| 1 | /*BHEADER**********************************************************************
|
|---|
| 2 | * Copyright (c) 2008, Lawrence Livermore National Security, LLC.
|
|---|
| 3 | * Produced at the Lawrence Livermore National Laboratory.
|
|---|
| 4 | * This file is part of HYPRE. See file COPYRIGHT for details.
|
|---|
| 5 | *
|
|---|
| 6 | * HYPRE is free software; you can redistribute it and/or modify it under the
|
|---|
| 7 | * terms of the GNU Lesser General Public License (as published by the Free
|
|---|
| 8 | * Software Foundation) version 2.1 dated February 1999.
|
|---|
| 9 | *
|
|---|
| 10 | * $Revision: 2.4 $
|
|---|
| 11 | ***********************************************************************EHEADER*/
|
|---|
| 12 |
|
|---|
| 13 |
|
|---|
| 14 | /******************************************************************************
|
|---|
| 15 | *
|
|---|
| 16 | * Routines for doing timing.
|
|---|
| 17 | *
|
|---|
| 18 | *****************************************************************************/
|
|---|
| 19 |
|
|---|
| 20 | #define HYPRE_TIMING_GLOBALS
|
|---|
| 21 | #include "utilities.h"
|
|---|
| 22 | #include "timing.h"
|
|---|
| 23 |
|
|---|
| 24 |
|
|---|
| 25 | /*-------------------------------------------------------
|
|---|
| 26 | * Timing macros
|
|---|
| 27 | *-------------------------------------------------------*/
|
|---|
| 28 |
|
|---|
| 29 | #define hypre_StartTiming() \
|
|---|
| 30 | hypre_TimingWallCount -= time_getWallclockSeconds();\
|
|---|
| 31 | hypre_TimingCPUCount -= time_getCPUSeconds()
|
|---|
| 32 |
|
|---|
| 33 | #define hypre_StopTiming() \
|
|---|
| 34 | hypre_TimingWallCount += time_getWallclockSeconds();\
|
|---|
| 35 | hypre_TimingCPUCount += time_getCPUSeconds()
|
|---|
| 36 |
|
|---|
| 37 | #ifndef HYPRE_USE_PTHREADS
|
|---|
| 38 | #define hypre_global_timing_ref(index,field) hypre_global_timing->field
|
|---|
| 39 | #else
|
|---|
| 40 | #define hypre_global_timing_ref(index,field) \
|
|---|
| 41 | hypre_global_timing[index].field
|
|---|
| 42 | #endif
|
|---|
| 43 |
|
|---|
| 44 | /*--------------------------------------------------------------------------
|
|---|
| 45 | * hypre_InitializeTiming
|
|---|
| 46 | *--------------------------------------------------------------------------*/
|
|---|
| 47 |
|
|---|
| 48 | int
|
|---|
| 49 | hypre_InitializeTiming( const char *name )
|
|---|
| 50 | {
|
|---|
| 51 | int time_index;
|
|---|
| 52 |
|
|---|
| 53 | double *old_wall_time;
|
|---|
| 54 | double *old_cpu_time;
|
|---|
| 55 | double *old_flops;
|
|---|
| 56 | char **old_name;
|
|---|
| 57 | int *old_state;
|
|---|
| 58 | int *old_num_regs;
|
|---|
| 59 |
|
|---|
| 60 | int new_name;
|
|---|
| 61 | int i;
|
|---|
| 62 | #ifdef HYPRE_USE_PTHREADS
|
|---|
| 63 | int threadid = hypre_GetThreadID();
|
|---|
| 64 | #endif
|
|---|
| 65 |
|
|---|
| 66 | /*-------------------------------------------------------
|
|---|
| 67 | * Allocate global TimingType structure if needed
|
|---|
