| 1 | [ |
| 2 | { |
| 3 | "key": { |
| 4 | "intermediate": 2048, |
| 5 | "numtokens": 256 |
| 6 | }, |
| 7 | "config": { |
| 8 | "block_sizes": [ |
| 9 | 64, |
| 10 | 32 |
| 11 | ], |
| 12 | "loop_orders": [ |
| 13 | [ |
| 14 | 0, |
| 15 | 1 |
| 16 | ] |
| 17 | ], |
| 18 | "flatten_loops": [ |
| 19 | true |
| 20 | ], |
| 21 | "l2_groupings": [ |
| 22 | 1 |
| 23 | ], |
| 24 | "range_unroll_factors": [ |
| 25 | 0 |
| 26 | ], |
| 27 | "range_warp_specializes": [], |
| 28 | "range_num_stages": [ |
| 29 | 0 |
| 30 | ], |
| 31 | "range_multi_buffers": [ |
| 32 | null |
| 33 | ], |
| 34 | "range_flattens": [ |
| 35 | null |
| 36 | ], |
| 37 | "load_eviction_policies": [ |
| 38 | "", |
| 39 | "", |
| 40 | "" |
| 41 | ], |
| 42 | "num_warps": 8, |
| 43 | "num_stages": 1, |
| 44 | "indexing": [ |
| 45 | "pointer", |
| 46 | "pointer", |
| 47 | "pointer", |
| 48 | "pointer" |
| 49 | ], |
| 50 | "pid_type": "flat" |
| 51 | } |
| 52 | }, |
| 53 | { |
| 54 | "key": { |
| 55 | "intermediate": 4096, |
| 56 | "numtokens": 256 |
| 57 | }, |
| 58 | "config": { |
| 59 | "block_sizes": [ |
| 60 | 256, |
| 61 | 32 |
| 62 | ], |
| 63 | "loop_orders": [ |
| 64 | [ |
| 65 | 0, |
| 66 | 1 |
| 67 | ] |
| 68 | ], |
| 69 | "flatten_loops": [ |
| 70 | true |
| 71 | ], |
| 72 | "l2_groupings": [ |
| 73 | 1 |
| 74 | ], |
| 75 | "range_unroll_factors": [ |
| 76 | 0 |
| 77 | ], |
| 78 | "range_warp_specializes": [], |
| 79 | "range_num_stages": [ |
| 80 | 0 |
| 81 | ], |
| 82 | "range_multi_buffers": [ |
| 83 | null |
| 84 | ], |
| 85 | "range_flattens": [ |
| 86 | null |
| 87 | ], |
| 88 | "load_eviction_policies": [ |
| 89 | "", |
| 90 | "", |
| 91 | "" |
| 92 | ], |
| 93 | "num_warps": 8, |
| 94 | "num_stages": 1, |
| 95 | "indexing": [ |
| 96 | "pointer", |
| 97 | "pointer", |
| 98 | "pointer", |
| 99 | "pointer" |
| 100 | ], |
| 101 | "pid_type": "flat" |
| 102 | } |
| 103 | }, |
| 104 | { |
| 105 | "key": {}, |
| 106 | "config": { |
| 107 | "block_sizes": [ |
| 108 | 1, |
| 109 | 512 |
| 110 | ], |
| 111 | "loop_orders": [ |
| 112 | [ |
| 113 | 1, |
| 114 | 0 |
| 115 | ] |
| 116 | ], |
| 117 | "flatten_loops": [ |
| 118 | false |
| 119 | ], |
| 120 | "l2_groupings": [ |
| 121 | 4 |
| 122 | ], |
| 123 | "range_unroll_factors": [ |
| 124 | 0 |
| 125 | ], |
| 126 | "range_warp_specializes": [], |
| 127 | "range_num_stages": [ |
| 128 | 0 |
| 129 | ], |
| 130 | "range_multi_buffers": [ |
| 131 | null |
| 132 | ], |
| 133 | "range_flattens": [ |
| 134 | null |
| 135 | ], |
| 136 | "load_eviction_policies": [ |
| 137 | "", |
| 138 | "first", |
| 139 | "" |
| 140 | ], |
| 141 | "num_warps": 8, |
| 142 | "num_stages": 2, |
| 143 | "indexing": [ |
| 144 | "tensor_descriptor", |
| 145 | "tensor_descriptor", |
| 146 | "tensor_descriptor", |
| 147 | "pointer" |
| 148 | ], |
| 149 | "pid_type": "flat" |
| 150 | } |
| 151 | }, |
| 152 | { |
| 153 | "key": { |
| 154 | "intermediate": 8192, |
| 155 | "numtokens": 256 |
| 156 | }, |
| 157 | "config": { |
| 158 | "block_sizes": [ |
| 159 | 256, |
| 160 | 64 |
| 161 | ], |
| 162 | "loop_orders": [ |
| 163 | [ |
| 164 | 0, |
| 165 | 1 |
| 166 | ] |
| 167 | ], |
| 168 | "flatten_loops": [ |
| 169 | true |
| 170 | ], |
| 171 | "l2_groupings": [ |
| 172 | 1 |
| 173 | ], |
| 174 | "range_unroll_factors": [ |
| 175 | 0 |
| 176 | ], |
| 177 | "range_warp_specializes": [], |
| 178 | "range_num_stages": [ |
| 179 | 0 |
| 180 | ], |
| 181 | "range_multi_buffers": [ |
| 182 | null |
| 183 | ], |
| 184 | "range_flattens": [ |
| 185 | null |
| 186 | ], |
| 187 | "load_eviction_policies": [ |
| 188 | "", |
| 189 | "", |
| 190 | "" |
| 191 | ], |
| 192 | "num_warps": 8, |
| 193 | "num_stages": 1, |
| 194 | "indexing": [ |
| 195 | "pointer", |
| 196 | "pointer", |
| 197 | "pointer", |
| 198 | "pointer" |
| 199 | ], |
| 200 | "pid_type": "flat" |
| 201 | } |
| 202 | }, |
| 203 | { |
| 204 | "key": { |
| 205 | "intermediate": 14336, |
| 206 | "numtokens": 256 |
| 207 | }, |
| 208 | "config": { |
| 209 | "block_sizes": [ |
| 210 | 8, |
| 211 | 4096 |
| 212 | ], |
| 213 | "loop_orders": [ |
| 214 | [ |
| 215 | 0, |
| 216 | 1 |
| 217 | ] |
| 218 | ], |
| 219 | "flatten_loops": [ |
| 220 | true |
| 221 | ], |
| 222 | "l2_groupings": [ |
| 223 | 1 |
| 224 | ], |
| 225 | "range_unroll_factors": [ |
| 226 | 0 |
| 227 | ], |
| 228 | "range_warp_specializes": [], |
| 229 | "range_num_stages": [ |
| 230 | 0 |
| 231 | ], |
| 232 | "range_multi_buffers": [ |
| 233 | null |
| 234 | ], |
| 235 | "range_flattens": [ |
| 236 | null |
| 237 | ], |
| 238 | "load_eviction_policies": [ |
| 239 | "", |
| 240 | "", |
| 241 | "" |
| 242 | ], |
| 243 | "num_warps": 8, |
| 244 | "num_stages": 1, |
| 245 | "indexing": [ |
| 246 | "pointer", |
| 247 | "pointer", |
| 248 | "pointer", |
| 249 | "pointer" |
| 250 | ], |
| 251 | "pid_type": "flat" |
| 252 | } |
| 253 | }, |
| 254 | { |
| 255 | "key": { |
| 256 | "intermediate": 11008, |
| 257 | "numtokens": 256 |
| 258 | }, |
| 259 | "config": { |
| 260 | "block_sizes": [ |
| 261 | 64, |
| 262 | 32 |
| 263 | ], |
| 264 | "loop_orders": [ |
| 265 | [ |
| 266 | 0, |
| 267 | 1 |
| 268 | ] |
| 269 | ], |
| 270 | "flatten_loops": [ |
| 271 | true |
| 272 | ], |
| 273 | "l2_groupings": [ |
| 274 | 1 |
| 275 | ], |
| 276 | "range_unroll_factors": [ |
| 277 | 0 |
| 278 | ], |
| 279 | "range_warp_specializes": [], |
| 280 | "range_num_stages": [ |
| 281 | 0 |
| 282 | ], |
| 283 | "range_multi_buffers": [ |
| 284 | null |
| 285 | ], |
| 286 | "range_flattens": [ |
| 287 | null |
| 288 | ], |
| 289 | "load_eviction_policies": [ |
| 290 | "", |
| 291 | "", |
| 292 | "" |
| 293 | ], |
| 294 | "num_warps": 2, |
| 295 | "num_stages": 1, |
| 296 | "indexing": [ |
| 297 | "pointer", |
| 298 | "pointer", |
| 299 | "pointer", |
| 300 | "pointer" |
| 301 | ], |
| 302 | "pid_type": "flat" |
| 303 | } |
| 304 | }, |
| 305 | { |
| 306 | "key": { |
| 307 | "intermediate": 7688, |
| 308 | "numtokens": 256 |
| 309 | }, |
| 310 | "config": { |
| 311 | "block_sizes": [ |
| 312 | 32, |
| 313 | 512 |
| 314 | ], |
| 315 | "loop_orders": [ |
| 316 | [ |
| 317 | 0, |
| 318 | 1 |
| 319 | ] |
| 320 | ], |
| 321 | "flatten_loops": [ |
| 322 | true |
| 323 | ], |
| 324 | "l2_groupings": [ |
| 325 | 1 |
| 326 | ], |
| 327 | "range_unroll_factors": [ |
| 328 | 0 |
| 329 | ], |
| 330 | "range_warp_specializes": [], |
| 331 | "range_num_stages": [ |
| 332 | 0 |
| 333 | ], |
| 334 | "range_multi_buffers": [ |
| 335 | null |
| 336 | ], |
| 337 | "range_flattens": [ |
| 338 | null |
| 339 | ], |
| 340 | "load_eviction_policies": [ |
| 341 | "", |
| 342 | "", |
| 343 | "" |
| 344 | ], |
| 345 | "num_warps": 8, |
| 346 | "num_stages": 1, |
| 347 | "indexing": [ |
| 348 | "pointer", |
| 349 | "pointer", |
| 350 | "pointer", |
| 351 | "pointer" |
| 352 | ], |
| 353 | "pid_type": "flat" |
| 354 | } |
| 355 | }, |
| 356 | { |
| 357 | "key": { |
| 358 | "intermediate": 2880, |
| 359 | "numtokens": 256 |
| 360 | }, |
| 361 | "config": { |
| 362 | "block_sizes": [ |
| 363 | 32, |
| 364 | 64 |
| 365 | ], |
| 366 | "loop_orders": [ |
| 367 | [ |
| 368 | 0, |
| 369 | 1 |
| 370 | ] |
| 371 | ], |
| 372 | "flatten_loops": [ |
| 373 | true |
| 374 | ], |
| 375 | "l2_groupings": [ |
| 376 | 2 |
| 377 | ], |
| 378 | "range_unroll_factors": [ |
| 379 | 0 |
| 380 | ], |
| 381 | "range_warp_specializes": [], |
| 382 | "range_num_stages": [ |
| 383 | 0 |
| 384 | ], |
| 385 | "range_multi_buffers": [ |
| 386 | null |
| 387 | ], |
| 388 | "range_flattens": [ |
| 389 | null |
| 390 | ], |
| 391 | "load_eviction_policies": [ |
| 392 | "", |
| 393 | "", |
| 394 | "" |
| 395 | ], |
| 396 | "num_warps": 8, |
| 397 | "num_stages": 1, |
| 398 | "indexing": [ |
| 399 | "pointer", |
| 400 | "pointer", |
| 401 | "pointer", |
| 402 | "pointer" |
| 403 | ], |
| 404 | "pid_type": "flat" |
| 405 | } |
| 406 | }, |
| 407 | { |
| 408 | "key": { |
| 409 | "intermediate": 2048, |
| 410 | "numtokens": 1 |
| 411 | }, |
| 412 | "config": { |
| 413 | "block_sizes": [ |
| 414 | 1, |
| 415 | 256 |
| 416 | ], |
| 417 | "loop_orders": [ |
| 418 | [ |
| 419 | 0, |
| 420 | 1 |
| 421 | ] |
| 422 | ], |
| 423 | "flatten_loops": [ |
| 424 | true |
| 425 | ], |
| 426 | "l2_groupings": [ |
| 427 | 1 |
| 428 | ], |
| 429 | "range_unroll_factors": [ |
| 430 | 0 |
| 431 | ], |
| 432 | "range_warp_specializes": [], |
| 433 | "range_num_stages": [ |
| 434 | 0 |
| 435 | ], |
| 436 | "range_multi_buffers": [ |
| 437 | null |
| 438 | ], |
| 439 | "range_flattens": [ |
| 440 | null |
| 441 | ], |
| 442 | "load_eviction_policies": [ |
| 443 | "", |
| 444 | "", |
| 445 | "" |
| 446 | ], |
| 447 | "num_warps": 2, |
| 448 | "num_stages": 1, |
| 449 | "indexing": [ |
| 450 | "pointer", |
| 451 | "pointer", |
| 452 | "pointer", |
| 453 | "pointer" |
| 454 | ], |
| 455 | "pid_type": "flat" |
| 456 | } |
| 457 | }, |
| 458 | { |
| 459 | "key": { |
| 460 | "intermediate": 2880, |
| 461 | "numtokens": 1 |
| 462 | }, |
| 463 | "config": { |
| 464 | "block_sizes": [ |
| 465 | 1, |
| 466 | 128 |
| 467 | ], |
| 468 | "loop_orders": [ |
| 469 | [ |
| 470 | 0, |
| 471 | 1 |
| 472 | ] |
| 473 | ], |
| 474 | "flatten_loops": [ |
| 475 | true |
| 476 | ], |
| 477 | "l2_groupings": [ |
| 478 | 1 |
| 479 | ], |
| 480 | "range_unroll_factors": [ |
| 481 | 0 |
| 482 | ], |
| 483 | "range_warp_specializes": [], |
| 484 | "range_num_stages": [ |
| 485 | 0 |
| 486 | ], |
| 487 | "range_multi_buffers": [ |
| 488 | null |
| 489 | ], |
| 490 | "range_flattens": [ |
| 491 | null |
| 492 | ], |
| 493 | "load_eviction_policies": [ |
| 494 | "", |
| 495 | "", |
| 496 | "" |
| 497 | ], |
| 498 | "num_warps": 4, |
| 499 | "num_stages": 1, |
| 500 | "indexing": [ |
| 501 | "pointer", |
| 502 | "pointer", |
| 503 | "pointer", |
| 504 | "pointer" |
| 505 | ], |
| 506 | "pid_type": "flat" |
| 507 | } |
| 508 | }, |
| 509 | { |
| 510 | "key": { |
| 511 | "intermediate": 4096, |
| 512 | "numtokens": 1 |
| 513 | }, |
| 514 | "config": { |
| 515 | "block_sizes": [ |
| 516 | 1, |
| 517 | 256 |
| 518 | ], |
| 519 | "loop_orders": [ |
| 520 | [ |
| 521 | 0, |
| 522 | 1 |
| 523 | ] |
| 524 | ], |
| 525 | "flatten_loops": [ |
| 526 | true |
| 527 | ], |
| 528 | "l2_groupings": [ |
| 529 | 1 |
| 530 | ], |
| 531 | "range_unroll_factors": [ |
| 532 | 0 |
| 533 | ], |
| 534 | "range_warp_specializes": [], |
| 535 | "range_num_stages": [ |
| 536 | 0 |
| 537 | ], |
| 538 | "range_multi_buffers": [ |
| 539 | null |
| 540 | ], |
| 541 | "range_flattens": [ |
| 542 | null |
| 543 | ], |
| 544 | "load_eviction_policies": [ |
| 545 | "", |
| 546 | "", |
| 547 | "" |
| 548 | ], |
| 549 | "num_warps": 2, |
| 550 | "num_stages": 1, |
| 551 | "indexing": [ |
| 552 | "pointer", |
| 553 | "pointer", |
| 554 | "pointer", |
| 555 | "pointer" |
| 556 | ], |
| 557 | "pid_type": "flat" |
| 558 | } |
| 559 | }, |
| 560 | { |
| 561 | "key": { |
| 562 | "intermediate": 8192, |
| 563 | "numtokens": 1 |
| 564 | }, |
| 565 | "config": { |
| 566 | "block_sizes": [ |
| 567 | 1, |
| 568 | 256 |
| 569 | ], |
| 570 | "loop_orders": [ |
| 571 | [ |
| 572 | 0, |
| 573 | 1 |
| 574 | ] |
| 575 | ], |
| 576 | "flatten_loops": [ |
| 577 | true |
| 578 | ], |
| 579 | "l2_groupings": [ |
| 580 | 1 |
| 581 | ], |
| 582 | "range_unroll_factors": [ |
| 583 | 0 |
| 584 | ], |
| 585 | "range_warp_specializes": [], |
| 586 | "range_num_stages": [ |
| 587 | 0 |
| 588 | ], |
| 589 | "range_multi_buffers": [ |
| 590 | null |
| 591 | ], |
| 592 | "range_flattens": [ |
| 593 | null |
| 594 | ], |
| 595 | "load_eviction_policies": [ |
| 596 | "", |
| 597 | "", |
| 598 | "" |
| 599 | ], |
| 600 | "num_warps": 2, |
| 601 | "num_stages": 1, |
| 602 | "indexing": [ |
| 603 | "pointer", |
| 604 | "pointer", |
| 605 | "pointer", |
| 606 | "pointer" |
| 607 | ], |
| 608 | "pid_type": "flat" |
| 609 | } |
| 610 | }, |
| 611 | { |
| 612 | "key": { |
| 613 | "intermediate": 11008, |
| 614 | "numtokens": 1 |
| 615 | }, |
| 616 | "config": { |
| 617 | "block_sizes": [ |
| 618 | 1, |
| 619 | 256 |
| 620 | ], |
| 621 | "loop_orders": [ |
| 622 | [ |
| 623 | 0, |
| 624 | 1 |
| 625 | ] |
| 626 | ], |
| 627 | "flatten_loops": [ |
| 628 | true |
| 629 | ], |
| 630 | "l2_groupings": [ |
| 631 | 1 |
| 632 | ], |
| 633 | "range_unroll_factors": [ |
| 634 | 0 |
| 635 | ], |
| 636 | "range_warp_specializes": [], |
| 637 | "range_num_stages": [ |
| 638 | 0 |
| 639 | ], |
| 640 | "range_multi_buffers": [ |
| 641 | null |
| 642 | ], |
| 643 | "range_flattens": [ |
| 644 | null |
| 645 | ], |
| 646 | "load_eviction_policies": [ |
| 647 | "", |
| 648 | "", |
| 649 | "" |
| 650 | ], |
| 651 | "num_warps": 2, |
| 652 | "num_stages": 1, |
| 653 | "indexing": [ |
| 654 | "pointer", |
| 655 | "pointer", |
| 656 | "pointer", |
| 657 | "pointer" |
| 658 | ], |
| 659 | "pid_type": "flat" |
| 660 | } |
| 661 | }, |
| 662 | { |
