Coverage for onnxcustom/utils/nvprof2json.py: 99%

1"""

2@file

3@brief Converts traces from :epkg:`nvprof`.

4The source comes from `nvprof2json <https://github.com/ezyang/nvprof2json>`_.

5"""

7import sqlite3

8import enum

9import json

10import copy

11import io

12import os

13import zipfile

14import cxxfilt

15import pandas

18def convert_trace_to_json(filename, output=None, temporary_file=None,

19 verbose=0, fLOG=None):

20 """

21 Converts traces produced by :epkg:`nvprof` and saved with

22 format *sqlite3* (extension `.sql`). The output format

23 follows `Trace Event Format

24 <https://docs.google.com/document/d/

25 1CvAClvFfyA5R-PhYUmn5OOQtYMH4h6I0nSsKchNAySU/preview>`_.

27 :param filename: filename

28 :param output: output file or None

29 :param temporary_file: if the file needs to be unzipped,

30 this file will be created to be the unzipped file,

31 it is not cleaned after the unzipping.

32 :param verbose: verbosity

33 :param fLOG: logging function

34 :return: json (if output is None, the list of events otherwise)

36 This file, if not too big, can be viewed with `chrome-tracing`.

37 The traces are usually generated by using a command line similar to:

39 ::

41 nvprof -o gpu_profile.sql python plot_gpu_training.py

42 """

43 ext = os.path.splitext(filename)[-1]

44 if ext == ".zip":

45 if temporary_file in (None, ''):

46 temporary_file = filename + ".unzipped"

47 if os.path.exists(temporary_file):

48 if verbose > 0 and fLOG is not None:

49 fLOG( # pragma: no cover

50 "[convert_trace_to_json] %r already unzipped into %r"

51 "." % (filename, temporary_file))

52 else:

53 if verbose > 0 and fLOG is not None:

54 fLOG( # pragma: no cover

55 f"[convert_trace_to_json] unzipping to file {temporary_file!r}.")

56 zipf = zipfile.ZipFile(filename)

57 names = zipf.namelist()

58 if len(names) != 1:

59 raise RuntimeError( # pragma: no cover

60 f"More than one file is stored in zip file {filename!r}.")

61 stream = zipf.open(names[0], "r")

62 with open(temporary_file, "wb") as f:

63 while True:

64 data = stream.read(65536)

65 if len(data) == 0:

66 break

67 f.write(data)

68 zipf.close()

69 filename = temporary_file

71 conn = sqlite3.connect(filename)

72 conn.row_factory = sqlite3.Row

74 strings = {}

75 for r in conn.execute("SELECT _id_ as id, value FROM StringTable"):

76 strings[r["id"]] = _demangle(r["value"])

78 traceEvents = []

80 # """

81 # _id_: 11625

82 # cbid: 17

83 # start: 1496933427584362152

84 # end: 1496933427584362435

85 # processId: 1317533

86 # threadId: 1142654784

87 # correlationId: 13119

88 # returnValue: 0

89 # """

90 if verbose > 0 and fLOG is not None:

91 fLOG("[convert_trace_to_json] step 1 begin.")

92 for row in conn.execute("SELECT * FROM CUPTI_ACTIVITY_KIND_RUNTIME"):

93 try:

94 cbid = Cbids(row["cbid"]).name

95 except ValueError: # pragma: no cover

96 cbid = str(row["cbid"])

97 if verbose > 0 and fLOG is None:

98 fLOG(f"[convert_trace_to_json] unrecognized cbid {cbid!r}.")

99 event = {

100 "name": cbid,

101 "ph": "X", # Complete Event (Begin + End event)

102 "cat": "cuda",

103 "ts": _munge_time(row["start"]),

104 "dur": _munge_time(row["end"] - row["start"]),

105 "tid": f"Thread {row['threadId']}: Runtime API",

106 "pid": f"[{row['processId']}] Process",

107 "args": {

108 # ...

