1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315
|
//===- LowerVectorContract.cpp - Lower 'vector.contract' operation --------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements target-independent rewrites and utilities to lower the
// 'vector.contract' operation.
//
//===----------------------------------------------------------------------===//
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Arith/Utils/Utils.h"
#include "mlir/Dialect/Linalg/IR/Linalg.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/SCF/IR/SCF.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/Dialect/Utils/IndexingUtils.h"
#include "mlir/Dialect/Utils/StructuredOpsUtils.h"
#include "mlir/Dialect/Vector/IR/VectorOps.h"
#include "mlir/Dialect/Vector/Transforms/LoweringPatterns.h"
#include "mlir/Dialect/Vector/Utils/VectorUtils.h"
#include "mlir/IR/BuiltinAttributeInterfaces.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/ImplicitLocOpBuilder.h"
#include "mlir/IR/Location.h"
#include "mlir/IR/Matchers.h"
#include "mlir/IR/PatternMatch.h"
#include "mlir/IR/TypeUtilities.h"
#include "mlir/Interfaces/VectorInterfaces.h"
#include "mlir/Support/LogicalResult.h"
#define DEBUG_TYPE "vector-contract-lowering"
using namespace mlir;
using namespace mlir::vector;
//===----------------------------------------------------------------------===//
// Helper functions
//===----------------------------------------------------------------------===//
// Helper to find an index in an affine map.
static std::optional<int64_t> getResultIndex(AffineMap map, int64_t index) {
for (int64_t i = 0, e = map.getNumResults(); i < e; ++i) {
int64_t idx = map.getDimPosition(i);
if (idx == index)
return i;
}
return std::nullopt;
}
// Helper to construct iterator types with one index removed.
static SmallVector<Attribute> adjustIter(ArrayAttr iteratorTypes,
int64_t index) {
SmallVector<Attribute> results;
for (const auto &it : llvm::enumerate(iteratorTypes)) {
int64_t idx = it.index();
if (idx == index)
continue;
results.push_back(it.value());
}
return results;
}
// Helper to construct an affine map with one index removed.
static AffineMap adjustMap(AffineMap map, int64_t index,
PatternRewriter &rewriter) {
auto *ctx = rewriter.getContext();
SmallVector<AffineExpr> results;
for (int64_t i = 0, e = map.getNumResults(); i < e; ++i) {
int64_t idx = map.getDimPosition(i);
if (idx == index)
continue;
// Re-insert remaining indices, but renamed when occurring
// after the removed index.
auto targetExpr = getAffineDimExpr(idx < index ? idx : idx - 1, ctx);
results.push_back(targetExpr);
}
return AffineMap::get(map.getNumDims() - 1, 0, results, ctx);
}
// Helper method to possibly drop a dimension in a load.
// TODO
static Value reshapeLoad(Location loc, Value val, VectorType type,
int64_t index, int64_t pos,
PatternRewriter &rewriter) {
if (index == -1)
return val;
Type lowType = VectorType::Builder(type).dropDim(0);
// At extraction dimension?
if (index == 0) {
auto posAttr = rewriter.getI64ArrayAttr(pos);
return rewriter.create<vector::ExtractOp>(loc, lowType, val, posAttr);
}
// Unroll leading dimensions.
VectorType vType = cast<VectorType>(lowType);
Type resType = VectorType::Builder(type).dropDim(index);
auto resVectorType = cast<VectorType>(resType);
Value result = rewriter.create<arith::ConstantOp>(
loc, resVectorType, rewriter.getZeroAttr(resVectorType));
for (int64_t d = 0, e = resVectorType.getDimSize(0); d < e; d++) {
auto posAttr = rewriter.getI64ArrayAttr(d);
Value ext = rewriter.create<vector::ExtractOp>(loc, vType, val, posAttr);
Value load = reshapeLoad(loc, ext, vType, index - 1, pos, rewriter);
result = rewriter.create<vector::InsertOp>(loc, resVectorType, load, result,
posAttr);
}
return result;
}
// Helper method to possibly drop a dimension in a store.
// TODO
static Value reshapeStore(Location loc, Value val, Value result,
VectorType type, int64_t index, int64_t pos,
PatternRewriter &rewriter) {
// Unmodified?
if (index == -1)
return val;
// At insertion dimension?
if (index == 0) {
auto posAttr = rewriter.getI64ArrayAttr(pos);
return rewriter.create<vector::InsertOp>(loc, type, val, result, posAttr);
}
// Unroll leading dimensions.
Type lowType = VectorType::Builder(type).dropDim(0);
VectorType vType = cast<VectorType>(lowType);
Type insType = VectorType::Builder(vType).dropDim(0);
for (int64_t d = 0, e = type.getDimSize(0); d < e; d++) {
auto posAttr = rewriter.getI64ArrayAttr(d);
Value ext = rewriter.create<vector::ExtractOp>(loc, vType, result, posAttr);
Value ins = rewriter.create<vector::ExtractOp>(loc, insType, val, posAttr);
Value sto = reshapeStore(loc, ins, ext, vType, index - 1, pos, rewriter);
result = rewriter.create<vector::InsertOp>(loc, type, sto, result, posAttr);
}
return result;
}
/// Helper to create arithmetic operation associated with a kind of contraction.
static std::optional<Value>
createContractArithOp(Location loc, Value x, Value y, Value acc,
vector::CombiningKind kind, PatternRewriter &rewriter,
bool isInt, Value mask = Value()) {
using vector::CombiningKind;
Value mul;
if (isInt) {
if (kind == CombiningKind::MINF || kind == CombiningKind::MAXF)
// Only valid for floating point types.
return std::nullopt;
mul = rewriter.create<arith::MulIOp>(loc, x, y);
} else {
// Float case.
if (kind == CombiningKind::AND || kind == CombiningKind::MINUI ||
kind == CombiningKind::MINSI || kind == CombiningKind::MAXUI ||
kind == CombiningKind::MAXSI || kind == CombiningKind::OR ||
kind == CombiningKind::XOR)
// Only valid for integer types.
return std::nullopt;
// Special case for fused multiply-add.
if (acc && isa<VectorType>(acc.getType()) && kind == CombiningKind::ADD) {
Value fma = rewriter.create<vector::FMAOp>(loc, x, y, acc);
if (mask)
// The fma op doesn't need explicit masking. However, fma ops used in
// reductions must preserve previous 'acc' values for masked-out lanes.
fma = selectPassthru(rewriter, mask, fma, acc);
return fma;
}
mul = rewriter.create<arith::MulFOp>(loc, x, y);
}
if (!acc)
return std::optional<Value>(mul);
return makeArithReduction(rewriter, loc, kind, mul, acc, mask);
}
/// Return the positions of the reductions in the given map.