| 663 | "key": { |
| 664 | "intermediate": 14336, |
| 665 | "numtokens": 1 |
| 666 | }, |
| 667 | "config": { |
| 668 | "block_sizes": [ |
| 669 | 1, |
| 670 | 256 |
| 671 | ], |
| 672 | "loop_orders": [ |
| 673 | [ |
| 674 | 0, |
| 675 | 1 |
| 676 | ] |
| 677 | ], |
| 678 | "flatten_loops": [ |
| 679 | true |
| 680 | ], |
| 681 | "l2_groupings": [ |
| 682 | 1 |
| 683 | ], |
| 684 | "range_unroll_factors": [ |
| 685 | 0 |
| 686 | ], |
| 687 | "range_warp_specializes": [], |
| 688 | "range_num_stages": [ |
| 689 | 0 |
| 690 | ], |
| 691 | "range_multi_buffers": [ |
| 692 | null |
| 693 | ], |
| 694 | "range_flattens": [ |
| 695 | null |
| 696 | ], |
| 697 | "load_eviction_policies": [ |
| 698 | "", |
| 699 | "", |
| 700 | "" |
| 701 | ], |
| 702 | "num_warps": 2, |
| 703 | "num_stages": 1, |
| 704 | "indexing": [ |
| 705 | "pointer", |
| 706 | "pointer", |
| 707 | "pointer", |
| 708 | "pointer" |
| 709 | ], |
| 710 | "pid_type": "flat" |
| 711 | } |
| 712 | }, |
| 713 | { |
| 714 | "key": { |
| 715 | "intermediate": 2048, |
| 716 | "numtokens": 2 |
| 717 | }, |
| 718 | "config": { |
| 719 | "block_sizes": [ |
| 720 | 2, |
| 721 | 128 |
| 722 | ], |
| 723 | "loop_orders": [ |
| 724 | [ |
| 725 | 0, |
| 726 | 1 |
| 727 | ] |
| 728 | ], |
| 729 | "flatten_loops": [ |
| 730 | true |
| 731 | ], |
| 732 | "l2_groupings": [ |
| 733 | 1 |
| 734 | ], |
| 735 | "range_unroll_factors": [ |
| 736 | 0 |
| 737 | ], |
| 738 | "range_warp_specializes": [], |
| 739 | "range_num_stages": [ |
| 740 | 0 |
| 741 | ], |
| 742 | "range_multi_buffers": [ |
| 743 | null |
| 744 | ], |
| 745 | "range_flattens": [ |
| 746 | null |
| 747 | ], |
| 748 | "load_eviction_policies": [ |
| 749 | "", |
| 750 | "", |
| 751 | "" |
| 752 | ], |
| 753 | "num_warps": 2, |
| 754 | "num_stages": 1, |
| 755 | "indexing": [ |
| 756 | "pointer", |
| 757 | "pointer", |
| 758 | "pointer", |
| 759 | "pointer" |
| 760 | ], |
| 761 | "pid_type": "flat" |
| 762 | } |
| 763 | }, |
| 764 | { |
| 765 | "key": { |
| 766 | "intermediate": 2880, |
| 767 | "numtokens": 2 |
| 768 | }, |
| 769 | "config": { |
| 770 | "block_sizes": [ |
| 771 | 2, |
| 772 | 256 |
| 773 | ], |
| 774 | "loop_orders": [ |
| 775 | [ |
| 776 | 0, |
| 777 | 1 |
| 778 | ] |
| 779 | ], |
| 780 | "flatten_loops": [ |
| 781 | true |
| 782 | ], |
| 783 | "l2_groupings": [ |
| 784 | 1 |
| 785 | ], |
| 786 | "range_unroll_factors": [ |
| 787 | 0 |
| 788 | ], |
| 789 | "range_warp_specializes": [], |
| 790 | "range_num_stages": [ |
| 791 | 0 |
| 792 | ], |
| 793 | "range_multi_buffers": [ |
| 794 | null |
| 795 | ], |
| 796 | "range_flattens": [ |
| 797 | null |
| 798 | ], |
| 799 | "load_eviction_policies": [ |
| 800 | "", |
| 801 | "", |
| 802 | "" |
| 803 | ], |
| 804 | "num_warps": 4, |
| 805 | "num_stages": 1, |
| 806 | "indexing": [ |
| 807 | "pointer", |
| 808 | "pointer", |
| 809 | "pointer", |
| 810 | "pointer" |
| 811 | ], |
| 812 | "pid_type": "flat" |
| 813 | } |
| 814 | }, |
| 815 | { |
| 816 | "key": { |
| 817 | "intermediate": 4096, |
| 818 | "numtokens": 2 |
| 819 | }, |
| 820 | "config": { |
| 821 | "block_sizes": [ |
| 822 | 2, |
| 823 | 128 |
| 824 | ], |
| 825 | "loop_orders": [ |
| 826 | [ |
| 827 | 0, |
| 828 | 1 |
| 829 | ] |
| 830 | ], |
| 831 | "flatten_loops": [ |
| 832 | true |
| 833 | ], |
| 834 | "l2_groupings": [ |
| 835 | 1 |
| 836 | ], |
| 837 | "range_unroll_factors": [ |
| 838 | 0 |
| 839 | ], |
| 840 | "range_warp_specializes": [], |
| 841 | "range_num_stages": [ |
| 842 | 0 |
| 843 | ], |
| 844 | "range_multi_buffers": [ |
| 845 | null |
| 846 | ], |
| 847 | "range_flattens": [ |
| 848 | null |
| 849 | ], |
| 850 | "load_eviction_policies": [ |
| 851 | "", |
| 852 | "", |
| 853 | "" |
| 854 | ], |
| 855 | "num_warps": 2, |
| 856 | "num_stages": 1, |
| 857 | "indexing": [ |
| 858 | "pointer", |
| 859 | "pointer", |
| 860 | "pointer", |
| 861 | "pointer" |
| 862 | ], |
| 863 | "pid_type": "flat" |
| 864 | } |
| 865 | }, |
| 866 | { |
| 867 | "key": { |
| 868 | "intermediate": 8192, |
| 869 | "numtokens": 2 |
| 870 | }, |
| 871 | "config": { |
| 872 | "block_sizes": [ |
| 873 | 2, |
| 874 | 128 |
| 875 | ], |
| 876 | "loop_orders": [ |
| 877 | [ |
| 878 | 0, |
| 879 | 1 |
| 880 | ] |
| 881 | ], |
| 882 | "flatten_loops": [ |
| 883 | true |
| 884 | ], |
| 885 | "l2_groupings": [ |
| 886 | 1 |
| 887 | ], |
| 888 | "range_unroll_factors": [ |
| 889 | 0 |
| 890 | ], |
| 891 | "range_warp_specializes": [], |
| 892 | "range_num_stages": [ |
| 893 | 0 |
| 894 | ], |
| 895 | "range_multi_buffers": [ |
| 896 | null |
| 897 | ], |
| 898 | "range_flattens": [ |
| 899 | null |
| 900 | ], |
| 901 | "load_eviction_policies": [ |
| 902 | "", |
| 903 | "", |
| 904 | "" |
| 905 | ], |
| 906 | "num_warps": 2, |
| 907 | "num_stages": 1, |
| 908 | "indexing": [ |
| 909 | "pointer", |
| 910 | "pointer", |
| 911 | "pointer", |
| 912 | "pointer" |
| 913 | ], |
| 914 | "pid_type": "flat" |
| 915 | } |
| 916 | }, |
| 917 | { |
| 918 | "key": { |
| 919 | "intermediate": 11008, |
| 920 | "numtokens": 2 |
| 921 | }, |
| 922 | "config": { |
| 923 | "block_sizes": [ |
| 924 | 1, |
| 925 | 16384 |
| 926 | ], |
| 927 | "loop_orders": [ |
| 928 | [ |
| 929 | 1, |
| 930 | 0 |
| 931 | ] |
| 932 | ], |
| 933 | "flatten_loops": [ |
| 934 | true |
| 935 | ], |
| 936 | "l2_groupings": [ |
| 937 | 16 |
| 938 | ], |
| 939 | "range_unroll_factors": [ |
| 940 | 0 |
| 941 | ], |
| 942 | "range_warp_specializes": [], |
| 943 | "range_num_stages": [ |
| 944 | 0 |
| 945 | ], |
| 946 | "range_multi_buffers": [ |
| 947 | null |
| 948 | ], |
| 949 | "range_flattens": [ |
| 950 | null |
| 951 | ], |
| 952 | "load_eviction_policies": [ |
| 953 | "last", |
| 954 | "first", |
| 955 | "last" |
| 956 | ], |
| 957 | "num_warps": 32, |
| 958 | "num_stages": 3, |
| 959 | "indexing": [ |
| 960 | "tensor_descriptor", |
| 961 | "pointer", |
| 962 | "pointer", |
| 963 | "tensor_descriptor" |
| 964 | ], |
| 965 | "pid_type": "xyz" |
| 966 | } |
| 967 | }, |
| 968 | { |
| 969 | "key": { |
| 970 | "intermediate": 14336, |
| 971 | "numtokens": 2 |
| 972 | }, |
| 973 | "config": { |
| 974 | "block_sizes": [ |
| 975 | 2, |
| 976 | 128 |
| 977 | ], |
| 978 | "loop_orders": [ |
| 979 | [ |
| 980 | 0, |
| 981 | 1 |
| 982 | ] |
| 983 | ], |
| 984 | "flatten_loops": [ |
| 985 | true |
| 986 | ], |
| 987 | "l2_groupings": [ |
| 988 | 1 |
| 989 | ], |
| 990 | "range_unroll_factors": [ |
| 991 | 0 |
| 992 | ], |
| 993 | "range_warp_specializes": [], |
| 994 | "range_num_stages": [ |
| 995 | 0 |
| 996 | ], |
| 997 | "range_multi_buffers": [ |
| 998 | null |
| 999 | ], |
| 1000 | "range_flattens": [ |
| 1001 | null |
| 1002 | ], |
| 1003 | "load_eviction_policies": [ |
| 1004 | "", |
| 1005 | "", |
| 1006 | "" |
| 1007 | ], |
| 1008 | "num_warps": 2, |
| 1009 | "num_stages": 1, |
| 1010 | "indexing": [ |
| 1011 | "pointer", |
| 1012 | "pointer", |
| 1013 | "pointer", |
| 1014 | "pointer" |
| 1015 | ], |
| 1016 | "pid_type": "flat" |
| 1017 | } |
| 1018 | }, |
| 1019 | { |
| 1020 | "key": { |
| 1021 | "intermediate": 2048, |
| 1022 | "numtokens": 4 |
| 1023 | }, |
| 1024 | "config": { |
| 1025 | "block_sizes": [ |
| 1026 | 4, |
| 1027 | 64 |
| 1028 | ], |
| 1029 | "loop_orders": [ |
| 1030 | [ |
| 1031 | 0, |
| 1032 | 1 |
| 1033 | ] |
| 1034 | ], |
| 1035 | "flatten_loops": [ |
| 1036 | true |
| 1037 | ], |
| 1038 | "l2_groupings": [ |
| 1039 | 1 |
| 1040 | ], |
| 1041 | "range_unroll_factors": [ |
| 1042 | 0 |
| 1043 | ], |
| 1044 | "range_warp_specializes": [], |
| 1045 | "range_num_stages": [ |
| 1046 | 0 |
| 1047 | ], |
| 1048 | "range_multi_buffers": [ |
| 1049 | null |
| 1050 | ], |
| 1051 | "range_flattens": [ |
| 1052 | null |
| 1053 | ], |
| 1054 | "load_eviction_policies": [ |
| 1055 | "", |
| 1056 | "", |
| 1057 | "" |
| 1058 | ], |
| 1059 | "num_warps": 2, |
| 1060 | "num_stages": 1, |
| 1061 | "indexing": [ |
| 1062 | "pointer", |
| 1063 | "pointer", |
| 1064 | "pointer", |
| 1065 | "pointer" |
| 1066 | ], |
| 1067 | "pid_type": "flat" |
| 1068 | } |
| 1069 | }, |
| 1070 | { |
| 1071 | "key": { |
| 1072 | "intermediate": 2880, |
| 1073 | "numtokens": 4 |
| 1074 | }, |
| 1075 | "config": { |
| 1076 | "block_sizes": [ |
| 1077 | 4, |
| 1078 | 64 |
| 1079 | ], |
| 1080 | "loop_orders": [ |
| 1081 | [ |
| 1082 | 0, |
| 1083 | 1 |
| 1084 | ] |
| 1085 | ], |
| 1086 | "flatten_loops": [ |
| 1087 | true |
| 1088 | ], |
| 1089 | "l2_groupings": [ |
| 1090 | 64 |
| 1091 | ], |
| 1092 | "range_unroll_factors": [ |
| 1093 | 0 |
| 1094 | ], |
| 1095 | "range_warp_specializes": [], |
| 1096 | "range_num_stages": [ |
| 1097 | 0 |
| 1098 | ], |
| 1099 | "range_multi_buffers": [ |
| 1100 | null |
| 1101 | ], |
| 1102 | "range_flattens": [ |
| 1103 | null |
| 1104 | ], |
| 1105 | "load_eviction_policies": [ |
| 1106 | "", |
| 1107 | "last", |
| 1108 | "first" |
| 1109 | ], |
| 1110 | "num_warps": 8, |
| 1111 | "num_stages": 7, |
| 1112 | "indexing": [ |
| 1113 | "pointer", |
| 1114 | "tensor_descriptor", |
| 1115 | "tensor_descriptor", |
| 1116 | "pointer" |
| 1117 | ], |
| 1118 | "pid_type": "xyz" |
| 1119 | } |
| 1120 | }, |
| 1121 | { |
| 1122 | "key": { |
| 1123 | "intermediate": 4096, |
| 1124 | "numtokens": 4 |
| 1125 | }, |
| 1126 | "config": { |
| 1127 | "block_sizes": [ |
| 1128 | 4, |
| 1129 | 64 |
| 1130 | ], |
| 1131 | "loop_orders": [ |
| 1132 | [ |
| 1133 | 0, |
| 1134 | 1 |
| 1135 | ] |
| 1136 | ], |
| 1137 | "flatten_loops": [ |
| 1138 | true |
| 1139 | ], |
| 1140 | "l2_groupings": [ |
| 1141 | 1 |
| 1142 | ], |
| 1143 | "range_unroll_factors": [ |
| 1144 | 0 |
| 1145 | ], |
| 1146 | "range_warp_specializes": [], |
| 1147 | "range_num_stages": [ |
| 1148 | 0 |
| 1149 | ], |
| 1150 | "range_multi_buffers": [ |
| 1151 | null |
| 1152 | ], |
| 1153 | "range_flattens": [ |
| 1154 | null |
| 1155 | ], |
| 1156 | "load_eviction_policies": [ |
| 1157 | "", |
| 1158 | "", |
| 1159 | "" |
| 1160 | ], |
| 1161 | "num_warps": 2, |
| 1162 | "num_stages": 1, |
| 1163 | "indexing": [ |
| 1164 | "pointer", |
| 1165 | "pointer", |
| 1166 | "pointer", |
| 1167 | "pointer" |
| 1168 | ], |
| 1169 | "pid_type": "flat" |
| 1170 | } |
| 1171 | }, |
| 1172 | { |
| 1173 | "key": { |
| 1174 | "intermediate": 8192, |
| 1175 | "numtokens": 4 |
| 1176 | }, |
| 1177 | "config": { |
| 1178 | "block_sizes": [ |
| 1179 | 1, |
| 1180 | 2048 |
| 1181 | ], |
| 1182 | "loop_orders": [ |
| 1183 | [ |
| 1184 | 0, |
| 1185 | 1 |
| 1186 | ] |
| 1187 | ], |
| 1188 | "flatten_loops": [ |
| 1189 | false |
| 1190 | ], |
| 1191 | "l2_groupings": [ |
| 1192 | 1 |
| 1193 | ], |
| 1194 | "range_unroll_factors": [ |
| 1195 | 0 |
| 1196 | ], |
| 1197 | "range_warp_specializes": [], |
| 1198 | "range_num_stages": [ |
| 1199 | 0 |
| 1200 | ], |
| 1201 | "range_multi_buffers": [ |
| 1202 | null |
| 1203 | ], |
| 1204 | "range_flattens": [ |
| 1205 | null |
| 1206 | ], |
| 1207 | "load_eviction_policies": [ |
| 1208 | "", |
| 1209 | "", |
| 1210 | "" |
| 1211 | ], |
| 1212 | "num_warps": 8, |
| 1213 | "num_stages": 8, |
| 1214 | "indexing": [ |
| 1215 | "pointer", |
| 1216 | "tensor_descriptor", |
| 1217 | "pointer", |
| 1218 | "pointer" |
| 1219 | ], |
| 1220 | "pid_type": "flat" |
| 1221 | } |
| 1222 | }, |
| 1223 | { |
| 1224 | "key": { |
| 1225 | "intermediate": 11008, |
| 1226 | "numtokens": 4 |
| 1227 | }, |
| 1228 | "config": { |
| 1229 | "block_sizes": [ |
| 1230 | 4, |
| 1231 | 2048 |
| 1232 | ], |
| 1233 | "loop_orders": [ |
| 1234 | [ |
| 1235 | 1, |
| 1236 | 0 |
| 1237 | ] |
| 1238 | ], |
| 1239 | "flatten_loops": [ |
| 1240 | true |
| 1241 | ], |
| 1242 | "l2_groupings": [ |
| 1243 | 4 |
| 1244 | ], |
| 1245 | "range_unroll_factors": [ |
| 1246 | 0 |
| 1247 | ], |
| 1248 | "range_warp_specializes": [], |
| 1249 | "range_num_stages": [ |
| 1250 | 0 |
| 1251 | ], |
| 1252 | "range_multi_buffers": [ |
| 1253 | null |
| 1254 | ], |
| 1255 | "range_flattens": [ |
| 1256 | null |
| 1257 | ], |
| 1258 | "load_eviction_policies": [ |
| 1259 | "first", |
| 1260 | "last", |
| 1261 | "last" |
| 1262 | ], |
| 1263 | "num_warps": 8, |
| 1264 | "num_stages": 6, |
| 1265 | "indexing": [ |
| 1266 | "pointer", |
| 1267 | "tensor_descriptor", |
| 1268 | "tensor_descriptor", |
| 1269 | "tensor_descriptor" |
| 1270 | ], |
| 1271 | "pid_type": "xyz" |
| 1272 | } |
| 1273 | }, |
| 1274 | { |
| 1275 | "key": { |
| 1276 | "intermediate": 14336, |
| 1277 | "numtokens": 4 |
| 1278 | }, |
| 1279 | "config": { |
| 1280 | "block_sizes": [ |
| 1281 | 4, |
| 1282 | 256 |
| 1283 | ], |
| 1284 | "loop_orders": [ |
| 1285 | [ |
| 1286 | 0, |
| 1287 | 1 |
| 1288 | ] |
| 1289 | ], |
| 1290 | "flatten_loops": [ |
| 1291 | true |
| 1292 | ], |
| 1293 | "l2_groupings": [ |
| 1294 | 64 |
| 1295 | ], |
| 1296 | "range_unroll_factors": [ |
| 1297 | 0 |
| 1298 | ], |
| 1299 | "range_warp_specializes": [], |