109 },

110 }

111 traceEvents.append(event)

112

113 # DRIVER?

114

115 # """

116 # _id_: 1

117 # flags: 2

118 # timestamp: 1496844806028263989

119 # id: 1

120 # objectKind: 2

121 # objectId: b'\xe5\xc0\x16\x00@\xe7\x10J\x00\x00\x00\x00'

122 # name: 3

123 # domain: 0

124 # """

125 if verbose > 0 and fLOG is not None:

126 fLOG("[convert_trace_to_json] step 2 begin.")

127 for row in conn.execute(" ".join([

128 "SELECT",

129 ",".join([

130 "start.name AS name",

131 "start.timestamp AS start_time",

132 "end.timestamp AS end_time"

133 ]),

134 "FROM",

135 "(SELECT * FROM CUPTI_ACTIVITY_KIND_MARKER WHERE name != 0) "

136 "AS start",

137 "LEFT JOIN",

138 "(SELECT * FROM CUPTI_ACTIVITY_KIND_MARKER WHERE name = 0) "

139 "AS end",

140 "ON start.id = end.id"])):

141 event = {

142 "name": strings[row["name"]],

143 "cat": "cuda",

144 "ts": _munge_time(row["start_time"]),

145 # Weirdly, these don't seem to be associated with a

146 # CPU/GPU. I guess there's no CUDA Context available

147 # when you run these, so it makes sense. But nvvp

148 # associates these with a GPU strangely enough

149 "tid": "Markers and Ranges",

150 "pid": "Markers and Ranges",

151 # parse objectId?

152 "args": {

153 # ...

154 },

155 }

156 if row["end_time"] is None:

157 event["ph"] = "I"

158 else:

159 event["ph"] = "X"

160 event["dur"] = _munge_time(row["end_time"] - row["start_time"])

161 traceEvents.append(event)

162

163 # """

164 # _id_: 1

165 # copyKind: 1

166 # srcKind: 1

167 # dstKind: 3

168 # flags: 0

169 # bytes: 7436640

170 # start: 1496933426915778221

171 # end: 1496933426916558424

172 # deviceId: 0

173 # contextId: 1

174 # streamId: 7

175 # correlationId: 809

176 # runtimeCorrelationId: 0

177 # """

178 if verbose > 0 and fLOG is not None:

179 fLOG("[convert_trace_to_json] step 3 begin.")

180 for row in conn.execute("SELECT * FROM CUPTI_ACTIVITY_KIND_MEMCPY"):

181 # copyKind:

182 # 1 - Memcpy HtoD

183 # 2 - Memcpy DtoH

184 # 8 - Memcpy DtoD

185 # flags: ???

186 # 0 - Sync

187 # 1 - Async

188 # srcKind/dstKind

189 # 1 - Pageable

190 # 2 - Page-locked ???

191 # 3 - Device

192 # eprintRow(row)

193 if row["copyKind"] == 1:

194 copyKind = "HtoD"

195 elif row["copyKind"] == 2:

196 copyKind = "DtoH"

197 elif row["copyKind"] == 8:

198 copyKind = "DtoD"

199 else:

200 copyKind = str(row["copyKind"])

201 if row["flags"] == 0:

202 flags = "sync"

203 elif row["flags"] == 1:

204 flags = "async"

205 else:

206 flags = str(row["flags"])

207 event = {

208 "name": f"Memcpy {copyKind} [{flags}]",

209 "ph": "X", # Complete Event (Begin + End event)

210 "cat": "cuda",

211 "ts": _munge_time(row["start"]),

212 "dur": _munge_time(row["end"] - row["start"]),

213 "tid": f"MemCpy ({copyKind})",

214 # lookup GPU name. This is tored in CUPTI_ACTIVITY_KIND_DEVICE

215 "pid": f"[{row['deviceId']}:{row['contextId']}] Overview",

216 "args": {

217 "Size": _sizeof_fmt(row["bytes"]),

218 },

219 }

220 traceEvents.append(event)

221

222 # name: index into StringTable

223 # What is thed difference between end and completed?