static SmallVector<int64_t> getReductionIndex(AffineMap map,
ArrayAttr iteratorTypes) {
SmallVector<int64_t> dimsIdx;
for (unsigned i = 0, e = map.getNumResults(); i < e; i++) {
if (isReductionIterator(iteratorTypes[map.getDimPosition(i)]))
dimsIdx.push_back(i);
}
return dimsIdx;
}
/// Look for a given dimension in an affine map and return its position. Return
/// std::nullopt if the dimension is not in the map results.
static std::optional<unsigned> getDimPosition(AffineMap map, unsigned dim) {
for (unsigned i = 0, e = map.getNumResults(); i < e; i++) {
if (map.getDimPosition(i) == dim)
return i;
}
return std::nullopt;
}
/// Creates an AddIOp if `isInt` is true otherwise create an arith::AddFOp using
/// operands `x` and `y`.
static Value createAdd(Location loc, Value x, Value y, bool isInt,
PatternRewriter &rewriter) {
if (isInt)
return rewriter.create<arith::AddIOp>(loc, x, y);
return rewriter.create<arith::AddFOp>(loc, x, y);
}
/// Creates a MulIOp if `isInt` is true otherwise create an MulFOp using
/// operands `x and `y`.
static Value createMul(Location loc, Value x, Value y, bool isInt,
PatternRewriter &rewriter) {
if (isInt)
return rewriter.create<arith::MulIOp>(loc, x, y);
return rewriter.create<arith::MulFOp>(loc, x, y);
}
namespace {
/// Progressive lowering of a `vector.contract %a, %b, %c` with row-major matmul
/// semantics to:
/// ```
/// %flattened_a = vector.shape_cast %a
/// %flattened_b = vector.shape_cast %b
/// %flattened_d = vector.matmul %flattened_a, %flattened_b
/// %d = vector.shape_cast %%flattened_d
/// %e = add %c, %d
/// ```
/// `vector.matmul` later lowers to `llvm.matrix.multiply`.
//
/// This only kicks in when VectorTransformsOptions is set to OuterProduct and
/// the vector.contract op is a row-major matrix multiply.
class ContractionOpToMatmulOpLowering
: public OpRewritePattern<vector::ContractionOp> {
public:
using OpRewritePattern::OpRewritePattern;
using FilterConstraintType =
std::function<LogicalResult(vector::ContractionOp op)>;
static LogicalResult defaultFilter(vector::ContractionOp op) {
return success();
}
ContractionOpToMatmulOpLowering(
vector::VectorTransformsOptions vectorTransformOptions,
MLIRContext *context, PatternBenefit benefit = 1,
FilterConstraintType constraint = defaultFilter)
: OpRewritePattern<vector::ContractionOp>(context, benefit),
vectorTransformOptions(vectorTransformOptions),
filter(std::move(constraint)) {}
LogicalResult matchAndRewrite(vector::ContractionOp op,
PatternRewriter &rewriter) const override;
private:
/// Options to control the vector patterns.
vector::VectorTransformsOptions vectorTransformOptions;
FilterConstraintType filter;
};
/// Progressive lowering of a `vector.contract %a, %b, %c` with row-major matmul
/// semantics to a reduction_size-unrolled sequence:
/// ```
/// %at = vector.transpose %a, [1, 0]
/// %bRow0 = vector.extract %b[0]
/// %atRow0 = vector.extract %at[0]
/// %c0 = vector.outerproduct %atRow0, %bRow0, %c
/// ...
/// %bRowK = vector.extract %b[K]
/// %atRowK = vector.extract %at[K]
/// %cK = vector.outerproduct %atRowK, %bRowK, %cK-1
/// ```
///
/// This only kicks in when VectorTransformsOptions is set to OuterProduct and
/// the vector.contract op is a row-major matrix multiply.
class ContractionOpToOuterProductOpLowering
: public OpRewritePattern<vector::ContractionOp> {
public:
using OpRewritePattern::OpRewritePattern;
using FilterConstraintType =
std::function<LogicalResult(vector::ContractionOp op)>;
static LogicalResult defaultFilter(vector::ContractionOp op) {
return success();
}
ContractionOpToOuterProductOpLowering(
vector::VectorTransformsOptions vectorTransformOptions,
MLIRContext *context, PatternBenefit benefit = 1,
FilterConstraintType constraint = defaultFilter)
: OpRewritePattern<vector::ContractionOp>(context, benefit),
vectorTransformOptions(vectorTransformOptions),
filter(std::move(constraint)) {}
LogicalResult matchAndRewrite(vector::ContractionOp op,
PatternRewriter &rewriter) const override;
private:
/// Options to control the vector patterns.
vector::VectorTransformsOptions vectorTransformOptions;
FilterConstraintType filter;
};
/// Progressive lowering of a `vector.contract %a, %b, %c` with row-major matmul
/// semantics to an output-size-unrolled sequence:
/// ```
/// %out = arith.constant ... : vector<MxNxelt_type>
/// %bt = vector.transpose %b, [1, 0]
/// %aRow0 = vector.extract %a[0]
/// %btRow0 = vector.extract %bt[0]
/// %c00 = vector.reduce %atRow0, %bRow0
/// %out00 = vector.insert %c00, %out[0, 0]
/// ...
/// %aRowLast = vector.extract %at[M-1]
/// %btRowLast = vector.extract %b[N-1]
/// %cLastLast = vector.reduce %atRowLast, %bRowLast
/// %outcLastLast = vector.insert %cLastLast, %out[M-1, N-1]
/// ```
///
/// This only kicks in when VectorTransformsOptions is set to Dot and
/// the vector.contract op is a row-major matmul or matvec.
class ContractionOpToDotLowering
: public OpRewritePattern<vector::ContractionOp> {
public:
using OpRewritePattern::OpRewritePattern;
using FilterConstraintType =
std::function<LogicalResult(vector::ContractionOp op)>;
static LogicalResult defaultFilter(vector::ContractionOp op) {
return success();
}
ContractionOpToDotLowering(
vector::VectorTransformsOptions vectorTransformOptions,
MLIRContext *context, PatternBenefit benefit = 1,
const FilterConstraintType &constraint = defaultFilter)
: OpRewritePattern<vector::ContractionOp>(context, benefit),
vectorTransformOptions(vectorTransformOptions), filter(defaultFilter) {}
LogicalResult matchAndRewrite(vector::ContractionOp op,
PatternRewriter &rewriter) const override;
private:
/// Options to control the vector patterns.
vector::VectorTransformsOptions vectorTransformOptions;
FilterConstraintType filter;
};
/// Progressive lowering of ContractionOp.
///
/// One:
/// %x = vector.contract with at least one free/batch dimension
/// is replaced by:
/// %a = vector.contract with one less free/batch dimension
/// %b = vector.contract with one less free/batch dimension
/// ..
/// %x = combine %a %b ..