| 1300 | "range_num_stages": [ |
| 1301 | 0 |
| 1302 | ], |
| 1303 | "range_multi_buffers": [ |
| 1304 | null |
| 1305 | ], |
| 1306 | "range_flattens": [ |
| 1307 | null |
| 1308 | ], |
| 1309 | "load_eviction_policies": [ |
| 1310 | "", |
| 1311 | "", |
| 1312 | "" |
| 1313 | ], |
| 1314 | "num_warps": 2, |
| 1315 | "num_stages": 2, |
| 1316 | "indexing": [ |
| 1317 | "pointer", |
| 1318 | "pointer", |
| 1319 | "pointer", |
| 1320 | "pointer" |
| 1321 | ], |
| 1322 | "pid_type": "xyz" |
| 1323 | } |
| 1324 | }, |
| 1325 | { |
| 1326 | "key": { |
| 1327 | "intermediate": 2048, |
| 1328 | "numtokens": 8 |
| 1329 | }, |
| 1330 | "config": { |
| 1331 | "block_sizes": [ |
| 1332 | 8, |
| 1333 | 32 |
| 1334 | ], |
| 1335 | "loop_orders": [ |
| 1336 | [ |
| 1337 | 0, |
| 1338 | 1 |
| 1339 | ] |
| 1340 | ], |
| 1341 | "flatten_loops": [ |
| 1342 | true |
| 1343 | ], |
| 1344 | "l2_groupings": [ |
| 1345 | 1 |
| 1346 | ], |
| 1347 | "range_unroll_factors": [ |
| 1348 | 0 |
| 1349 | ], |
| 1350 | "range_warp_specializes": [], |
| 1351 | "range_num_stages": [ |
| 1352 | 0 |
| 1353 | ], |
| 1354 | "range_multi_buffers": [ |
| 1355 | null |
| 1356 | ], |
| 1357 | "range_flattens": [ |
| 1358 | null |
| 1359 | ], |
| 1360 | "load_eviction_policies": [ |
| 1361 | "", |
| 1362 | "", |
| 1363 | "" |
| 1364 | ], |
| 1365 | "num_warps": 2, |
| 1366 | "num_stages": 1, |
| 1367 | "indexing": [ |
| 1368 | "pointer", |
| 1369 | "pointer", |
| 1370 | "pointer", |
| 1371 | "pointer" |
| 1372 | ], |
| 1373 | "pid_type": "flat" |
| 1374 | } |
| 1375 | }, |
| 1376 | { |
| 1377 | "key": { |
| 1378 | "intermediate": 2880, |
| 1379 | "numtokens": 8 |
| 1380 | }, |
| 1381 | "config": { |
| 1382 | "block_sizes": [ |
| 1383 | 4, |
| 1384 | 2048 |
| 1385 | ], |
| 1386 | "loop_orders": [ |
| 1387 | [ |
| 1388 | 0, |
| 1389 | 1 |
| 1390 | ] |
| 1391 | ], |
| 1392 | "flatten_loops": [ |
| 1393 | true |
| 1394 | ], |
| 1395 | "l2_groupings": [ |
| 1396 | 4 |
| 1397 | ], |
| 1398 | "range_unroll_factors": [ |
| 1399 | 0 |
| 1400 | ], |
| 1401 | "range_warp_specializes": [], |
| 1402 | "range_num_stages": [ |
| 1403 | 0 |
| 1404 | ], |
| 1405 | "range_multi_buffers": [ |
| 1406 | null |
| 1407 | ], |
| 1408 | "range_flattens": [ |
| 1409 | null |
| 1410 | ], |
| 1411 | "load_eviction_policies": [ |
| 1412 | "last", |
| 1413 | "last", |
| 1414 | "" |
| 1415 | ], |
| 1416 | "num_warps": 16, |
| 1417 | "num_stages": 5, |
| 1418 | "indexing": [ |
| 1419 | "pointer", |
| 1420 | "tensor_descriptor", |
| 1421 | "tensor_descriptor", |
| 1422 | "pointer" |
| 1423 | ], |
| 1424 | "pid_type": "xyz" |
| 1425 | } |
| 1426 | }, |
| 1427 | { |
| 1428 | "key": { |
| 1429 | "intermediate": 4096, |
| 1430 | "numtokens": 8 |
| 1431 | }, |
| 1432 | "config": { |
| 1433 | "block_sizes": [ |
| 1434 | 8, |
| 1435 | 32 |
| 1436 | ], |
| 1437 | "loop_orders": [ |
| 1438 | [ |
| 1439 | 0, |
| 1440 | 1 |
| 1441 | ] |
| 1442 | ], |
| 1443 | "flatten_loops": [ |
| 1444 | true |
| 1445 | ], |
| 1446 | "l2_groupings": [ |
| 1447 | 1 |
| 1448 | ], |
| 1449 | "range_unroll_factors": [ |
| 1450 | 0 |
| 1451 | ], |
| 1452 | "range_warp_specializes": [], |
| 1453 | "range_num_stages": [ |
| 1454 | 0 |
| 1455 | ], |
| 1456 | "range_multi_buffers": [ |
| 1457 | null |
| 1458 | ], |
| 1459 | "range_flattens": [ |
| 1460 | null |
| 1461 | ], |
| 1462 | "load_eviction_policies": [ |
| 1463 | "", |
| 1464 | "", |
| 1465 | "" |
| 1466 | ], |
| 1467 | "num_warps": 2, |
| 1468 | "num_stages": 1, |
| 1469 | "indexing": [ |
| 1470 | "pointer", |
| 1471 | "pointer", |
| 1472 | "pointer", |
| 1473 | "pointer" |
| 1474 | ], |
| 1475 | "pid_type": "flat" |
| 1476 | } |
| 1477 | }, |
| 1478 | { |
| 1479 | "key": { |
| 1480 | "intermediate": 8192, |
| 1481 | "numtokens": 8 |
| 1482 | }, |
| 1483 | "config": { |
| 1484 | "block_sizes": [ |
| 1485 | 2, |
| 1486 | 1024 |
| 1487 | ], |
| 1488 | "loop_orders": [ |
| 1489 | [ |
| 1490 | 0, |
| 1491 | 1 |
| 1492 | ] |
| 1493 | ], |
| 1494 | "flatten_loops": [ |
| 1495 | true |
| 1496 | ], |
| 1497 | "l2_groupings": [ |
| 1498 | 16 |
| 1499 | ], |
| 1500 | "range_unroll_factors": [ |
| 1501 | 0 |
| 1502 | ], |
| 1503 | "range_warp_specializes": [], |
| 1504 | "range_num_stages": [ |
| 1505 | 0 |
| 1506 | ], |
| 1507 | "range_multi_buffers": [ |
| 1508 | null |
| 1509 | ], |
| 1510 | "range_flattens": [ |
| 1511 | null |
| 1512 | ], |
| 1513 | "load_eviction_policies": [ |
| 1514 | "first", |
| 1515 | "last", |
| 1516 | "" |
| 1517 | ], |
| 1518 | "num_warps": 1, |
| 1519 | "num_stages": 8, |
| 1520 | "indexing": [ |
| 1521 | "tensor_descriptor", |
| 1522 | "pointer", |
| 1523 | "tensor_descriptor", |
| 1524 | "pointer" |
| 1525 | ], |
| 1526 | "pid_type": "flat" |
| 1527 | } |
| 1528 | }, |
| 1529 | { |
| 1530 | "key": { |
| 1531 | "intermediate": 11008, |
| 1532 | "numtokens": 8 |
| 1533 | }, |
| 1534 | "config": { |
| 1535 | "block_sizes": [ |
| 1536 | 1, |
| 1537 | 1024 |
| 1538 | ], |
| 1539 | "loop_orders": [ |
| 1540 | [ |
| 1541 | 1, |
| 1542 | 0 |
| 1543 | ] |
| 1544 | ], |
| 1545 | "flatten_loops": [ |
| 1546 | true |
| 1547 | ], |
| 1548 | "l2_groupings": [ |
| 1549 | 4 |
| 1550 | ], |
| 1551 | "range_unroll_factors": [ |
| 1552 | 0 |
| 1553 | ], |
| 1554 | "range_warp_specializes": [], |
| 1555 | "range_num_stages": [ |
| 1556 | 0 |
| 1557 | ], |
| 1558 | "range_multi_buffers": [ |
| 1559 | null |
| 1560 | ], |
| 1561 | "range_flattens": [ |
| 1562 | null |
| 1563 | ], |
| 1564 | "load_eviction_policies": [ |
| 1565 | "last", |
| 1566 | "", |
| 1567 | "first" |
| 1568 | ], |
| 1569 | "num_warps": 2, |
| 1570 | "num_stages": 5, |
| 1571 | "indexing": [ |
| 1572 | "tensor_descriptor", |
| 1573 | "tensor_descriptor", |
| 1574 | "tensor_descriptor", |
| 1575 | "tensor_descriptor" |
| 1576 | ], |
| 1577 | "pid_type": "flat" |
| 1578 | } |
| 1579 | }, |
| 1580 | { |
| 1581 | "key": { |
| 1582 | "intermediate": 14336, |
| 1583 | "numtokens": 8 |
| 1584 | }, |
| 1585 | "config": { |
| 1586 | "block_sizes": [ |
| 1587 | 8, |
| 1588 | 128 |
| 1589 | ], |
| 1590 | "loop_orders": [ |
| 1591 | [ |
| 1592 | 0, |
| 1593 | 1 |
| 1594 | ] |
| 1595 | ], |
| 1596 | "flatten_loops": [ |
| 1597 | true |
| 1598 | ], |
| 1599 | "l2_groupings": [ |
| 1600 | 1 |
| 1601 | ], |
| 1602 | "range_unroll_factors": [ |
| 1603 | 0 |
| 1604 | ], |
| 1605 | "range_warp_specializes": [], |
| 1606 | "range_num_stages": [ |
| 1607 | 0 |
| 1608 | ], |
| 1609 | "range_multi_buffers": [ |
| 1610 | null |
| 1611 | ], |
| 1612 | "range_flattens": [ |
| 1613 | null |
| 1614 | ], |
| 1615 | "load_eviction_policies": [ |
| 1616 | "", |
| 1617 | "", |
| 1618 | "" |
| 1619 | ], |
| 1620 | "num_warps": 4, |
| 1621 | "num_stages": 1, |
| 1622 | "indexing": [ |
| 1623 | "pointer", |
| 1624 | "pointer", |
| 1625 | "pointer", |
| 1626 | "pointer" |
| 1627 | ], |
| 1628 | "pid_type": "flat" |
| 1629 | } |
| 1630 | }, |
| 1631 | { |
| 1632 | "key": { |
| 1633 | "intermediate": 2048, |
| 1634 | "numtokens": 16 |
| 1635 | }, |
| 1636 | "config": { |
| 1637 | "block_sizes": [ |
| 1638 | 8, |
| 1639 | 512 |
| 1640 | ], |
| 1641 | "loop_orders": [ |
| 1642 | [ |
| 1643 | 0, |
| 1644 | 1 |
| 1645 | ] |
| 1646 | ], |
| 1647 | "flatten_loops": [ |
| 1648 | true |
| 1649 | ], |
| 1650 | "l2_groupings": [ |
| 1651 | 64 |
| 1652 | ], |
| 1653 | "range_unroll_factors": [ |
| 1654 | 0 |
| 1655 | ], |
| 1656 | "range_warp_specializes": [], |
| 1657 | "range_num_stages": [ |
| 1658 | 0 |
| 1659 | ], |
| 1660 | "range_multi_buffers": [ |
| 1661 | null |
| 1662 | ], |
| 1663 | "range_flattens": [ |
| 1664 | null |
| 1665 | ], |
| 1666 | "load_eviction_policies": [ |
| 1667 | "last", |
| 1668 | "last", |
| 1669 | "first" |
| 1670 | ], |
| 1671 | "num_warps": 16, |
| 1672 | "num_stages": 2, |
| 1673 | "indexing": [ |
| 1674 | "pointer", |
| 1675 | "pointer", |
| 1676 | "tensor_descriptor", |
| 1677 | "pointer" |
| 1678 | ], |
| 1679 | "pid_type": "xyz" |
| 1680 | } |
| 1681 | }, |
| 1682 | { |
| 1683 | "key": { |
| 1684 | "intermediate": 2880, |
| 1685 | "numtokens": 16 |
| 1686 | }, |
| 1687 | "config": { |
| 1688 | "block_sizes": [ |
| 1689 | 2, |
| 1690 | 256 |
| 1691 | ], |
| 1692 | "loop_orders": [ |
| 1693 | [ |
| 1694 | 0, |
| 1695 | 1 |
| 1696 | ] |
| 1697 | ], |
| 1698 | "flatten_loops": [ |
| 1699 | true |
| 1700 | ], |
| 1701 | "l2_groupings": [ |
| 1702 | 1 |
| 1703 | ], |
| 1704 | "range_unroll_factors": [ |
| 1705 | 0 |
| 1706 | ], |
| 1707 | "range_warp_specializes": [], |
| 1708 | "range_num_stages": [ |
| 1709 | 0 |
| 1710 | ], |
| 1711 | "range_multi_buffers": [ |
| 1712 | null |
| 1713 | ], |
| 1714 | "range_flattens": [ |
| 1715 | null |
| 1716 | ], |
| 1717 | "load_eviction_policies": [ |
| 1718 | "", |
| 1719 | "", |
| 1720 | "" |
| 1721 | ], |
| 1722 | "num_warps": 4, |
| 1723 | "num_stages": 1, |
| 1724 | "indexing": [ |
| 1725 | "pointer", |
| 1726 | "pointer", |
| 1727 | "pointer", |
| 1728 | "pointer" |
| 1729 | ], |
| 1730 | "pid_type": "flat" |
| 1731 | } |
| 1732 | }, |
| 1733 | { |
| 1734 | "key": { |
| 1735 | "intermediate": 4096, |
| 1736 | "numtokens": 16 |
| 1737 | }, |
| 1738 | "config": { |
| 1739 | "block_sizes": [ |
| 1740 | 16, |
| 1741 | 256 |
| 1742 | ], |
| 1743 | "loop_orders": [ |
| 1744 | [ |
| 1745 | 0, |
| 1746 | 1 |
| 1747 | ] |
| 1748 | ], |
| 1749 | "flatten_loops": [ |
| 1750 | true |
| 1751 | ], |
| 1752 | "l2_groupings": [ |
| 1753 | 16 |
| 1754 | ], |
| 1755 | "range_unroll_factors": [ |
| 1756 | 0 |
| 1757 | ], |
| 1758 | "range_warp_specializes": [], |
| 1759 | "range_num_stages": [ |
| 1760 | 0 |
| 1761 | ], |
| 1762 | "range_multi_buffers": [ |
| 1763 | null |
| 1764 | ], |
| 1765 | "range_flattens": [ |
| 1766 | null |
| 1767 | ], |
| 1768 | "load_eviction_policies": [ |
| 1769 | "last", |
| 1770 | "", |
| 1771 | "last" |
| 1772 | ], |
| 1773 | "num_warps": 32, |
| 1774 | "num_stages": 1, |
| 1775 | "indexing": [ |
| 1776 | "pointer", |
| 1777 | "pointer", |
| 1778 | "pointer", |
| 1779 | "pointer" |
| 1780 | ], |
| 1781 | "pid_type": "flat" |
| 1782 | } |
| 1783 | }, |
| 1784 | { |
| 1785 | "key": { |
| 1786 | "intermediate": 8192, |
| 1787 | "numtokens": 16 |
| 1788 | }, |
| 1789 | "config": { |
| 1790 | "block_sizes": [ |
| 1791 | 16, |
| 1792 | 64 |
| 1793 | ], |
| 1794 | "loop_orders": [ |
| 1795 | [ |
| 1796 | 0, |
| 1797 | 1 |
| 1798 | ] |
| 1799 | ], |
| 1800 | "flatten_loops": [ |
| 1801 | true |
| 1802 | ], |
| 1803 | "l2_groupings": [ |
| 1804 | 1 |
| 1805 | ], |
| 1806 | "range_unroll_factors": [ |
| 1807 | 0 |
| 1808 | ], |
| 1809 | "range_warp_specializes": [], |
| 1810 | "range_num_stages": [ |
| 1811 | 0 |
| 1812 | ], |
| 1813 | "range_multi_buffers": [ |
| 1814 | null |
| 1815 | ], |
| 1816 | "range_flattens": [ |
| 1817 | null |
| 1818 | ], |
| 1819 | "load_eviction_policies": [ |
| 1820 | "", |
| 1821 | "", |
| 1822 | "" |
| 1823 | ], |
| 1824 | "num_warps": 4, |
| 1825 | "num_stages": 1, |
| 1826 | "indexing": [ |
| 1827 | "pointer", |
| 1828 | "pointer", |
| 1829 | "pointer", |
| 1830 | "pointer" |
| 1831 | ], |
| 1832 | "pid_type": "flat" |
| 1833 | } |
| 1834 | }, |
| 1835 | { |
| 1836 | "key": { |
| 1837 | "intermediate": 11008, |
| 1838 | "numtokens": 16 |
| 1839 | }, |
| 1840 | "config": { |
| 1841 | "block_sizes": [ |
| 1842 | 1, |
| 1843 | 2048 |
| 1844 | ], |
| 1845 | "loop_orders": [ |
| 1846 | [ |
| 1847 | 0, |
| 1848 | 1 |
| 1849 | ] |
| 1850 | ], |
| 1851 | "flatten_loops": [ |
| 1852 | true |
| 1853 | ], |
| 1854 | "l2_groupings": [ |
| 1855 | 1 |
| 1856 | ], |
| 1857 | "range_unroll_factors": [ |
| 1858 | 0 |
| 1859 | ], |
| 1860 | "range_warp_specializes": [], |
| 1861 | "range_num_stages": [ |
| 1862 | 0 |
| 1863 | ], |
| 1864 | "range_multi_buffers": [ |
| 1865 | null |
| 1866 | ], |
| 1867 | "range_flattens": [ |
| 1868 | null |
| 1869 | ], |
| 1870 | "load_eviction_policies": [ |
| 1871 | "", |
| 1872 | "", |
| 1873 | "" |
| 1874 | ], |
| 1875 | "num_warps": 8, |
| 1876 | "num_stages": 1, |
| 1877 | "indexing": [ |
| 1878 | "pointer", |
| 1879 | "pointer", |
| 1880 | "pointer", |
| 1881 | "pointer" |
| 1882 | ], |
| 1883 | "pid_type": "flat" |
| 1884 | } |
| 1885 | }, |
| 1886 | { |
| 1887 | "key": { |
| 1888 | "intermediate": 14336, |
| 1889 | "numtokens": 16 |
| 1890 | }, |
| 1891 | "config": { |
| 1892 | "block_sizes": [ |
| 1893 | 2, |
| 1894 | 256 |
| 1895 | ], |
| 1896 | "loop_orders": [ |
| 1897 | [ |
| 1898 | 1, |
| 1899 | 0 |
| 1900 | ] |
| 1901 | ], |
| 1902 | "flatten_loops": [ |
| 1903 | true |
| 1904 | ], |
| 1905 | "l2_groupings": [ |
| 1906 | 64 |
| 1907 | ], |
| 1908 | "range_unroll_factors": [ |
| 1909 | 0 |
| 1910 | ], |
| 1911 | "range_warp_specializes": [], |
| 1912 | "range_num_stages": [ |
| 1913 | 0 |
| 1914 | ], |
| 1915 | "range_multi_buffers": [ |
| 1916 | null |