224 # """

225 # _id_: 1

226 # cacheConfig: b'\x00'

227 # sharedMemoryConfig: 1

228 # registersPerThread: 32

229 # partitionedGlobalCacheRequested: 2

230 # partitionedGlobalCacheExecuted: 2

231 # start: 1496844806032514222

232 # end: 1496844806032531694

233 # completed: 1496844806032531694

234 # deviceId: 0

235 # contextId: 1

236 # streamId: 7

237 # gridX: 57

238 # gridY: 1

239 # gridZ: 1

240 # blockX: 128

241 # blockY: 1

242 # blockZ: 1

243 # staticSharedMemory: 0

244 # dynamicSharedMemory: 0

245 # localMemoryPerThread: 0

246 # localMemoryTotal: 78643200

247 # correlationId: 487

248 # gridId: 669

249 # name: 5

250 # """

251 if verbose > 0 and fLOG is not None:

252 fLOG("[convert_trace_to_json] step 4 begin.")

253 for row in conn.execute(

254 "SELECT * FROM CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL"):

255 # eprint(strings[row["name"]])

256 # eprintRow(row)

257 event = {

258 "name": strings[row["name"]],

259 "ph": "X", # Complete Event (Begin + End event)

260 "cat": "cuda",

261 "ts": _munge_time(row["start"]),

262 "dur": _munge_time(row["end"] - row["start"]),

263 "tid": "Compute",

264 # lookup GPU name?

265 "pid": f"[{row['deviceId']}:{row['contextId']}] Overview",

266 "args": {

267 "Grid size": f"[ {row['gridX']}, {row['gridY']}, {row['gridZ']} ]",

268 "Block size": f"[ {row['blockX']}, {row['blockY']}, {row['blockZ']} ]",

269 # ...

270 },

271 }

272 alt_event = copy.deepcopy(event)

273 alt_event["tid"] = alt_event["name"]

274 alt_event["pid"] = f"[{row['deviceId']}:{row['contextId']}] Compute"

275 traceEvents.append(event)

276 traceEvents.append(alt_event)

277

278 if output not in (None, ''):

279 if verbose > 0 and fLOG is not None:

280 fLOG(

281 f"[convert_trace_to_json] converting into json in {output!r}.")

282 with open(output, "w", encoding="utf-8") as f:

283 json.dump(traceEvents, f, separators=(',\n', ':'))

284 f.write('\n')

285 if verbose > 0 and fLOG is not None:

286 fLOG("[convert_trace_to_json] done.")

287 return traceEvents

288 else:

289 if verbose > 0 and fLOG is not None:

290 fLOG( # pragma: no cover

291 "[convert_trace_to_json] converting into json.")

292 st = io.StringIO()

293 json.dump(traceEvents, st, separators=(',\n', ':'))

294 st.write('\n')

295 if verbose > 0 and fLOG is not None:

296 fLOG("[convert_trace_to_json] done.") # pragma: no cover

297 fLOG(st.getvalue()) # pragma: no cover

298 return st.getvalue()

299

300

301def _munge_time(t):

302 """Take a time from nvprof and convert it into a chrome://tracing time."""

303 # For strict correctness, divide by 1000, but this reduces accuracy.

304 return t # / 1000.

305

306

307def _demangle(name):

308 """Demangle a C++ identifier using c++filt"""

309 try:

310 return cxxfilt.demangle(name)

311 except cxxfilt.LibraryNotFound: # pragma: no cover

312 # One library is missing.

313 return name

314

315

316class Cbids(enum.IntEnum):

317 "List of events."

