Optimiser une boucle à l'aide d'instructions simd

Bonjour,

Je cherche à optimiser une boucle classique, à savoir la résolution 2D de l’équation de la chaleur. Le calcul est très simple : on calcul U1 à partir de U0, puis U0 à partir de U1, et on recommence niter fois pour que le calcul converge. Il suffit de respecter le critère de stabilité pour que le calcul fonctionne.

Il s'agit d'une boucle utilisant des tableaux à deux dimensions et faisant appel à des instructions simples.

Je cherche à améliorer les performances en utilisant les instructions SIMD des processeurs, notamment les AVX des nouveaux Intels qui permettent de gérer 4 double à la fois. Il y a aussi l'instruction d'AMD permettant de faire une multiplication et une addition en 1 cycle.

Voici mon programme de test actuel, je cherche à pousser le compilateur à faire le travail lui même, mais ca en me dérange pas de mettre les mains dans le cambouis. Le raisonnement est très simple : je décompose de plus en plus les boucles pour étudier le comportement :

Code:

---

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362

#include <stdio.h> #include <time.h> /*#include<emmintrin.h>*/   #define n1 102 #define n2 102 #define niter 20000   double U0[n1][n2]; double U1[n1][n2]; double U2[n1][n2]; /* tampon */   /* simd */   /*__m128d vU0[n1][n2]; __m128d vU1[n1][n2]; __m128d vU2[n1][n2]; *//* tampon */     int i,j,t; double Dx,Dy,Lx,Ly,InvDxDx,InvDyDy,Dt,alpha,totaltime,Stab,DtAlpha,DxDx,DyDy; clock_t time0,time1; FILE *f1;   int main() {   /* ---- GENERAL ---- */   alpha = 0.4; totaltime = 1.0; Dt = totaltime/((niter-1)*1.0); Lx = 1.0; Ly = 1.0; Dx = Lx/((n1-1)*1.0); Dy = Ly/((n2-1)*1.0); InvDxDx = 1.0/(Dx*Dx); InvDyDy = 1.0/(Dy*Dy); DxDx = Dx*Dx; DyDy = Dy*Dy; Stab = alpha*Dt*(InvDxDx+InvDyDy); DtAlpha = Dt*alpha;   /* Stability if result <= 0.5 */ printf("Stability factor : %f \n",Stab);     /* +----------------+ */ /* | LOOP 0 | */ /* +----------------+ */   /* Init */ for( i = 0; i < n1; i++) { for( j = 0; j < n2; j++) { U0[i][j] = 0.0; U1[i][j] = 0.0; } }   for( i = 0; i < n1; i++) { U0[i][0] = 1.0; U1[i][0] = 1.0; }   printf("Init OK \n");   /* Core */   time0=clock();   for( t = 0; t < niter; t++) {   /* even */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { U1[i][j] = U0[i][j] + DtAlpha*( (U0[i+1][j]-2.0*U0[i][j]+U0[i-1][j])*InvDxDx + (U0[i][j+1]-2.0*U0[i][j]+U0[i][j-1])*InvDyDy); } }   /* odd */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { U0[i][j] = U1[i][j] + DtAlpha*( (U1[i+1][j]-2.0*U1[i][j]+U1[i-1][j])*InvDxDx + (U1[i][j+1]-2.0*U1[i][j]+U1[i][j-1])*InvDyDy); } }   }   time1=clock();   printf("Loop 0, total time : %f \n", (double) time1-time0);     f1 = fopen ("out0.dat", "wt"); for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { fprintf (f1, "%d\t%d\t%f\n", i, j, U0[i][j]); } }         /* +----------------+ */ /* | LOOP 1 | */ /* +----------------+ */   /* Init */ for( i = 0; i < n1; i++) { for( j = 0; j < n2; j++) { U0[i][j] = 0.0; U1[i][j] = 0.0; } }   for( i = 0; i < n1; i++) { U0[i][0] = 1.0; U1[i][0] = 1.0; }   printf("Init OK \n");   /* Core */   time0=clock();   for( t = 0; t < niter; t++) {   /* even */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { /* U1[i][j] = U0[i][j] + Dt*alpha*( (U0[i+1][j]-2.0*U0[i][j]+U0[i-1][j])*InvDxDx + (U0[i][j+1]-2.0*U0[i][j]+U0[i][j-1])*InvDyDy); */   U1[i][j] = -2.0*U0[i][j];   