| 1917 | ], |
| 1918 | "range_flattens": [ |
| 1919 | null |
| 1920 | ], |
| 1921 | "load_eviction_policies": [ |
| 1922 | "last", |
| 1923 | "last", |
| 1924 | "last" |
| 1925 | ], |
| 1926 | "num_warps": 16, |
| 1927 | "num_stages": 7, |
| 1928 | "indexing": [ |
| 1929 | "tensor_descriptor", |
| 1930 | "pointer", |
| 1931 | "tensor_descriptor", |
| 1932 | "tensor_descriptor" |
| 1933 | ], |
| 1934 | "pid_type": "flat" |
| 1935 | } |
| 1936 | }, |
| 1937 | { |
| 1938 | "key": { |
| 1939 | "intermediate": 2048, |
| 1940 | "numtokens": 24 |
| 1941 | }, |
| 1942 | "config": { |
| 1943 | "block_sizes": [ |
| 1944 | 1, |
| 1945 | 1024 |
| 1946 | ], |
| 1947 | "loop_orders": [ |
| 1948 | [ |
| 1949 | 0, |
| 1950 | 1 |
| 1951 | ] |
| 1952 | ], |
| 1953 | "flatten_loops": [ |
| 1954 | true |
| 1955 | ], |
| 1956 | "l2_groupings": [ |
| 1957 | 64 |
| 1958 | ], |
| 1959 | "range_unroll_factors": [ |
| 1960 | 0 |
| 1961 | ], |
| 1962 | "range_warp_specializes": [], |
| 1963 | "range_num_stages": [ |
| 1964 | 0 |
| 1965 | ], |
| 1966 | "range_multi_buffers": [ |
| 1967 | null |
| 1968 | ], |
| 1969 | "range_flattens": [ |
| 1970 | null |
| 1971 | ], |
| 1972 | "load_eviction_policies": [ |
| 1973 | "", |
| 1974 | "", |
| 1975 | "first" |
| 1976 | ], |
| 1977 | "num_warps": 4, |
| 1978 | "num_stages": 8, |
| 1979 | "indexing": [ |
| 1980 | "tensor_descriptor", |
| 1981 | "pointer", |
| 1982 | "pointer", |
| 1983 | "tensor_descriptor" |
| 1984 | ], |
| 1985 | "pid_type": "flat" |
| 1986 | } |
| 1987 | }, |
| 1988 | { |
| 1989 | "key": { |
| 1990 | "intermediate": 2880, |
| 1991 | "numtokens": 24 |
| 1992 | }, |
| 1993 | "config": { |
| 1994 | "block_sizes": [ |
| 1995 | 4, |
| 1996 | 1024 |
| 1997 | ], |
| 1998 | "loop_orders": [ |
| 1999 | [ |
| 2000 | 0, |
| 2001 | 1 |
| 2002 | ] |
| 2003 | ], |
| 2004 | "flatten_loops": [ |
| 2005 | true |
| 2006 | ], |
| 2007 | "l2_groupings": [ |
| 2008 | 16 |
| 2009 | ], |
| 2010 | "range_unroll_factors": [ |
| 2011 | 0 |
| 2012 | ], |
| 2013 | "range_warp_specializes": [], |
| 2014 | "range_num_stages": [ |
| 2015 | 0 |
| 2016 | ], |
| 2017 | "range_multi_buffers": [ |
| 2018 | null |
| 2019 | ], |
| 2020 | "range_flattens": [ |
| 2021 | null |
| 2022 | ], |
| 2023 | "load_eviction_policies": [ |
| 2024 | "first", |
| 2025 | "", |
| 2026 | "first" |
| 2027 | ], |
| 2028 | "num_warps": 16, |
| 2029 | "num_stages": 3, |
| 2030 | "indexing": [ |
| 2031 | "tensor_descriptor", |
| 2032 | "pointer", |
| 2033 | "tensor_descriptor", |
| 2034 | "pointer" |
| 2035 | ], |
| 2036 | "pid_type": "flat" |
| 2037 | } |
| 2038 | }, |
| 2039 | { |
| 2040 | "key": { |
| 2041 | "intermediate": 4096, |
| 2042 | "numtokens": 24 |
| 2043 | }, |
| 2044 | "config": { |
| 2045 | "block_sizes": [ |
| 2046 | 16, |
| 2047 | 64 |
| 2048 | ], |
| 2049 | "loop_orders": [ |
| 2050 | [ |
| 2051 | 0, |
| 2052 | 1 |
| 2053 | ] |
| 2054 | ], |
| 2055 | "flatten_loops": [ |
| 2056 | true |
| 2057 | ], |
| 2058 | "l2_groupings": [ |
| 2059 | 1 |
| 2060 | ], |
| 2061 | "range_unroll_factors": [ |
| 2062 | 0 |
| 2063 | ], |
| 2064 | "range_warp_specializes": [], |
| 2065 | "range_num_stages": [ |
| 2066 | 0 |
| 2067 | ], |
| 2068 | "range_multi_buffers": [ |
| 2069 | null |
| 2070 | ], |
| 2071 | "range_flattens": [ |
| 2072 | null |
| 2073 | ], |
| 2074 | "load_eviction_policies": [ |
| 2075 | "", |
| 2076 | "", |
| 2077 | "" |
| 2078 | ], |
| 2079 | "num_warps": 4, |
| 2080 | "num_stages": 1, |
| 2081 | "indexing": [ |
| 2082 | "pointer", |
| 2083 | "pointer", |
| 2084 | "pointer", |
| 2085 | "pointer" |
| 2086 | ], |
| 2087 | "pid_type": "flat" |
| 2088 | } |
| 2089 | }, |
| 2090 | { |
| 2091 | "key": { |
| 2092 | "intermediate": 8192, |
| 2093 | "numtokens": 24 |
| 2094 | }, |
| 2095 | "config": { |
| 2096 | "block_sizes": [ |
| 2097 | 1, |
| 2098 | 2048 |
| 2099 | ], |
| 2100 | "loop_orders": [ |
| 2101 | [ |
| 2102 | 0, |
| 2103 | 1 |
| 2104 | ] |
| 2105 | ], |
| 2106 | "flatten_loops": [ |
| 2107 | true |
| 2108 | ], |
| 2109 | "l2_groupings": [ |
| 2110 | 32 |
| 2111 | ], |
| 2112 | "range_unroll_factors": [ |
| 2113 | 0 |
| 2114 | ], |
| 2115 | "range_warp_specializes": [], |
| 2116 | "range_num_stages": [ |
| 2117 | 0 |
| 2118 | ], |
| 2119 | "range_multi_buffers": [ |
| 2120 | null |
| 2121 | ], |
| 2122 | "range_flattens": [ |
| 2123 | null |
| 2124 | ], |
| 2125 | "load_eviction_policies": [ |
| 2126 | "", |
| 2127 | "last", |
| 2128 | "last" |
| 2129 | ], |
| 2130 | "num_warps": 32, |
| 2131 | "num_stages": 5, |
| 2132 | "indexing": [ |
| 2133 | "pointer", |
| 2134 | "pointer", |
| 2135 | "pointer", |
| 2136 | "tensor_descriptor" |
| 2137 | ], |
| 2138 | "pid_type": "flat" |
| 2139 | } |
| 2140 | }, |
| 2141 | { |
| 2142 | "key": { |
| 2143 | "intermediate": 11008, |
| 2144 | "numtokens": 24 |
| 2145 | }, |
| 2146 | "config": { |
| 2147 | "block_sizes": [ |
| 2148 | 1, |
| 2149 | 1024 |
| 2150 | ], |
| 2151 | "loop_orders": [ |
| 2152 | [ |
| 2153 | 1, |
| 2154 | 0 |
| 2155 | ] |
| 2156 | ], |
| 2157 | "flatten_loops": [ |
| 2158 | false |
| 2159 | ], |
| 2160 | "l2_groupings": [ |
| 2161 | 64 |
| 2162 | ], |
| 2163 | "range_unroll_factors": [ |
| 2164 | 0 |
| 2165 | ], |
| 2166 | "range_warp_specializes": [], |
| 2167 | "range_num_stages": [ |
| 2168 | 0 |
| 2169 | ], |
| 2170 | "range_multi_buffers": [ |
| 2171 | null |
| 2172 | ], |
| 2173 | "range_flattens": [ |
| 2174 | null |
| 2175 | ], |
| 2176 | "load_eviction_policies": [ |
| 2177 | "last", |
| 2178 | "", |
| 2179 | "" |
| 2180 | ], |
| 2181 | "num_warps": 2, |
| 2182 | "num_stages": 2, |
| 2183 | "indexing": [ |
| 2184 | "pointer", |
| 2185 | "pointer", |
| 2186 | "pointer", |
| 2187 | "pointer" |
| 2188 | ], |
| 2189 | "pid_type": "flat" |
| 2190 | } |
| 2191 | }, |
| 2192 | { |
| 2193 | "key": { |
| 2194 | "intermediate": 14336, |
| 2195 | "numtokens": 24 |
| 2196 | }, |
| 2197 | "config": { |
| 2198 | "block_sizes": [ |
| 2199 | 8, |
| 2200 | 512 |
| 2201 | ], |
| 2202 | "loop_orders": [ |
| 2203 | [ |
| 2204 | 0, |
| 2205 | 1 |
| 2206 | ] |
| 2207 | ], |
| 2208 | "flatten_loops": [ |
| 2209 | true |
| 2210 | ], |
| 2211 | "l2_groupings": [ |
| 2212 | 4 |
| 2213 | ], |
| 2214 | "range_unroll_factors": [ |
| 2215 | 0 |
| 2216 | ], |
| 2217 | "range_warp_specializes": [], |
| 2218 | "range_num_stages": [ |
| 2219 | 0 |
| 2220 | ], |
| 2221 | "range_multi_buffers": [ |
| 2222 | null |
| 2223 | ], |
| 2224 | "range_flattens": [ |
| 2225 | null |
| 2226 | ], |
| 2227 | "load_eviction_policies": [ |
| 2228 | "last", |
| 2229 | "first", |
| 2230 | "first" |
| 2231 | ], |
| 2232 | "num_warps": 32, |
| 2233 | "num_stages": 3, |
| 2234 | "indexing": [ |
| 2235 | "pointer", |
| 2236 | "tensor_descriptor", |
| 2237 | "pointer", |
| 2238 | "tensor_descriptor" |
| 2239 | ], |
| 2240 | "pid_type": "flat" |
| 2241 | } |
| 2242 | }, |
| 2243 | { |
| 2244 | "key": { |
| 2245 | "intermediate": 2048, |
| 2246 | "numtokens": 32 |
| 2247 | }, |
| 2248 | "config": { |
| 2249 | "block_sizes": [ |
| 2250 | 32, |
| 2251 | 16 |
| 2252 | ], |
| 2253 | "loop_orders": [ |
| 2254 | [ |
| 2255 | 0, |
| 2256 | 1 |
| 2257 | ] |
| 2258 | ], |
| 2259 | "flatten_loops": [ |
| 2260 | true |
| 2261 | ], |
| 2262 | "l2_groupings": [ |
| 2263 | 1 |
| 2264 | ], |
| 2265 | "range_unroll_factors": [ |
| 2266 | 0 |
| 2267 | ], |
| 2268 | "range_warp_specializes": [], |
| 2269 | "range_num_stages": [ |
| 2270 | 0 |
| 2271 | ], |
| 2272 | "range_multi_buffers": [ |
| 2273 | null |
| 2274 | ], |
| 2275 | "range_flattens": [ |
| 2276 | null |
| 2277 | ], |
| 2278 | "load_eviction_policies": [ |
| 2279 | "", |
| 2280 | "", |
| 2281 | "" |
| 2282 | ], |
| 2283 | "num_warps": 4, |
| 2284 | "num_stages": 1, |
| 2285 | "indexing": [ |
| 2286 | "pointer", |
| 2287 | "pointer", |
| 2288 | "pointer", |
| 2289 | "pointer" |
| 2290 | ], |
| 2291 | "pid_type": "flat" |
| 2292 | } |
| 2293 | }, |
| 2294 | { |
| 2295 | "key": { |
| 2296 | "intermediate": 2880, |
| 2297 | "numtokens": 32 |
| 2298 | }, |
| 2299 | "config": { |
| 2300 | "block_sizes": [ |
| 2301 | 4, |
| 2302 | 256 |
| 2303 | ], |
| 2304 | "loop_orders": [ |
| 2305 | [ |
| 2306 | 0, |
| 2307 | 1 |
| 2308 | ] |
| 2309 | ], |
| 2310 | "flatten_loops": [ |
| 2311 | true |
| 2312 | ], |
| 2313 | "l2_groupings": [ |
| 2314 | 1 |
| 2315 | ], |
| 2316 | "range_unroll_factors": [ |
| 2317 | 0 |
| 2318 | ], |
| 2319 | "range_warp_specializes": [], |
| 2320 | "range_num_stages": [ |
| 2321 | 0 |
| 2322 | ], |
| 2323 | "range_multi_buffers": [ |
| 2324 | null |
| 2325 | ], |
| 2326 | "range_flattens": [ |
| 2327 | null |
| 2328 | ], |
| 2329 | "load_eviction_policies": [ |
| 2330 | "", |
| 2331 | "", |
| 2332 | "" |
| 2333 | ], |
| 2334 | "num_warps": 4, |
| 2335 | "num_stages": 1, |
| 2336 | "indexing": [ |
| 2337 | "pointer", |
| 2338 | "pointer", |
| 2339 | "pointer", |
| 2340 | "pointer" |
| 2341 | ], |
| 2342 | "pid_type": "flat" |
| 2343 | } |
| 2344 | }, |
| 2345 | { |
| 2346 | "key": { |
| 2347 | "intermediate": 4096, |
| 2348 | "numtokens": 32 |
| 2349 | }, |
| 2350 | "config": { |
| 2351 | "block_sizes": [ |
| 2352 | 4, |
| 2353 | 4096 |
| 2354 | ], |
| 2355 | "loop_orders": [ |
| 2356 | [ |
| 2357 | 0, |
| 2358 | 1 |
| 2359 | ] |
| 2360 | ], |
| 2361 | "flatten_loops": [ |
| 2362 | true |
| 2363 | ], |
| 2364 | "l2_groupings": [ |
| 2365 | 16 |
| 2366 | ], |
| 2367 | "range_unroll_factors": [ |
| 2368 | 0 |
| 2369 | ], |
| 2370 | "range_warp_specializes": [], |
| 2371 | "range_num_stages": [ |
| 2372 | 0 |
| 2373 | ], |
| 2374 | "range_multi_buffers": [ |
| 2375 | null |
| 2376 | ], |
| 2377 | "range_flattens": [ |
| 2378 | null |
| 2379 | ], |
| 2380 | "load_eviction_policies": [ |
| 2381 | "first", |
| 2382 | "last", |
| 2383 | "" |
| 2384 | ], |
| 2385 | "num_warps": 32, |
| 2386 | "num_stages": 5, |
| 2387 | "indexing": [ |
| 2388 | "pointer", |
| 2389 | "tensor_descriptor", |
| 2390 | "pointer", |
| 2391 | "pointer" |
| 2392 | ], |
| 2393 | "pid_type": "flat" |
| 2394 | } |
| 2395 | }, |
| 2396 | { |
| 2397 | "key": { |
| 2398 | "intermediate": 8192, |
| 2399 | "numtokens": 32 |
| 2400 | }, |
| 2401 | "config": { |
| 2402 | "block_sizes": [ |
| 2403 | 4, |
| 2404 | 1024 |
| 2405 | ], |
| 2406 | "loop_orders": [ |
| 2407 | [ |
| 2408 | 0, |
| 2409 | 1 |
| 2410 | ] |
| 2411 | ], |
| 2412 | "flatten_loops": [ |
| 2413 | true |
| 2414 | ], |
| 2415 | "l2_groupings": [ |
| 2416 | 2 |
| 2417 | ], |
| 2418 | "range_unroll_factors": [ |
| 2419 | 0 |
| 2420 | ], |
| 2421 | "range_warp_specializes": [], |
| 2422 | "range_num_stages": [ |
| 2423 | 0 |
| 2424 | ], |
| 2425 | "range_multi_buffers": [ |
| 2426 | null |
| 2427 | ], |
| 2428 | "range_flattens": [ |
| 2429 | null |
| 2430 | ], |
| 2431 | "load_eviction_policies": [ |
| 2432 | "", |
| 2433 | "last", |
| 2434 | "" |
| 2435 | ], |
| 2436 | "num_warps": 2, |
| 2437 | "num_stages": 3, |
| 2438 | "indexing": [ |
| 2439 | "tensor_descriptor", |
| 2440 | "tensor_descriptor", |
| 2441 | "pointer", |
| 2442 | "tensor_descriptor" |
| 2443 | ], |
| 2444 | "pid_type": "flat" |
| 2445 | } |
| 2446 | }, |
| 2447 | { |
| 2448 | "key": { |
| 2449 | "intermediate": 11008, |
| 2450 | "numtokens": 32 |
| 2451 | }, |
| 2452 | "config": { |
| 2453 | "block_sizes": [ |
| 2454 | 2, |
| 2455 | 2048 |
| 2456 | ], |
| 2457 | "loop_orders": [ |
| 2458 | [ |
| 2459 | 0, |
| 2460 | 1 |
| 2461 | ] |
| 2462 | ], |
| 2463 | "flatten_loops": [ |
| 2464 | true |
| 2465 | ], |
| 2466 | "l2_groupings": [ |
| 2467 | 1 |
| 2468 | ], |
| 2469 | "range_unroll_factors": [ |
| 2470 | 0 |
| 2471 | ], |
| 2472 | "range_warp_specializes": [], |
| 2473 | "range_num_stages": [ |
| 2474 | 0 |
| 2475 | ], |
| 2476 | "range_multi_buffers": [ |
| 2477 | null |
| 2478 | ], |
| 2479 | "range_flattens": [ |
| 2480 | null |
| 2481 | ], |
| 2482 | "load_eviction_policies": [ |
| 2483 | "", |
| 2484 | "", |
| 2485 | "" |
| 2486 | ], |
| 2487 | "num_warps": 8, |
| 2488 | "num_stages": 1, |
| 2489 | "indexing": [ |
| 2490 | "pointer", |
| 2491 | "pointer", |
| 2492 | "pointer", |
| 2493 | "pointer" |
| 2494 | ], |
| 2495 | "pid_type": "flat" |
| 2496 | } |
| 2497 | }, |
| 2498 | { |
| 2499 | "key": { |
| 2500 | "intermediate": 14336, |
| 2501 | "numtokens": 32 |
| 2502 | }, |
| 2503 | "config": { |
| 2504 | "block_sizes": [ |
| 2505 | 1, |
| 2506 | 512 |
| 2507 | ], |
| 2508 | "loop_orders": [ |
| 2509 | [ |
| 2510 | 0, |
| 2511 | 1 |
| 2512 | ] |
| 2513 | ], |
| 2514 | "flatten_loops": [ |
| 2515 | true |
| 2516 | ], |
| 2517 | "l2_groupings": [ |
| 2518 | 64 |
| 2519 | ], |
| 2520 | "range_unroll_factors": [ |
| 2521 | 0 |
| 2522 | ], |
| 2523 | "range_warp_specializes": [], |
| 2524 | "range_num_stages": [ |
| 2525 | 0 |
| 2526 | ], |
| 2527 | "range_multi_buffers": [ |
| 2528 | null |
| 2529 | ], |
| 2530 | "range_flattens": [ |