318 INVALID = 0

319 cudaDriverGetVersion = 1

320 cudaRuntimeGetVersion = 2

321 cudaGetDeviceCount = 3

322 cudaGetDeviceProperties = 4

323 cudaChooseDevice = 5

324 cudaGetChannelDesc = 6

325 cudaCreateChannelDesc = 7

326 cudaConfigureCall = 8

327 cudaSetupArgument = 9

328 cudaGetLastError = 10

329 cudaPeekAtLastError = 11

330 cudaGetErrorString = 12

331 cudaLaunch = 13

332 cudaFuncSetCacheConfig = 14

333 cudaFuncGetAttributes = 15

334 cudaSetDevice = 16

335 cudaGetDevice = 17

336 cudaSetValidDevices = 18

337 cudaSetDeviceFlags = 19

338 cudaMalloc = 20

339 cudaMallocPitch = 21

340 cudaFree = 22

341 cudaMallocArray = 23

342 cudaFreeArray = 24

343 cudaMallocHost = 25

344 cudaFreeHost = 26

345 cudaHostAlloc = 27

346 cudaHostGetDevicePointer = 28

347 cudaHostGetFlags = 29

348 cudaMemGetInfo = 30

349 cudaMemcpy = 31

350 cudaMemcpy2D = 32

351 cudaMemcpyToArray = 33

352 cudaMemcpy2DToArray = 34

353 cudaMemcpyFromArray = 35

354 cudaMemcpy2DFromArray = 36

355 cudaMemcpyArrayToArray = 37

356 cudaMemcpy2DArrayToArray = 38

357 cudaMemcpyToSymbol = 39

358 cudaMemcpyFromSymbol = 40

359 cudaMemcpyAsync = 41

360 cudaMemcpyToArrayAsync = 42

361 cudaMemcpyFromArrayAsync = 43

362 cudaMemcpy2DAsync = 44

363 cudaMemcpy2DToArrayAsync = 45

364 cudaMemcpy2DFromArrayAsync = 46

365 cudaMemcpyToSymbolAsync = 47

366 cudaMemcpyFromSymbolAsync = 48

367 cudaMemset = 49

368 cudaMemset2D = 50

369 cudaMemsetAsync = 51

370 cudaMemset2DAsync = 52

371 cudaGetSymbolAddress = 53

372 cudaGetSymbolSize = 54

373 cudaBindTexture = 55

374 cudaBindTexture2D = 56

375 cudaBindTextureToArray = 57

376 cudaUnbindTexture = 58

377 cudaGetTextureAlignmentOffset = 59

378 cudaGetTextureReference = 60

379 cudaBindSurfaceToArray = 61

380 cudaGetSurfaceReference = 62

381 cudaGLSetGLDevice = 63

382 cudaGLRegisterBufferObject = 64

383 cudaGLMapBufferObject = 65

384 cudaGLUnmapBufferObject = 66

385 cudaGLUnregisterBufferObject = 67

386 cudaGLSetBufferObjectMapFlags = 68

387 cudaGLMapBufferObjectAsync = 69

388 cudaGLUnmapBufferObjectAsync = 70

389 cudaWGLGetDevice = 71

390 cudaGraphicsGLRegisterImage = 72

391 cudaGraphicsGLRegisterBuffer = 73

392 cudaGraphicsUnregisterResource = 74

393 cudaGraphicsResourceSetMapFlags = 75

394 cudaGraphicsMapResources = 76

395 cudaGraphicsUnmapResources = 77

396 cudaGraphicsResourceGetMappedPointer = 78

397 cudaGraphicsSubResourceGetMappedArray = 79

398 cudaVDPAUGetDevice = 80

399 cudaVDPAUSetVDPAUDevice = 81

400 cudaGraphicsVDPAURegisterVideoSurface = 82

401 cudaGraphicsVDPAURegisterOutputSurface = 83

402 cudaD3D11GetDevice = 84

403 cudaD3D11GetDevices = 85

404 cudaD3D11SetDirect3DDevice = 86

405 cudaGraphicsD3D11RegisterResource = 87

406 cudaD3D10GetDevice = 88

