| 1 | //===- SparseReinterpretMap.cpp - reinterpret sparse tensor maps ----------===/ |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | |
| 9 | #include "Utils/CodegenUtils.h" |
| 10 | #include "Utils/IterationGraphSorter.h" |
| 11 | |
| 12 | #include "mlir/Dialect/Bufferization/IR/Bufferization.h" |
| 13 | #include "mlir/Dialect/Linalg/IR/Linalg.h" |
| 14 | #include "mlir/Dialect/Linalg/Utils/Utils.h" |
| 15 | #include "mlir/Dialect/SparseTensor/IR/SparseTensor.h" |
| 16 | #include "mlir/Dialect/SparseTensor/IR/SparseTensorType.h" |
| 17 | #include "mlir/Dialect/SparseTensor/Transforms/Passes.h" |
| 18 | #include "mlir/Dialect/Tensor/IR/Tensor.h" |
| 19 | #include "mlir/IR/AffineExprVisitor.h" |
| 20 | #include "mlir/IR/AffineMap.h" |
| 21 | |
| 22 | using namespace mlir; |
| 23 | using namespace mlir::sparse_tensor; |
| 24 | |
| 25 | namespace { |
| 26 | |
| 27 | //===----------------------------------------------------------------------===// |
| 28 | // File Local Helper classes. |
| 29 | //===----------------------------------------------------------------------===// |
| 30 | |
| 31 | // CRTP to help implementing a rewriter that demaps all its inputs. |
| 32 | template <typename SubClass, typename SourceOp> |
| 33 | struct DemapInsRewriter : public OpRewritePattern<SourceOp> { |
| 34 | using OpRewritePattern<SourceOp>::OpRewritePattern; |
| 35 | using OpAdaptor = typename SourceOp::Adaptor; |
| 36 | |
| 37 | LogicalResult matchAndRewrite(SourceOp op, |
| 38 | PatternRewriter &rewriter) const override { |
| 39 | Location loc = op.getLoc(); |
| 40 | |
| 41 | // Demaps non-trivial inputs. |
| 42 | bool changed = false; |
| 43 | SmallVector<Value> deMappedIns(op->getOperands()); |
| 44 | for (Value &in : deMappedIns) { |
| 45 | if (auto stt = tryGetSparseTensorType(val: in); stt && !stt->isIdentity()) { |
| 46 | in = rewriter.create<ReinterpretMapOp>(location: loc, args: stt->getDemappedType(), args&: in); |
| 47 | changed = true; |
| 48 | } |
| 49 | } |
| 50 | |
| 51 | // CRTP call. |
| 52 | OpAdaptor adaptor(deMappedIns, op); |
| 53 | LogicalResult status = |
| 54 | static_cast<const SubClass *>(this)->rewriteOp(op, adaptor, rewriter); |
| 55 | return changed ? success() : status; |
| 56 | } |
| 57 | }; |
| 58 | |
| 59 | // Flattens an affine expression into a list of AffineDimExprs. |
| 60 | struct AffineDimCollector : public AffineExprVisitor<AffineDimCollector> { |
| 61 | explicit AffineDimCollector(unsigned dimNum) : dims(dimNum){}; |
| 62 | void visitDimExpr(AffineDimExpr expr) { dims.set(expr.getPosition()); } |
| 63 | BitVector dims; |
| 64 | }; |
| 65 | |
| 66 | // Flattens an affine expression into a list of AffineDimExprs. |
| 67 | struct AffineExprAdmissibleVisitor |
| 68 | : public AffineExprVisitor<AffineExprAdmissibleVisitor> { |
| 69 | explicit AffineExprAdmissibleVisitor(bool isOutput) |
| 70 | : admissible(true), isOutput(isOutput){}; |
| 71 | |
| 72 | // We only allow AffineDimExpr on output. |
| 73 | void visitAddExpr(AffineBinaryOpExpr expr) { |
| 74 | if (isOutput) |
| 75 | admissible = false; |
| 76 | } |
| 77 | void visitMulExpr(AffineBinaryOpExpr expr) { |
| 78 | if (isOutput) |
| 79 | admissible = false; |
| 80 | } |
| 81 | |
| 82 | // We disallow mod, floor div and ceil div on inputs. |
| 83 | void visitModExpr(AffineBinaryOpExpr expr) { admissible = false; } |
| 84 | void visitFloorDivExpr(AffineBinaryOpExpr expr) { admissible = false; } |
| 85 | void visitCeilDivExpr(AffineBinaryOpExpr expr) { admissible = false; } |
| 86 | operator bool() { return admissible; } |
| 87 | |
| 88 | private: |
| 89 | bool admissible; |
| 90 | bool isOutput; |
| 91 | }; |
| 92 | |
| 93 | // The first BitVector stores levels where inadmissible exprs are used. |
| 94 | // The second BitVector stores the AffineDimExp that are used by the |
| 95 | // inadmissible expressions. |
| 96 | using InadmissInfo = std::pair<BitVector, BitVector>; |
| 97 | |
| 98 | } // namespace |
| 99 | |
| 100 | //===----------------------------------------------------------------------===// |