| 2531 | null |
| 2532 | ], |
| 2533 | "load_eviction_policies": [ |
| 2534 | "first", |
| 2535 | "", |
| 2536 | "" |
| 2537 | ], |
| 2538 | "num_warps": 4, |
| 2539 | "num_stages": 3, |
| 2540 | "indexing": [ |
| 2541 | "pointer", |
| 2542 | "tensor_descriptor", |
| 2543 | "pointer", |
| 2544 | "pointer" |
| 2545 | ], |
| 2546 | "pid_type": "flat" |
| 2547 | } |
| 2548 | }, |
| 2549 | { |
| 2550 | "key": { |
| 2551 | "intermediate": 2048, |
| 2552 | "numtokens": 40 |
| 2553 | }, |
| 2554 | "config": { |
| 2555 | "block_sizes": [ |
| 2556 | 32, |
| 2557 | 32 |
| 2558 | ], |
| 2559 | "loop_orders": [ |
| 2560 | [ |
| 2561 | 0, |
| 2562 | 1 |
| 2563 | ] |
| 2564 | ], |
| 2565 | "flatten_loops": [ |
| 2566 | true |
| 2567 | ], |
| 2568 | "l2_groupings": [ |
| 2569 | 1 |
| 2570 | ], |
| 2571 | "range_unroll_factors": [ |
| 2572 | 0 |
| 2573 | ], |
| 2574 | "range_warp_specializes": [], |
| 2575 | "range_num_stages": [ |
| 2576 | 0 |
| 2577 | ], |
| 2578 | "range_multi_buffers": [ |
| 2579 | null |
| 2580 | ], |
| 2581 | "range_flattens": [ |
| 2582 | null |
| 2583 | ], |
| 2584 | "load_eviction_policies": [ |
| 2585 | "", |
| 2586 | "", |
| 2587 | "" |
| 2588 | ], |
| 2589 | "num_warps": 4, |
| 2590 | "num_stages": 1, |
| 2591 | "indexing": [ |
| 2592 | "pointer", |
| 2593 | "pointer", |
| 2594 | "pointer", |
| 2595 | "pointer" |
| 2596 | ], |
| 2597 | "pid_type": "flat" |
| 2598 | } |
| 2599 | }, |
| 2600 | { |
| 2601 | "key": { |
| 2602 | "intermediate": 2880, |
| 2603 | "numtokens": 40 |
| 2604 | }, |
| 2605 | "config": { |
| 2606 | "block_sizes": [ |
| 2607 | 1, |
| 2608 | 4096 |
| 2609 | ], |
| 2610 | "loop_orders": [ |
| 2611 | [ |
| 2612 | 1, |
| 2613 | 0 |
| 2614 | ] |
| 2615 | ], |
| 2616 | "flatten_loops": [ |
| 2617 | false |
| 2618 | ], |
| 2619 | "l2_groupings": [ |
| 2620 | 32 |
| 2621 | ], |
| 2622 | "range_unroll_factors": [ |
| 2623 | 0 |
| 2624 | ], |
| 2625 | "range_warp_specializes": [], |
| 2626 | "range_num_stages": [ |
| 2627 | 0 |
| 2628 | ], |
| 2629 | "range_multi_buffers": [ |
| 2630 | null |
| 2631 | ], |
| 2632 | "range_flattens": [ |
| 2633 | null |
| 2634 | ], |
| 2635 | "load_eviction_policies": [ |
| 2636 | "", |
| 2637 | "last", |
| 2638 | "last" |
| 2639 | ], |
| 2640 | "num_warps": 8, |
| 2641 | "num_stages": 4, |
| 2642 | "indexing": [ |
| 2643 | "tensor_descriptor", |
| 2644 | "pointer", |
| 2645 | "pointer", |
| 2646 | "pointer" |
| 2647 | ], |
| 2648 | "pid_type": "flat" |
| 2649 | } |
| 2650 | }, |
| 2651 | { |
| 2652 | "key": { |
| 2653 | "intermediate": 4096, |
| 2654 | "numtokens": 40 |
| 2655 | }, |
| 2656 | "config": { |
| 2657 | "block_sizes": [ |
| 2658 | 32, |
| 2659 | 64 |
| 2660 | ], |
| 2661 | "loop_orders": [ |
| 2662 | [ |
| 2663 | 0, |
| 2664 | 1 |
| 2665 | ] |
| 2666 | ], |
| 2667 | "flatten_loops": [ |
| 2668 | true |
| 2669 | ], |
| 2670 | "l2_groupings": [ |
| 2671 | 1 |
| 2672 | ], |
| 2673 | "range_unroll_factors": [ |
| 2674 | 0 |
| 2675 | ], |
| 2676 | "range_warp_specializes": [], |
| 2677 | "range_num_stages": [ |
| 2678 | 0 |
| 2679 | ], |
| 2680 | "range_multi_buffers": [ |
| 2681 | null |
| 2682 | ], |
| 2683 | "range_flattens": [ |
| 2684 | null |
| 2685 | ], |
| 2686 | "load_eviction_policies": [ |
| 2687 | "", |
| 2688 | "", |
| 2689 | "" |
| 2690 | ], |
| 2691 | "num_warps": 8, |
| 2692 | "num_stages": 1, |
| 2693 | "indexing": [ |
| 2694 | "pointer", |
| 2695 | "pointer", |
| 2696 | "pointer", |
| 2697 | "pointer" |
| 2698 | ], |
| 2699 | "pid_type": "flat" |
| 2700 | } |
| 2701 | }, |
| 2702 | { |
| 2703 | "key": { |
| 2704 | "intermediate": 8192, |
| 2705 | "numtokens": 40 |
| 2706 | }, |
| 2707 | "config": { |
| 2708 | "block_sizes": [ |
| 2709 | 2, |
| 2710 | 256 |
| 2711 | ], |
| 2712 | "loop_orders": [ |
| 2713 | [ |
| 2714 | 0, |
| 2715 | 1 |
| 2716 | ] |
| 2717 | ], |
| 2718 | "flatten_loops": [ |
| 2719 | true |
| 2720 | ], |
| 2721 | "l2_groupings": [ |
| 2722 | 16 |
| 2723 | ], |
| 2724 | "range_unroll_factors": [ |
| 2725 | 0 |
| 2726 | ], |
| 2727 | "range_warp_specializes": [], |
| 2728 | "range_num_stages": [ |
| 2729 | 0 |
| 2730 | ], |
| 2731 | "range_multi_buffers": [ |
| 2732 | null |
| 2733 | ], |
| 2734 | "range_flattens": [ |
| 2735 | null |
| 2736 | ], |
| 2737 | "load_eviction_policies": [ |
| 2738 | "", |
| 2739 | "last", |
| 2740 | "last" |
| 2741 | ], |
| 2742 | "num_warps": 2, |
| 2743 | "num_stages": 2, |
| 2744 | "indexing": [ |
| 2745 | "pointer", |
| 2746 | "pointer", |
| 2747 | "pointer", |
| 2748 | "tensor_descriptor" |
| 2749 | ], |
| 2750 | "pid_type": "flat" |
| 2751 | } |
| 2752 | }, |
| 2753 | { |
| 2754 | "key": { |
| 2755 | "intermediate": 11008, |
| 2756 | "numtokens": 40 |
| 2757 | }, |
| 2758 | "config": { |
| 2759 | "block_sizes": [ |
| 2760 | 16, |
| 2761 | 256 |
| 2762 | ], |
| 2763 | "loop_orders": [ |
| 2764 | [ |
| 2765 | 0, |
| 2766 | 1 |
| 2767 | ] |
| 2768 | ], |
| 2769 | "flatten_loops": [ |
| 2770 | true |
| 2771 | ], |
| 2772 | "l2_groupings": [ |
| 2773 | 1 |
| 2774 | ], |
| 2775 | "range_unroll_factors": [ |
| 2776 | 0 |
| 2777 | ], |
| 2778 | "range_warp_specializes": [], |
| 2779 | "range_num_stages": [ |
| 2780 | 0 |
| 2781 | ], |
| 2782 | "range_multi_buffers": [ |
| 2783 | null |
| 2784 | ], |
| 2785 | "range_flattens": [ |
| 2786 | null |
| 2787 | ], |
| 2788 | "load_eviction_policies": [ |
| 2789 | "", |
| 2790 | "", |
| 2791 | "" |
| 2792 | ], |
| 2793 | "num_warps": 8, |
| 2794 | "num_stages": 1, |
| 2795 | "indexing": [ |
| 2796 | "pointer", |
| 2797 | "pointer", |
| 2798 | "pointer", |
| 2799 | "pointer" |
| 2800 | ], |
| 2801 | "pid_type": "flat" |
| 2802 | } |
| 2803 | }, |
| 2804 | { |
| 2805 | "key": { |
| 2806 | "intermediate": 14336, |
| 2807 | "numtokens": 40 |
| 2808 | }, |
| 2809 | "config": { |
| 2810 | "block_sizes": [ |
| 2811 | 4, |
| 2812 | 1024 |
| 2813 | ], |
| 2814 | "loop_orders": [ |
| 2815 | [ |
| 2816 | 0, |
| 2817 | 1 |
| 2818 | ] |
| 2819 | ], |
| 2820 | "flatten_loops": [ |
| 2821 | true |
| 2822 | ], |
| 2823 | "l2_groupings": [ |
| 2824 | 1 |
| 2825 | ], |
| 2826 | "range_unroll_factors": [ |
| 2827 | 0 |
| 2828 | ], |
| 2829 | "range_warp_specializes": [], |
| 2830 | "range_num_stages": [ |
| 2831 | 1 |
| 2832 | ], |
| 2833 | "range_multi_buffers": [ |
| 2834 | null |
| 2835 | ], |
| 2836 | "range_flattens": [ |
| 2837 | null |
| 2838 | ], |
| 2839 | "load_eviction_policies": [ |
| 2840 | "first", |
| 2841 | "", |
| 2842 | "last" |
| 2843 | ], |
| 2844 | "num_warps": 16, |
| 2845 | "num_stages": 5, |
| 2846 | "indexing": [ |
| 2847 | "tensor_descriptor", |
| 2848 | "tensor_descriptor", |
| 2849 | "tensor_descriptor", |
| 2850 | "tensor_descriptor" |
| 2851 | ], |
| 2852 | "pid_type": "persistent_interleaved", |
| 2853 | "num_sm_multiplier": 32, |
| 2854 | "maxnreg": 32 |
| 2855 | } |
| 2856 | }, |
| 2857 | { |
| 2858 | "key": { |
| 2859 | "intermediate": 2048, |
| 2860 | "numtokens": 48 |
| 2861 | }, |
| 2862 | "config": { |
| 2863 | "block_sizes": [ |
| 2864 | 32, |
| 2865 | 32 |
| 2866 | ], |
| 2867 | "loop_orders": [ |
| 2868 | [ |
| 2869 | 0, |
| 2870 | 1 |
| 2871 | ] |
| 2872 | ], |
| 2873 | "flatten_loops": [ |
| 2874 | true |
| 2875 | ], |
| 2876 | "l2_groupings": [ |
| 2877 | 1 |
| 2878 | ], |
| 2879 | "range_unroll_factors": [ |
| 2880 | 0 |
| 2881 | ], |
| 2882 | "range_warp_specializes": [], |
| 2883 | "range_num_stages": [ |
| 2884 | 0 |
| 2885 | ], |
| 2886 | "range_multi_buffers": [ |
| 2887 | null |
| 2888 | ], |
| 2889 | "range_flattens": [ |
| 2890 | null |
| 2891 | ], |
| 2892 | "load_eviction_policies": [ |
| 2893 | "", |
| 2894 | "", |
| 2895 | "" |
| 2896 | ], |
| 2897 | "num_warps": 4, |
| 2898 | "num_stages": 1, |
| 2899 | "indexing": [ |
| 2900 | "pointer", |
| 2901 | "pointer", |
| 2902 | "pointer", |
| 2903 | "pointer" |
| 2904 | ], |
| 2905 | "pid_type": "flat" |
| 2906 | } |
| 2907 | }, |
| 2908 | { |
| 2909 | "key": { |
| 2910 | "intermediate": 2880, |
| 2911 | "numtokens": 48 |
| 2912 | }, |
| 2913 | "config": { |
| 2914 | "block_sizes": [ |
| 2915 | 16, |
| 2916 | 64 |
| 2917 | ], |
| 2918 | "loop_orders": [ |
| 2919 | [ |
| 2920 | 0, |
| 2921 | 1 |
| 2922 | ] |
| 2923 | ], |
| 2924 | "flatten_loops": [ |
| 2925 | true |
| 2926 | ], |
| 2927 | "l2_groupings": [ |
| 2928 | 1 |
| 2929 | ], |
| 2930 | "range_unroll_factors": [ |
| 2931 | 0 |
| 2932 | ], |
| 2933 | "range_warp_specializes": [], |
| 2934 | "range_num_stages": [ |
| 2935 | 0 |
| 2936 | ], |
| 2937 | "range_multi_buffers": [ |
| 2938 | null |
| 2939 | ], |
| 2940 | "range_flattens": [ |
| 2941 | null |
| 2942 | ], |
| 2943 | "load_eviction_policies": [ |
| 2944 | "", |
| 2945 | "", |
| 2946 | "" |
| 2947 | ], |
| 2948 | "num_warps": 4, |
| 2949 | "num_stages": 1, |
| 2950 | "indexing": [ |
| 2951 | "pointer", |
| 2952 | "pointer", |
| 2953 | "pointer", |
| 2954 | "pointer" |
| 2955 | ], |
| 2956 | "pid_type": "flat" |
| 2957 | } |
| 2958 | }, |
| 2959 | { |
| 2960 | "key": { |
| 2961 | "intermediate": 4096, |
| 2962 | "numtokens": 48 |
| 2963 | }, |
| 2964 | "config": { |
| 2965 | "block_sizes": [ |
| 2966 | 32, |
| 2967 | 64 |
| 2968 | ], |
| 2969 | "loop_orders": [ |
| 2970 | [ |
| 2971 | 0, |
| 2972 | 1 |
| 2973 | ] |
| 2974 | ], |
| 2975 | "flatten_loops": [ |
| 2976 | true |
| 2977 | ], |
| 2978 | "l2_groupings": [ |
| 2979 | 1 |
| 2980 | ], |
| 2981 | "range_unroll_factors": [ |
| 2982 | 0 |
| 2983 | ], |
| 2984 | "range_warp_specializes": [], |
| 2985 | "range_num_stages": [ |
| 2986 | 0 |
| 2987 | ], |
| 2988 | "range_multi_buffers": [ |
| 2989 | null |
| 2990 | ], |
| 2991 | "range_flattens": [ |
| 2992 | null |
| 2993 | ], |
| 2994 | "load_eviction_policies": [ |
| 2995 | "", |
| 2996 | "", |
| 2997 | "" |
| 2998 | ], |
| 2999 | "num_warps": 8, |
| 3000 | "num_stages": 1, |
| 3001 | "indexing": [ |
| 3002 | "pointer", |
| 3003 | "pointer", |
| 3004 | "pointer", |
| 3005 | "pointer" |
| 3006 | ], |
| 3007 | "pid_type": "flat" |
| 3008 | } |
| 3009 | }, |
| 3010 | { |
| 3011 | "key": { |
| 3012 | "intermediate": 8192, |
| 3013 | "numtokens": 48 |
| 3014 | }, |
| 3015 | "config": { |
| 3016 | "block_sizes": [ |
| 3017 | 1, |
| 3018 | 4096 |
| 3019 | ], |
| 3020 | "loop_orders": [ |
| 3021 | [ |
| 3022 | 0, |
| 3023 | 1 |
| 3024 | ] |
| 3025 | ], |
| 3026 | "flatten_loops": [ |
| 3027 | true |
| 3028 | ], |
| 3029 | "l2_groupings": [ |
| 3030 | 2 |
| 3031 | ], |
| 3032 | "range_unroll_factors": [ |
| 3033 | 0 |
| 3034 | ], |
| 3035 | "range_warp_specializes": [], |
| 3036 | "range_num_stages": [ |
| 3037 | 0 |
| 3038 | ], |
| 3039 | "range_multi_buffers": [ |
| 3040 | null |
| 3041 | ], |
| 3042 | "range_flattens": [ |
| 3043 | null |
| 3044 | ], |
| 3045 | "load_eviction_policies": [ |
| 3046 | "first", |
| 3047 | "", |
| 3048 | "last" |
| 3049 | ], |
| 3050 | "num_warps": 4, |
| 3051 | "num_stages": 2, |
| 3052 | "indexing": [ |
| 3053 | "pointer", |
| 3054 | "tensor_descriptor", |
| 3055 | "tensor_descriptor", |
| 3056 | "pointer" |
| 3057 | ], |
| 3058 | "pid_type": "flat" |
| 3059 | } |
| 3060 | }, |
| 3061 | { |
| 3062 | "key": { |
| 3063 | "intermediate": 11008, |
| 3064 | "numtokens": 48 |
| 3065 | }, |
| 3066 | "config": { |
| 3067 | "block_sizes": [ |
| 3068 | 16, |
| 3069 | 256 |
| 3070 | ], |
| 3071 | "loop_orders": [ |
| 3072 | [ |
| 3073 | 0, |
| 3074 | 1 |
| 3075 | ] |
| 3076 | ], |
| 3077 | "flatten_loops": [ |
| 3078 | true |
| 3079 | ], |
| 3080 | "l2_groupings": [ |
| 3081 | 1 |
| 3082 | ], |
| 3083 | "range_unroll_factors": [ |
| 3084 | 0 |
| 3085 | ], |
| 3086 | "range_warp_specializes": [], |
| 3087 | "range_num_stages": [ |
| 3088 | 0 |
| 3089 | ], |
| 3090 | "range_multi_buffers": [ |
| 3091 | null |
| 3092 | ], |
| 3093 | "range_flattens": [ |
| 3094 | null |
| 3095 | ], |
| 3096 | "load_eviction_policies": [ |
| 3097 | "", |
| 3098 | "", |
| 3099 | "" |
| 3100 | ], |
| 3101 | "num_warps": 8, |
| 3102 | "num_stages": 1, |
| 3103 | "indexing": [ |
| 3104 | "pointer", |
| 3105 | "pointer", |
| 3106 | "pointer", |
| 3107 | "pointer" |
| 3108 | ], |
| 3109 | "pid_type": "flat" |
| 3110 | } |
| 3111 | }, |
| 3112 | { |
| 3113 | "key": { |
| 3114 | "intermediate": 14336, |
| 3115 | "numtokens": 48 |
| 3116 | }, |
| 3117 | "config": { |
| 3118 | "block_sizes": [ |
| 3119 | 32, |
| 3120 | 256 |
| 3121 | ], |
| 3122 | "loop_orders": [ |
| 3123 | [ |
| 3124 | 0, |
| 3125 | 1 |
| 3126 | ] |
| 3127 | ], |
| 3128 | "flatten_loops": [ |
| 3129 | true |
| 3130 | ], |
| 3131 | "l2_groupings": [ |
| 3132 | 1 |
| 3133 | ], |
| 3134 | "range_unroll_factors": [ |
| 3135 | 0 |
| 3136 | ], |
| 3137 | "range_warp_specializes": [], |
| 3138 | "range_num_stages": [ |
| 3139 | 0 |
| 3140 | ], |
| 3141 | "range_multi_buffers": [ |
| 3142 | null |
| 3143 | ], |
| 3144 | "range_flattens": [ |
| 3145 | null |
| 3146 | ], |