U1[i][j] = U1[i][j] + U0[i+1][j];   U1[i][j] = U1[i][j] + U0[i-1][j];   U1[i][j] = U1[i][j] * InvDxDx;   U2[i][j] = -2.0*U0[i][j];   U2[i][j] = U2[i][j] + U0[i][j+1];   U2[i][j] = U2[i][j] + U0[i][j-1];   U2[i][j] = U2[i][j] * InvDyDy;   U1[i][j] = U1[i][j] + U2[i][j];   U1[i][j] = U1[i][j] * DtAlpha;   U1[i][j] = U1[i][j] + U0[i][j];   } }   /* odd */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) {   U0[i][j] = -2.0*U1[i][j];   U0[i][j] = U0[i][j] + U1[i+1][j];   U0[i][j] = U0[i][j] + U1[i-1][j];   U0[i][j] = U0[i][j] * InvDxDx;   U2[i][j] = -2.0*U1[i][j];   U2[i][j] = U2[i][j] + U1[i][j+1];   U2[i][j] = U2[i][j] + U1[i][j-1];   U2[i][j] = U2[i][j] * InvDyDy;   U0[i][j] = U0[i][j] + U2[i][j];   U0[i][j] = U0[i][j] * DtAlpha;   U0[i][j] = U0[i][j] + U1[i][j];   } }   }   time1=clock();   printf("Loop 1, total time : %f \n", (double) time1-time0);       /* End */   f1 = fopen ("out1.dat", "wt"); for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { fprintf (f1, "%d\t%d\t%f\n", i, j, U0[i][j]); } }         /* +----------------+ */ /* | LOOP 2 | */ /* +----------------+ */   /* Init */ for( i = 0; i < n1; i++) { for( j = 0; j < n2; j++) { U0[i][j] = 0.0; U1[i][j] = 0.0; } }   for( i = 0; i < n1; i++) { U0[i][0] = 1.0; U1[i][0] = 1.0; }   printf("Init OK \n");   /* Core */   time0=clock();   for( t = 0; t < niter; t++) {   /* even */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j=j+2) { /* U1[i][j] = U0[i][j] + Dt*alpha*( (U0[i+1][j]-2.0*U0[i][j]+U0[i-1][j])*InvDxDx + (U0[i][j+1]-2.0*U0[i][j]+U0[i][j-1])*InvDyDy); */   U1[i][j] = -2.0*U0[i][j]; U1[i][j+1] = -2.0*U0[i][j+1];   U1[i][j] = U1[i][j] + U0[i+1][j]; U1[i][j+1] = U1[i][j+1] + U0[i+1][j+1];   U1[i][j] = U1[i][j] + U0[i-1][j]; U1[i][j+1] = U1[i][j+1] + U0[i-1][j+1];   U1[i][j] = U1[i][j] * InvDxDx; U1[i][j+1] = U1[i][j+1] * InvDxDx;   U2[i][j] = -2.0*U0[i][j]; U2[i][j+1] = -2.0*U0[i][j+1];   U2[i][j] = U2[i][j] + U0[i][j+1]; U2[i][j+1] = U2[i][j+1] + U0[i][j+2];   U2[i][j] = U2[i][j] + U0[i][j-1]; U2[i][j+1] = U2[i][j+1] + U0[i][j];   U2[i][j] = U2[i][j] * InvDyDy; U2[i][j+1] = U2[i][j+1] * InvDyDy;   U1[i][j] = U1[i][j] + U2[i][j]; U1[i][j+1] = U1[i][j+1] + U2[i][j+1];   U1[i][j] = U1[i][j] * DtAlpha; U1[i][j+1] = U1[i][j+1] * DtAlpha;   U1[i][j] = U1[i][j] + U0[i][j]; U1[i][j+1] = U1[i][j+1] + U0[i][j+1];   } }   /* odd */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j=j+2) {   U0[i][j] = -2.0*U1[i][j]; U0[i][j+1] = -2.0*U1[i][j+1];   U0[i][j] = U0[i][j] + U1[i+1][j]; U0[i][j+1] = U0[i][j+1] + U1[i+1][j+1];   U0[i][j] = U0[i][j] + U1[i-1][j]; U0[i][j+1] = U0[i][j+1] + U1[i-1][j+1];   U0[i][j] = U0[i][j] * InvDxDx; U0[i][j+1] = U0[i][j+1] * InvDxDx;   U2[i][j] = -2.0*U1[i][j]; U2[i][j+1] = -2.0*U1[i][j+1];   U2[i][j] = U2[i][j] + U1[i][j+1]; U2[i][j+1] = U2[i][j+1] + U1[i][j+2];   U2[i][j] = U2[i][j] + U1[i][j-1]; U2[i][j+1] = U2[i][j+1] + U1[i][j];   U2[i][j] = U2[i][j] * InvDyDy; U2[i][j+1] = U2[i][j+1] * InvDyDy;   U0[i][j] = U0[i][j] + U2[i][j]; U0[i][j+1] = U0[i][j+1] + U2[i][j+1];   U0[i][j] = U0[i][j] * DtAlpha; U0[i][j+1] = U0[i][j+1] * DtAlpha;   U0[i][j] = U0[i][j] + U1[i][j]; U0[i][j+1] = U0[i][j+1] + U1[i][j+1];   } }   }   time1=clock();   printf("Loop 2, total time : %f \n", (double) time1-time0);       /* End */   f1 = fopen ("out2.dat", "wt"); for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { fprintf (f1, "%d\t%d\t%f\n", i, j, U0[i][j]); } }     }