/// until a pure contraction is reached (no free/batch dimensions),
/// which is replaced by a dot-product.
///
/// This only kicks in when either VectorTransformsOptions is set
/// to Dot or when other contraction patterns fail.
class ContractionOpLowering : public OpRewritePattern<vector::ContractionOp> {
public:
using OpRewritePattern::OpRewritePattern;
using FilterConstraintType =
std::function<LogicalResult(vector::ContractionOp op)>;
static LogicalResult defaultFilter(vector::ContractionOp op) {
return success();
}
ContractionOpLowering(vector::VectorTransformsOptions vectorTransformOptions,
MLIRContext *context, PatternBenefit benefit = 1,
FilterConstraintType constraint = defaultFilter)
: OpRewritePattern<vector::ContractionOp>(context, benefit),
vectorTransformOptions(vectorTransformOptions),
filter(std::move(constraint)) {}
LogicalResult matchAndRewrite(vector::ContractionOp op,
PatternRewriter &rewriter) const override;
private:
/// Options to control the vector patterns.
vector::VectorTransformsOptions vectorTransformOptions;
FilterConstraintType filter;
// Lower one parallel dimension.
FailureOr<Value> lowerParallel(PatternRewriter &rewriter,
vector::ContractionOp op, int64_t lhsIndex,
int64_t rhsIndex, Value mask) const;
// Lower one reduction dimension.
FailureOr<Value> lowerReduction(PatternRewriter &rewriter,
vector::ContractionOp op, Value mask) const;
};
/// Generate a vector implementation for matmat, matvec and tmatvec.
/// This unrolls outer-products along the reduction dimension.
struct UnrolledOuterProductGenerator
: public StructuredGenerator<vector::ContractionOp, vector::IteratorType> {
UnrolledOuterProductGenerator(RewriterBase &b, vector::ContractionOp op)
: StructuredGenerator<vector::ContractionOp, vector::IteratorType>(b, op),
kind(op.getKind()), lhs(op.getLhs()), rhs(op.getRhs()),
res(op.getAcc()), lhsType(op.getLhsType()) {
auto maskableOp = cast<MaskableOpInterface>(op.getOperation());
if (maskableOp.isMasked())
mask = maskableOp.getMaskingOp().getMask();
}
Value t(Value v, ArrayRef<int64_t> perm = {1, 0}) {
if (!v)
return v;
return rewriter.create<vector::TransposeOp>(loc, v, perm);
}
Value promote(Value v, Type dstElementType) {
Type elementType = v.getType();
auto vecType = dyn_cast<VectorType>(elementType);
if (vecType)
elementType = vecType.getElementType();
if (elementType == dstElementType)
return v;
Type promotedType = dstElementType;
if (vecType)
promotedType = VectorType::get(vecType.getShape(), promotedType);
if (isa<FloatType>(dstElementType))
return rewriter.create<arith::ExtFOp>(loc, promotedType, v);
return rewriter.create<arith::ExtSIOp>(loc, promotedType, v);
}
FailureOr<Value> outerProd(Value lhs, Value rhs, Value res, int reductionSize,
std::optional<Value> maybeMask = std::nullopt) {
assert(reductionSize > 0);
// Incremental support for masking.
if (mask && !maybeMask.has_value())
return failure();
Type resElementType = cast<VectorType>(res.getType()).getElementType();
for (int64_t k = 0; k < reductionSize; ++k) {
Value extractA = rewriter.create<vector::ExtractOp>(loc, lhs, k);
Value extractB = rewriter.create<vector::ExtractOp>(loc, rhs, k);
extractA = promote(extractA, resElementType);
extractB = promote(extractB, resElementType);
Value extractMask;
if (maybeMask.has_value() && maybeMask.value())
extractMask =
rewriter.create<vector::ExtractOp>(loc, maybeMask.value(), k);
Operation *outerProdOp = rewriter.create<vector::OuterProductOp>(
loc, res.getType(), extractA, extractB, res, kind);
res = maskOperation(rewriter, outerProdOp, extractMask)->getResult(0);
}
return res;
}
/// Two outer parallel, one inner reduction (matmat flavor).
FailureOr<Value> matmat() {
if (!iters({Par(), Par(), Red()}))
return failure();
// Set up the parallel/reduction structure in the right form.
AffineExpr m, n, k;
bindDims(rewriter.getContext(), m, n, k);
// Classical row-major matmul: Just permute the lhs.
if (layout({{m, k}, {k, n}, {m, n}}))
return outerProd(t(lhs), rhs, res, lhsType.getDimSize(1),
t(mask, {2, 0, 1}));
// TODO: may be better to fail and use some vector<k> -> scalar reduction.
if (layout({{m, k}, {n, k}, {m, n}})) {
Value tlhs = t(lhs);
return outerProd(tlhs, t(rhs), res, lhsType.getDimSize(1));
}
// No need to permute anything.
if (layout({{k, m}, {k, n}, {m, n}}))
return outerProd(lhs, rhs, res, lhsType.getDimSize(0));
// Just permute the rhs.
if (layout({{k, m}, {n, k}, {m, n}}))
return outerProd(lhs, t(rhs), res, lhsType.getDimSize(0));
// Transposed output: swap RHS and LHS.
// Classical row-major matmul: permute the lhs.
if (layout({{m, k}, {k, n}, {n, m}}))
return outerProd(rhs, t(lhs), res, lhsType.getDimSize(1));
// TODO: may be better to fail and use some vector<k> -> scalar reduction.
if (layout({{m, k}, {n, k}, {n, m}})) {
Value trhs = t(rhs);
return outerProd(trhs, t(lhs), res, lhsType.getDimSize(1));
}
if (layout({{k, m}, {k, n}, {n, m}}))
return outerProd(rhs, lhs, res, lhsType.getDimSize(0));
if (layout({{k, m}, {n, k}, {n, m}}))
return outerProd(t(rhs), lhs, res, lhsType.getDimSize(0));
return failure();
}
/// One outer parallel, one inner reduction (matvec flavor)
FailureOr<Value> matvec() {
if (!iters({Par(), Red()}))
return failure();
AffineExpr m, k;
bindDims(rewriter.getContext(), m, k);
// Case mat-vec: transpose.
if (layout({{m, k}, {k}, {m}}))
return outerProd(t(lhs), rhs, res, lhsType.getDimSize(1), t(mask));
// Case mat-trans-vec: ready to go.
if (layout({{k, m}, {k}, {m}}))
return outerProd(lhs, rhs, res, lhsType.getDimSize(0));
// Case vec-mat: swap and transpose.
if (layout({{k}, {m, k}, {m}}))
return outerProd(t(rhs), lhs, res, lhsType.getDimSize(0));
// Case vec-mat-trans: swap and ready to go.