| 3147 | "load_eviction_policies": [ |
| 3148 | "", |
| 3149 | "", |
| 3150 | "" |
| 3151 | ], |
| 3152 | "num_warps": 8, |
| 3153 | "num_stages": 1, |
| 3154 | "indexing": [ |
| 3155 | "pointer", |
| 3156 | "pointer", |
| 3157 | "pointer", |
| 3158 | "pointer" |
| 3159 | ], |
| 3160 | "pid_type": "flat" |
| 3161 | } |
| 3162 | }, |
| 3163 | { |
| 3164 | "key": { |
| 3165 | "intermediate": 2048, |
| 3166 | "numtokens": 56 |
| 3167 | }, |
| 3168 | "config": { |
| 3169 | "block_sizes": [ |
| 3170 | 32, |
| 3171 | 32 |
| 3172 | ], |
| 3173 | "loop_orders": [ |
| 3174 | [ |
| 3175 | 0, |
| 3176 | 1 |
| 3177 | ] |
| 3178 | ], |
| 3179 | "flatten_loops": [ |
| 3180 | true |
| 3181 | ], |
| 3182 | "l2_groupings": [ |
| 3183 | 1 |
| 3184 | ], |
| 3185 | "range_unroll_factors": [ |
| 3186 | 0 |
| 3187 | ], |
| 3188 | "range_warp_specializes": [], |
| 3189 | "range_num_stages": [ |
| 3190 | 0 |
| 3191 | ], |
| 3192 | "range_multi_buffers": [ |
| 3193 | null |
| 3194 | ], |
| 3195 | "range_flattens": [ |
| 3196 | null |
| 3197 | ], |
| 3198 | "load_eviction_policies": [ |
| 3199 | "", |
| 3200 | "", |
| 3201 | "" |
| 3202 | ], |
| 3203 | "num_warps": 4, |
| 3204 | "num_stages": 1, |
| 3205 | "indexing": [ |
| 3206 | "pointer", |
| 3207 | "pointer", |
| 3208 | "pointer", |
| 3209 | "pointer" |
| 3210 | ], |
| 3211 | "pid_type": "flat" |
| 3212 | } |
| 3213 | }, |
| 3214 | { |
| 3215 | "key": { |
| 3216 | "intermediate": 2880, |
| 3217 | "numtokens": 56 |
| 3218 | }, |
| 3219 | "config": { |
| 3220 | "block_sizes": [ |
| 3221 | 1, |
| 3222 | 2048 |
| 3223 | ], |
| 3224 | "loop_orders": [ |
| 3225 | [ |
| 3226 | 0, |
| 3227 | 1 |
| 3228 | ] |
| 3229 | ], |
| 3230 | "flatten_loops": [ |
| 3231 | true |
| 3232 | ], |
| 3233 | "l2_groupings": [ |
| 3234 | 1 |
| 3235 | ], |
| 3236 | "range_unroll_factors": [ |
| 3237 | 0 |
| 3238 | ], |
| 3239 | "range_warp_specializes": [], |
| 3240 | "range_num_stages": [ |
| 3241 | 0 |
| 3242 | ], |
| 3243 | "range_multi_buffers": [ |
| 3244 | null |
| 3245 | ], |
| 3246 | "range_flattens": [ |
| 3247 | null |
| 3248 | ], |
| 3249 | "load_eviction_policies": [ |
| 3250 | "", |
| 3251 | "", |
| 3252 | "" |
| 3253 | ], |
| 3254 | "num_warps": 8, |
| 3255 | "num_stages": 1, |
| 3256 | "indexing": [ |
| 3257 | "pointer", |
| 3258 | "pointer", |
| 3259 | "pointer", |
| 3260 | "pointer" |
| 3261 | ], |
| 3262 | "pid_type": "flat" |
| 3263 | } |
| 3264 | }, |
| 3265 | { |
| 3266 | "key": { |
| 3267 | "intermediate": 4096, |
| 3268 | "numtokens": 56 |
| 3269 | }, |
| 3270 | "config": { |
| 3271 | "block_sizes": [ |
| 3272 | 32, |
| 3273 | 64 |
| 3274 | ], |
| 3275 | "loop_orders": [ |
| 3276 | [ |
| 3277 | 0, |
| 3278 | 1 |
| 3279 | ] |
| 3280 | ], |
| 3281 | "flatten_loops": [ |
| 3282 | true |
| 3283 | ], |
| 3284 | "l2_groupings": [ |
| 3285 | 1 |
| 3286 | ], |
| 3287 | "range_unroll_factors": [ |
| 3288 | 0 |
| 3289 | ], |
| 3290 | "range_warp_specializes": [], |
| 3291 | "range_num_stages": [ |
| 3292 | 0 |
| 3293 | ], |
| 3294 | "range_multi_buffers": [ |
| 3295 | null |
| 3296 | ], |
| 3297 | "range_flattens": [ |
| 3298 | null |
| 3299 | ], |
| 3300 | "load_eviction_policies": [ |
| 3301 | "", |
| 3302 | "", |
| 3303 | "" |
| 3304 | ], |
| 3305 | "num_warps": 8, |
| 3306 | "num_stages": 1, |
| 3307 | "indexing": [ |
| 3308 | "pointer", |
| 3309 | "pointer", |
| 3310 | "pointer", |
| 3311 | "pointer" |
| 3312 | ], |
| 3313 | "pid_type": "flat" |
| 3314 | } |
| 3315 | }, |
| 3316 | { |
| 3317 | "key": { |
| 3318 | "intermediate": 8192, |
| 3319 | "numtokens": 56 |
| 3320 | }, |
| 3321 | "config": { |
| 3322 | "block_sizes": [ |
| 3323 | 32, |
| 3324 | 128 |
| 3325 | ], |
| 3326 | "loop_orders": [ |
| 3327 | [ |
| 3328 | 0, |
| 3329 | 1 |
| 3330 | ] |
| 3331 | ], |
| 3332 | "flatten_loops": [ |
| 3333 | true |
| 3334 | ], |
| 3335 | "l2_groupings": [ |
| 3336 | 1 |
| 3337 | ], |
| 3338 | "range_unroll_factors": [ |
| 3339 | 0 |
| 3340 | ], |
| 3341 | "range_warp_specializes": [], |
| 3342 | "range_num_stages": [ |
| 3343 | 0 |
| 3344 | ], |
| 3345 | "range_multi_buffers": [ |
| 3346 | null |
| 3347 | ], |
| 3348 | "range_flattens": [ |
| 3349 | null |
| 3350 | ], |
| 3351 | "load_eviction_policies": [ |
| 3352 | "", |
| 3353 | "", |
| 3354 | "" |
| 3355 | ], |
| 3356 | "num_warps": 8, |
| 3357 | "num_stages": 1, |
| 3358 | "indexing": [ |
| 3359 | "pointer", |
| 3360 | "pointer", |
| 3361 | "pointer", |
| 3362 | "pointer" |
| 3363 | ], |
| 3364 | "pid_type": "flat" |
| 3365 | } |
| 3366 | }, |
| 3367 | { |
| 3368 | "key": { |
| 3369 | "intermediate": 11008, |
| 3370 | "numtokens": 56 |
| 3371 | }, |
| 3372 | "config": { |
| 3373 | "block_sizes": [ |
| 3374 | 1, |
| 3375 | 8192 |
| 3376 | ], |
| 3377 | "loop_orders": [ |
| 3378 | [ |
| 3379 | 0, |
| 3380 | 1 |
| 3381 | ] |
| 3382 | ], |
| 3383 | "flatten_loops": [ |
| 3384 | true |
| 3385 | ], |
| 3386 | "l2_groupings": [ |
| 3387 | 1 |
| 3388 | ], |
| 3389 | "range_unroll_factors": [ |
| 3390 | 0 |
| 3391 | ], |
| 3392 | "range_warp_specializes": [], |
| 3393 | "range_num_stages": [ |
| 3394 | 0 |
| 3395 | ], |
| 3396 | "range_multi_buffers": [ |
| 3397 | null |
| 3398 | ], |
| 3399 | "range_flattens": [ |
| 3400 | null |
| 3401 | ], |
| 3402 | "load_eviction_policies": [ |
| 3403 | "", |
| 3404 | "", |
| 3405 | "" |
| 3406 | ], |
| 3407 | "num_warps": 8, |
| 3408 | "num_stages": 1, |
| 3409 | "indexing": [ |
| 3410 | "pointer", |
| 3411 | "pointer", |
| 3412 | "pointer", |
| 3413 | "pointer" |
| 3414 | ], |
| 3415 | "pid_type": "flat" |
| 3416 | } |
| 3417 | }, |
| 3418 | { |
| 3419 | "key": { |
| 3420 | "intermediate": 14336, |
| 3421 | "numtokens": 56 |
| 3422 | }, |
| 3423 | "config": { |
| 3424 | "block_sizes": [ |
| 3425 | 2, |
| 3426 | 4096 |
| 3427 | ], |
| 3428 | "loop_orders": [ |
| 3429 | [ |
| 3430 | 0, |
| 3431 | 1 |
| 3432 | ] |
| 3433 | ], |
| 3434 | "flatten_loops": [ |
| 3435 | false |
| 3436 | ], |
| 3437 | "l2_groupings": [ |
| 3438 | 32 |
| 3439 | ], |
| 3440 | "range_unroll_factors": [ |
| 3441 | 0 |
| 3442 | ], |
| 3443 | "range_warp_specializes": [], |
| 3444 | "range_num_stages": [ |
| 3445 | 0 |
| 3446 | ], |
| 3447 | "range_multi_buffers": [ |
| 3448 | null |
| 3449 | ], |
| 3450 | "range_flattens": [ |
| 3451 | null |
| 3452 | ], |
| 3453 | "load_eviction_policies": [ |
| 3454 | "first", |
| 3455 | "first", |
| 3456 | "" |
| 3457 | ], |
| 3458 | "num_warps": 2, |
| 3459 | "num_stages": 4, |
| 3460 | "indexing": [ |
| 3461 | "tensor_descriptor", |
| 3462 | "pointer", |
| 3463 | "tensor_descriptor", |
| 3464 | "tensor_descriptor" |
| 3465 | ], |
| 3466 | "pid_type": "flat" |
| 3467 | } |
| 3468 | }, |
| 3469 | { |
| 3470 | "key": { |
| 3471 | "intermediate": 2048, |
| 3472 | "numtokens": 64 |
| 3473 | }, |
| 3474 | "config": { |
| 3475 | "block_sizes": [ |
| 3476 | 64, |
| 3477 | 32 |
| 3478 | ], |
| 3479 | "loop_orders": [ |
| 3480 | [ |
| 3481 | 0, |
| 3482 | 1 |
| 3483 | ] |
| 3484 | ], |
| 3485 | "flatten_loops": [ |
| 3486 | true |
| 3487 | ], |
| 3488 | "l2_groupings": [ |
| 3489 | 1 |
| 3490 | ], |
| 3491 | "range_unroll_factors": [ |
| 3492 | 0 |
| 3493 | ], |
| 3494 | "range_warp_specializes": [], |
| 3495 | "range_num_stages": [ |
| 3496 | 0 |
| 3497 | ], |
| 3498 | "range_multi_buffers": [ |
| 3499 | null |
| 3500 | ], |
| 3501 | "range_flattens": [ |
| 3502 | null |
| 3503 | ], |
| 3504 | "load_eviction_policies": [ |
| 3505 | "", |
| 3506 | "", |
| 3507 | "" |
| 3508 | ], |
| 3509 | "num_warps": 8, |
| 3510 | "num_stages": 1, |
| 3511 | "indexing": [ |
| 3512 | "pointer", |
| 3513 | "pointer", |
| 3514 | "pointer", |
| 3515 | "pointer" |
| 3516 | ], |
| 3517 | "pid_type": "flat" |
| 3518 | } |
| 3519 | }, |
| 3520 | { |
| 3521 | "key": { |
| 3522 | "intermediate": 2880, |
| 3523 | "numtokens": 64 |
| 3524 | }, |
| 3525 | "config": { |
| 3526 | "block_sizes": [ |
| 3527 | 1, |
| 3528 | 2048 |
| 3529 | ], |
| 3530 | "loop_orders": [ |
| 3531 | [ |
| 3532 | 0, |
| 3533 | 1 |
| 3534 | ] |
| 3535 | ], |
| 3536 | "flatten_loops": [ |
| 3537 | true |
| 3538 | ], |
| 3539 | "l2_groupings": [ |
| 3540 | 1 |
| 3541 | ], |
| 3542 | "range_unroll_factors": [ |
| 3543 | 0 |
| 3544 | ], |
| 3545 | "range_warp_specializes": [], |
| 3546 | "range_num_stages": [ |
| 3547 | 0 |
| 3548 | ], |
| 3549 | "range_multi_buffers": [ |
| 3550 | null |
| 3551 | ], |
| 3552 | "range_flattens": [ |
| 3553 | null |
| 3554 | ], |
| 3555 | "load_eviction_policies": [ |
| 3556 | "", |
| 3557 | "", |
| 3558 | "" |
| 3559 | ], |
| 3560 | "num_warps": 8, |
| 3561 | "num_stages": 1, |
| 3562 | "indexing": [ |
| 3563 | "pointer", |
| 3564 | "pointer", |
| 3565 | "pointer", |
| 3566 | "pointer" |
| 3567 | ], |
| 3568 | "pid_type": "flat" |
| 3569 | } |
| 3570 | }, |
| 3571 | { |
| 3572 | "key": { |
| 3573 | "intermediate": 4096, |
| 3574 | "numtokens": 64 |
| 3575 | }, |
| 3576 | "config": { |
| 3577 | "block_sizes": [ |
| 3578 | 64, |
| 3579 | 32 |
| 3580 | ], |
| 3581 | "loop_orders": [ |
| 3582 | [ |
| 3583 | 0, |
| 3584 | 1 |
| 3585 | ] |
| 3586 | ], |
| 3587 | "flatten_loops": [ |
| 3588 | true |
| 3589 | ], |
| 3590 | "l2_groupings": [ |
| 3591 | 1 |
| 3592 | ], |
| 3593 | "range_unroll_factors": [ |
| 3594 | 0 |
| 3595 | ], |
| 3596 | "range_warp_specializes": [], |
| 3597 | "range_num_stages": [ |
| 3598 | 0 |
| 3599 | ], |
| 3600 | "range_multi_buffers": [ |
| 3601 | null |
| 3602 | ], |
| 3603 | "range_flattens": [ |
| 3604 | null |
| 3605 | ], |
| 3606 | "load_eviction_policies": [ |
| 3607 | "", |
| 3608 | "", |
| 3609 | "" |
| 3610 | ], |
| 3611 | "num_warps": 8, |
| 3612 | "num_stages": 1, |
| 3613 | "indexing": [ |
| 3614 | "pointer", |
| 3615 | "pointer", |
| 3616 | "pointer", |
| 3617 | "pointer" |
| 3618 | ], |
| 3619 | "pid_type": "flat" |
| 3620 | } |
| 3621 | }, |
| 3622 | { |
| 3623 | "key": { |
| 3624 | "intermediate": 8192, |
| 3625 | "numtokens": 64 |
| 3626 | }, |
| 3627 | "config": { |
| 3628 | "block_sizes": [ |
| 3629 | 64, |
| 3630 | 64 |
| 3631 | ], |
| 3632 | "loop_orders": [ |
| 3633 | [ |
| 3634 | 0, |
| 3635 | 1 |
| 3636 | ] |
| 3637 | ], |
| 3638 | "flatten_loops": [ |
| 3639 | true |
| 3640 | ], |
| 3641 | "l2_groupings": [ |
| 3642 | 1 |
| 3643 | ], |
| 3644 | "range_unroll_factors": [ |
| 3645 | 0 |
| 3646 | ], |
| 3647 | "range_warp_specializes": [], |
| 3648 | "range_num_stages": [ |
| 3649 | 0 |
| 3650 | ], |
| 3651 | "range_multi_buffers": [ |
| 3652 | null |
| 3653 | ], |
| 3654 | "range_flattens": [ |
| 3655 | null |
| 3656 | ], |
| 3657 | "load_eviction_policies": [ |
| 3658 | "", |
| 3659 | "", |
| 3660 | "" |
| 3661 | ], |
| 3662 | "num_warps": 8, |
| 3663 | "num_stages": 1, |
| 3664 | "indexing": [ |
| 3665 | "pointer", |
| 3666 | "pointer", |
| 3667 | "pointer", |
| 3668 | "pointer" |
| 3669 | ], |
| 3670 | "pid_type": "flat" |
| 3671 | } |
| 3672 | }, |
| 3673 | { |
| 3674 | "key": { |
| 3675 | "intermediate": 11008, |
| 3676 | "numtokens": 64 |
| 3677 | }, |
| 3678 | "config": { |
| 3679 | "block_sizes": [ |
| 3680 | 1, |
| 3681 | 8192 |
| 3682 | ], |
| 3683 | "loop_orders": [ |
| 3684 | [ |
| 3685 | 0, |
| 3686 | 1 |
| 3687 | ] |
| 3688 | ], |
| 3689 | "flatten_loops": [ |
| 3690 | true |
| 3691 | ], |
| 3692 | "l2_groupings": [ |
| 3693 | 1 |
| 3694 | ], |
| 3695 | "range_unroll_factors": [ |
| 3696 | 0 |
| 3697 | ], |
| 3698 | "range_warp_specializes": [], |
| 3699 | "range_num_stages": [ |
| 3700 | 0 |
| 3701 | ], |
| 3702 | "range_multi_buffers": [ |
| 3703 | null |
| 3704 | ], |
| 3705 | "range_flattens": [ |
| 3706 | null |
| 3707 | ], |
| 3708 | "load_eviction_policies": [ |
| 3709 | "", |
| 3710 | "", |
| 3711 | "" |
| 3712 | ], |
| 3713 | "num_warps": 8, |
| 3714 | "num_stages": 1, |
| 3715 | "indexing": [ |
| 3716 | "pointer", |
| 3717 | "pointer", |
| 3718 | "pointer", |
| 3719 | "pointer" |
| 3720 | ], |
| 3721 | "pid_type": "flat" |
| 3722 | } |
| 3723 | }, |
| 3724 | { |
| 3725 | "key": { |
| 3726 | "intermediate": 14336, |
| 3727 | "numtokens": 64 |
| 3728 | }, |
| 3729 | "config": { |
| 3730 | "block_sizes": [ |
| 3731 | 16, |
| 3732 | 512 |
| 3733 | ], |
| 3734 | "loop_orders": [ |
| 3735 | [ |
| 3736 | 0, |
| 3737 | 1 |
| 3738 | ] |
| 3739 | ], |
| 3740 | "flatten_loops": [ |
| 3741 | true |
| 3742 | ], |
| 3743 | "l2_groupings": [ |
| 3744 | 16 |
| 3745 | ], |
| 3746 | "range_unroll_factors": [ |
| 3747 | 0 |
| 3748 | ], |
| 3749 | "range_warp_specializes": [], |
| 3750 | "range_num_stages": [ |
| 3751 | 0 |
| 3752 | ], |
| 3753 | "range_multi_buffers": [ |
| 3754 | null |
| 3755 | ], |
| 3756 | "range_flattens": [ |
| 3757 | null |
| 3758 | ], |
| 3759 | "load_eviction_policies": [ |
| 3760 | "", |
| 3761 | "first", |
| 3762 | "" |
| 3763 | ], |
| 3764 | "num_warps": 4, |
| 3765 | "num_stages": 1, |
| 3766 | "indexing": [ |
| 3767 | "pointer", |