| 101 | // File Local Helper methods. |
| 102 | //===----------------------------------------------------------------------===// |
| 103 | |
| 104 | // Collects the inadmissible affine expression imposed on levels. |
| 105 | static InadmissInfo collectInadmissInfo(AffineMap map, bool isOutput) { |
| 106 | auto ret = std::make_pair(x: BitVector(map.getNumResults()), |
| 107 | y: BitVector(map.getNumDims())); |
| 108 | AffineDimCollector collector(map.getNumDims()); |
| 109 | for (unsigned lvl = 0, e = map.getNumResults(); lvl < e; lvl++) { |
| 110 | AffineExprAdmissibleVisitor admissible(isOutput); |
| 111 | admissible.walkPostOrder(expr: map.getResult(idx: lvl)); |
| 112 | if (!admissible) { |
| 113 | // Record the inadmissible level. |
| 114 | ret.first.set(lvl); |
| 115 | // Record the AffineDimExpr that is used in the inadmissible expr. |
| 116 | collector.walkPostOrder(expr: map.getResult(idx: lvl)); |
| 117 | } |
| 118 | } |
| 119 | ret.second = collector.dims; |
| 120 | return ret; |
| 121 | } |
| 122 | |
| 123 | // Builds the AffineMap to replace the idx in idxMap to lvl such that all tht |
| 124 | // inadmissible affine expressions can be eliminated. |
| 125 | // For example, we can rewrite |
| 126 | // idxMap = (d0, d1) -> (d0 floordiv 2, d1 floordiv 3, d0 mod 2, d1 mod 3) |
| 127 | // to |
| 128 | // idxMap = (l0, l1, l2, l3) -> (l0, l1, l2, l3) |
| 129 | // by composing inverse(idxMap), that is |
| 130 | // inverse(idxMap) . idxMap = (l0, l1, l2, l3) -> (l0 * 2 + l2, l1 * 3 + l3) |
| 131 | // -> ((l0 * 2 + l2) floordiv 2, |
| 132 | // (l1 * 3 + l3) floordiv 3, |
| 133 | // (l0 * 2 + l2) mod 2, |
| 134 | // (l1 * 3 + l3) mod 3) = (l0, l1, l2, l3) |
| 135 | // |
| 136 | // This function builds the inverse(idxMap) that replace every dimensions used |
| 137 | // in `info` to levels, and updates the iterator type array `itTps` for the new |
| 138 | // index variable introduced. |
| 139 | // |
| 140 | // Note that the returned affine map does not retain the order of the input |
| 141 | // affine map. Instead, it always uses the first `info.inAdlvls.count()` for the |
| 142 | // replaced levels, and remaining ones for unused dimensions. |
| 143 | // For example, to handle |
| 144 | // idxMap = (d0, d1) -> (d0, d1 floordiv 4, d2 mod 4) |
| 145 | // which is a typical map for block_2to4. The function returns: |
| 146 | // inverse(idxMap) = (l0, l1, d0) -> (d0, l0 * 4 + l1) |
| 147 | // in which, (l0, l1) together replaces `d1`, yet they appear |
| 148 | // before `d0` in the resulting affine map. |
| 149 | // The index (loop) order can later be canonicalized by a topo sort. |
| 150 | static AffineMap |
| 151 | genReplaceDimToLvlMap(const InadmissInfo &info, AffineMap idxMap, |
| 152 | SmallVector<utils::IteratorType> &itTps) { |
| 153 | MLIRContext *ctx = idxMap.getContext(); |
| 154 | auto [inAdLvls, usedDims] = info; |
| 155 | // Note that idxMap does not equal to dim2Lvl map, it is computed by |
| 156 | // composing idx2Dim(dim2Lvl). They are only equal when idx2Dim is an |
| 157 | // ID map. |
| 158 | // TODO: we might fail here, in those case we should really return |
| 159 | // failure instead of assertion error. |
| 160 | auto lvl2Idx = inferLvlToDim(dimToLvl: idxMap, context: ctx); |
| 161 | |
| 162 | assert(lvl2Idx.getNumResults() <= idxMap.getNumDims()); |
| 163 | if (lvl2Idx.getNumResults() != idxMap.getNumDims()) { |
| 164 | // This could happen when some dimensions are projected. |
| 165 | // E.g., idx2Lvl = (*i*, j, k) -> (j, k) |
| 166 | // ==> lvl2Idx = (j, k) -> (j, k) |
| 167 | // In this case, we append the unused dimesion at the end. |
| 168 | // ==> lvl2Idx = (j, k, *i*) -> (*i*, j, k) |
| 169 | SmallVector<AffineExpr> results; |
| 170 | AffineDimCollector usedInLvl(idxMap.getNumDims()); |
| 171 | for (auto e : idxMap.getResults()) |
| 172 | usedInLvl.walkPostOrder(expr: e); |
| 173 | |
| 174 | unsigned curUsedDimID = 0; |
| 175 | unsigned curUnusedDimID = lvl2Idx.getNumDims(); |