if (layout({{k}, {k, m}, {m}}))
return outerProd(rhs, lhs, res, lhsType.getDimSize(0));
return failure();
}
//
// One outer reduction, one inner parallel (tmatvec flavor)
//
FailureOr<Value> tmatvec() {
if (!iters({Red(), Par()}))
return failure();
AffineExpr k, m;
bindDims(rewriter.getContext(), k, m);
// Case mat-vec: transpose.
if (layout({{m, k}, {k}, {m}}))
return outerProd(t(lhs), rhs, res, lhsType.getDimSize(1));
// Case mat-trans-vec: ready to go.
if (layout({{k, m}, {k}, {m}}))
return outerProd(lhs, rhs, res, lhsType.getDimSize(0));
// Case vec-mat: swap and transpose.
if (layout({{k}, {m, k}, {m}}))
return outerProd(t(rhs), lhs, res, lhsType.getDimSize(0));
// Case vec-mat-trans: swap and ready to go.
if (layout({{k}, {k, m}, {m}}))
return outerProd(rhs, lhs, res, lhsType.getDimSize(0));
return failure();
}
private:
vector::CombiningKind kind;
Value lhs, rhs, res, mask;
VectorType lhsType;
};
/// Progressively lower a `vector.contract %a, %b, %c` with row-major matmul
/// semantics to a reduction_size-unrolled sequence:
/// ```
/// %at = vector.transpose %a, [1, 0]
/// %bRow0 = vector.extract %b[0]
/// %atRow0 = vector.extract %at[0]
/// %c0 = vector.outerproduct %atRow0, %bRow0, %c
/// ...
/// %bRowK = vector.extract %b[K]
/// %atRowK = vector.extract %at[K]
/// %cK = vector.outerproduct %atRowK, %bRowK, %cK-1
/// ```
///
/// This only kicks in when VectorTransformsOptions is set to OuterProduct but
/// otherwise supports any layout permutation of the matrix-multiply.
LogicalResult ContractionOpToOuterProductOpLowering::matchAndRewrite(
vector::ContractionOp op, PatternRewriter &rewriter) const {
if (vectorTransformOptions.vectorContractLowering !=
vector::VectorContractLowering::OuterProduct)
return failure();
if (failed(filter(op)))
return failure();
// Vector mask setup.
OpBuilder::InsertionGuard guard(rewriter);
auto maskableOp = cast<vector::MaskableOpInterface>(op.getOperation());
Operation *rootOp;
if (maskableOp.isMasked()) {
rewriter.setInsertionPoint(maskableOp.getMaskingOp());
rootOp = maskableOp.getMaskingOp();
} else {
rootOp = op;
}
UnrolledOuterProductGenerator e(rewriter, op);
FailureOr<Value> matmatRes = e.matmat();
if (succeeded(matmatRes)) {
rewriter.replaceOp(rootOp, *matmatRes);
return success();
}
FailureOr<Value> matvecRes = e.matvec();
if (succeeded(matvecRes)) {
rewriter.replaceOp(rootOp, *matvecRes);
return success();
}
FailureOr<Value> tmatvecRes = e.tmatvec();
if (succeeded(tmatvecRes)) {
rewriter.replaceOp(rootOp, *tmatvecRes);
return success();
}
return failure();
}
LogicalResult
ContractionOpToDotLowering::matchAndRewrite(vector::ContractionOp op,
PatternRewriter &rewriter) const {
// TODO: Support vector.mask.
auto maskableOp = cast<MaskableOpInterface>(op.getOperation());
if (maskableOp.isMasked())
return failure();
if (failed(filter(op)))
return failure();
if (vectorTransformOptions.vectorContractLowering !=
vector::VectorContractLowering::Dot)
return failure();
auto iteratorTypes = op.getIteratorTypes().getValue();
static constexpr std::array<int64_t, 2> perm = {1, 0};
Location loc = op.getLoc();
Value lhs = op.getLhs(), rhs = op.getRhs();
using MapList = ArrayRef<ArrayRef<AffineExpr>>;
auto infer = [](MapList m) { return AffineMap::inferFromExprList(m); };
AffineExpr m, n, k;
bindDims(rewriter.getContext(), m, n, k);
SmallVector<AffineMap> maps = op.getIndexingMapsArray();
//
// In the following we wish to make the reduction dimension innermost so we
// can load vectors and just fmul + reduce into a scalar.
//
if (isParallelIterator(iteratorTypes[0]) &&
isParallelIterator(iteratorTypes[1]) &&
isReductionIterator(iteratorTypes[2])) {
//
// Two outer parallel, one inner reduction (matmat flavor).
//
if (maps == infer({{m, k}, {k, n}, {m, n}})) {
rhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm);
} else if (maps == infer({{m, k}, {n, k}, {m, n}})) {
// No need to permute anything.
} else if (maps == infer({{k, m}, {k, n}, {m, n}})) {
lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm);
rhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm);
} else if (maps == infer({{k, m}, {n, k}, {m, n}})) {
lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm);
} else if (maps == infer({{m, k}, {k, n}, {n, m}})) {
// This is the classical row-major matmul. Just permute the lhs.
Value tmp = lhs;
lhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm);
rhs = tmp;
} else if (maps == infer({{m, k}, {n, k}, {n, m}})) {
std::swap(lhs, rhs);
} else if (maps == infer({{k, m}, {k, n}, {n, m}})) {
Value tmp = lhs;
lhs = rewriter.create<vector::TransposeOp>(loc, rhs, perm);
rhs = rewriter.create<vector::TransposeOp>(loc, tmp, perm);
} else if (maps == infer({{k, m}, {n, k}, {n, m}})) {
Value tmp = rhs;
rhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm);
lhs = tmp;
} else {
return failure();
}
} else if (isParallelIterator(iteratorTypes[0]) &&
isReductionIterator(iteratorTypes[1])) {
//
// One outer parallel, one inner reduction (matvec flavor)
//
if (maps == infer({{m, n}, {n}, {m}})) {
// No need to permute anything.
} else if (maps == infer({{n, m}, {n}, {m}})) {
lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm);
} else if (maps == infer({{n}, {m, n}, {m}})) {
std::swap(lhs, rhs);
} else if (maps == infer({{n}, {n, m}, {m}})) {
std::swap(lhs, rhs);
lhs = rewriter.create<vector::TransposeOp>(loc, lhs, perm);
} else {
return failure();
}
} else {
return failure();
}
VectorType dstType = cast<VectorType>(op.getResultType());
assert(dstType.getRank() >= 1 && dstType.getRank() <= 2 &&
"Expected dst type of rank 1 or 2");
unsigned rank = dstType.getRank();
unsigned dstRows = dstType.getShape()[0];
unsigned dstColumns = rank == 1 ? 1 : dstType.getShape()[1];
// ExtractOp does not allow dynamic indexing, we must unroll explicitly.