| 3768 | "tensor_descriptor", |
| 3769 | "tensor_descriptor", |
| 3770 | "tensor_descriptor" |
| 3771 | ], |
| 3772 | "pid_type": "flat" |
| 3773 | } |
| 3774 | }, |
| 3775 | { |
| 3776 | "key": { |
| 3777 | "intermediate": 2048, |
| 3778 | "numtokens": 72 |
| 3779 | }, |
| 3780 | "config": { |
| 3781 | "block_sizes": [ |
| 3782 | 64, |
| 3783 | 32 |
| 3784 | ], |
| 3785 | "loop_orders": [ |
| 3786 | [ |
| 3787 | 0, |
| 3788 | 1 |
| 3789 | ] |
| 3790 | ], |
| 3791 | "flatten_loops": [ |
| 3792 | true |
| 3793 | ], |
| 3794 | "l2_groupings": [ |
| 3795 | 1 |
| 3796 | ], |
| 3797 | "range_unroll_factors": [ |
| 3798 | 0 |
| 3799 | ], |
| 3800 | "range_warp_specializes": [], |
| 3801 | "range_num_stages": [ |
| 3802 | 0 |
| 3803 | ], |
| 3804 | "range_multi_buffers": [ |
| 3805 | null |
| 3806 | ], |
| 3807 | "range_flattens": [ |
| 3808 | null |
| 3809 | ], |
| 3810 | "load_eviction_policies": [ |
| 3811 | "", |
| 3812 | "", |
| 3813 | "" |
| 3814 | ], |
| 3815 | "num_warps": 8, |
| 3816 | "num_stages": 1, |
| 3817 | "indexing": [ |
| 3818 | "pointer", |
| 3819 | "pointer", |
| 3820 | "pointer", |
| 3821 | "pointer" |
| 3822 | ], |
| 3823 | "pid_type": "flat" |
| 3824 | } |
| 3825 | }, |
| 3826 | { |
| 3827 | "key": { |
| 3828 | "intermediate": 2880, |
| 3829 | "numtokens": 72 |
| 3830 | }, |
| 3831 | "config": { |
| 3832 | "block_sizes": [ |
| 3833 | 32, |
| 3834 | 64 |
| 3835 | ], |
| 3836 | "loop_orders": [ |
| 3837 | [ |
| 3838 | 0, |
| 3839 | 1 |
| 3840 | ] |
| 3841 | ], |
| 3842 | "flatten_loops": [ |
| 3843 | true |
| 3844 | ], |
| 3845 | "l2_groupings": [ |
| 3846 | 1 |
| 3847 | ], |
| 3848 | "range_unroll_factors": [ |
| 3849 | 0 |
| 3850 | ], |
| 3851 | "range_warp_specializes": [], |
| 3852 | "range_num_stages": [ |
| 3853 | 0 |
| 3854 | ], |
| 3855 | "range_multi_buffers": [ |
| 3856 | null |
| 3857 | ], |
| 3858 | "range_flattens": [ |
| 3859 | null |
| 3860 | ], |
| 3861 | "load_eviction_policies": [ |
| 3862 | "", |
| 3863 | "", |
| 3864 | "" |
| 3865 | ], |
| 3866 | "num_warps": 8, |
| 3867 | "num_stages": 1, |
| 3868 | "indexing": [ |
| 3869 | "pointer", |
| 3870 | "pointer", |
| 3871 | "pointer", |
| 3872 | "pointer" |
| 3873 | ], |
| 3874 | "pid_type": "flat" |
| 3875 | } |
| 3876 | }, |
| 3877 | { |
| 3878 | "key": { |
| 3879 | "intermediate": 4096, |
| 3880 | "numtokens": 72 |
| 3881 | }, |
| 3882 | "config": { |
| 3883 | "block_sizes": [ |
| 3884 | 1, |
| 3885 | 1024 |
| 3886 | ], |
| 3887 | "loop_orders": [ |
| 3888 | [ |
| 3889 | 0, |
| 3890 | 1 |
| 3891 | ] |
| 3892 | ], |
| 3893 | "flatten_loops": [ |
| 3894 | true |
| 3895 | ], |
| 3896 | "l2_groupings": [ |
| 3897 | 4 |
| 3898 | ], |
| 3899 | "range_unroll_factors": [ |
| 3900 | 0 |
| 3901 | ], |
| 3902 | "range_warp_specializes": [], |
| 3903 | "range_num_stages": [ |
| 3904 | 0 |
| 3905 | ], |
| 3906 | "range_multi_buffers": [ |
| 3907 | null |
| 3908 | ], |
| 3909 | "range_flattens": [ |
| 3910 | null |
| 3911 | ], |
| 3912 | "load_eviction_policies": [ |
| 3913 | "", |
| 3914 | "", |
| 3915 | "first" |
| 3916 | ], |
| 3917 | "num_warps": 16, |
| 3918 | "num_stages": 2, |
| 3919 | "indexing": [ |
| 3920 | "tensor_descriptor", |
| 3921 | "pointer", |
| 3922 | "pointer", |
| 3923 | "tensor_descriptor" |
| 3924 | ], |
| 3925 | "pid_type": "flat" |
| 3926 | } |
| 3927 | }, |
| 3928 | { |
| 3929 | "key": { |
| 3930 | "intermediate": 8192, |
| 3931 | "numtokens": 72 |
| 3932 | }, |
| 3933 | "config": { |
| 3934 | "block_sizes": [ |
| 3935 | 64, |
| 3936 | 128 |
| 3937 | ], |
| 3938 | "loop_orders": [ |
| 3939 | [ |
| 3940 | 0, |
| 3941 | 1 |
| 3942 | ] |
| 3943 | ], |
| 3944 | "flatten_loops": [ |
| 3945 | true |
| 3946 | ], |
| 3947 | "l2_groupings": [ |
| 3948 | 1 |
| 3949 | ], |
| 3950 | "range_unroll_factors": [ |
| 3951 | 0 |
| 3952 | ], |
| 3953 | "range_warp_specializes": [], |
| 3954 | "range_num_stages": [ |
| 3955 | 0 |
| 3956 | ], |
| 3957 | "range_multi_buffers": [ |
| 3958 | null |
| 3959 | ], |
| 3960 | "range_flattens": [ |
| 3961 | null |
| 3962 | ], |
| 3963 | "load_eviction_policies": [ |
| 3964 | "", |
| 3965 | "", |
| 3966 | "" |
| 3967 | ], |
| 3968 | "num_warps": 8, |
| 3969 | "num_stages": 1, |
| 3970 | "indexing": [ |
| 3971 | "pointer", |
| 3972 | "pointer", |
| 3973 | "pointer", |
| 3974 | "pointer" |
| 3975 | ], |
| 3976 | "pid_type": "flat" |
| 3977 | } |
| 3978 | }, |
| 3979 | { |
| 3980 | "key": { |
| 3981 | "intermediate": 11008, |
| 3982 | "numtokens": 72 |
| 3983 | }, |
| 3984 | "config": { |
| 3985 | "block_sizes": [ |
| 3986 | 4, |
| 3987 | 256 |
| 3988 | ], |
| 3989 | "loop_orders": [ |
| 3990 | [ |
| 3991 | 0, |
| 3992 | 1 |
| 3993 | ] |
| 3994 | ], |
| 3995 | "flatten_loops": [ |
| 3996 | true |
| 3997 | ], |
| 3998 | "l2_groupings": [ |
| 3999 | 2 |
| 4000 | ], |
| 4001 | "range_unroll_factors": [ |
| 4002 | 0 |
| 4003 | ], |
| 4004 | "range_warp_specializes": [], |
| 4005 | "range_num_stages": [ |
| 4006 | 0 |
| 4007 | ], |
| 4008 | "range_multi_buffers": [ |
| 4009 | null |
| 4010 | ], |
| 4011 | "range_flattens": [ |
| 4012 | null |
| 4013 | ], |
| 4014 | "load_eviction_policies": [ |
| 4015 | "first", |
| 4016 | "last", |
| 4017 | "last" |
| 4018 | ], |
| 4019 | "num_warps": 32, |
| 4020 | "num_stages": 5, |
| 4021 | "indexing": [ |
| 4022 | "pointer", |
| 4023 | "pointer", |
| 4024 | "tensor_descriptor", |
| 4025 | "tensor_descriptor" |
| 4026 | ], |
| 4027 | "pid_type": "flat" |
| 4028 | } |
| 4029 | }, |
| 4030 | { |
| 4031 | "key": { |
| 4032 | "intermediate": 14336, |
| 4033 | "numtokens": 72 |
| 4034 | }, |
| 4035 | "config": { |
| 4036 | "block_sizes": [ |
| 4037 | 32, |
| 4038 | 32 |
| 4039 | ], |
| 4040 | "loop_orders": [ |
| 4041 | [ |
| 4042 | 0, |
| 4043 | 1 |
| 4044 | ] |
| 4045 | ], |
| 4046 | "flatten_loops": [ |
| 4047 | true |
| 4048 | ], |
| 4049 | "l2_groupings": [ |
| 4050 | 2 |
| 4051 | ], |
| 4052 | "range_unroll_factors": [ |
| 4053 | 0 |
| 4054 | ], |
| 4055 | "range_warp_specializes": [], |
| 4056 | "range_num_stages": [ |
| 4057 | 0 |
| 4058 | ], |
| 4059 | "range_multi_buffers": [ |
| 4060 | null |
| 4061 | ], |
| 4062 | "range_flattens": [ |
| 4063 | null |
| 4064 | ], |
| 4065 | "load_eviction_policies": [ |
| 4066 | "last", |
| 4067 | "", |
| 4068 | "last" |
| 4069 | ], |
| 4070 | "num_warps": 8, |
| 4071 | "num_stages": 3, |
| 4072 | "indexing": [ |
| 4073 | "pointer", |
| 4074 | "pointer", |
| 4075 | "tensor_descriptor", |
| 4076 | "pointer" |
| 4077 | ], |
| 4078 | "pid_type": "flat" |
| 4079 | } |
| 4080 | }, |
| 4081 | { |
| 4082 | "key": { |
| 4083 | "intermediate": 2048, |
| 4084 | "numtokens": 80 |
| 4085 | }, |
| 4086 | "config": { |
| 4087 | "block_sizes": [ |
| 4088 | 64, |
| 4089 | 32 |
| 4090 | ], |
| 4091 | "loop_orders": [ |
| 4092 | [ |
| 4093 | 0, |
| 4094 | 1 |
| 4095 | ] |
| 4096 | ], |
| 4097 | "flatten_loops": [ |
| 4098 | true |
| 4099 | ], |
| 4100 | "l2_groupings": [ |
| 4101 | 1 |
| 4102 | ], |
| 4103 | "range_unroll_factors": [ |
| 4104 | 0 |
| 4105 | ], |
| 4106 | "range_warp_specializes": [], |
| 4107 | "range_num_stages": [ |
| 4108 | 0 |
| 4109 | ], |
| 4110 | "range_multi_buffers": [ |
| 4111 | null |
| 4112 | ], |
| 4113 | "range_flattens": [ |
| 4114 | null |
| 4115 | ], |
| 4116 | "load_eviction_policies": [ |
| 4117 | "", |
| 4118 | "", |
| 4119 | "" |
| 4120 | ], |
| 4121 | "num_warps": 8, |
| 4122 | "num_stages": 1, |
| 4123 | "indexing": [ |
| 4124 | "pointer", |
| 4125 | "pointer", |
| 4126 | "pointer", |
| 4127 | "pointer" |
| 4128 | ], |
| 4129 | "pid_type": "flat" |
| 4130 | } |
| 4131 | }, |
| 4132 | { |
| 4133 | "key": { |
| 4134 | "intermediate": 2880, |
| 4135 | "numtokens": 80 |
| 4136 | }, |
| 4137 | "config": { |
| 4138 | "block_sizes": [ |
| 4139 | 32, |
| 4140 | 64 |
| 4141 | ], |
| 4142 | "loop_orders": [ |
| 4143 | [ |
| 4144 | 0, |
| 4145 | 1 |
| 4146 | ] |
| 4147 | ], |
| 4148 | "flatten_loops": [ |
| 4149 | true |
| 4150 | ], |
| 4151 | "l2_groupings": [ |
| 4152 | 1 |
| 4153 | ], |
| 4154 | "range_unroll_factors": [ |
| 4155 | 0 |
| 4156 | ], |
| 4157 | "range_warp_specializes": [], |
| 4158 | "range_num_stages": [ |
| 4159 | 0 |
| 4160 | ], |
| 4161 | "range_multi_buffers": [ |
| 4162 | null |
| 4163 | ], |
| 4164 | "range_flattens": [ |
| 4165 | null |
| 4166 | ], |
| 4167 | "load_eviction_policies": [ |
| 4168 | "", |
| 4169 | "", |
| 4170 | "" |
| 4171 | ], |
| 4172 | "num_warps": 8, |
| 4173 | "num_stages": 1, |
| 4174 | "indexing": [ |
| 4175 | "pointer", |
| 4176 | "pointer", |
| 4177 | "pointer", |
| 4178 | "pointer" |
| 4179 | ], |
| 4180 | "pid_type": "flat" |
| 4181 | } |
| 4182 | }, |
| 4183 | { |
| 4184 | "key": { |
| 4185 | "intermediate": 4096, |
| 4186 | "numtokens": 80 |
| 4187 | }, |
| 4188 | "config": { |
| 4189 | "block_sizes": [ |
| 4190 | 64, |
| 4191 | 64 |
| 4192 | ], |
| 4193 | "loop_orders": [ |
| 4194 | [ |
| 4195 | 0, |
| 4196 | 1 |
| 4197 | ] |
| 4198 | ], |
| 4199 | "flatten_loops": [ |
| 4200 | true |
| 4201 | ], |
| 4202 | "l2_groupings": [ |
| 4203 | 1 |
| 4204 | ], |
| 4205 | "range_unroll_factors": [ |
| 4206 | 0 |
| 4207 | ], |
| 4208 | "range_warp_specializes": [], |
| 4209 | "range_num_stages": [ |
| 4210 | 0 |
| 4211 | ], |
| 4212 | "range_multi_buffers": [ |
| 4213 | null |
| 4214 | ], |
| 4215 | "range_flattens": [ |
| 4216 | null |
| 4217 | ], |
| 4218 | "load_eviction_policies": [ |
| 4219 | "", |
| 4220 | "", |
| 4221 | "" |
| 4222 | ], |
| 4223 | "num_warps": 8, |
| 4224 | "num_stages": 1, |
| 4225 | "indexing": [ |
| 4226 | "pointer", |
| 4227 | "pointer", |
| 4228 | "pointer", |
| 4229 | "pointer" |
| 4230 | ], |
| 4231 | "pid_type": "flat" |
| 4232 | } |
| 4233 | }, |
| 4234 | { |
| 4235 | "key": { |
| 4236 | "intermediate": 8192, |
| 4237 | "numtokens": 80 |
| 4238 | }, |
| 4239 | "config": { |
| 4240 | "block_sizes": [ |
| 4241 | 4, |
| 4242 | 128 |
| 4243 | ], |
| 4244 | "loop_orders": [ |
| 4245 | [ |
| 4246 | 0, |
| 4247 | 1 |
| 4248 | ] |
| 4249 | ], |
| 4250 | "flatten_loops": [ |
| 4251 | true |
| 4252 | ], |
| 4253 | "l2_groupings": [ |
| 4254 | 4 |
| 4255 | ], |
| 4256 | "range_unroll_factors": [ |
| 4257 | 0 |
| 4258 | ], |
| 4259 | "range_warp_specializes": [], |
| 4260 | "range_num_stages": [ |
| 4261 | 0 |
| 4262 | ], |
| 4263 | "range_multi_buffers": [ |
| 4264 | null |
| 4265 | ], |
| 4266 | "range_flattens": [ |
| 4267 | null |
| 4268 | ], |
| 4269 | "load_eviction_policies": [ |
| 4270 | "last", |
| 4271 | "last", |
| 4272 | "last" |
| 4273 | ], |
| 4274 | "num_warps": 16, |
| 4275 | "num_stages": 2, |
| 4276 | "indexing": [ |
| 4277 | "pointer", |
| 4278 | "pointer", |
| 4279 | "pointer", |
| 4280 | "pointer" |
| 4281 | ], |
| 4282 | "pid_type": "flat" |
| 4283 | } |
| 4284 | }, |
| 4285 | { |
| 4286 | "key": { |
| 4287 | "intermediate": 11008, |
| 4288 | "numtokens": 80 |
| 4289 | }, |
| 4290 | "config": { |
| 4291 | "block_sizes": [ |
| 4292 | 1, |
| 4293 | 4096 |
| 4294 | ], |
| 4295 | "loop_orders": [ |
| 4296 | [ |
| 4297 | 0, |
| 4298 | 1 |
| 4299 | ] |
| 4300 | ], |
| 4301 | "flatten_loops": [ |
| 4302 | true |
| 4303 | ], |
| 4304 | "l2_groupings": [ |
| 4305 | 32 |
| 4306 | ], |
| 4307 | "range_unroll_factors": [ |
| 4308 | 0 |
| 4309 | ], |
| 4310 | "range_warp_specializes": [], |
| 4311 | "range_num_stages": [ |
| 4312 | 0 |
| 4313 | ], |
| 4314 | "range_multi_buffers": [ |
| 4315 | null |
| 4316 | ], |
| 4317 | "range_flattens": [ |
| 4318 | null |
| 4319 | ], |
| 4320 | "load_eviction_policies": [ |
| 4321 | "last", |
| 4322 | "last", |
| 4323 | "" |
| 4324 | ], |
| 4325 | "num_warps": 16, |
| 4326 | "num_stages": 2, |
| 4327 | "indexing": [ |
| 4328 | "pointer", |
| 4329 | "tensor_descriptor", |
| 4330 | "tensor_descriptor", |
| 4331 | "tensor_descriptor" |
| 4332 | ], |
| 4333 | "pid_type": "flat" |
| 4334 | } |
| 4335 | }, |
| 4336 | { |
| 4337 | "key": { |
| 4338 | "intermediate": 14336, |
| 4339 | "numtokens": 80 |
| 4340 | }, |
| 4341 | "config": { |
| 4342 | "block_sizes": [ |
| 4343 | 2, |
| 4344 | 2048 |
| 4345 | ], |
| 4346 | "loop_orders": [ |
| 4347 | [ |
| 4348 | 0, |
| 4349 | 1 |
| 4350 | ] |
| 4351 | ], |
| 4352 | "flatten_loops": [ |
| 4353 | true |
| 4354 | ], |
| 4355 | "l2_groupings": [ |
| 4356 | 4 |
| 4357 | ], |
| 4358 | "range_unroll_factors": [ |
| 4359 | 0 |
| 4360 | ], |
| 4361 | "range_warp_specializes": [], |
| 4362 | "range_num_stages": [ |
| 4363 | 0 |
| 4364 | ], |
| 4365 | "range_multi_buffers": [ |
| 4366 | null |
| 4367 | ], |
| 4368 | "range_flattens": [ |
| 4369 | null |
| 4370 | ], |
| 4371 | "load_eviction_policies": [ |
| 4372 | "last", |
| 4373 | "last", |
| 4374 | "" |
| 4375 | ], |
| 4376 | "num_warps": 4, |
| 4377 | "num_stages": 6, |
| 4378 | "indexing": [ |
| 4379 | "tensor_descriptor", |
| 4380 | "pointer", |
| 4381 | "pointer", |
| 4382 | "tensor_descriptor" |
| 4383 | ], |