| 68 | *-------------------------------------------------------*/
|
|---|
| 69 |
|
|---|
| 70 | if (hypre_global_timing == NULL)
|
|---|
| 71 | {
|
|---|
| 72 | #ifndef HYPRE_USE_PTHREADS
|
|---|
| 73 | hypre_global_timing = hypre_CTAlloc(hypre_TimingType, 1);
|
|---|
| 74 | #else
|
|---|
| 75 | hypre_global_timing = hypre_CTAlloc(hypre_TimingType,
|
|---|
| 76 | hypre_NumThreads + 1);
|
|---|
| 77 | #endif
|
|---|
| 78 | }
|
|---|
| 79 |
|
|---|
| 80 | /*-------------------------------------------------------
|
|---|
| 81 | * Check to see if name has already been registered
|
|---|
| 82 | *-------------------------------------------------------*/
|
|---|
| 83 |
|
|---|
| 84 | new_name = 1;
|
|---|
| 85 | for (i = 0; i < (hypre_global_timing_ref(threadid, size)); i++)
|
|---|
| 86 | {
|
|---|
| 87 | if (hypre_TimingNumRegs(i) > 0)
|
|---|
| 88 | {
|
|---|
| 89 | if (strcmp(name, hypre_TimingName(i)) == 0)
|
|---|
| 90 | {
|
|---|
| 91 | new_name = 0;
|
|---|
| 92 | time_index = i;
|
|---|
| 93 | hypre_TimingNumRegs(time_index) ++;
|
|---|
| 94 | break;
|
|---|
| 95 | }
|
|---|
| 96 | }
|
|---|
| 97 | }
|
|---|
| 98 |
|
|---|
| 99 | if (new_name)
|
|---|
| 100 | {
|
|---|
| 101 | for (i = 0; i < hypre_global_timing_ref(threadid ,size); i++)
|
|---|
| 102 | {
|
|---|
| 103 | if (hypre_TimingNumRegs(i) == 0)
|
|---|
| 104 | {
|
|---|
| 105 | break;
|
|---|
| 106 | }
|
|---|
| 107 | }
|
|---|
| 108 | time_index = i;
|
|---|
| 109 | }
|
|---|
| 110 |
|
|---|
| 111 | /*-------------------------------------------------------
|
|---|
| 112 | * Register the new timing name
|
|---|
| 113 | *-------------------------------------------------------*/
|
|---|
| 114 |
|
|---|
| 115 | if (new_name)
|
|---|
| 116 | {
|
|---|
| 117 | if (time_index == (hypre_global_timing_ref(threadid, size)))
|
|---|
| 118 | {
|
|---|
| 119 | old_wall_time = (hypre_global_timing_ref(threadid, wall_time));
|
|---|
| 120 | old_cpu_time = (hypre_global_timing_ref(threadid, cpu_time));
|
|---|
| 121 | old_flops = (hypre_global_timing_ref(threadid, flops));
|
|---|
| 122 | old_name = (hypre_global_timing_ref(threadid, name));
|
|---|
| 123 | old_state = (hypre_global_timing_ref(threadid, state));
|
|---|
| 124 | old_num_regs = (hypre_global_timing_ref(threadid, num_regs));
|
|---|
| 125 |
|
|---|
| 126 | (hypre_global_timing_ref(threadid, wall_time)) =
|
|---|
| 127 | hypre_CTAlloc(double, (time_index+1));
|
|---|
| 128 | (hypre_global_timing_ref(threadid, cpu_time)) =
|
|---|
| 129 | hypre_CTAlloc(double, (time_index+1));
|
|---|
| 130 | (hypre_global_timing_ref(threadid, flops)) =
|
|---|
| 131 | hypre_CTAlloc(double, (time_index+1));
|
|---|
| 132 | (hypre_global_timing_ref(threadid, name)) =
|
|---|
| 133 | hypre_CTAlloc(char *, (time_index+1));
|
|---|
| 134 | (hypre_global_timing_ref(threadid, state)) =
|
|---|
| 135 | hypre_CTAlloc(int, (time_index+1));
|
|---|
| 136 | (hypre_global_timing_ref(threadid, num_regs)) =
|
|---|
| 137 | hypre_CTAlloc(int, (time_index+1));
|
|---|