Value res = rewriter.create<arith::ConstantOp>(loc, dstType,
rewriter.getZeroAttr(dstType));
bool isInt = isa<IntegerType>(dstType.getElementType());
for (unsigned r = 0; r < dstRows; ++r) {
Value a = rewriter.create<vector::ExtractOp>(op.getLoc(), lhs, r);
for (unsigned c = 0; c < dstColumns; ++c) {
Value b = rank == 1
? rhs
: rewriter.create<vector::ExtractOp>(op.getLoc(), rhs, c);
Value m = createMul(op.getLoc(), a, b, isInt, rewriter);
Value reduced = rewriter.create<vector::ReductionOp>(
op.getLoc(), vector::CombiningKind::ADD, m);
SmallVector<int64_t, 2> pos = rank == 1 ? SmallVector<int64_t, 2>{r}
: SmallVector<int64_t, 2>{r, c};
res = rewriter.create<vector::InsertOp>(op.getLoc(), reduced, res, pos);
}
}
if (auto acc = op.getAcc())
res = createAdd(op.getLoc(), res, acc, isInt, rewriter);
rewriter.replaceOp(op, res);
return success();
}
/// Lower vector.contract with all size one reduction dimensions to
/// elementwise ops when possible.
struct ContractOpToElementwise
: public OpRewritePattern<vector::ContractionOp> {
using OpRewritePattern::OpRewritePattern;
using FilterConstraintType =
std::function<LogicalResult(vector::ContractionOp op)>;
static LogicalResult defaultFilter(vector::ContractionOp op) {
return success();
}
ContractOpToElementwise(
vector::VectorTransformsOptions vectorTransformOptions,
MLIRContext *context, PatternBenefit benefit = 1,
const FilterConstraintType &constraint = defaultFilter)
: OpRewritePattern<vector::ContractionOp>(context, benefit),
vectorTransformOptions(vectorTransformOptions), filter(defaultFilter) {}
LogicalResult matchAndRewrite(vector::ContractionOp contractOp,
PatternRewriter &rewriter) const override {
// TODO: Support vector.mask.
auto maskableOp = cast<MaskableOpInterface>(contractOp.getOperation());
if (maskableOp.isMasked())
return failure();
if (failed(filter(contractOp)))
return failure();
if (vectorTransformOptions.vectorContractLowering !=
vector::VectorContractLowering::ParallelArith)
return failure();
ArrayRef<int64_t> lhsShape = contractOp.getLhsType().getShape();
ArrayRef<int64_t> rhsShape = contractOp.getRhsType().getShape();
AffineMap lhsMap = contractOp.getIndexingMapsArray()[0];
AffineMap rhsMap = contractOp.getIndexingMapsArray()[1];
SmallVector<int64_t> lhsReductionDims =
getReductionIndex(lhsMap, contractOp.getIteratorTypes());
SmallVector<int64_t> rhsReductionDims =
getReductionIndex(rhsMap, contractOp.getIteratorTypes());
// All the reduction dimensions must be a size 1.
for (int64_t dim : lhsReductionDims) {
if (lhsShape[dim] != 1)
return failure();
}
for (int64_t dim : rhsReductionDims) {
if (rhsShape[dim] != 1)
return failure();
}
AffineMap accMap = contractOp.getIndexingMapsArray()[2];
unsigned numParallelDims = accMap.getNumResults();
unsigned numLhsDimToBroadcast =
numParallelDims - (lhsMap.getNumResults() - lhsReductionDims.size());
unsigned numRhsDimToBroadcast =
numParallelDims - (rhsMap.getNumResults() - rhsReductionDims.size());
SmallVector<int64_t> lhsDims;
SmallVector<int64_t> lhsTranspose;
SmallVector<int64_t> rhsDims;
SmallVector<int64_t> rhsTranspose;
for (int64_t dim : lhsReductionDims)
lhsTranspose.push_back(numLhsDimToBroadcast + dim);
for (int64_t dim : rhsReductionDims)
rhsTranspose.push_back(numRhsDimToBroadcast + dim);
// Loop through the parallel dimensions to calculate the dimensions to
// broadcast and to permute in order to extract only parallel dimensions.
for (unsigned i = 0; i < numParallelDims; i++) {
std::optional<unsigned> lhsDim =
getDimPosition(lhsMap, accMap.getDimPosition(i));
if (lhsDim) {
lhsTranspose.push_back(numLhsDimToBroadcast + *lhsDim);
} else {
// If the parallel dimension doesn't exist we will have to broadcast it.
lhsDims.push_back(
cast<VectorType>(contractOp.getResultType()).getDimSize(i));
lhsTranspose.push_back(lhsDims.size() - 1);
}
std::optional<unsigned> rhsDim =
getDimPosition(rhsMap, accMap.getDimPosition(i));
if (rhsDim) {
rhsTranspose.push_back(numRhsDimToBroadcast + *rhsDim);
} else {
// If the parallel dimension doesn't exist we will have to broadcast it.
rhsDims.push_back(
cast<VectorType>(contractOp.getResultType()).getDimSize(i));
rhsTranspose.push_back(rhsDims.size() - 1);
}
}
Value newLhs = contractOp.getLhs();
Value newRhs = contractOp.getRhs();
Location loc = contractOp.getLoc();
if (!lhsDims.empty()) {
lhsDims.append(lhsShape.begin(), lhsShape.end());
auto expandedType =
VectorType::get(lhsDims, contractOp.getLhsType().getElementType());
newLhs = rewriter.create<vector::BroadcastOp>(loc, expandedType, newLhs);
}
if (!rhsDims.empty()) {
rhsDims.append(rhsShape.begin(), rhsShape.end());
auto expandedType =
VectorType::get(rhsDims, contractOp.getRhsType().getElementType());
newRhs = rewriter.create<vector::BroadcastOp>(loc, expandedType, newRhs);
}
bool isInt = contractOp.getLhsType().getElementType().isIntOrIndex();
newLhs = rewriter.create<vector::TransposeOp>(loc, newLhs, lhsTranspose);
newRhs = rewriter.create<vector::TransposeOp>(loc, newRhs, rhsTranspose);
SmallVector<int64_t> lhsOffsets(lhsReductionDims.size(), 0);
SmallVector<int64_t> rhsOffsets(rhsReductionDims.size(), 0);
newLhs = rewriter.create<vector::ExtractOp>(
loc, newLhs, rewriter.getI64ArrayAttr(lhsOffsets));
newRhs = rewriter.create<vector::ExtractOp>(
loc, newRhs, rewriter.getI64ArrayAttr(rhsOffsets));
std::optional<Value> result =
createContractArithOp(loc, newLhs, newRhs, contractOp.getAcc(),
contractOp.getKind(), rewriter, isInt);
rewriter.replaceOp(contractOp, {*result});
return success();
}
private:
/// Options to control the vector patterns.
vector::VectorTransformsOptions vectorTransformOptions;
FilterConstraintType filter;
};
/// Progressive lowering of ContractionOp.