| 4384 | "pid_type": "flat" |
| 4385 | } |
| 4386 | }, |
| 4387 | { |
| 4388 | "key": { |
| 4389 | "intermediate": 2048, |
| 4390 | "numtokens": 88 |
| 4391 | }, |
| 4392 | "config": { |
| 4393 | "block_sizes": [ |
| 4394 | 64, |
| 4395 | 32 |
| 4396 | ], |
| 4397 | "loop_orders": [ |
| 4398 | [ |
| 4399 | 0, |
| 4400 | 1 |
| 4401 | ] |
| 4402 | ], |
| 4403 | "flatten_loops": [ |
| 4404 | true |
| 4405 | ], |
| 4406 | "l2_groupings": [ |
| 4407 | 1 |
| 4408 | ], |
| 4409 | "range_unroll_factors": [ |
| 4410 | 0 |
| 4411 | ], |
| 4412 | "range_warp_specializes": [], |
| 4413 | "range_num_stages": [ |
| 4414 | 0 |
| 4415 | ], |
| 4416 | "range_multi_buffers": [ |
| 4417 | null |
| 4418 | ], |
| 4419 | "range_flattens": [ |
| 4420 | null |
| 4421 | ], |
| 4422 | "load_eviction_policies": [ |
| 4423 | "", |
| 4424 | "", |
| 4425 | "" |
| 4426 | ], |
| 4427 | "num_warps": 8, |
| 4428 | "num_stages": 1, |
| 4429 | "indexing": [ |
| 4430 | "pointer", |
| 4431 | "pointer", |
| 4432 | "pointer", |
| 4433 | "pointer" |
| 4434 | ], |
| 4435 | "pid_type": "flat" |
| 4436 | } |
| 4437 | }, |
| 4438 | { |
| 4439 | "key": { |
| 4440 | "intermediate": 2880, |
| 4441 | "numtokens": 88 |
| 4442 | }, |
| 4443 | "config": { |
| 4444 | "block_sizes": [ |
| 4445 | 8, |
| 4446 | 256 |
| 4447 | ], |
| 4448 | "loop_orders": [ |
| 4449 | [ |
| 4450 | 0, |
| 4451 | 1 |
| 4452 | ] |
| 4453 | ], |
| 4454 | "flatten_loops": [ |
| 4455 | true |
| 4456 | ], |
| 4457 | "l2_groupings": [ |
| 4458 | 1 |
| 4459 | ], |
| 4460 | "range_unroll_factors": [ |
| 4461 | 0 |
| 4462 | ], |
| 4463 | "range_warp_specializes": [], |
| 4464 | "range_num_stages": [ |
| 4465 | 0 |
| 4466 | ], |
| 4467 | "range_multi_buffers": [ |
| 4468 | null |
| 4469 | ], |
| 4470 | "range_flattens": [ |
| 4471 | null |
| 4472 | ], |
| 4473 | "load_eviction_policies": [ |
| 4474 | "", |
| 4475 | "", |
| 4476 | "" |
| 4477 | ], |
| 4478 | "num_warps": 8, |
| 4479 | "num_stages": 1, |
| 4480 | "indexing": [ |
| 4481 | "pointer", |
| 4482 | "pointer", |
| 4483 | "pointer", |
| 4484 | "pointer" |
| 4485 | ], |
| 4486 | "pid_type": "flat" |
| 4487 | } |
| 4488 | }, |
| 4489 | { |
| 4490 | "key": { |
| 4491 | "intermediate": 4096, |
| 4492 | "numtokens": 88 |
| 4493 | }, |
| 4494 | "config": { |
| 4495 | "block_sizes": [ |
| 4496 | 64, |
| 4497 | 64 |
| 4498 | ], |
| 4499 | "loop_orders": [ |
| 4500 | [ |
| 4501 | 0, |
| 4502 | 1 |
| 4503 | ] |
| 4504 | ], |
| 4505 | "flatten_loops": [ |
| 4506 | true |
| 4507 | ], |
| 4508 | "l2_groupings": [ |
| 4509 | 1 |
| 4510 | ], |
| 4511 | "range_unroll_factors": [ |
| 4512 | 0 |
| 4513 | ], |
| 4514 | "range_warp_specializes": [], |
| 4515 | "range_num_stages": [ |
| 4516 | 0 |
| 4517 | ], |
| 4518 | "range_multi_buffers": [ |
| 4519 | null |
| 4520 | ], |
| 4521 | "range_flattens": [ |
| 4522 | null |
| 4523 | ], |
| 4524 | "load_eviction_policies": [ |
| 4525 | "", |
| 4526 | "", |
| 4527 | "" |
| 4528 | ], |
| 4529 | "num_warps": 8, |
| 4530 | "num_stages": 1, |
| 4531 | "indexing": [ |
| 4532 | "pointer", |
| 4533 | "pointer", |
| 4534 | "pointer", |
| 4535 | "pointer" |
| 4536 | ], |
| 4537 | "pid_type": "flat" |
| 4538 | } |
| 4539 | }, |
| 4540 | { |
| 4541 | "key": { |
| 4542 | "intermediate": 8192, |
| 4543 | "numtokens": 88 |
| 4544 | }, |
| 4545 | "config": { |
| 4546 | "block_sizes": [ |
| 4547 | 64, |
| 4548 | 128 |
| 4549 | ], |
| 4550 | "loop_orders": [ |
| 4551 | [ |
| 4552 | 0, |
| 4553 | 1 |
| 4554 | ] |
| 4555 | ], |
| 4556 | "flatten_loops": [ |
| 4557 | true |
| 4558 | ], |
| 4559 | "l2_groupings": [ |
| 4560 | 1 |
| 4561 | ], |
| 4562 | "range_unroll_factors": [ |
| 4563 | 0 |
| 4564 | ], |
| 4565 | "range_warp_specializes": [], |
| 4566 | "range_num_stages": [ |
| 4567 | 0 |
| 4568 | ], |
| 4569 | "range_multi_buffers": [ |
| 4570 | null |
| 4571 | ], |
| 4572 | "range_flattens": [ |
| 4573 | null |
| 4574 | ], |
| 4575 | "load_eviction_policies": [ |
| 4576 | "", |
| 4577 | "", |
| 4578 | "" |
| 4579 | ], |
| 4580 | "num_warps": 8, |
| 4581 | "num_stages": 1, |
| 4582 | "indexing": [ |
| 4583 | "pointer", |
| 4584 | "pointer", |
| 4585 | "pointer", |
| 4586 | "pointer" |
| 4587 | ], |
| 4588 | "pid_type": "flat" |
| 4589 | } |
| 4590 | }, |
| 4591 | { |
| 4592 | "key": { |
| 4593 | "intermediate": 11008, |
| 4594 | "numtokens": 88 |
| 4595 | }, |
| 4596 | "config": { |
| 4597 | "block_sizes": [ |
| 4598 | 16, |
| 4599 | 2048 |
| 4600 | ], |
| 4601 | "loop_orders": [ |
| 4602 | [ |
| 4603 | 0, |
| 4604 | 1 |
| 4605 | ] |
| 4606 | ], |
| 4607 | "flatten_loops": [ |
| 4608 | true |
| 4609 | ], |
| 4610 | "l2_groupings": [ |
| 4611 | 2 |
| 4612 | ], |
| 4613 | "range_unroll_factors": [ |
| 4614 | 0 |
| 4615 | ], |
| 4616 | "range_warp_specializes": [], |
| 4617 | "range_num_stages": [ |
| 4618 | 0 |
| 4619 | ], |
| 4620 | "range_multi_buffers": [ |
| 4621 | null |
| 4622 | ], |
| 4623 | "range_flattens": [ |
| 4624 | null |
| 4625 | ], |
| 4626 | "load_eviction_policies": [ |
| 4627 | "first", |
| 4628 | "first", |
| 4629 | "" |
| 4630 | ], |
| 4631 | "num_warps": 32, |
| 4632 | "num_stages": 2, |
| 4633 | "indexing": [ |
| 4634 | "tensor_descriptor", |
| 4635 | "tensor_descriptor", |
| 4636 | "pointer", |
| 4637 | "pointer" |
| 4638 | ], |
| 4639 | "pid_type": "flat" |
| 4640 | } |
| 4641 | }, |
| 4642 | { |
| 4643 | "key": { |
| 4644 | "intermediate": 14336, |
| 4645 | "numtokens": 88 |
| 4646 | }, |
| 4647 | "config": { |
| 4648 | "block_sizes": [ |
| 4649 | 4, |
| 4650 | 512 |
| 4651 | ], |
| 4652 | "loop_orders": [ |
| 4653 | [ |
| 4654 | 0, |
| 4655 | 1 |
| 4656 | ] |
| 4657 | ], |
| 4658 | "flatten_loops": [ |
| 4659 | true |
| 4660 | ], |
| 4661 | "l2_groupings": [ |
| 4662 | 16 |
| 4663 | ], |
| 4664 | "range_unroll_factors": [ |
| 4665 | 0 |
| 4666 | ], |
| 4667 | "range_warp_specializes": [], |
| 4668 | "range_num_stages": [ |
| 4669 | 0 |
| 4670 | ], |
| 4671 | "range_multi_buffers": [ |
| 4672 | null |
| 4673 | ], |
| 4674 | "range_flattens": [ |
| 4675 | null |
| 4676 | ], |
| 4677 | "load_eviction_policies": [ |
| 4678 | "", |
| 4679 | "", |
| 4680 | "" |
| 4681 | ], |
| 4682 | "num_warps": 32, |
| 4683 | "num_stages": 5, |
| 4684 | "indexing": [ |
| 4685 | "pointer", |
| 4686 | "pointer", |
| 4687 | "pointer", |
| 4688 | "pointer" |
| 4689 | ], |
| 4690 | "pid_type": "flat" |
| 4691 | } |
| 4692 | }, |
| 4693 | { |
| 4694 | "key": { |
| 4695 | "intermediate": 2048, |
| 4696 | "numtokens": 96 |
| 4697 | }, |
| 4698 | "config": { |
| 4699 | "block_sizes": [ |
| 4700 | 64, |
| 4701 | 32 |
| 4702 | ], |
| 4703 | "loop_orders": [ |
| 4704 | [ |
| 4705 | 0, |
| 4706 | 1 |
| 4707 | ] |
| 4708 | ], |
| 4709 | "flatten_loops": [ |
| 4710 | true |
| 4711 | ], |
| 4712 | "l2_groupings": [ |
| 4713 | 1 |
| 4714 | ], |
| 4715 | "range_unroll_factors": [ |
| 4716 | 0 |
| 4717 | ], |
| 4718 | "range_warp_specializes": [], |
| 4719 | "range_num_stages": [ |
| 4720 | 0 |
| 4721 | ], |
| 4722 | "range_multi_buffers": [ |
| 4723 | null |
| 4724 | ], |
| 4725 | "range_flattens": [ |
| 4726 | null |
| 4727 | ], |
| 4728 | "load_eviction_policies": [ |
| 4729 | "", |
| 4730 | "", |
| 4731 | "" |
| 4732 | ], |
| 4733 | "num_warps": 8, |
| 4734 | "num_stages": 1, |
| 4735 | "indexing": [ |
| 4736 | "pointer", |
| 4737 | "pointer", |
| 4738 | "pointer", |
| 4739 | "pointer" |
| 4740 | ], |
| 4741 | "pid_type": "flat" |
| 4742 | } |
| 4743 | }, |
| 4744 | { |
| 4745 | "key": { |
| 4746 | "intermediate": 2880, |
| 4747 | "numtokens": 96 |
| 4748 | }, |
| 4749 | "config": { |
| 4750 | "block_sizes": [ |
| 4751 | 32, |
| 4752 | 64 |
| 4753 | ], |
| 4754 | "loop_orders": [ |
| 4755 | [ |
| 4756 | 0, |
| 4757 | 1 |
| 4758 | ] |
| 4759 | ], |
| 4760 | "flatten_loops": [ |
| 4761 | true |
| 4762 | ], |
| 4763 | "l2_groupings": [ |
| 4764 | 1 |
| 4765 | ], |
| 4766 | "range_unroll_factors": [ |
| 4767 | 0 |
| 4768 | ], |
| 4769 | "range_warp_specializes": [], |
| 4770 | "range_num_stages": [ |
| 4771 | 0 |
| 4772 | ], |
| 4773 | "range_multi_buffers": [ |
| 4774 | null |
| 4775 | ], |
| 4776 | "range_flattens": [ |
| 4777 | null |
| 4778 | ], |
| 4779 | "load_eviction_policies": [ |
| 4780 | "", |
| 4781 | "", |
| 4782 | "" |
| 4783 | ], |
| 4784 | "num_warps": 8, |
| 4785 | "num_stages": 1, |
| 4786 | "indexing": [ |
| 4787 | "pointer", |
| 4788 | "pointer", |
| 4789 | "pointer", |
| 4790 | "pointer" |
| 4791 | ], |
| 4792 | "pid_type": "flat" |
| 4793 | } |
| 4794 | }, |
| 4795 | { |
| 4796 | "key": { |
| 4797 | "intermediate": 4096, |
| 4798 | "numtokens": 96 |
| 4799 | }, |
| 4800 | "config": { |
| 4801 | "block_sizes": [ |
| 4802 | 1, |
| 4803 | 4096 |
| 4804 | ], |
| 4805 | "loop_orders": [ |
| 4806 | [ |
| 4807 | 1, |
| 4808 | 0 |
| 4809 | ] |
| 4810 | ], |
| 4811 | "flatten_loops": [ |
| 4812 | false |
| 4813 | ], |
| 4814 | "l2_groupings": [ |
| 4815 | 64 |
| 4816 | ], |
| 4817 | "range_unroll_factors": [ |
| 4818 | 0 |
| 4819 | ], |
| 4820 | "range_warp_specializes": [], |
| 4821 | "range_num_stages": [ |
| 4822 | 0 |
| 4823 | ], |
| 4824 | "range_multi_buffers": [ |
| 4825 | null |
| 4826 | ], |
| 4827 | "range_flattens": [ |
| 4828 | null |
| 4829 | ], |
| 4830 | "load_eviction_policies": [ |
| 4831 | "first", |
| 4832 | "last", |
| 4833 | "last" |
| 4834 | ], |
| 4835 | "num_warps": 4, |
| 4836 | "num_stages": 6, |
| 4837 | "indexing": [ |
| 4838 | "pointer", |
| 4839 | "pointer", |
| 4840 | "tensor_descriptor", |
| 4841 | "pointer" |
| 4842 | ], |
| 4843 | "pid_type": "flat" |
| 4844 | } |
| 4845 | }, |
| 4846 | { |
| 4847 | "key": { |
| 4848 | "intermediate": 8192, |
| 4849 | "numtokens": 96 |
| 4850 | }, |
| 4851 | "config": { |
| 4852 | "block_sizes": [ |
| 4853 | 64, |
| 4854 | 128 |
| 4855 | ], |
| 4856 | "loop_orders": [ |
| 4857 | [ |
| 4858 | 0, |
| 4859 | 1 |
| 4860 | ] |
| 4861 | ], |
| 4862 | "flatten_loops": [ |
| 4863 | true |
| 4864 | ], |
| 4865 | "l2_groupings": [ |
| 4866 | 1 |
| 4867 | ], |
| 4868 | "range_unroll_factors": [ |
| 4869 | 0 |
| 4870 | ], |
| 4871 | "range_warp_specializes": [], |
| 4872 | "range_num_stages": [ |
| 4873 | 0 |
| 4874 | ], |
| 4875 | "range_multi_buffers": [ |
| 4876 | null |
| 4877 | ], |
| 4878 | "range_flattens": [ |
| 4879 | null |
| 4880 | ], |
| 4881 | "load_eviction_policies": [ |
| 4882 | "", |
| 4883 | "", |
| 4884 | "" |
| 4885 | ], |
| 4886 | "num_warps": 8, |
| 4887 | "num_stages": 1, |
| 4888 | "indexing": [ |
| 4889 | "pointer", |
| 4890 | "pointer", |
| 4891 | "pointer", |
| 4892 | "pointer" |
| 4893 | ], |
| 4894 | "pid_type": "flat" |
| 4895 | } |
| 4896 | }, |
| 4897 | { |
| 4898 | "key": { |
| 4899 | "intermediate": 11008, |
| 4900 | "numtokens": 96 |
| 4901 | }, |
| 4902 | "config": { |
| 4903 | "block_sizes": [ |
| 4904 | 1, |
| 4905 | 2048 |
| 4906 | ], |
| 4907 | "loop_orders": [ |
| 4908 | [ |
| 4909 | 1, |
| 4910 | 0 |
| 4911 | ] |
| 4912 | ], |
| 4913 | "flatten_loops": [ |
| 4914 | false |
| 4915 | ], |
| 4916 | "l2_groupings": [ |
| 4917 | 64 |
| 4918 | ], |
| 4919 | "range_unroll_factors": [ |
| 4920 | 0 |
| 4921 | ], |
| 4922 | "range_warp_specializes": [], |
| 4923 | "range_num_stages": [ |
| 4924 | 0 |
| 4925 | ], |
| 4926 | "range_multi_buffers": [ |
| 4927 | null |
| 4928 | ], |
| 4929 | "range_flattens": [ |
| 4930 | null |
| 4931 | ], |
| 4932 | "load_eviction_policies": [ |
| 4933 | "first", |
| 4934 | "", |
| 4935 | "" |
| 4936 | ], |
| 4937 | "num_warps": 32, |
| 4938 | "num_stages": 2, |
| 4939 | "indexing": [ |
| 4940 | "tensor_descriptor", |
| 4941 | "pointer", |
| 4942 | "pointer", |
| 4943 | "tensor_descriptor" |
| 4944 | ], |
| 4945 | "pid_type": "flat" |
| 4946 | } |
| 4947 | }, |
| 4948 | { |
| 4949 | "key": { |
| 4950 | "intermediate": 14336, |
| 4951 | "numtokens": 96 |
| 4952 | }, |
| 4953 | "config": { |
| 4954 | "block_sizes": [ |
| 4955 | 4, |
| 4956 | 4096 |
| 4957 | ], |
| 4958 | "loop_orders": [ |
| 4959 | [ |
| 4960 | 0, |
| 4961 | 1 |
| 4962 | ] |
| 4963 | ], |
| 4964 | "flatten_loops": [ |
| 4965 | false |
| 4966 | ], |
| 4967 | "l2_groupings": [ |
| 4968 | 32 |
| 4969 | ], |
| 4970 | "range_unroll_factors": [ |
| 4971 | 0 |
| 4972 | ], |
| 4973 | "range_warp_specializes": [], |
| 4974 | "range_num_stages": [ |
| 4975 | 0 |
| 4976 | ], |
| 4977 | "range_multi_buffers": [ |
| 4978 | null |
| 4979 | ], |
| 4980 | "range_flattens": [ |
| 4981 | null |
| 4982 | ], |
| 4983 | "load_eviction_policies": [ |
| 4984 | "last", |
| 4985 | "first", |
| 4986 | "" |
| 4987 | ], |
| 4988 | "num_warps": 4, |
| 4989 | "num_stages": 1, |
| 4990 | "indexing": [ |
| 4991 | "tensor_descriptor", |
| 4992 | "tensor_descriptor", |
| 4993 | "pointer", |
| 4994 | "tensor_descriptor" |
| 4995 | ], |
| 4996 | "pid_type": "flat" |
| 4997 | } |
| 4998 | }, |
| 4999 | { |
| 5000 | "key": { |
Showing first 5,000 of 15,711 lines.
View raw