407 cudaD3D10GetDevices = 89

408 cudaD3D10SetDirect3DDevice = 90

409 cudaGraphicsD3D10RegisterResource = 91

410 cudaD3D10RegisterResource = 92

411 cudaD3D10UnregisterResource = 93

412 cudaD3D10MapResources = 94

413 cudaD3D10UnmapResources = 95

414 cudaD3D10ResourceSetMapFlags = 96

415 cudaD3D10ResourceGetSurfaceDimensions = 97

416 cudaD3D10ResourceGetMappedArray = 98

417 cudaD3D10ResourceGetMappedPointer = 99

418 cudaD3D10ResourceGetMappedSize = 100

419 cudaD3D10ResourceGetMappedPitch = 101

420 cudaD3D9GetDevice = 102

421 cudaD3D9GetDevices = 103

422 cudaD3D9SetDirect3DDevice = 104

423 cudaD3D9GetDirect3DDevice = 105

424 cudaGraphicsD3D9RegisterResource = 106

425 cudaD3D9RegisterResource = 107

426 cudaD3D9UnregisterResource = 108

427 cudaD3D9MapResources = 109

428 cudaD3D9UnmapResources = 110

429 cudaD3D9ResourceSetMapFlags = 111

430 cudaD3D9ResourceGetSurfaceDimensions = 112

431 cudaD3D9ResourceGetMappedArray = 113

432 cudaD3D9ResourceGetMappedPointer = 114

433 cudaD3D9ResourceGetMappedSize = 115

434 cudaD3D9ResourceGetMappedPitch = 116

435 cudaD3D9Begin = 117

436 cudaD3D9End = 118

437 cudaD3D9RegisterVertexBuffer = 119

438 cudaD3D9UnregisterVertexBuffer = 120

439 cudaD3D9MapVertexBuffer = 121

440 cudaD3D9UnmapVertexBuffer = 122

441 cudaThreadExit = 123

442 cudaSetDoubleForDevice = 124

443 cudaSetDoubleForHost = 125

444 cudaThreadSynchronize = 126

445 cudaThreadGetLimit = 127

446 cudaThreadSetLimit = 128

447 cudaStreamCreate = 129

448 cudaStreamDestroy = 130

449 cudaStreamSynchronize = 131

450 cudaStreamQuery = 132

451 cudaEventCreate = 133

452 cudaEventCreateWithFlags = 134

453 cudaEventRecord = 135

454 cudaEventDestroy = 136

455 cudaEventSynchronize = 137

456 cudaEventQuery = 138

457 cudaEventElapsedTime = 139

458 cudaMalloc3D = 140

459 cudaMalloc3DArray = 141

460 cudaMemset3D = 142

461 cudaMemset3DAsync = 143

462 cudaMemcpy3D = 144

463 cudaMemcpy3DAsync = 145

464 cudaThreadSetCacheConfig = 146

465 cudaStreamWaitEvent = 147

466 cudaD3D11GetDirect3DDevice = 148

467 cudaD3D10GetDirect3DDevice = 149

468 cudaThreadGetCacheConfig = 150

469 cudaPointerGetAttributes = 151

470 cudaHostRegister = 152

471 cudaHostUnregister = 153

472 cudaDeviceCanAccessPeer = 154

473 cudaDeviceEnablePeerAccess = 155

474 cudaDeviceDisablePeerAccess = 156

475 cudaPeerRegister = 157

476 cudaPeerUnregister = 158

477 cudaPeerGetDevicePointer = 159

478 cudaMemcpyPeer = 160

479 cudaMemcpyPeerAsync = 161

480 cudaMemcpy3DPeer = 162

481 cudaMemcpy3DPeerAsync = 163

482 cudaDeviceReset = 164

483 cudaDeviceSynchronize = 165

484 cudaDeviceGetLimit = 166

485 cudaDeviceSetLimit = 167

486 cudaDeviceGetCacheConfig = 168

487 cudaDeviceSetCacheConfig = 169

488 cudaProfilerInitialize = 170