/// One:
/// %x = vector.contract with at least one free/batch dimension
/// is replaced by:
/// %a = vector.contract with one less free/batch dimension
/// %b = vector.contract with one less free/batch dimension
/// ..
/// %x = combine %a %b ..
/// until a pure contraction is reached (no free/batch dimensions),
/// which is replaced by a dot-product.
///
/// This only kicks in when either VectorTransformsOptions is set
/// to DOT or when other contraction patterns fail.
//
// TODO: break down into transpose/reshape/cast ops
// when they become available to avoid code dup
// TODO: investigate lowering order impact on performance
LogicalResult
ContractionOpLowering::matchAndRewrite(vector::ContractionOp op,
PatternRewriter &rewriter) const {
if (failed(filter(op)))
return failure();
// TODO: support mixed mode contract lowering.
if (op.getLhsType().getElementType() !=
getElementTypeOrSelf(op.getAccType()) ||
op.getRhsType().getElementType() != getElementTypeOrSelf(op.getAccType()))
return failure();
// TODO: the code below assumes the default contraction, make sure it supports
// other kinds before enabling this lowering.
if (op.getKind() != vector::CombiningKind::ADD) {
return rewriter.notifyMatchFailure(
op, "contractions other than 'add' not supported");
}
// TODO: implement benefits, cost models.
MLIRContext *ctx = op.getContext();
ContractionOpToMatmulOpLowering pat1(vectorTransformOptions, ctx);
if (succeeded(pat1.matchAndRewrite(op, rewriter)))
return success();
ContractionOpToOuterProductOpLowering pat2(vectorTransformOptions, ctx);
if (succeeded(pat2.matchAndRewrite(op, rewriter)))
return success();
ContractionOpToDotLowering pat3(vectorTransformOptions, ctx);
if (succeeded(pat3.matchAndRewrite(op, rewriter)))
return success();
ContractOpToElementwise pat4(vectorTransformOptions, ctx);
if (succeeded(pat4.matchAndRewrite(op, rewriter)))
return success();
// Vector mask setup.
OpBuilder::InsertionGuard guard(rewriter);
Operation *rootOp = op;
Value mask;
if (op.isMasked()) {
rewriter.setInsertionPoint(op.getMaskingOp());
rootOp = op.getMaskingOp();
mask = op.getMaskingOp().getMask();
}
// Find first batch dimension in LHS/RHS, and lower when found.
std::vector<std::pair<int64_t, int64_t>> batchDimMap = op.getBatchDimMap();
if (!batchDimMap.empty()) {
int64_t lhsIndex = batchDimMap[0].first;
int64_t rhsIndex = batchDimMap[0].second;
auto newOp = lowerParallel(rewriter, op, lhsIndex, rhsIndex, mask);
if (failed(newOp))
return failure();
rewriter.replaceOp(rootOp, *newOp);
return success();
}
// Collect contracting dimensions.
std::vector<std::pair<int64_t, int64_t>> contractingDimMap =
op.getContractingDimMap();
DenseSet<int64_t> lhsContractingDimSet;
DenseSet<int64_t> rhsContractingDimSet;
for (auto &dimPair : contractingDimMap) {
lhsContractingDimSet.insert(dimPair.first);
rhsContractingDimSet.insert(dimPair.second);
}
// Find first free dimension in LHS, and lower when found.
VectorType lhsType = op.getLhsType();
for (int64_t lhsIndex = 0, e = lhsType.getRank(); lhsIndex < e; ++lhsIndex) {
if (lhsContractingDimSet.count(lhsIndex) == 0) {
auto newOp = lowerParallel(rewriter, op, lhsIndex, /*rhsIndex=*/-1, mask);
if (failed(newOp))
return failure();
rewriter.replaceOp(rootOp, *newOp);
return success();
}
}
// Find first free dimension in RHS, and lower when found.
VectorType rhsType = op.getRhsType();
for (int64_t rhsIndex = 0, e = rhsType.getRank(); rhsIndex < e; ++rhsIndex) {
if (rhsContractingDimSet.count(rhsIndex) == 0) {
auto newOp = lowerParallel(rewriter, op, /*lhsIndex=*/-1, rhsIndex, mask);
if (failed(newOp))
return failure();
rewriter.replaceOp(rootOp, *newOp);
return success();
}
}
// Lower the first remaining reduction dimension.
if (!contractingDimMap.empty()) {
auto newOp = lowerReduction(rewriter, op, mask);
if (failed(newOp))
return failure();
rewriter.replaceOp(rootOp, *newOp);
return success();
}
return failure();
}
// Lower one parallel dimension.
// Incidentally also tolerates unit-size (hence trivial) reduction dimensions.
// TODO: consider reusing existing contract unrolling
FailureOr<Value> ContractionOpLowering::lowerParallel(PatternRewriter &rewriter,
vector::ContractionOp op,
int64_t lhsIndex,
int64_t rhsIndex,
Value mask) const {
VectorType lhsType = op.getLhsType();
VectorType rhsType = op.getRhsType();
VectorType resType = cast<VectorType>(op.getResultType());
// Find the iterator type index and result index.
SmallVector<AffineMap> iMap = op.getIndexingMapsArray();
int64_t iterIndex = -1;
int64_t dimSize = -1;
if (lhsIndex >= 0) {
iterIndex = iMap[0].getDimPosition(lhsIndex);
if (rhsIndex >= 0 && iterIndex != iMap[1].getDimPosition(rhsIndex))
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
diag << "expected lhsIndex=" << lhsIndex << " and rhsIndex=" << rhsIndex
<< " to map to the same dimension";
});
dimSize = lhsType.getDimSize(lhsIndex);
} else if (rhsIndex >= 0) {
iterIndex = iMap[1].getDimPosition(rhsIndex);
dimSize = rhsType.getDimSize(rhsIndex);
}
if (iterIndex < 0)
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
diag << "expected either lhsIndex=" << lhsIndex
<< " or rhsIndex=" << rhsIndex << " to be nonnegative";
});
// value_or(-1) means that we tolerate a dimension not appearing
// in the result map. That can't happen for actual parallel iterators, but
// the caller ContractionOpLowering::matchAndRewrite is currently calling
// lowerParallel also for the case of unit-size reduction dims appearing only
// on one of LHS or RHS, not both. At the moment, such cases are created by
// CastAwayContractionLeadingOneDim, so we need to either support that or
// modify that pattern.
int64_t resIndex = getResultIndex(iMap[2], iterIndex).value_or(-1);
if (resIndex == -1 && dimSize != 1)
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
diag << "expected the dimension for iterIndex=" << iterIndex
<< " to either appear in the result map, or to be a unit dimension";
});
// Construct new iterator types and affine map array attribute.
std::array<AffineMap, 3> lowIndexingMaps = {
adjustMap(iMap[0], iterIndex, rewriter),
adjustMap(iMap[1], iterIndex, rewriter),
adjustMap(iMap[2], iterIndex, rewriter)};
auto lowAffine = rewriter.getAffineMapArrayAttr(lowIndexingMaps);
auto lowIter =
rewriter.getArrayAttr(adjustIter(op.getIteratorTypes(), iterIndex));
// Unroll into a series of lower dimensional vector.contract ops.
