newdims[nk - 1];

}

} else {

newstrides[nj-1] = a->strides[oj-1];

for (nk = nj-1; nk > ni; nk--) {

newstrides[nk-1] = newstrides[nk]*newdims[nk];

}

ni = nj++;

oi = oj++;

}

/* Fixup trailing ones */

if (ord == GA_F_ORDER) {

for (i = nj-1; i < nd; i++) {

newstrides[i] = newstrides[i-1] * newdims[i-1];

}

} else {

for (i = nj-1; i < nd; i++) {

newstrides[i] = gpuarray_get_elsize(a->typecode);

}

/* We can reuse newstrides since it was allocated in this function.

Can't do the same with newdims (which is a parameter). */

tmpdims = calloc(nd, sizeof(size_t));

if (tmpdims == NULL) {

return GA_MEMORY_ERROR;

}

memcpy(tmpdims, newdims, nd*sizeof(size_t));

a->nd = nd;

free(a->dimensions);

free(a->strides);

a->dimensions = tmpdims;

a->strides = newstrides;

goto fix_flags;

need_copy:

free(newstrides);

return GA_COPY_ERROR;

do_final_copy:

tmpdims = calloc(nd, sizeof(size_t));

newstrides = calloc(nd, sizeof(ssize_t));

if (tmpdims == NULL || newstrides == NULL) {

free(tmpdims);

free(newstrides);

return GA_MEMORY_ERROR;

}

memcpy(tmpdims, newdims, nd*sizeof(size_t));

if (nd > 0) {

if (ord == GA_F_ORDER) {

newstrides[0] = gpuarray_get_elsize(a->typecode);

for (i = 1; i < nd; i++) {

newstrides[i] = newstrides[i-1] * tmpdims[i-1];

}

} else {

newstrides[nd-1] = gpuarray_get_elsize(a->typecode);

for (i = nd-1; i > 0; i--) {

newstrides[i-1] = newstrides[i] * tmpdims[i];

}

free(a->dimensions);

free(a->strides);

a->nd = nd;

a->dimensions = tmpdims;

a->strides = newstrides;

fix_flags:

GpuArray_fix_flags(a);

return GA_NO_ERROR;

}

int GpuArray_transpose(GpuArray *res, const GpuArray *a,

const unsigned int *new_axes) {

int err;

err = GpuArray_view(res, a);

if (err != GA_NO_ERROR) return err;

err = GpuArray_transpose_inplace(res, new_axes);

if (err != GA_NO_ERROR) GpuArray_clear(res);

return err;

}

int GpuArray_transpose_inplace(GpuArray *a, const unsigned int *new_axes) {

size_t *newdims;

ssize_t *newstrs;

unsigned int i;

unsigned int j;

unsigned int k;

newdims = calloc(a->nd, sizeof(size_t));

newstrs = calloc(a->nd, sizeof(ssize_t));

if (newdims == NULL || newstrs == NULL) {

free(newdims);

free(newstrs);

return GA_MEMORY_ERROR;

}

for (i = 0; i < a->nd; i++) {

if (new_axes == NULL) {

j = a->nd - i - 1;

} else {

j = new_axes[i];

// Repeated axes will lead to a broken output

for (k = 0; k < i; k++)

if (j == new_axes[k]) {

free(newdims);

free(newstrs);

return GA_VALUE_ERROR;

}

newdims[i] = a->dimensions[j];

newstrs[i] = a->strides[j];

}

free(a->dimensions);

free(a->strides);

a->dimensions = newdims;

a->strides = newstrs;

GpuArray_fix_flags(a);

return GA_NO_ERROR;

}

void GpuArray_clear(GpuArray *a) {

if (a->data)

gpudata_release(a->data);

free(a->dimensions);

free(a->strides);

memset(a, 0, sizeof(*a));

}

int GpuArray_share(const GpuArray *a, const GpuArray *b) {

if (a->data != b->data) return 0;

/* XXX: redefine buffer_share to mean: is it possible to share?

and use offset to make sure */

return gpudata_share(a->data, b->data, NULL);

}

gpucontext *GpuArray_context(const GpuArray *a) {

return gpudata_context(a->data);

}

int GpuArray_move(GpuArray *dst, const GpuArray *src) {

size_t sz;

unsigned int i;

if (!GpuArray_ISWRITEABLE(dst))

return GA_VALUE_ERROR;

if (!GpuArray_ISALIGNED(src) || !GpuArray_ISALIGNED(dst))

return GA_UNALIGNED_ERROR;

if (src->nd != dst->nd)

return GA_VALUE_ERROR;

for (i = 0; i < src->nd; i++) {

if (src->dimensions[i] != dst->dimensions[i])

return GA_VALUE_ERROR;

}

if (!GpuArray_ISONESEGMENT(dst) || !GpuArray_ISONESEGMENT(src) ||

GpuArray_ISFORTRAN(dst) != GpuArray_ISFORTRAN(src) ||

dst->typecode != src->typecode) {

return ga_extcopy(dst, src);