| 176 | |
| 177 | BitVector unused = usedInLvl.dims.flip(); |
| 178 | for (unsigned i = 0; i < idxMap.getNumDims(); i++) { |
| 179 | if (unused.test(Idx: i)) |
| 180 | results.push_back(Elt: getAffineDimExpr(position: curUnusedDimID++, context: ctx)); |
| 181 | else |
| 182 | results.push_back(Elt: lvl2Idx.getResult(idx: curUsedDimID++)); |
| 183 | } |
| 184 | lvl2Idx = |
| 185 | AffineMap::get(dimCount: lvl2Idx.getNumDims() + unused.count(), symbolCount: 0, results, context: ctx); |
| 186 | } |
| 187 | assert(lvl2Idx.getNumResults() == idxMap.getNumDims()); |
| 188 | |
| 189 | // We do not need to replace the DimExpr that is not used in inadmissible |
| 190 | // level expressions. We use the first inAdLvl.count() dim to represent the |
| 191 | // replaced level, the remainings are reserved for unchanged ones. |
| 192 | // Note that results from the inverse map computed previously does not follow |
| 193 | // the convention we used, and we need to fix the mismatch below. |
| 194 | unsigned curRepID = 0; |
| 195 | unsigned curOriID = inAdLvls.count(); |
| 196 | SmallVector<AffineExpr> results; |
| 197 | SmallVector<AffineExpr> dimRep(idxMap.getNumResults(), AffineExpr()); |
| 198 | SmallVector<utils::IteratorType> transItTps; |
| 199 | |
| 200 | for (unsigned l : inAdLvls.set_bits()) { |
| 201 | // By our convention, the inadmissible level `l` always appears in the |
| 202 | // leading part (accumulated by curRepID) of the affine map's parameter |
| 203 | // list. Record the mapping so that we can replace all the uses of `l` to |
| 204 | // the correct position after the translation. |
| 205 | dimRep[l] = getAffineDimExpr(position: curRepID++, context: ctx); |
| 206 | // A new index variable is introduced for the inadmissible level, inherit |
| 207 | // the iterator type. E.g., if l0 = d0 floordiv 2, the |
| 208 | // iterator type of l0 equals to the iterator type of d0. |
| 209 | AffineExpr lvlExp = idxMap.getResult(idx: l); |
| 210 | AffineDimCollector collector(idxMap.getNumDims()); |
| 211 | collector.walkPostOrder(expr: lvlExp); |
| 212 | // We assumes a level can only be derived from one dimension. |
| 213 | assert(collector.dims.count() == 1); |
| 214 | transItTps.push_back(Elt: itTps[collector.dims.find_first()]); |
| 215 | } |
| 216 | |
| 217 | for (unsigned d = 0, e = idxMap.getNumDims(); d < e; d++) { |
| 218 | if (usedDims.test(Idx: d)) { |
| 219 | // The dimension is used in some of the inadmissible levels, and it need |
| 220 | // to be inversed. Get the inversion from the inverse map, and fix the |
| 221 | // mismatch captured by the above loop. |
| 222 | results.push_back(Elt: lvl2Idx.getResult(idx: d).replaceDims(dimReplacements: dimRep)); |
| 223 | } else { |
| 224 | // The dimension is not used in any of the inadmissible levels, and it |
| 225 | // does not need to be inversed. Fix the mismatch by mapping it to the |
| 226 | // trailing part of the affine map (accumulated by curOriID). |
| 227 | results.push_back(Elt: getAffineDimExpr(position: curOriID++, context: ctx)); |
| 228 | transItTps.push_back(Elt: itTps[d]); |
| 229 | } |
| 230 | } |
| 231 | unsigned numDim = idxMap.getNumDims() - usedDims.count() + inAdLvls.count(); |
| 232 | // Update iterator type. |
| 233 | itTps.assign(in_start: transItTps.begin(), in_end: transItTps.end()); |
| 234 | return AffineMap::get(dimCount: numDim, symbolCount: 0, results, context: ctx); |
| 235 | } |
| 236 | |
| 237 | // Translates the index map in the linalg::GenericOp from idx->dim map to |
| 238 | // idx->lvl map. Returns failure if the index map can not be translated to an |
| 239 | // admissible form. |
| 240 | // Returns the translated index map array and the iterator type array. |
| 241 | static std::optional<std::pair<ArrayAttr, ArrayAttr>> |
| 242 | translateMap(linalg::GenericOp op, PatternRewriter &rewriter) { |
| 243 | // idxMap is a idx2dim map before reinterpretation. |
| 244 | MLIRContext *ctx = op.getContext(); |
| 245 | SmallVector<AffineMap> idxMapArray = op.getIndexingMapsArray(); |