489 cudaProfilerStart = 171

490 cudaProfilerStop = 172

491 cudaDeviceGetByPCIBusId = 173

492 cudaDeviceGetPCIBusId = 174

493 cudaGLGetDevices = 175

494 cudaIpcGetEventHandle = 176

495 cudaIpcOpenEventHandle = 177

496 cudaIpcGetMemHandle = 178

497 cudaIpcOpenMemHandle = 179

498 cudaIpcCloseMemHandle = 180

499 cudaArrayGetInfo = 181

500 cudaFuncSetSharedMemConfig = 182

501 cudaDeviceGetSharedMemConfig = 183

502 cudaDeviceSetSharedMemConfig = 184

503 cudaCreateTextureObject = 185

504 cudaDestroyTextureObject = 186

505 cudaGetTextureObjectResourceDesc = 187

506 cudaGetTextureObjectTextureDesc = 188

507 cudaCreateSurfaceObject = 189

508 cudaDestroySurfaceObject = 190

509 cudaGetSurfaceObjectResourceDesc = 191

510 cudaMallocMipmappedArray = 192

511 cudaGetMipmappedArrayLevel = 193

512 cudaFreeMipmappedArray = 194

513 cudaBindTextureToMipmappedArray = 195

514 cudaGraphicsResourceGetMappedMipmappedArray = 196

515 cudaStreamAddCallback = 197

516 cudaStreamCreateWithFlags = 198

517 cudaGetTextureObjectResourceViewDesc = 199

518 cudaDeviceGetAttribute = 200

519 cudaStreamDestroy_v5050 = 201

520 cudaStreamCreateWithPriority = 202

521 cudaStreamGetPriority = 203

522 cudaStreamGetFlags = 204

523 cudaDeviceGetStreamPriorityRange = 205

524 cudaMallocManaged = 206

525 cudaOccupancyMaxActiveBlocksPerMultiprocessor_v6000 = 207

526 cudaStreamAttachMemAsync = 208

527 cudaGetErrorName = 209

528 cudaOccupancyMaxActiveBlocksPerMultiprocessor_v6050 = 210

529 cudaLaunchKernel = 211

530 cudaGetDeviceFlags = 212

531 cudaLaunch_ptsz = 213

532 cudaLaunchKernel_ptsz = 214

533 cudaMemcpy_ptds = 215

534 cudaMemcpy2D_ptds = 216

535 cudaMemcpyToArray_ptds = 217

536 cudaMemcpy2DToArray_ptds = 218

537 cudaMemcpyFromArray_ptds = 219

538 cudaMemcpy2DFromArray_ptds = 220

539 cudaMemcpyArrayToArray_ptds = 221

540 cudaMemcpy2DArrayToArray_ptds = 222

541 cudaMemcpyToSymbol_ptds = 223

542 cudaMemcpyFromSymbol_ptds = 224

543 cudaMemcpyAsync_ptsz = 225

544 cudaMemcpyToArrayAsync_ptsz = 226

545 cudaMemcpyFromArrayAsync_ptsz = 227

546 cudaMemcpy2DAsync_ptsz = 228

547 cudaMemcpy2DToArrayAsync_ptsz = 229

548 cudaMemcpy2DFromArrayAsync_ptsz = 230

549 cudaMemcpyToSymbolAsync_ptsz = 231

550 cudaMemcpyFromSymbolAsync_ptsz = 232

551 cudaMemset_ptds = 233

552 cudaMemset2D_ptds = 234

553 cudaMemsetAsync_ptsz = 235

554 cudaMemset2DAsync_ptsz = 236

555 cudaStreamGetPriority_ptsz = 237

556 cudaStreamGetFlags_ptsz = 238

557 cudaStreamSynchronize_ptsz = 239

558 cudaStreamQuery_ptsz = 240

559 cudaStreamAttachMemAsync_ptsz = 241

560 cudaEventRecord_ptsz = 242

561 cudaMemset3D_ptds = 243

562 cudaMemset3DAsync_ptsz = 244

563 cudaMemcpy3D_ptds = 245

564 cudaMemcpy3DAsync_ptsz = 246

565 cudaStreamWaitEvent_ptsz = 247