| 138 | (hypre_global_timing_ref(threadid, size)) ++;
|
|---|
| 139 |
|
|---|
| 140 | for (i = 0; i < time_index; i++)
|
|---|
| 141 | {
|
|---|
| 142 | hypre_TimingWallTime(i) = old_wall_time[i];
|
|---|
| 143 | hypre_TimingCPUTime(i) = old_cpu_time[i];
|
|---|
| 144 | hypre_TimingFLOPS(i) = old_flops[i];
|
|---|
| 145 | hypre_TimingName(i) = old_name[i];
|
|---|
| 146 | hypre_TimingState(i) = old_state[i];
|
|---|
| 147 | hypre_TimingNumRegs(i) = old_num_regs[i];
|
|---|
| 148 | }
|
|---|
| 149 |
|
|---|
| 150 | hypre_TFree(old_wall_time);
|
|---|
| 151 | hypre_TFree(old_cpu_time);
|
|---|
| 152 | hypre_TFree(old_flops);
|
|---|
| 153 | hypre_TFree(old_name);
|
|---|
| 154 | hypre_TFree(old_state);
|
|---|
| 155 | hypre_TFree(old_num_regs);
|
|---|
| 156 | }
|
|---|
| 157 |
|
|---|
| 158 | hypre_TimingName(time_index) = hypre_CTAlloc(char, 80);
|
|---|
| 159 | strncpy(hypre_TimingName(time_index), name, 79);
|
|---|
| 160 | hypre_TimingState(time_index) = 0;
|
|---|
| 161 | hypre_TimingNumRegs(time_index) = 1;
|
|---|
| 162 | (hypre_global_timing_ref(threadid, num_names)) ++;
|
|---|
| 163 | }
|
|---|
| 164 |
|
|---|
| 165 | return time_index;
|
|---|
| 166 | }
|
|---|
| 167 |
|
|---|
| 168 | /*--------------------------------------------------------------------------
|
|---|
| 169 | * hypre_FinalizeTiming
|
|---|
| 170 | *--------------------------------------------------------------------------*/
|
|---|
| 171 |
|
|---|
| 172 | int
|
|---|
| 173 | hypre_FinalizeTiming( int time_index )
|
|---|
| 174 | {
|
|---|
| 175 | int ierr = 0;
|
|---|
| 176 | int i;
|
|---|
| 177 | #ifdef HYPRE_USE_PTHREADS
|
|---|
| 178 | int threadid = hypre_GetThreadID();
|
|---|
| 179 | int free_global_timing;
|
|---|
| 180 | #endif
|
|---|
| 181 |
|
|---|
| 182 | if (hypre_global_timing == NULL)
|
|---|
| 183 | return ierr;
|
|---|
| 184 |
|
|---|
| 185 | if (time_index < (hypre_global_timing_ref(threadid, size)))
|
|---|
| 186 | {
|
|---|
| 187 | if (hypre_TimingNumRegs(time_index) > 0)
|
|---|
| 188 | {
|
|---|
| 189 | hypre_TimingNumRegs(time_index) --;
|
|---|
| 190 | }
|
|---|
| 191 |
|
|---|
| 192 | if (hypre_TimingNumRegs(time_index) == 0)
|
|---|
| 193 | {
|
|---|
| 194 | hypre_TFree(hypre_TimingName(time_index));
|
|---|
| 195 | (hypre_global_timing_ref(threadid, num_names)) --;
|
|---|
| 196 | }
|
|---|
| 197 | }
|
|---|
| 198 |
|
|---|
| 199 | #ifdef HYPRE_USE_PTHREADS
|
|---|
| 200 |
|
|---|
| 201 | free_global_timing = 1;
|
|---|
| 202 | for (i = 0; i <= hypre_NumThreads; i++)
|
|---|
| 203 | {
|
|---|
| 204 | if (hypre_global_timing_ref(i, num_names))
|
|---|
| 205 | {
|
|---|
| 206 | free_global_timing = 0;
|
|---|
| 207 | break;
|
|---|
| 208 | }
|
|---|
| 209 | }
|
|---|
| 210 |
|
|---|
| 211 | if (free_global_timing)
|
|---|
| 212 | {
|
|---|
| 213 | pthread_mutex_lock(&time_mtx);
|
|---|
| 214 | if(hypre_global_timing)
|
|---|
| 215 | {
|
|---|
| 216 | for (i = 0; i <= hypre_NumThreads; i++)
|
|---|
| 217 |
|
|---|
| 218 | {
|
|---|