| 246 | SmallVector<utils::IteratorType> itTps = op.getIteratorTypesArray(); |
| 247 | for (unsigned i = 0, e = idxMapArray.size(); i < e; i++) { |
| 248 | Value tensor = op->getOpOperand(idx: i).get(); |
| 249 | auto stt = tryGetSparseTensorType(val: tensor); |
| 250 | if (stt && !stt->isIdentity()) { |
| 251 | AffineMap dim2Lvl = stt->getDimToLvl(); |
| 252 | // By composing the idx2dim(dim2lvl), we got a idx2lvl Map |
| 253 | idxMapArray[i] = dim2Lvl.compose(map: idxMapArray[i]); |
| 254 | } |
| 255 | } |
| 256 | |
| 257 | // A naive way to handle common constant expressions that arise during dim2lvl |
| 258 | // translation. |
| 259 | auto populateCstMapping = [ctx](DenseMap<AffineExpr, AffineExpr> &cstMapping, |
| 260 | unsigned pos, int64_t lvlSz) { |
| 261 | if (ShapedType::isStatic(dValue: lvlSz)) { |
| 262 | auto c0 = getAffineConstantExpr(constant: 0, context: ctx); |
| 263 | auto lvlExp = getAffineDimExpr(position: pos, context: ctx); |
| 264 | auto szExp = getAffineConstantExpr(constant: lvlSz, context: ctx); |
| 265 | |
| 266 | // lvl floordiv lvlSz = 0 |
| 267 | auto divExp = |
| 268 | getAffineBinaryOpExpr(kind: AffineExprKind::FloorDiv, lhs: lvlExp, rhs: szExp); |
| 269 | cstMapping.try_emplace(Key: divExp, Args&: c0); |
| 270 | |
| 271 | // lvl mod lvlSz = lvl |
| 272 | auto modExp = getAffineBinaryOpExpr(kind: AffineExprKind::Mod, lhs: lvlExp, rhs: szExp); |
| 273 | cstMapping.try_emplace(Key: modExp, Args&: lvlExp); |
| 274 | } |
| 275 | }; |
| 276 | |
| 277 | unsigned boundedNum = 0; |
| 278 | // A fixed-point algorithm. |
| 279 | bool changed = true; |
| 280 | while (changed) { |
| 281 | changed = false; |
| 282 | for (OpOperand &operand : op->getOpOperands()) { |
| 283 | auto stt = tryGetSparseTensorType(val: operand.get()); |
| 284 | // Skip on dense operands. |
| 285 | if (!stt || !stt->getEncoding()) |
| 286 | continue; |
| 287 | |
| 288 | unsigned tid = operand.getOperandNumber(); |
| 289 | bool isOutput = &operand == op.getDpsInitOperand(i: 0); |
| 290 | AffineMap idxMap = idxMapArray[tid]; |
| 291 | InadmissInfo inAdInfo = collectInadmissInfo(map: idxMap, isOutput); |
| 292 | auto [inAdLvls, dimExprs] = inAdInfo; |
| 293 | for (unsigned d : dimExprs.set_bits()) { |
| 294 | // The first `boundedNum` used in the AffineMap is introduced to |
| 295 | // resolve previous inadmissible expressions. We can not replace them |
| 296 | // as it might bring back the inadmissible expressions. |
| 297 | if (d < boundedNum) |
| 298 | return std::nullopt; |
| 299 | } |
| 300 | |
| 301 | if (inAdLvls.count() != 0) { |
| 302 | // Naive constant progagation, should be sufficient to handle block |
| 303 | // sparsity in our cases. |
| 304 | SmallVector<int64_t> lvlShape = stt->getLvlShape(); |
| 305 | DenseMap<AffineExpr, AffineExpr> cstMapping; |
| 306 | unsigned position = 0; |
| 307 | for (unsigned lvl : inAdLvls.set_bits()) { |
| 308 | int64_t lvlSz = lvlShape[lvl]; |
| 309 | populateCstMapping(cstMapping, position, lvlSz); |
| 310 | position++; |
| 311 | } |
| 312 | |
| 313 | AffineMap lvl2Idx = genReplaceDimToLvlMap(info: inAdInfo, idxMap, itTps); |
| 314 | // Compose the lvl2Idx Map to all AffineIdxMap to eliminate |
| 315 | // inadmissible expressions. |
| 316 | for (unsigned tid = 0, e = idxMapArray.size(); tid < e; tid++) { |
| 317 | AffineMap transMap = idxMapArray[tid].compose(map: lvl2Idx); |
| 318 | idxMapArray[tid] = transMap.replace( |
| 319 | map: cstMapping, /*numResultDims=*/transMap.getNumDims(), |
| 320 | /*numResultSyms=*/0); |
| 321 | } |
| 322 | changed = true; |
| 323 | boundedNum += inAdLvls.count(); |
| 324 | } |
| 325 | } |
| 326 | }; |
| 327 | |
| 328 | SmallVector<Attribute> iterAttr = |
| 329 | llvm::map_to_vector(C&: itTps, F: [ctx](auto itTp) -> Attribute { |
| 330 | return linalg::IteratorTypeAttr::get(context: ctx, value: itTp); |
| 331 | }); |
| 332 | |
| 333 | return std::make_pair(x: rewriter.getAffineMapArrayAttr(values: idxMapArray), |
| 334 | y: rewriter.getArrayAttr(value: iterAttr)); |