---

Les résultats sont pour le moment les mêmes avec gcc et icc : la première boucle est sensiblement moins rapide que la seconde, et la dernière est à la ramasse (presque deux fois plus lent). Je cherchai à le pousser à utiliser les SIMD par paquets de deux. (4 sur les Snady/Ivy Bridge)

Avez vous des idées sur la façon d'optimiser tout ça, et de pousser le compilateur à utiliser les SIMD pour accélérer le tout ?

Moomba

Code mis à jours. Les instructions sont codés en dur, mais c'est encore plus lent...

Code:

---

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622

#include <stdio.h> #include <time.h> #include<emmintrin.h>   #define n1 102 #define n2 102 #define niter 20000   double U0[n1][n2]; double U1[n1][n2]; double U2[n1][n2]; /* tampon */   double *cU0[n1][n2];   /* simd */   __m128d vU0[n1][n2]; __m128d vU1[n1][n2]; __m128d vU2[n1][n2]; /* tampon */   int i,j,t; double Dx,Dy,Lx,Ly,InvDxDx,InvDyDy,Dt,alpha,totaltime,Stab,DtAlpha,DxDx,DyDy; clock_t time0,time1; FILE *f1;   int main() {   /* ---- GENERAL ---- */   alpha = 0.4; totaltime = 1.0; Dt = totaltime/((niter-1)*1.0); Lx = 1.0; Ly = 1.0; Dx = Lx/((n1-1)*1.0); Dy = Ly/((n2-1)*1.0); InvDxDx = 1.0/(Dx*Dx); InvDyDy = 1.0/(Dy*Dy); DxDx = Dx*Dx; DyDy = Dy*Dy; Stab = alpha*Dt*(InvDxDx+InvDyDy); DtAlpha = Dt*alpha;   /* Stability if result <= 0.5 */ printf("Stability factor : %f \n",Stab);     /* +----------------+ */ /* | LOOP 0 | */ /* +----------------+ */   /* Init */ for( i = 0; i < n1; i++) { for( j = 0; j < n2; j++) { U0[i][j] = 0.0; U1[i][j] = 0.0; } }   for( i = 0; i < n1; i++) { U0[i][0] = 1.0; U1[i][0] = 1.0; }   printf("Init OK \n");   /* Core */   time0=clock();   for( t = 0; t < niter; t++) {   /* even */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { U1[i][j] = U0[i][j] + DtAlpha*( (U0[i+1][j]-2.0*U0[i][j]+U0[i-1][j])*InvDxDx + (U0[i][j+1]-2.0*U0[i][j]+U0[i][j-1])*InvDyDy); } }   /* odd */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { U0[i][j] = U1[i][j] + DtAlpha*( (U1[i+1][j]-2.0*U1[i][j]+U1[i-1][j])*InvDxDx + (U1[i][j+1]-2.0*U1[i][j]+U1[i][j-1])*InvDyDy); } }   }   time1=clock();   printf("Loop 0, total time : %f \n", (double) time1-time0);     f1 = fopen ("out0.dat", "wt"); for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { fprintf (f1, "%d\t%d\t%f\n", i, j, U0[i][j]); } }         /* +----------------+ */ /* | LOOP 1 | */ /* +----------------+ */   /* Init */ for( i = 0; i < n1; i++) { for( j = 0; j < n2; j++) { U0[i][j] = 0.0; U1[i][j] = 0.0; } }   for( i = 0; i < n1; i++) { U0[i][0] = 1.0; U1[i][0] = 1.0; }   printf("Init OK \n");   /* Core */   time0=clock();   for( t = 0; t < niter; t++) {   /* even */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { /* U1[i][j] = U0[i][j] + Dt*alpha*( (U0[i+1][j]-2.0*U0[i][j]+U0[i-1][j])*InvDxDx + (U0[i][j+1]-2.0*U0[i][j]+U0[i][j-1])*InvDyDy); */   U1[i][j] = -2.0*U0[i][j];   U1[i][j] = U1[i][j] + U0[i+1][j];   U1[i][j] = U1[i][j] + U0[i-1][j];   U1[i][j] = U1[i][j] * InvDxDx;   U2[i][j] = -2.0*U0[i][j];   U2[i][j] = U2[i][j] + U0[i][j+1];   U2[i][j] = U2[i][j] + U0[i][j-1];   U2[i][j] = U2[i][j] * InvDyDy;   U1[i][j] = U1[i][j] + U2[i][j];   U1[i][j] = U1[i][j] * DtAlpha;   U1[i][j] = U1[i][j] + U0[i][j];   } }   /* odd */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) {   U0[i][j] = -2.0*U1[i][j];   U0[i][j] = U0[i][j] + U1[i+1][j];   U0[i][j] = U0[i][j] + U1[i-1][j];   U0[i][j] = U0[i][j] * InvDxDx;   U2[i][j] = -2.0*U1[i][j];   U2[i][j] = U2[i][j] + U1[i][j+1];   U2[i][j] = U2[i][j] + U1[i][j-1];   U2[i][j] = U2[i][j] * InvDyDy;   U0[i][j] = U0[i][j] + U2[i][j];   U0[i][j] = U0[i][j] * DtAlpha;   U0[i][j] = U0[i][j] + U1[i][j];   } }   }   time1=clock();   printf("Loop 1, total time : %f \n", (double) time1-time0);       /* End */   f1 = fopen ("out1.dat", "wt"); for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { fprintf (f1, "%d\t%d\t%f\n", i, j, U0[i][j]); } }         /* +----------------+ */ /* | LOOP 2 | */ /* +----------------+ */   /* Init */ for( i = 0; i < n1; i++) { for( j = 0; j < n2; j++) { U0[i][j] = 0.0; U1[i][j] = 0.0; } }   for( i = 0; i < n1; i++) { U0[i][0] = 1.0; U1[i][0] = 1.0; }   printf("Init OK \n");   /* Core */   time0=clock();   for( t = 0; t < niter; t++) {   /* even */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j=j+2) { /* U1[i][j] = U0[i][j] + Dt*alpha*( (U0[i+1][j]-2.0*U0[i][j]+U0[i-1][j])*InvDxDx + (U0[i][j+1]-2.0*U0[i][j]+U0[i][j-1])*InvDyDy); */   U1[i][j] = -2.0*U0[i][j]; U1[i][j+1] = -2.0*U0[i][j+1];   U1[i][j] = U1[i][j] + U0[i+1][j]; U1[i][j+1] = U1[i][j+1] + U0[i+1][j+1];   U1[i][j] = U1[i][j] + U0[i-1][j]; U1[i][j+1] = U1[i][j+1] + U0[i-1][j+1];   U1[i][j] = U1[i][j] * InvDxDx; U1[i][j+1] = U1[i][j+1] * InvDxDx;   U2[i][j] = -2.0*U0[i][j]; U2[i][j+1] = -2.0*U0[i][j+1];   U2[i][j] = U2[i][j] + U0[i][j+1]; U2[i][j+1] = U2[i][j+1] + U0[i][j+2];   