| 219 | hypre_TFree(hypre_global_timing_ref(i, wall_time));
|
|---|
| 220 | hypre_TFree(hypre_global_timing_ref(i, cpu_time));
|
|---|
| 221 | hypre_TFree(hypre_global_timing_ref(i, flops));
|
|---|
| 222 | hypre_TFree(hypre_global_timing_ref(i, name));
|
|---|
| 223 | hypre_TFree(hypre_global_timing_ref(i, state));
|
|---|
| 224 | hypre_TFree(hypre_global_timing_ref(i, num_regs));
|
|---|
| 225 | }
|
|---|
| 226 |
|
|---|
| 227 | hypre_TFree(hypre_global_timing);
|
|---|
| 228 | hypre_global_timing = NULL;
|
|---|
| 229 | }
|
|---|
| 230 | pthread_mutex_unlock(&time_mtx);
|
|---|
| 231 | }
|
|---|
| 232 |
|
|---|
| 233 | #else
|
|---|
| 234 |
|
|---|
| 235 | if ((hypre_global_timing -> num_names) == 0)
|
|---|
| 236 | {
|
|---|
| 237 | for (i = 0; i < (hypre_global_timing -> size); i++)
|
|---|
| 238 | {
|
|---|
| 239 | hypre_TFree(hypre_global_timing_ref(i, wall_time));
|
|---|
| 240 | hypre_TFree(hypre_global_timing_ref(i, cpu_time));
|
|---|
| 241 | hypre_TFree(hypre_global_timing_ref(i, flops));
|
|---|
| 242 | hypre_TFree(hypre_global_timing_ref(i, name));
|
|---|
| 243 | hypre_TFree(hypre_global_timing_ref(i, state));
|
|---|
| 244 | hypre_TFree(hypre_global_timing_ref(i, num_regs));
|
|---|
| 245 | }
|
|---|
| 246 |
|
|---|
| 247 | hypre_TFree(hypre_global_timing);
|
|---|
| 248 | hypre_global_timing = NULL;
|
|---|
| 249 | }
|
|---|
| 250 |
|
|---|
| 251 | #endif
|
|---|
| 252 |
|
|---|
| 253 | return ierr;
|
|---|
| 254 | }
|
|---|
| 255 |
|
|---|
| 256 | /*--------------------------------------------------------------------------
|
|---|
| 257 | * hypre_IncFLOPCount
|
|---|
| 258 | *--------------------------------------------------------------------------*/
|
|---|
| 259 |
|
|---|
| 260 | int
|
|---|
| 261 | hypre_IncFLOPCount( int inc )
|
|---|
| 262 | {
|
|---|
| 263 | int ierr = 0;
|
|---|
| 264 | #ifdef HYPRE_USE_PTHREADS
|
|---|
| 265 | int threadid = hypre_GetThreadID();
|
|---|
| 266 | #endif
|
|---|
| 267 |
|
|---|
| 268 | if (hypre_global_timing == NULL)
|
|---|
| 269 | return ierr;
|
|---|
| 270 |
|
|---|
| 271 | hypre_TimingFLOPCount += (double) (inc);
|
|---|
| 272 |
|
|---|
| 273 | #ifdef HYPRE_USE_PTHREADS
|
|---|
| 274 | if (threadid != hypre_NumThreads)
|
|---|
| 275 | {
|
|---|
| 276 | pthread_mutex_lock(&time_mtx);
|
|---|
| 277 | hypre_TimingAllFLOPS += (double) (inc);
|
|---|
| 278 | pthread_mutex_unlock(&time_mtx);
|
|---|
| 279 | }
|
|---|
| 280 | #endif
|
|---|
| 281 |
|
|---|
| 282 | return ierr;
|
|---|
| 283 | }
|
|---|
| 284 |
|
|---|
| 285 | /*--------------------------------------------------------------------------
|
|---|
| 286 | * hypre_BeginTiming
|
|---|
| 287 | *--------------------------------------------------------------------------*/
|
|---|
| 288 |
|
|---|
| 289 | int
|
|---|
| 290 | hypre_BeginTiming( int time_index )
|
|---|
| 291 | {
|
|---|
| 292 | int ierr = 0;
|
|---|
| 293 | #ifdef HYPRE_USE_PTHREADS
|
|---|
| 294 | int threadid = hypre_GetThreadID();
|
|---|
| 295 | #endif
|
|---|
| 296 |
|
|---|