| 335 | } |
| 336 | |
| 337 | // Generates a "de"mapping reinterpretation of the map. |
| 338 | static Value genDemap(OpBuilder &builder, SparseTensorEncodingAttr enc, |
| 339 | Value val) { |
| 340 | return builder.create<ReinterpretMapOp>(location: val.getLoc(), args: enc.withoutDimToLvl(), |
| 341 | args&: val); |
| 342 | } |
| 343 | |
| 344 | // Generates a "re"mapping reinterpretation of the map. |
| 345 | static Value genRemap(OpBuilder &builder, SparseTensorEncodingAttr enc, |
| 346 | Value val) { |
| 347 | return builder.create<ReinterpretMapOp>(location: val.getLoc(), args&: enc, args&: val); |
| 348 | } |
| 349 | |
| 350 | static SmallVector<Value> remapValueRange(OpBuilder &rewriter, TypeRange types, |
| 351 | ValueRange outs) { |
| 352 | SmallVector<Value> ret(outs); |
| 353 | assert(outs.size() == types.size()); |
| 354 | for (auto [r, t] : llvm::zip(t&: ret, u&: types)) |
| 355 | if (r.getType() != t) |
| 356 | r = rewriter.create<ReinterpretMapOp>(location: r.getLoc(), args&: t, args&: r); |
| 357 | return ret; |
| 358 | } |
| 359 | |
| 360 | namespace { |
| 361 | |
| 362 | //===----------------------------------------------------------------------===// |
| 363 | // Rewriting rules for linalg generic ops. |
| 364 | //===----------------------------------------------------------------------===// |
| 365 | |
| 366 | /// Sparse rewriting rule for the generic `linalg` operation. |
| 367 | struct GenericOpReinterpretMap |
| 368 | : public DemapInsRewriter<GenericOpReinterpretMap, linalg::GenericOp> { |
| 369 | public: |
| 370 | using DemapInsRewriter::DemapInsRewriter; |
| 371 | LogicalResult rewriteOp(linalg::GenericOp linalgOp, OpAdaptor adaptor, |
| 372 | PatternRewriter &rewriter) const { |
| 373 | // Only rewrite single output operations with pure (sparse) tensor |
| 374 | // semantics. |
| 375 | if (linalgOp.getNumDpsInits() != 1 || !linalgOp.hasPureTensorSemantics() || |
| 376 | !hasAnySparseOperandOrResult(op: linalgOp) || |
| 377 | !hasAnyNonIdentityOperandsOrResults(op: linalgOp)) |
| 378 | return failure(); |
| 379 | |
| 380 | // Try translating the index map. |
| 381 | auto transMap = translateMap(op: linalgOp, rewriter); |
| 382 | if (!transMap) |
| 383 | return rewriter.notifyMatchFailure( |
| 384 | arg&: linalgOp, msg: "the sparse kernel can not be sparsified." ); |
| 385 | |
| 386 | // On success, replace update the linalg operands and maps in place. |
| 387 | Value res = linalgOp.getResult(i: 0); |
| 388 | auto stt = tryGetSparseTensorType(val: res); |
| 389 | auto [idxMap, itTp] = *transMap; |
| 390 | |
| 391 | rewriter.startOpModification(op: linalgOp); |
| 392 | linalgOp.setIndexingMapsAttr(idxMap); |
| 393 | linalgOp.setIteratorTypesAttr(itTp); |
| 394 | // Use demapped arguments. |
| 395 | linalgOp.getInputsMutable().assign(values: adaptor.getInputs()); |
| 396 | linalgOp.getDpsInitsMutable().assign(values: adaptor.getOutputs()); |
| 397 | res.setType(adaptor.getOutputs()[0].getType()); |
| 398 | rewriter.finalizeOpModification(op: linalgOp); |
| 399 | |
| 400 | rewriter.setInsertionPointAfter(linalgOp); |
| 401 | if (stt && stt->hasEncoding()) { |
| 402 | Value t = genRemap(builder&: rewriter, enc: stt->getEncoding(), val: res); |
| 403 | rewriter.replaceAllUsesExcept(from: res, to: t, exceptedUser: t.getDefiningOp()); |
| 404 | } |
| 405 | return success(); |
| 406 | } |
| 407 | }; |
| 408 | |
| 409 | struct GenericOpScheduler : public OpRewritePattern<linalg::GenericOp> { |
| 410 | using OpRewritePattern::OpRewritePattern; |
| 411 | LogicalResult matchAndRewrite(linalg::GenericOp linalgOp, |
| 412 | PatternRewriter &rewriter) const override { |
| 413 | if (linalgOp.getNumDpsInits() != 1 || !linalgOp.hasPureTensorSemantics() || |
| 414 | hasAnyNonIdentityOperandsOrResults(op: linalgOp) || // need demap first |
| 415 | !hasAnySparseOperandOrResult(op: linalgOp)) { |
| 416 | return failure(); |
| 417 | } |
| 418 | |
| 419 | const StringRef sorted = "sorted" ; |