U2[i][j] = U2[i][j] + U0[i][j-1]; U2[i][j+1] = U2[i][j+1] + U0[i][j];   U2[i][j] = U2[i][j] * InvDyDy; U2[i][j+1] = U2[i][j+1] * InvDyDy;   U1[i][j] = U1[i][j] + U2[i][j]; U1[i][j+1] = U1[i][j+1] + U2[i][j+1];   U1[i][j] = U1[i][j] * DtAlpha; U1[i][j+1] = U1[i][j+1] * DtAlpha;   U1[i][j] = U1[i][j] + U0[i][j]; U1[i][j+1] = U1[i][j+1] + U0[i][j+1];   } }   /* odd */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j=j+2) {   U0[i][j] = -2.0*U1[i][j]; U0[i][j+1] = -2.0*U1[i][j+1];   U0[i][j] = U0[i][j] + U1[i+1][j]; U0[i][j+1] = U0[i][j+1] + U1[i+1][j+1];   U0[i][j] = U0[i][j] + U1[i-1][j]; U0[i][j+1] = U0[i][j+1] + U1[i-1][j+1];   U0[i][j] = U0[i][j] * InvDxDx; U0[i][j+1] = U0[i][j+1] * InvDxDx;   U2[i][j] = -2.0*U1[i][j]; U2[i][j+1] = -2.0*U1[i][j+1];   U2[i][j] = U2[i][j] + U1[i][j+1]; U2[i][j+1] = U2[i][j+1] + U1[i][j+2];   U2[i][j] = U2[i][j] + U1[i][j-1]; U2[i][j+1] = U2[i][j+1] + U1[i][j];   U2[i][j] = U2[i][j] * InvDyDy; U2[i][j+1] = U2[i][j+1] * InvDyDy;   U0[i][j] = U0[i][j] + U2[i][j]; U0[i][j+1] = U0[i][j+1] + U2[i][j+1];   U0[i][j] = U0[i][j] * DtAlpha; U0[i][j+1] = U0[i][j+1] * DtAlpha;   U0[i][j] = U0[i][j] + U1[i][j]; U0[i][j+1] = U0[i][j+1] + U1[i][j+1];   } }   }   time1=clock();   printf("Loop 2, total time : %f \n", (double) time1-time0);       /* End */   f1 = fopen ("out2.dat", "wt"); for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { fprintf (f1, "%d\t%d\t%f\n", i, j, U0[i][j]); } }               /* +----------------+ */ /* | LOOP 3 | */ /* +----------------+ */   /* Init */ for( i = 0; i < n1; i++) { for( j = 0; j < n2; j++) { vU0[i][j] = _mm_set1_pd(0.0); vU1[i][j] = _mm_set1_pd(0.0); } }   for( i = 0; i < n1; i++) { vU0[i][0] = _mm_set1_pd(1.0); vU1[i][0] = _mm_set1_pd(1.0); }   printf("Init OK \n");   /* Core */   time0=clock();   for( t = 0; t < niter; t++) {   /* even */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j+=2) { /* U1[i][j] = U0[i][j] + Dt*alpha*( (U0[i+1][j]-2.0*U0[i][j]+U0[i-1][j])*InvDxDx + (U0[i][j+1]-2.0*U0[i][j]+U0[i][j-1])*InvDyDy); */   /*va = _mm_load_pd(cU0[i][j]);   vb = __m128d -2.0;   vU1[i][j] = _mm_mul_pd (vb, va);*/     __m128d va = _mm_set1_pd(-2.0); /* U1[i][j] = -2.0*U0[i][j]; */ vU1[i][j] = _mm_mul_pd(va,vU0[i][j]);   vU1[i][j] = _mm_add_pd(vU1[i][j],vU0[i+1][j]); /* U1[i][j] = U1[i][j] + U0[i+1][j]; */   vU1[i][j] = _mm_add_pd(vU1[i][j],vU0[i-1][j]); /* U1[i][j] = U1[i][j] + U0[i-1][j]; */   __m128d vb = _mm_set1_pd(InvDxDx); /* U1[i][j] = U1[i][j] * InvDxDx; */ vU1[i][j] = _mm_mul_pd(vb,vU1[i][j]);   vU2[i][j] = _mm_mul_pd(va,vU0[i][j]); /* U2[i][j] = -2.0*U0[i][j]; */   vU2[i][j] = _mm_add_pd(vU2[i][j],vU0[i+1][j]); /* U2[i][j] = U2[i][j] + U0[i+1][j]; */   vU2[i][j] = _mm_add_pd(vU2[i][j],vU0[i-1][j]); /* U2[i][j] = U2[i][j] + U0[i-1][j]; */   __m128d vc = _mm_set1_pd(InvDyDy); /* U2[i][j] = U2[i][j] * InvDyDy; */ vU2[i][j] = _mm_mul_pd(vc,vU2[i][j]);   vU1[i][j] = _mm_add_pd(vU1[i][j],vU2[i][j]); /* U1[i][j] = U1[i][j] + U2[i][j]; */   __m128d vd = _mm_set1_pd(DtAlpha); /* U1[i][j] = U1[i][j] * DtAlpha; */ vU1[i][j] = _mm_mul_pd(vd,vU1[i][j]);   vU1[i][j] = _mm_add_pd(vU1[i][j],vU0[i][j]); /* U1[i][j] = U1[i][j] + U0[i][j]; */   } }   /* odd */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) { __m128d va = _mm_set1_pd(-2.0); /* U0[i][j] = -2.0*U1[i][j]; */ vU0[i][j] = _mm_mul_pd(va,vU1[i][j]);   vU0[i][j] = _mm_add_pd(vU0[i][j],vU1[i+1][j]); /* U0[i][j] = U0[i][j] + U1[i+1][j]; */   vU0[i][j] = _mm_add_pd(vU0[i][j],vU1[i-1][j]); /* U0[i][j] = U0[i][j] + U1[i-1][j]; */   __m128d vb = _mm_set1_pd(InvDxDx); /* U0[i][j] = U0[i][j] * InvDxDx; */ vU0[i][j] = _mm_mul_pd(vb,vU0[i][j]);   vU2[i][j] = _mm_mul_pd(va,vU1[i][j]); /* U2[i][j] = -2.0*U1[i][j]; */   vU2[i][j] = _mm_add_pd(vU2[i][j],vU1[i+1][j]); /* U2[i][j] = U2[i][j] + U1[i+1][j]; */   vU2[i][j] = _mm_add_pd(vU2[i][j],vU1[i-1][j]); /* U2[i][j] = U2[i][j] + U1[i-1][j]; */   __m128d vc = _mm_set1_pd(InvDyDy); /* U2[i][j] = U2[i][j] * InvDyDy; */ vU2[i][j] = _mm_mul_pd(vc,vU2[i][j]);   vU0[i][j] = _mm_add_pd(vU0[i][j],vU2[i][j]); /* U0[i][j] = U0[i][j] + U2[i][j]; */   __m128d vd = _mm_set1_pd(DtAlpha); /* U0[i][j] = U0[i][j] * DtAlpha; */ vU0[i][j] = _mm_mul_pd(vd,vU0[i][j]);   vU0[i][j] = _mm_add_pd(vU0[i][j],vU1[i][j]); /* U0[i][j] = U0[i][j] + U1[i][j]; */   } }   }   time1=clock();   printf("Loop 3, total time : %f \n", (double) time1-time0);       /* End */   f1 = fopen ("out3.dat", "wt"); for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j+=2) { fprintf (f1, "%d\t%d\t%f\n", i, j, vU0[i][j]); } }         /* +----------------+ */ /* | LOOP 4 | */ /* +----------------+ */   /* Init */ for( i = 0; i < n1; i++) { for( j = 0; j < n2; j++) { vU0[i][j] = _mm_set1_pd(0.0); vU1[i][j] = _mm_set1_pd(0.0); } }   for( i = 0; i < n1; i++) { vU0[i][0] = _mm_set1_pd(1.0); vU1[i][0] = _mm_set1_pd(1.0); }   printf("Init OK \n");   /* Core */   time0=clock();   __m128d va = _mm_set1_pd(-2.0); /* U1[i][j] = -2.0*U0[i][j]; */ __m128d vb = _mm_set1_pd(InvDxDx); /* U1[i][j] = U1[i][j] * InvDxDx; */ __m128d vc = _mm_set1_pd(InvDyDy); /* U2[i][j] = U2[i][j] * InvDyDy; */ __m128d vd = _mm_set1_pd(DtAlpha); /* U1[i][j] = U1[i][j] * DtAlpha; */   for( t = 0; t < niter; t++) {   /* even */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j+=2) { /* U1[i][j] = U0[i][j] + Dt*alpha*( (U0[i+1][j]-2.0*U0[i][j]+U0[i-1][j])*InvDxDx + (U0[i][j+1]-2.0*U0[i][j]+U0[i][j-1])*InvDyDy); */   /*va = _mm_load_pd(cU0[i][j]);   vb = __m128d -2.0;   vU1[i][j] = _mm_mul_pd (vb, va);*/       vU1[i][j] = _mm_mul_pd(va,vU0[i][j]);   vU1[i][j] = _mm_add_pd(vU1[i][j],vU0[i+1][j]); /* U1[i][j] = U1[i][j] + U0[i+1][j]; */   vU1[i][j] = _mm_add_pd(vU1[i][j],vU0[i-1][j]); /* U1[i][j] = U1[i][j] + U0[i-1][j]; */     vU1[i][j] = _mm_mul_pd(vb,vU1[i][j]);   vU2[i][j] = _mm_mul_pd(va,vU0[i][j]); /* U2[i][j] = -2.0*U0[i][j]; */   vU2[i][j] = _mm_add_pd(vU2[i][j],vU0[i+1][j]); /* U2[i][j] = U2[i][j] + U0[i+1][j]; */   vU2[i][j] = _mm_add_pd(vU2[i][j],vU0[i-1][j]); /* U2[i][j] = U2[i][j] + U0[i-1][j]; */     vU2[i][j] = _mm_mul_pd(vc,vU2[i][j]);   vU1[i][j] = _mm_add_pd(vU1[i][j],vU2[i][j]); /* U1[i][j] = U1[i][j] + U2[i][j]; */     vU1[i][j] = _mm_mul_pd(vd,vU1[i][j]);   vU1[i][j] = _mm_add_pd(vU1[i][j],vU0[i][j]); /* U1[i][j] = U1[i][j] + U0[i][j]; */   } }   /* odd */ for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j++) {   vU0[i][j] = _mm_mul_pd(va,vU1[i][j]);   vU0[i][j] = _mm_add_pd(vU0[i][j],vU1[i+1][j]); /* U0[i][j] = U0[i][j] + U1[i+1][j]; */   vU0[i][j] = _mm_add_pd(vU0[i][j],vU1[i-1][j]); /* U0[i][j] = U0[i][j] + U1[i-1][j]; */     vU0[i][j] = _mm_mul_pd(vb,vU0[i][j]);   vU2[i][j] = _mm_mul_pd(va,vU1[i][j]); /* U2[i][j] = -2.0*U1[i][j]; */   vU2[i][j] = _mm_add_pd(vU2[i][j],vU1[i+1][j]); /* U2[i][j] = U2[i][j] + U1[i+1][j]; */   vU2[i][j] = _mm_add_pd(vU2[i][j],vU1[i-1][j]); /* U2[i][j] = U2[i][j] + U1[i-1][j]; */     vU2[i][j] = _mm_mul_pd(vc,vU2[i][j]);   vU0[i][j] = _mm_add_pd(vU0[i][j],vU2[i][j]); /* U0[i][j] = U0[i][j] + U2[i][j]; */     vU0[i][j] = _mm_mul_pd(vd,vU0[i][j]);   vU0[i][j] = _mm_add_pd(vU0[i][j],vU1[i][j]); /* U0[i][j] = U0[i][j] + U1[i][j]; */   } }   }   time1=clock();   printf("Loop 4, total time : %f \n", (double) time1-time0);       /* End */   f1 = fopen ("out4.dat", "wt"); for( i = 1; i < n1-1; i++) { for( j = 1; j < n2-1; j+=2) { fprintf (f1, "%d\t%d\t%f\n", i, j, vU0[i][j]); } }   }