| 297 | if (hypre_global_timing == NULL)
|
|---|
| 298 | return ierr;
|
|---|
| 299 |
|
|---|
| 300 | if (hypre_TimingState(time_index) == 0)
|
|---|
| 301 | {
|
|---|
| 302 | hypre_StopTiming();
|
|---|
| 303 | hypre_TimingWallTime(time_index) -= hypre_TimingWallCount;
|
|---|
| 304 | hypre_TimingCPUTime(time_index) -= hypre_TimingCPUCount;
|
|---|
| 305 | #ifdef HYPRE_USE_PTHREADS
|
|---|
| 306 | if (threadid != hypre_NumThreads)
|
|---|
| 307 | hypre_TimingFLOPS(time_index) -= hypre_TimingFLOPCount;
|
|---|
| 308 | else
|
|---|
| 309 | hypre_TimingFLOPS(time_index) -= hypre_TimingAllFLOPS;
|
|---|
| 310 | #else
|
|---|
| 311 | hypre_TimingFLOPS(time_index) -= hypre_TimingFLOPCount;
|
|---|
| 312 | #endif
|
|---|
| 313 |
|
|---|
| 314 | hypre_StartTiming();
|
|---|
| 315 | }
|
|---|
| 316 | hypre_TimingState(time_index) ++;
|
|---|
| 317 |
|
|---|
| 318 | return ierr;
|
|---|
| 319 | }
|
|---|
| 320 |
|
|---|
| 321 | /*--------------------------------------------------------------------------
|
|---|
| 322 | * hypre_EndTiming
|
|---|
| 323 | *--------------------------------------------------------------------------*/
|
|---|
| 324 |
|
|---|
| 325 | int
|
|---|
| 326 | hypre_EndTiming( int time_index )
|
|---|
| 327 | {
|
|---|
| 328 | int ierr = 0;
|
|---|
| 329 | #ifdef HYPRE_USE_PTHREADS
|
|---|
| 330 | int threadid = hypre_GetThreadID();
|
|---|
| 331 | #endif
|
|---|
| 332 |
|
|---|
| 333 | if (hypre_global_timing == NULL)
|
|---|
| 334 | return ierr;
|
|---|
| 335 |
|
|---|
| 336 | hypre_TimingState(time_index) --;
|
|---|
| 337 | if (hypre_TimingState(time_index) == 0)
|
|---|
| 338 | {
|
|---|
| 339 | hypre_StopTiming();
|
|---|
| 340 | hypre_TimingWallTime(time_index) += hypre_TimingWallCount;
|
|---|
| 341 | hypre_TimingCPUTime(time_index) += hypre_TimingCPUCount;
|
|---|
| 342 | #ifdef HYPRE_USE_PTHREADS
|
|---|
| 343 | if (threadid != hypre_NumThreads)
|
|---|
| 344 | hypre_TimingFLOPS(time_index) += hypre_TimingFLOPCount;
|
|---|
| 345 | else
|
|---|
| 346 | hypre_TimingFLOPS(time_index) += hypre_TimingAllFLOPS;
|
|---|
| 347 | #else
|
|---|
| 348 | hypre_TimingFLOPS(time_index) += hypre_TimingFLOPCount;
|
|---|
| 349 | #endif
|
|---|
| 350 | hypre_StartTiming();
|
|---|
| 351 | }
|
|---|
| 352 |
|
|---|
| 353 | return ierr;
|
|---|
| 354 | }
|
|---|
| 355 |
|
|---|
| 356 | /*--------------------------------------------------------------------------
|
|---|
| 357 | * hypre_ClearTiming
|
|---|
| 358 | *--------------------------------------------------------------------------*/
|
|---|
| 359 |
|
|---|
| 360 | int
|
|---|
| 361 | hypre_ClearTiming( )
|
|---|
| 362 | {
|
|---|
| 363 | int ierr = 0;
|
|---|
| 364 | int i;
|
|---|
| 365 | #ifdef HYPRE_USE_PTHREADS
|
|---|
| 366 | int threadid = hypre_GetThreadID();
|
|---|
| 367 | #endif
|
|---|
| 368 |
|
|---|
| 369 | if (hypre_global_timing == NULL)
|
|---|
| 370 | return ierr;
|
|---|
| 371 |
|
|---|
| 372 | for (i = 0; i < (hypre_global_timing_ref(threadid,size)); i++)
|
|---|