566 cudaStreamAddCallback_ptsz = 248

567 cudaMemcpy3DPeer_ptds = 249

568 cudaMemcpy3DPeerAsync_ptsz = 250

569 cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags = 251

570 cudaMemPrefetchAsync = 252

571 cudaMemPrefetchAsync_ptsz = 253

572 cudaMemAdvise = 254

573 cudaDeviceGetP2PAttribute = 255

574 cudaGraphicsEGLRegisterImage = 256

575 cudaEGLStreamConsumerConnect = 257

576 cudaEGLStreamConsumerDisconnect = 258

577 cudaEGLStreamConsumerAcquireFrame = 259

578 cudaEGLStreamConsumerReleaseFrame = 260

579 cudaEGLStreamProducerConnect = 261

580 cudaEGLStreamProducerDisconnect = 262

581 cudaEGLStreamProducerPresentFrame = 263

582 cudaEGLStreamProducerReturnFrame = 264

583 cudaGraphicsResourceGetMappedEglFrame = 265

584 cudaMemRangeGetAttribute = 266

585 cudaMemRangeGetAttributes = 267

586 cudaEGLStreamConsumerConnectWithFlags = 268

587 cudaLaunchCooperativeKernel = 269

588 cudaLaunchCooperativeKernel_ptsz = 270

589 cudaEventCreateFromEGLSync = 271

590 cudaLaunchCooperativeKernelMultiDevice = 272

591 cudaFuncSetAttribute = 273

592 cudaImportExternalMemory = 274

593 cudaExternalMemoryGetMappedBuffer = 275

594 cudaExternalMemoryGetMappedMipmappedArray = 276

595 cudaDestroyExternalMemory = 277

596 cudaImportExternalSemaphore = 278

597 cudaSignalExternalSemaphoresAsync = 279

598 cudaSignalExternalSemaphoresAsync_ptsz = 280

599 cudaWaitExternalSemaphoresAsync = 281

600 cudaWaitExternalSemaphoresAsync_ptsz = 282

601 cudaDestroyExternalSemaphore = 283

602 cudaLaunchHostFunc = 284

603 cudaLaunchHostFunc_ptsz = 285

604 cudaGraphCreate = 286

605 cudaGraphKernelNodeGetParams = 287

606 cudaGraphKernelNodeSetParams = 288

607 cudaGraphAddKernelNode = 289

608 cudaGraphAddMemcpyNode = 290

609 cudaGraphMemcpyNodeGetParams = 291

610 cudaGraphMemcpyNodeSetParams = 292

611 cudaGraphAddMemsetNode = 293

612 cudaGraphMemsetNodeGetParams = 294

613 cudaGraphMemsetNodeSetParams = 295

614 cudaGraphAddHostNode = 296

615 cudaGraphHostNodeGetParams = 297

616 cudaGraphAddChildGraphNode = 298

617 cudaGraphChildGraphNodeGetGraph = 299

618 cudaGraphAddEmptyNode = 300

619 cudaGraphClone = 301

620 cudaGraphNodeFindInClone = 302

621 cudaGraphNodeGetType = 303

622 cudaGraphGetRootNodes = 304

623 cudaGraphNodeGetDependencies = 305

624 cudaGraphNodeGetDependentNodes = 306

625 cudaGraphAddDependencies = 307

626 cudaGraphRemoveDependencies = 308

627 cudaGraphDestroyNode = 309

628 cudaGraphInstantiate = 310

629 cudaGraphLaunch = 311

630 cudaGraphLaunch_ptsz = 312

631 cudaGraphExecDestroy = 313

632 cudaGraphDestroy = 314

633 cudaStreamBeginCapture = 315

634 cudaStreamBeginCapture_ptsz = 316

635 cudaStreamIsCapturing = 317

636 cudaStreamIsCapturing_ptsz = 318

637 cudaStreamEndCapture = 319