---

Salut,

Je doute qui tu puisse faire comprendre au compilo qu'il faut utiliser une instruction SIMD. je pense qu'il faut obligatoirement passer par de l'assembleur, mais je n'en mettrais pas ma main a couper...

Tu peux essayer de jouer avec les option de compilation
Avec gcc de tenterais le coup avec -O3 -march=corei7

SI ca ne marche pas , cela me confortera dans l'idée que les architectures CISC ca sert à rien, sinon à consommer + => polluer +.

Bon courage.

Les deux dernières boucles utilisent des instructions spécifiques aux SIMD, mais effectivement gcc ne semble pas en tirer partie.

Cependant, concernant les CISC je peut t'assurer qu'elles fonctionnent : j'ai réussi en Fortran 95 à faire comprendre au compilateur Ifort 12.x d'utiliser les SSE dans une boucle particulière (qui consomme presque 95% du temps de calcul du code qui est de plusieurs semaines tout de même sur 1024 processeurs). Le gain est de 1.73x environ pour du double précision, donc quand on arrive à les lancer ça sert, et pas qu'un peu. J'attends des résultats encore meilleurs avec l'AVX.

Mais il faut les activer, tout le problème est là :mrgreen:

As tu essayer avec l'option de compilation -mavx?

Bonne idée. Mais mon processeur ne supporte pas l'avx, c'est un xeon équivalent au i7 920, donc sur le socket 1366. et le super calculateur est sur westmer, donc pas d'avx non plus.

Je me demande si ce n'est pas due au fait que nos processeurs ont deux fpu, et que donc le sse en double précision ne sert à rien...