Location loc = op.getLoc();
Value result = rewriter.create<arith::ConstantOp>(
loc, resType, rewriter.getZeroAttr(resType));
for (int64_t d = 0; d < dimSize; ++d) {
auto lhs = reshapeLoad(loc, op.getLhs(), lhsType, lhsIndex, d, rewriter);
auto rhs = reshapeLoad(loc, op.getRhs(), rhsType, rhsIndex, d, rewriter);
auto acc = reshapeLoad(loc, op.getAcc(), resType, resIndex, d, rewriter);
Value lowMask;
if (mask)
lowMask = reshapeLoad(loc, mask, cast<VectorType>(mask.getType()),
iterIndex, d, rewriter);
Operation *lowContract = rewriter.create<vector::ContractionOp>(
loc, lhs, rhs, acc, lowAffine, lowIter);
lowContract = maskOperation(rewriter, lowContract, lowMask);
result = reshapeStore(loc, lowContract->getResult(0), result, resType,
resIndex, d, rewriter);
}
return result;
}
// Lower one reduction dimension.
FailureOr<Value> ContractionOpLowering::lowerReduction(
PatternRewriter &rewriter, vector::ContractionOp op, Value mask) const {
auto loc = op.getLoc();
VectorType lhsType = op.getLhsType();
VectorType rhsType = op.getRhsType();
Type resType = op.getResultType();
if (isa<VectorType>(resType))
return rewriter.notifyMatchFailure(op,
"did not expect a VectorType result");
bool isInt = isa<IntegerType>(resType);
// Use iterator index 0.
int64_t iterIndex = 0;
SmallVector<AffineMap> iMap = op.getIndexingMapsArray();
std::optional<int64_t> lookupLhs = getResultIndex(iMap[0], iterIndex);
std::optional<int64_t> lookupRhs = getResultIndex(iMap[1], iterIndex);
if (!lookupLhs.has_value())
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
diag << "expected iterIndex=" << iterIndex << "to map to a LHS dimension";
});
if (!lookupRhs.has_value())
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
diag << "expected iterIndex=" << iterIndex << "to map to a RHS dimension";
});
int64_t lhsIndex = *lookupLhs;
int64_t rhsIndex = *lookupRhs;
int64_t dimSize = lhsType.getDimSize(lhsIndex);
if (dimSize != rhsType.getDimSize(rhsIndex))
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
diag << "expect LHS dimension " << lhsIndex
<< " to have the same size as RHS dimension " << rhsIndex;
});
// Base case.
if (lhsType.getRank() == 1) {
if (rhsType.getRank() != 1)
return rewriter.notifyMatchFailure(
op, "When LHS has rank 1, expected also RHS to have rank 1");
Value m = createMul(loc, op.getLhs(), op.getRhs(), isInt, rewriter);
auto kind = vector::CombiningKind::ADD;
Value acc = op.getAcc();
Operation *reductionOp =
acc ? rewriter.create<vector::ReductionOp>(loc, kind, m, acc)
: rewriter.create<vector::ReductionOp>(loc, kind, m);
return maskOperation(rewriter, reductionOp, mask)->getResult(0);
}
// Construct new iterator types and affine map array attribute.
std::array<AffineMap, 3> lowIndexingMaps = {
adjustMap(iMap[0], iterIndex, rewriter),
adjustMap(iMap[1], iterIndex, rewriter),
adjustMap(iMap[2], iterIndex, rewriter)};
auto lowAffine = rewriter.getAffineMapArrayAttr(lowIndexingMaps);
auto lowIter =
rewriter.getArrayAttr(adjustIter(op.getIteratorTypes(), iterIndex));
// Unroll into a series of lower dimensional vector.contract ops.
// By feeding the initial accumulator into the first contraction,
// and the result of each contraction into the next, eventually
// the sum of all reductions is computed.
Value result = op.getAcc();
for (int64_t d = 0; d < dimSize; ++d) {
auto lhs = reshapeLoad(loc, op.getLhs(), lhsType, lhsIndex, d, rewriter);
auto rhs = reshapeLoad(loc, op.getRhs(), rhsType, rhsIndex, d, rewriter);
Value newMask;
if (mask)
newMask = reshapeLoad(loc, mask, cast<VectorType>(mask.getType()),
iterIndex, d, rewriter);
Operation *newContract = rewriter.create<vector::ContractionOp>(
loc, lhs, rhs, result, lowAffine, lowIter);
result = maskOperation(rewriter, newContract, newMask)->getResult(0);
}
return result;
}
/// Progressive lowering of OuterProductOp.
/// One:
/// %x = vector.outerproduct %lhs, %rhs, %acc
/// is replaced by:
/// %z = zero-result
/// %0 = vector.extract %lhs[0]
/// %1 = vector.broadcast %0
/// %2 = vector.extract %acc[0]
/// %3 = vector.fma %1, %rhs, %2
/// %4 = vector.insert %3, %z[0]
/// ..
/// %x = vector.insert %.., %..[N-1]
///
class OuterProductOpLowering : public OpRewritePattern<vector::OuterProductOp> {
public:
using OpRewritePattern::OpRewritePattern;
LogicalResult matchAndRewrite(vector::OuterProductOp op,
PatternRewriter &rewriter) const override {
auto loc = op.getLoc();
VectorType lhsType = op.getOperandVectorTypeLHS();
VectorType rhsType = dyn_cast<VectorType>(op.getOperandTypeRHS());
VectorType resType = op.getResultVectorType();
Type eltType = resType.getElementType();
bool isInt = isa<IntegerType, IndexType>(eltType);
Value acc = (op.getAcc().empty()) ? nullptr : op.getAcc()[0];
vector::CombiningKind kind = op.getKind();
// Vector mask setup.
OpBuilder::InsertionGuard guard(rewriter);
auto maskableOp = cast<vector::MaskableOpInterface>(op.getOperation());
Operation *rootOp;
Value mask;
if (maskableOp.isMasked()) {
rewriter.setInsertionPoint(maskableOp.getMaskingOp());
rootOp = maskableOp.getMaskingOp();
mask = maskableOp.getMaskingOp().getMask();
} else {
rootOp = op;
}
if (!rhsType) {
// Special case: AXPY operation.