| 373 | {
|
|---|
| 374 | hypre_TimingWallTime(i) = 0.0;
|
|---|
| 375 | hypre_TimingCPUTime(i) = 0.0;
|
|---|
| 376 | hypre_TimingFLOPS(i) = 0.0;
|
|---|
| 377 | }
|
|---|
| 378 |
|
|---|
| 379 | return ierr;
|
|---|
| 380 | }
|
|---|
| 381 |
|
|---|
| 382 | /*--------------------------------------------------------------------------
|
|---|
| 383 | * hypre_PrintTiming
|
|---|
| 384 | *--------------------------------------------------------------------------*/
|
|---|
| 385 |
|
|---|
| 386 | #ifndef HYPRE_USE_PTHREADS /* non-threaded version of hypre_PrintTiming */
|
|---|
| 387 |
|
|---|
| 388 | int
|
|---|
| 389 | hypre_PrintTiming( const char *heading,
|
|---|
| 390 | double *wall_time_ptr,
|
|---|
| 391 | MPI_Comm comm )
|
|---|
| 392 | {
|
|---|
| 393 | int ierr = 0;
|
|---|
| 394 |
|
|---|
| 395 | double local_wall_time;
|
|---|
| 396 | double local_cpu_time;
|
|---|
| 397 | double wall_time;
|
|---|
| 398 | double cpu_time;
|
|---|
| 399 | double wall_mflops;
|
|---|
| 400 | double cpu_mflops;
|
|---|
| 401 |
|
|---|
| 402 | int i;
|
|---|
| 403 | int myrank;
|
|---|
| 404 |
|
|---|
| 405 | if (hypre_global_timing == NULL)
|
|---|
| 406 | return ierr;
|
|---|
| 407 |
|
|---|
| 408 | MPI_Comm_rank(comm, &myrank );
|
|---|
| 409 |
|
|---|
| 410 | /* print heading */
|
|---|
| 411 | if (myrank == 0)
|
|---|
| 412 | {
|
|---|
| 413 | printf("=============================================\n");
|
|---|
| 414 | printf("%s:\n", heading);
|
|---|
| 415 | printf("=============================================\n");
|
|---|
| 416 | }
|
|---|
| 417 |
|
|---|
| 418 | for (i = 0; i < (hypre_global_timing -> size); i++)
|
|---|
| 419 | {
|
|---|
| 420 | if (hypre_TimingNumRegs(i) > 0)
|
|---|
| 421 | {
|
|---|
| 422 | local_wall_time = hypre_TimingWallTime(i);
|
|---|
| 423 | local_cpu_time = hypre_TimingCPUTime(i);
|
|---|
| 424 | MPI_Allreduce(&local_wall_time, &wall_time, 1,
|
|---|
| 425 | MPI_DOUBLE, MPI_MAX, comm);
|
|---|
| 426 | MPI_Allreduce(&local_cpu_time, &cpu_time, 1,
|
|---|
| 427 | MPI_DOUBLE, MPI_MAX, comm);
|
|---|
| 428 |
|
|---|
| 429 | if (myrank == 0)
|
|---|
| 430 | {
|
|---|
| 431 | printf("%s:\n", hypre_TimingName(i));
|
|---|
| 432 | *wall_time_ptr = wall_time;
|
|---|
| 433 | /* print wall clock info */
|
|---|
| 434 | printf("%s wall clock time = %f seconds\n", hypre_TimingName(i), wall_time);
|
|---|
| 435 | if (wall_time)
|
|---|
| 436 | wall_mflops = hypre_TimingFLOPS(i) / wall_time / 1.0E6;
|
|---|
| 437 | else
|
|---|
| 438 | wall_mflops = 0.0;
|
|---|
| 439 | /*printf(" wall MFLOPS = %f\n", wall_mflops);*/
|
|---|
| 440 |
|
|---|
| 441 | /* print CPU clock info */
|
|---|
| 442 | printf("%s cpu clock time = %f seconds\n", hypre_TimingName(i), cpu_time);
|
|---|
| 443 | if (cpu_time)
|
|---|
| 444 | cpu_mflops = hypre_TimingFLOPS(i) / cpu_time / 1.0E6;
|
|---|
| 445 | else
|
|---|
| 446 | cpu_mflops = 0.0;
|
|---|
| 447 | /*printf(" cpu MFLOPS = %f\n\n", cpu_mflops);*/
|
|---|
| 448 | }
|
|---|
| 449 | }
|
|---|