}

sz = gpuarray_get_elsize(dst->typecode);

for (i = 0; i < dst->nd; i++) sz *= dst->dimensions[i];

return gpudata_move(dst->data, dst->offset, src->data, src->offset, sz);

}

int GpuArray_write(GpuArray *dst, const void *src, size_t src_sz) {

if (!GpuArray_ISWRITEABLE(dst))

return GA_VALUE_ERROR;

if (!GpuArray_ISONESEGMENT(dst))

return GA_UNSUPPORTED_ERROR;

return gpudata_write(dst->data, dst->offset, src, src_sz);

}

int GpuArray_read(void *dst, size_t dst_sz, const GpuArray *src) {

if (!GpuArray_ISONESEGMENT(src))

return GA_UNSUPPORTED_ERROR;

return gpudata_read(dst, src->data, src->offset, dst_sz);

}

int GpuArray_memset(GpuArray *a, int data) {

if (!GpuArray_ISONESEGMENT(a))

return GA_UNSUPPORTED_ERROR;

return gpudata_memset(a->data, a->offset, data);

}

int GpuArray_copy(GpuArray *res, const GpuArray *a, ga_order order) {

int err;

err = GpuArray_empty(res, GpuArray_context(a), a->typecode,

a->nd, a->dimensions, order);

if (err != GA_NO_ERROR) return err;

err = GpuArray_move(res, a);

if (err != GA_NO_ERROR)

GpuArray_clear(res);

return err;

}

int GpuArray_transfer(GpuArray *res, const GpuArray *a) {

size_t sz;

unsigned int i;

if (!GpuArray_ISONESEGMENT(res))

return GA_UNSUPPORTED_ERROR;

if (!GpuArray_ISONESEGMENT(a))

return GA_UNSUPPORTED_ERROR;

if (res->typecode != a->typecode)

return GA_UNSUPPORTED_ERROR;

sz = gpuarray_get_elsize(a->typecode);

for (i = 0; i < a->nd; i++) sz *= a->dimensions[i];

return gpudata_transfer(res->data, res->offset, a->data, a->offset, sz);

}

int GpuArray_split(GpuArray **rs, const GpuArray *a, size_t n, size_t *p,

unsigned int axis) {

size_t i;

ssize_t *starts, *stops, *steps;

int err;

starts = calloc(a->nd, sizeof(ssize_t));

stops = calloc(a->nd, sizeof(ssize_t));

steps = calloc(a->nd, sizeof(ssize_t));

if (starts == NULL || stops == NULL || steps == NULL) {

free(starts);

free(stops);

free(steps);

return GA_MEMORY_ERROR;

}

for (i = 0; i < a->nd; i++) {

starts[i] = 0;

stops[i] = a->dimensions[i];

steps[i] = 1;

}

for (i = 0; i <= n; i++) {

if (i > 0)

starts[axis] = p[i-1];

else

starts[axis] = 0;

if (i < n)

stops[axis] = p[i];

else

stops[axis] = a->dimensions[axis];

err = GpuArray_index(rs[i], a, starts, stops, steps);

if (err != GA_NO_ERROR)

break;

}

free(starts);

free(stops);

free(steps);

if (err != GA_NO_ERROR) {

size_t ii;

for (ii = 0; ii < i; ii++)

GpuArray_clear(rs[ii]);

}

return err;

}

int GpuArray_concatenate(GpuArray *r, const GpuArray **as, size_t n,

unsigned int axis, int restype) {

size_t *dims, *res_dims;

size_t i, res_off;

unsigned int p;

int res_flags;

int err = GA_NO_ERROR;

if (axis >= as[0]->nd)

return GA_VALUE_ERROR;

dims = calloc(as[0]->nd, sizeof(size_t));

if (dims == NULL)

return GA_MEMORY_ERROR;

for (p = 0; p < as[0]->nd; p++) {

dims[p] = as[0]->dimensions[p];

}

if (!GpuArray_ISALIGNED(as[0])) {

err = GA_UNALIGNED_ERROR;

goto afterloop;

}

for (i = 1; i < n; i++) {

if (!GpuArray_ISALIGNED(as[i])) {

err = GA_UNALIGNED_ERROR;

goto afterloop;

}

if (as[i]->nd != as[0]->nd) {

err = GA_VALUE_ERROR;

goto afterloop;

}

for (p = 0; p < as[0]->nd; p++) {

if (p != axis && dims[p] != as[i]->dimensions[p]) {

err = GA_VALUE_ERROR;

goto afterloop;

} else if (p == axis) {

dims[p] += as[i]->dimensions[p];

}

afterloop:

if (err != GA_NO_ERROR) {

free(dims);

return err;

}

err = GpuArray_empty(r, GpuArray_context(as[0]), restype,

as[0]->nd, dims, GA_ANY_ORDER);

free(dims);

View remainder of file in raw view

libgpuarray/src/gpuarray_array.c at master · jrao1/libgpuarray · GitHub

Expand file tree

Breadcrumbs

gpuarray_array.c

Latest commit

History

Breadcrumbs

gpuarray_array.c

File metadata and controls