| 420 | if (linalgOp->hasAttr(name: sorted)) |
| 421 | return failure(); |
| 422 | |
| 423 | auto scheduler = IterationGraphSorter::fromGenericOp(genericOp: linalgOp); |
| 424 | bool isAdmissible = false; |
| 425 | AffineMap order; |
| 426 | // A const list of all masks that we used for iteration graph |
| 427 | // computation. Must be ordered from more strict to less strict. |
| 428 | // Ideally (though might not be guaranteed), the earlier a constraint mask |
| 429 | // can be satisfied, the faster the generated kernel will be. |
| 430 | const auto allMasks = {SortMask::kIncludeAll, SortMask::kIncludeDense, |
| 431 | SortMask::kIncludeDenseInput, |
| 432 | SortMask::kIncludeDenseOutput, |
| 433 | SortMask::kSparseOnly}; |
| 434 | for (const SortMask mask : allMasks) { |
| 435 | order = scheduler.sort(mask); |
| 436 | if (order) { |
| 437 | if (isAdmissibleOrder(linalgOp, order)) { |
| 438 | isAdmissible = true; |
| 439 | break; |
| 440 | } |
| 441 | // else try a set of less strict constraints. |
| 442 | } |
| 443 | } |
| 444 | |
| 445 | if (!order) { |
| 446 | // Cycles detected. |
| 447 | if (failed(Result: resolveCycle(scheduler, linalgOp, rewriter))) { |
| 448 | return rewriter.notifyMatchFailure( |
| 449 | arg&: linalgOp, msg: "the sparse kernel can not be scheduled: loop detected." ); |
| 450 | } |
| 451 | return success(); |
| 452 | } |
| 453 | |
| 454 | if (!isAdmissible) { |
| 455 | return rewriter.notifyMatchFailure( |
| 456 | arg&: linalgOp, msg: "the sparse kernel can not be scheduled." ); |
| 457 | } |
| 458 | |
| 459 | // Marks the GenericOp to avoid recursive matching. |
| 460 | rewriter.modifyOpInPlace(root: linalgOp, callable: [&]() { |
| 461 | linalgOp->setAttr(name: sorted, value: rewriter.getBoolAttr(value: true)); |
| 462 | }); |
| 463 | |
| 464 | // Already sorted. |
| 465 | if (order.isIdentity()) |
| 466 | return success(); |
| 467 | |
| 468 | assert(order.isPermutation()); |
| 469 | // `order` is orignial loop -> sorted loop map |
| 470 | ArrayAttr preItTypes = linalgOp.getIteratorTypesAttr(); |
| 471 | SmallVector<Attribute> curItTypes; |
| 472 | curItTypes.reserve(N: preItTypes.size()); |
| 473 | for (AffineExpr expr : order.getResults()) { |
| 474 | unsigned loopID = llvm::cast<AffineDimExpr>(Val&: expr).getPosition(); |
| 475 | curItTypes.push_back(Elt: preItTypes[loopID]); |
| 476 | } |
| 477 | |
| 478 | // Inverse `order` to get sorted loop -> original loop map |
| 479 | order = inversePermutation(map: order); |
| 480 | SmallVector<AffineMap> idxMaps = linalgOp.getIndexingMapsArray(); |
| 481 | for (AffineMap &idxMap : idxMaps) |
| 482 | idxMap = idxMap.compose(map: order); // sorted loop -> lvl map |
| 483 | |
| 484 | rewriter.startOpModification(op: linalgOp); |
| 485 | linalgOp.setIndexingMapsAttr(rewriter.getAffineMapArrayAttr(values: idxMaps)); |
| 486 | linalgOp.setIteratorTypesAttr(rewriter.getArrayAttr(value: curItTypes)); |
| 487 | rewriter.finalizeOpModification(op: linalgOp); |
| 488 | |
| 489 | return success(); |
| 490 | } |
| 491 | |
| 492 | private: |
| 493 | /// Whether the loop order is admissible by sparsification. |
| 494 | static bool isAdmissibleOrder(linalg::GenericOp linalgOp, AffineMap order) { |
| 495 | if (!hasAnySparseResult(op: linalgOp)) |
| 496 | return true; |
| 497 | |
| 498 | OpOperand *lhs = linalgOp.getDpsInitOperand(i: 0); |
| 499 | unsigned nest = 0; |
| 500 | const auto iteratorTypes = linalgOp.getIteratorTypesArray(); |
| 501 | for (const AffineExpr l : order.getResults()) { |
| 502 | unsigned loopId = llvm::cast<AffineDimExpr>(Val: l).getPosition(); |
| 503 | auto itTp = |
| 504 | cast<linalg::IteratorTypeAttr>(Val: linalgOp.getIteratorTypes()[loopId]); |
| 505 | if (linalg::isReductionIterator(iteratorType: itTp.getValue())) |
| 506 | break; // terminate at first reduction |
| 507 | nest++; |
| 508 | } |
| 509 | // Determine admissible dynamic insertion situations: |
| 510 | // (1) fully injective, since there are no reductions, |
| 511 | // (2) admissible 1-d expansion in innermost dimension. |