Value b = rewriter.create<vector::BroadcastOp>(loc, lhsType, op.getRhs());
std::optional<Value> mult = createContractArithOp(
loc, op.getLhs(), b, acc, kind, rewriter, isInt, mask);
if (!mult.has_value())
return failure();
rewriter.replaceOp(rootOp, *mult);
return success();
}
Value result = rewriter.create<arith::ConstantOp>(
loc, resType, rewriter.getZeroAttr(resType));
for (int64_t d = 0, e = resType.getDimSize(0); d < e; ++d) {
auto pos = rewriter.getI64ArrayAttr(d);
Value x = rewriter.create<vector::ExtractOp>(loc, op.getLhs(), pos);
Value a = rewriter.create<vector::BroadcastOp>(loc, rhsType, x);
Value r = nullptr;
if (acc)
r = rewriter.create<vector::ExtractOp>(loc, acc, pos);
Value extrMask;
if (mask)
extrMask = rewriter.create<vector::ExtractOp>(loc, mask, pos);
std::optional<Value> m = createContractArithOp(
loc, a, op.getRhs(), r, kind, rewriter, isInt, extrMask);
if (!m.has_value())
return failure();
result = rewriter.create<vector::InsertOp>(loc, resType, *m, result, pos);
}
rewriter.replaceOp(rootOp, result);
return success();
}
};
/// Progressively lower a `vector.contract %a, %b, %c` with row-major matmul
/// semantics to:
/// ```
/// %mta = maybe_transpose
/// %mtb = maybe_transpose
/// %flattened_a = vector.shape_cast %mta
/// %flattened_b = vector.shape_cast %mtb
/// %flattened_d = vector.matmul %flattened_a, %flattened_b
/// %mtd = vector.shape_cast %flattened_d
/// %d = maybe_untranspose %mtd
/// %e = add %c, %d
/// ```
/// `vector.matmul` later lowers to `llvm.matrix.multiply`.
//
/// This only kicks in when VectorTransformsOptions is set to `Matmul`.
/// vector.transpose operations are inserted if the vector.contract op is not a
/// row-major matrix multiply.
LogicalResult
ContractionOpToMatmulOpLowering::matchAndRewrite(vector::ContractionOp op,
PatternRewriter &rew) const {
// TODO: Support vector.mask.
auto maskableOp = cast<MaskableOpInterface>(op.getOperation());
if (maskableOp.isMasked())
return failure();
if (vectorTransformOptions.vectorContractLowering !=
vector::VectorContractLowering::Matmul)
return failure();
if (failed(filter(op)))
return failure();
auto iteratorTypes = op.getIteratorTypes().getValue();
if (!isParallelIterator(iteratorTypes[0]) ||
!isParallelIterator(iteratorTypes[1]) ||
!isReductionIterator(iteratorTypes[2]))
return failure();
Type elementType = op.getLhsType().getElementType();
if (!elementType.isIntOrFloat())
return failure();
Type dstElementType = op.getType();
if (auto vecType = dyn_cast<VectorType>(dstElementType))
dstElementType = vecType.getElementType();
if (elementType != dstElementType)
return failure();
// Perform lhs + rhs transpositions to conform to matmul row-major semantics.
// Bail out if the contraction cannot be put in this form.
MLIRContext *ctx = op.getContext();
Location loc = op.getLoc();
AffineExpr m, n, k;
bindDims(rew.getContext(), m, n, k);
// LHS must be A(m, k) or A(k, m).
Value lhs = op.getLhs();
auto lhsMap = op.getIndexingMapsArray()[0];
if (lhsMap == AffineMap::get(3, 0, {k, m}, ctx))
lhs = rew.create<vector::TransposeOp>(loc, lhs, ArrayRef<int64_t>{1, 0});
else if (lhsMap != AffineMap::get(3, 0, {m, k}, ctx))
return failure();
// RHS must be B(k, n) or B(n, k).
Value rhs = op.getRhs();
auto rhsMap = op.getIndexingMapsArray()[1];
if (rhsMap == AffineMap::get(3, 0, {n, k}, ctx))
rhs = rew.create<vector::TransposeOp>(loc, rhs, ArrayRef<int64_t>{1, 0});
else if (rhsMap != AffineMap::get(3, 0, {k, n}, ctx))
return failure();
// At this point lhs and rhs are in row-major.
VectorType lhsType = cast<VectorType>(lhs.getType());
VectorType rhsType = cast<VectorType>(rhs.getType());
int64_t lhsRows = lhsType.getDimSize(0);
int64_t lhsColumns = lhsType.getDimSize(1);
int64_t rhsColumns = rhsType.getDimSize(1);
Type flattenedLHSType =
VectorType::get(lhsType.getNumElements(), lhsType.getElementType());
lhs = rew.create<vector::ShapeCastOp>(loc, flattenedLHSType, lhs);
Type flattenedRHSType =
VectorType::get(rhsType.getNumElements(), rhsType.getElementType());
rhs = rew.create<vector::ShapeCastOp>(loc, flattenedRHSType, rhs);
Value mul = rew.create<vector::MatmulOp>(loc, lhs, rhs, lhsRows, lhsColumns,
rhsColumns);
mul = rew.create<vector::ShapeCastOp>(
loc,
VectorType::get({lhsRows, rhsColumns},
getElementTypeOrSelf(op.getAcc().getType())),
mul);
// ACC must be C(m, n) or C(n, m).
auto accMap = op.getIndexingMapsArray()[2];
if (accMap == AffineMap::get(3, 0, {n, m}, ctx))
mul = rew.create<vector::TransposeOp>(loc, mul, ArrayRef<int64_t>{1, 0});
else if (accMap != AffineMap::get(3, 0, {m, n}, ctx))
llvm_unreachable("invalid contraction semantics");
Value res =
isa<IntegerType>(elementType)
? static_cast<Value>(rew.create<arith::AddIOp>(loc, op.getAcc(), mul))
: static_cast<Value>(
rew.create<arith::AddFOp>(loc, op.getAcc(), mul));
rew.replaceOp(op, res);
return success();
}
} // namespace
void mlir::vector::populateVectorContractLoweringPatterns(
RewritePatternSet &patterns, VectorTransformsOptions options,
PatternBenefit benefit, bool disableOuterProductLowering) {
if (!disableOuterProductLowering)
patterns.add<OuterProductOpLowering>(patterns.getContext(), benefit);
patterns.add<ContractionOpLowering, ContractionOpToMatmulOpLowering,
ContractionOpToOuterProductOpLowering>(
options, patterns.getContext(), benefit);
}
void mlir::vector::populateVectorOuterProductLoweringPatterns(
RewritePatternSet &patterns, PatternBenefit benefit) {
patterns.add<OuterProductOpLowering>(patterns.getContext(), benefit);
}
|