| 450 | }
|
|---|
| 451 |
|
|---|
| 452 | return ierr;
|
|---|
| 453 | }
|
|---|
| 454 |
|
|---|
| 455 | #else /* threaded version of hypre_PrintTiming */
|
|---|
| 456 |
|
|---|
| 457 | #ifdef MPI_Comm_rank
|
|---|
| 458 | #undef MPI_Comm_rank
|
|---|
| 459 | #endif
|
|---|
| 460 | #ifdef MPI_Allreduce
|
|---|
| 461 | #undef MPI_Allreduce
|
|---|
| 462 | #endif
|
|---|
| 463 |
|
|---|
| 464 | int
|
|---|
| 465 | hypre_PrintTiming( const char *heading,
|
|---|
| 466 | MPI_Comm comm )
|
|---|
| 467 | {
|
|---|
| 468 | int ierr = 0;
|
|---|
| 469 |
|
|---|
| 470 | double local_wall_time;
|
|---|
| 471 | double local_cpu_time;
|
|---|
| 472 | double wall_time;
|
|---|
| 473 | double cpu_time;
|
|---|
| 474 | double wall_mflops;
|
|---|
| 475 | double cpu_mflops;
|
|---|
| 476 |
|
|---|
| 477 | int i, j, index;
|
|---|
| 478 | int myrank;
|
|---|
| 479 | int my_thread = hypre_GetThreadID();
|
|---|
| 480 | int threadid;
|
|---|
| 481 | int max_size;
|
|---|
| 482 | int num_regs;
|
|---|
| 483 |
|
|---|
| 484 | char target_name[32];
|
|---|
| 485 |
|
|---|
| 486 | if (my_thread == hypre_NumThreads)
|
|---|
| 487 | {
|
|---|
| 488 | if (hypre_global_timing == NULL)
|
|---|
| 489 | return ierr;
|
|---|
| 490 |
|
|---|
| 491 | MPI_Comm_rank(comm, &myrank );
|
|---|
| 492 |
|
|---|
| 493 | /* print heading */
|
|---|
| 494 | if (myrank == 0)
|
|---|
| 495 | {
|
|---|
| 496 | printf("=============================================\n");
|
|---|
| 497 | printf("%s:\n", heading);
|
|---|
| 498 | printf("=============================================\n");
|
|---|
| 499 | }
|
|---|
| 500 |
|
|---|
| 501 | for (i = 0; i < 7; i++)
|
|---|
| 502 | {
|
|---|
| 503 | switch (i)
|
|---|
| 504 | {
|
|---|
| 505 | case 0:
|
|---|
| 506 | threadid = my_thread;
|
|---|
| 507 | strcpy(target_name, hypre_TimingName(i));
|
|---|
| 508 | break;
|
|---|
| 509 | case 1:
|
|---|
| 510 | strcpy(target_name, "SMG");
|
|---|
| 511 | break;
|
|---|
| 512 | case 2:
|
|---|
| 513 | strcpy(target_name, "SMGRelax");
|
|---|
| 514 | break;
|
|---|
| 515 | case 3:
|
|---|
| 516 | strcpy(target_name, "SMGResidual");
|
|---|
| 517 | break;
|
|---|
| 518 | case 4:
|
|---|
| 519 | strcpy(target_name, "CyclicReduction");
|
|---|
| 520 | break;
|
|---|
| 521 | case 5:
|
|---|
| 522 | strcpy(target_name, "SMGIntAdd");
|
|---|
| 523 | break;
|
|---|
| 524 | case 6:
|
|---|
| 525 | strcpy(target_name, "SMGRestrict");
|
|---|
| 526 | break;
|
|---|
| 527 | }
|
|---|
| 528 |
|
|---|
| 529 | threadid = 0;
|
|---|
| 530 | for (j = 0; j < hypre_global_timing[threadid].size; j++)
|
|---|
| 531 | {
|
|---|
| 532 | if (strcmp(target_name, hypre_TimingName(j)) == 0)
|
|---|
| 533 | {
|
|---|
| 534 | index = j;
|
|---|
| 535 | break;
|
|---|
| 536 | }
|
|---|
| 537 | else
|
|---|
| 538 | index = -1;
|
|---|
| 539 | }
|
|---|
| 540 |
|
|---|
| 541 | if (i < hypre_global_timing[my_thread].size)
|
|---|
| 542 | {
|
|---|
| 543 | threadid = my_thread;
|
|---|
| 544 | num_regs = hypre_TimingNumRegs(i);
|
|---|