638 cudaStreamEndCapture_ptsz = 320

639 cudaGraphHostNodeSetParams = 321

640 cudaGraphGetNodes = 322

641 cudaGraphGetEdges = 323

642 cudaStreamGetCaptureInfo = 324

643 cudaStreamGetCaptureInfo_ptsz = 325

644 cudaGraphExecKernelNodeSetParams = 326

645 cudaThreadExchangeStreamCaptureMode = 327

646 cudaDeviceGetNvSciSyncAttributes = 328

647 cudaOccupancyAvailableDynamicSMemPerBlock = 329

648 cudaStreamSetFlags = 330

649 cudaStreamSetFlags_ptsz = 331

650 cudaGraphExecMemcpyNodeSetParams = 332

651 cudaGraphExecMemsetNodeSetParams = 333

652 cudaGraphExecHostNodeSetParams = 334

653 cudaGraphExecUpdate = 335

654 SIZE = 336

655 FORCE_INT = 0x7FFFFFFF

656

657

658def _sizeof_fmt(num, suffix='B'):

659 """Format size with metric units (like nvvp)"""

660 for unit in ['', 'K', 'M', 'G', 'T', 'P', 'E', 'Z']:

661 if abs(num) < 1000.0:

662 return f"{num:3.1f}{unit}{suffix}"

663 num /= 1000.0 # pragma: no cover

664 return f"{num:.1f}{'Y'}{suffix}" # pragma: no cover

665

666

667def json_to_dataframe(js):

668 """

669 Converts a json dump obtained with function

670 @see fn convert_trace_to_json

671 to a dataframe.

672

673 :param js: a filename, a json string, a stream containing json

674 :return: a dataframe

675 """

676 if isinstance(js, str) and os.path.exists(js):

677 if len(js) < 5000:

678 df = pandas.read_json(js)

679 else: # pragma: no cover

680 st = io.StringIO(js)

681 df = pandas.read_json(st)

682 else:

683 df = pandas.read_json(js)

684

685 df['ts_sec'] = df['ts'].apply(lambda t: t / 1e9)

686 return df

687

688

689def json_to_dataframe_streaming(js, chunksize=100000, flatten=False, **kwargs):

690 """

691 Converts a big json dump (from @see fn convert_trace_to_json)

692 to a dataframe. The function processes the data by streaming to avoid

693 loading huge data in memory.

694 Returns an iterator on dataframes.

695 The function relies on :epkg:`pandas_streaming`.

696

697 :param js: a filename, a json string, a stream containing json

698 :param chunksize:

699 see :func:`pandas_streaming.df.StreamingDataFrame.read_json`

700 :param flatten:

701 see :func:`pandas_streaming.df.StreamingDataFrame.read_json`

702 :param kwargs:

703 see :func:`pandas_streaming.df.StreamingDataFrame.read_json`

704 :return: a dataframe

705 """

706 from pandas_streaming.df import StreamingDataFrame # pylint: disable=C0415

707 if isinstance(js, str):

708 if len(js) < 5000 and os.path.exists(js):

709 sdf = StreamingDataFrame.read_json(js)

710 else:

711 raise RuntimeError(

712 "Use a stream or function json_to_dataframe instead of "

713 "the streaming version.")

714 else:

715 sdf = StreamingDataFrame.read_json(js)

716

717 sdf['ts_sec'] = sdf['ts'].apply(lambda t: t / 1e9)

718 return sdf