| 512 | return static_cast<int64_t>(nest) >= linalgOp.getRank(opOperand: lhs) - 1; |
| 513 | }; |
| 514 | |
| 515 | // Last resort cycle resolution. |
| 516 | static LogicalResult resolveCycle(IterationGraphSorter &scheduler, |
| 517 | linalg::LinalgOp linalgOp, |
| 518 | PatternRewriter &rewriter) { |
| 519 | // Compute topological sort while leaving out every sparse input tensor in |
| 520 | // succession until an acylic iteration graph results. |
| 521 | for (OpOperand *t : linalgOp.getDpsInputOperands()) { |
| 522 | Value tval = t->get(); |
| 523 | auto srcEnc = getSparseTensorEncoding(type: tval.getType()); |
| 524 | // The constraints introduced by compound index expression are |
| 525 | // complicated. Skip them. |
| 526 | AffineMap idxMap = linalgOp.getMatchingIndexingMap(opOperand: t); |
| 527 | bool hasCompExpr = llvm::any_of(Range: idxMap.getResults(), P: [](AffineExpr exp) { |
| 528 | return !llvm::isa<AffineDimExpr>(Val: exp); |
| 529 | }); |
| 530 | if (!srcEnc || hasCompExpr) |
| 531 | continue; |
| 532 | |
| 533 | // Try scheduling loop without constraints from `tval`. |
| 534 | AffineMap order = scheduler.sort(mask: SortMask::kSparseOnly, ignored: tval); |
| 535 | if (!order) // still cyclic |
| 536 | continue; |
| 537 | |
| 538 | // Found an input tensor that resolves the cycle by inserting a |
| 539 | // conversion into a sparse tensor that adheres to the iteration |
| 540 | // graph order. |
| 541 | auto stt = getSparseTensorType(val: tval); |
| 542 | assert(stt.isIdentity()); |
| 543 | order = inversePermutation(map: order); |
| 544 | // sorted loop -> lvl map. |
| 545 | idxMap = idxMap.compose(map: order); |
| 546 | |
| 547 | // Found a permutation such that the results in `idxMap` is sorted. |
| 548 | // For example, |
| 549 | // (d0, d1, d2, d3) -> (d2, d1, d0) |
| 550 | // loops are scheduled in order of d0->d1->d2->d3, to resolve the cycle, |
| 551 | // we find a permutation, perm(d2, d1, d0) -> (d0, d1, d2), such that the |
| 552 | // transposed tensor's levels are visited in the same order as the loop |
| 553 | // scheduling order. |
| 554 | SmallVector<std::pair<unsigned, unsigned>> lvlSeq; |
| 555 | for (AffineExpr expr : idxMap.getResults()) { |
| 556 | unsigned lvl = llvm::cast<AffineDimExpr>(Val&: expr).getPosition(); |
| 557 | lvlSeq.push_back(Elt: std::make_pair(x&: lvl, y: lvlSeq.size())); |
| 558 | } |
| 559 | llvm::sort(C&: lvlSeq, Comp: llvm::less_first()); |
| 560 | SmallVector<unsigned> perm = |
| 561 | llvm::to_vector(Range: llvm::make_second_range(c&: lvlSeq)); |
| 562 | auto dimToLvl = AffineMap::getPermutationMap(permutation: perm, context: linalgOp.getContext()); |
| 563 | // The result of the idxMap must be unsorted. |
| 564 | assert(!dimToLvl.isIdentity()); |
| 565 | |
| 566 | // Inserting the transpose |
| 567 | rewriter.setInsertionPoint(linalgOp); |
| 568 | RankedTensorType dstTp = stt.withDimToLvl(dimToLvl).getRankedTensorType(); |
| 569 | Value dst = rewriter.create<ConvertOp>(location: tval.getLoc(), args&: dstTp, args&: tval); |
| 570 | rewriter.modifyOpInPlace(root: linalgOp, callable: [&]() { |
| 571 | linalgOp->setOperand(idx: t->getOperandNumber(), value: dst); |
| 572 | }); |
| 573 | |
| 574 | // Release the transposed form afterwards. |
| 575 | // TODO: CSE when used in more than one following op? |
| 576 | rewriter.setInsertionPointAfter(linalgOp); |
| 577 | rewriter.create<bufferization::DeallocTensorOp>(location: dst.getLoc(), args&: dst); |
| 578 | |
| 579 | return success(); |
| 580 | } |
| 581 | // Cannot be resolved with a single conversion. |
| 582 | // TODO: convert more than one? |
| 583 | return failure(); |
| 584 | } |
| 585 | }; |
| 586 | |
| 587 | //===----------------------------------------------------------------------===// |
| 588 | // Reinterpret Map Rewriters for operations other than linalg.generics |
| 589 | //===----------------------------------------------------------------------===// |
| 590 | |
| 591 | template <typename AllocOp> |