| 545 | }
|
|---|
| 546 | else
|
|---|
| 547 | num_regs = hypre_TimingNumRegs(index);
|
|---|
| 548 |
|
|---|
| 549 | if (num_regs > 0)
|
|---|
| 550 | {
|
|---|
| 551 | local_wall_time = 0.0;
|
|---|
| 552 | local_cpu_time = 0.0;
|
|---|
| 553 | if (index >= 0)
|
|---|
| 554 | {
|
|---|
| 555 | for (threadid = 0; threadid < hypre_NumThreads; threadid++)
|
|---|
| 556 | {
|
|---|
| 557 | local_wall_time =
|
|---|
| 558 | hypre_max(local_wall_time, hypre_TimingWallTime(index));
|
|---|
| 559 | local_cpu_time =
|
|---|
| 560 | hypre_max(local_cpu_time, hypre_TimingCPUTime(index));
|
|---|
| 561 | }
|
|---|
| 562 | }
|
|---|
| 563 |
|
|---|
| 564 | if (i < hypre_global_timing[my_thread].size)
|
|---|
| 565 | {
|
|---|
| 566 | threadid = my_thread;
|
|---|
| 567 | local_wall_time += hypre_TimingWallTime(i);
|
|---|
| 568 | local_cpu_time += hypre_TimingCPUTime(i);
|
|---|
| 569 | }
|
|---|
| 570 |
|
|---|
| 571 | MPI_Allreduce(&local_wall_time, &wall_time, 1,
|
|---|
| 572 | MPI_DOUBLE, MPI_MAX, comm);
|
|---|
| 573 | MPI_Allreduce(&local_cpu_time, &cpu_time, 1,
|
|---|
| 574 | MPI_DOUBLE, MPI_MAX, comm);
|
|---|
| 575 |
|
|---|
| 576 | if (myrank == 0)
|
|---|
| 577 | {
|
|---|
| 578 | printf("%s:\n", target_name);
|
|---|
| 579 |
|
|---|
| 580 | /* print wall clock info */
|
|---|
| 581 | printf(" wall clock time = %f seconds\n", wall_time);
|
|---|
| 582 | wall_mflops = 0.0;
|
|---|
| 583 | if (wall_time)
|
|---|
| 584 | {
|
|---|
| 585 | if (index >= 0)
|
|---|
| 586 | {
|
|---|
| 587 | for (threadid = 0; threadid < hypre_NumThreads; threadid++)
|
|---|
| 588 | {
|
|---|
| 589 | wall_mflops +=
|
|---|
| 590 | hypre_TimingFLOPS(index) / wall_time / 1.0E6;
|
|---|
| 591 | }
|
|---|
| 592 | }
|
|---|
| 593 | if (i < hypre_global_timing[my_thread].size)
|
|---|
| 594 | {
|
|---|
| 595 | threadid = my_thread;
|
|---|
| 596 | wall_mflops += hypre_TimingFLOPS(i) / wall_time / 1.0E6;
|
|---|
| 597 | }
|
|---|
| 598 | }
|
|---|
| 599 |
|
|---|
| 600 | /*printf(" wall MFLOPS = %f\n", wall_mflops);*/
|
|---|
| 601 |
|
|---|
| 602 | /* print CPU clock info */
|
|---|
| 603 | printf(" cpu clock time = %f seconds\n", cpu_time);
|
|---|
| 604 | cpu_mflops = 0.0;
|
|---|
| 605 | if (cpu_time)
|
|---|
| 606 | {
|
|---|
| 607 | if (index >= 0)
|
|---|
| 608 | {
|
|---|
| 609 | for (threadid = 0; threadid < hypre_NumThreads; threadid++)
|
|---|
| 610 | {
|
|---|
| 611 | cpu_mflops +=
|
|---|
| 612 | hypre_TimingFLOPS(index) / cpu_time / 1.0E6;
|
|---|
| 613 | }
|
|---|
| 614 | }
|
|---|
| 615 | if (i < hypre_global_timing[my_thread].size)
|
|---|
| 616 | {
|
|---|
| 617 | threadid = my_thread;
|
|---|
| 618 | cpu_mflops += hypre_TimingFLOPS(i) / cpu_time / 1.0E6;
|
|---|
| 619 | }
|
|---|
| 620 | }
|
|---|
| 621 |
|
|---|
| 622 | /*printf(" cpu MFLOPS = %f\n\n", cpu_mflops);*/
|
|---|
| 623 | }
|
|---|
| 624 | }
|
|---|
| 625 | }
|
|---|
| 626 | }
|
|---|
| 627 |
|
|---|
| 628 | return ierr;
|
|---|
| 629 | }
|
|---|
| 630 |
|
|---|
| 631 | #endif
|
|---|