| 592 | struct TensorAllocDemapper : public OpRewritePattern<AllocOp> { |
| 593 | using OpRewritePattern<AllocOp>::OpRewritePattern; |
| 594 | LogicalResult matchAndRewrite(AllocOp op, |
| 595 | PatternRewriter &rewriter) const override { |
| 596 | if (!hasAnyNonIdentityOperandsOrResults(op)) |
| 597 | return failure(); |
| 598 | |
| 599 | Location loc = op.getLoc(); |
| 600 | auto stt = getSparseTensorType(op.getResult()); |
| 601 | |
| 602 | SmallVector<Value> maxDimCrds; |
| 603 | maxDimCrds.reserve(N: stt.getDimRank()); |
| 604 | ValueRange dynSz = op.getDynamicSizes(); |
| 605 | for (int64_t dimSz : stt.getDimShape()) { |
| 606 | if (ShapedType::isDynamic(dValue: dimSz)) { |
| 607 | Value maxCrd = rewriter.create<arith::SubIOp>( |
| 608 | location: loc, args: dynSz.front(), args: constantIndex(builder&: rewriter, loc, i: 1)); |
| 609 | maxDimCrds.push_back(Elt: maxCrd); |
| 610 | dynSz = dynSz.drop_front(); |
| 611 | } else { |
| 612 | maxDimCrds.push_back(Elt: constantIndex(builder&: rewriter, loc, i: dimSz - 1)); |
| 613 | } |
| 614 | } |
| 615 | |
| 616 | ValueRange maxLvlCrds = stt.translateCrds(rewriter, loc, maxDimCrds, |
| 617 | CrdTransDirectionKind::dim2lvl); |
| 618 | auto lvlShape = stt.getLvlShape(); |
| 619 | SmallVector<Value> dynLvlSzs; |
| 620 | for (unsigned i = 0, e = lvlShape.size(); i < e; i++) { |
| 621 | if (ShapedType::isDynamic(dValue: lvlShape[i])) { |
| 622 | Value sz = rewriter.create<arith::AddIOp>( |
| 623 | location: loc, args: maxLvlCrds[i], args: constantIndex(builder&: rewriter, loc, i: 1)); |
| 624 | dynLvlSzs.push_back(Elt: sz); |
| 625 | } |
| 626 | } |
| 627 | |
| 628 | assert(dynSz.empty()); // should have consumed all. |
| 629 | rewriter.startOpModification(op); |
| 630 | op->setOperands(dynLvlSzs); |
| 631 | op.getResult().setType(stt.getDemappedType()); |
| 632 | rewriter.finalizeOpModification(op); |
| 633 | rewriter.setInsertionPointAfter(op); |
| 634 | |
| 635 | Value t = genRemap(rewriter, stt.getEncoding(), op.getResult()); |
| 636 | rewriter.replaceAllUsesExcept(op.getResult(), t, t.getDefiningOp()); |
| 637 | return success(); |
| 638 | } |
| 639 | }; |
| 640 | |
| 641 | struct TensorInsertDemapper |
| 642 | : public DemapInsRewriter<TensorInsertDemapper, tensor::InsertOp> { |
| 643 | using DemapInsRewriter::DemapInsRewriter; |
| 644 | LogicalResult rewriteOp(tensor::InsertOp op, OpAdaptor adaptor, |
| 645 | PatternRewriter &rewriter) const { |
| 646 | if (!hasAnySparseResult(op) || !hasAnyNonIdentityOperandsOrResults(op)) |
| 647 | return failure(); |
| 648 | |
| 649 | Location loc = op.getLoc(); |
| 650 | auto stt = getSparseTensorType(val: op.getResult()); |
| 651 | ValueRange lvlCrd = stt.translateCrds(builder&: rewriter, loc, crds: op.getIndices(), |
| 652 | dir: CrdTransDirectionKind::dim2lvl); |
| 653 | auto insertOp = rewriter.create<tensor::InsertOp>( |
| 654 | location: loc, args: op.getScalar(), args: adaptor.getDest(), args&: lvlCrd); |
| 655 | |
| 656 | Value out = genRemap(builder&: rewriter, enc: stt.getEncoding(), val: insertOp.getResult()); |
| 657 | rewriter.replaceOp(op, newValues: out); |
| 658 | return success(); |
| 659 | } |
| 660 | }; |
| 661 | |
| 662 | struct SparseAssembleDemapper : public OpRewritePattern<AssembleOp> { |
| 663 | using OpRewritePattern::OpRewritePattern; |
| 664 | LogicalResult matchAndRewrite(AssembleOp op, |
| 665 | PatternRewriter &rewriter) const override { |
| 666 | if (!hasAnyNonIdentityOperandsOrResults(op)) |
| 667 | return failure(); |
| 668 | |
| 669 | assert(hasAnySparseResult(op)); |
| 670 | auto stt = getSparseTensorType(val: op.getResult()); |
| 671 | rewriter.modifyOpInPlace( |
| 672 | root: op, callable: [&op, &stt]() { op.getResult().setType(stt.getDemappedType()); }); |
| 673 | rewriter.setInsertionPointAfter(op); |
| 674 | Value out = genRemap(builder&: rewriter, enc: stt.getEncoding(), val: op.getResult()); |
| 675 | rewriter.replaceAllUsesExcept(from: op, to: out, exceptedUser: out.getDefiningOp()); |
| 676 | return success(); |
| 677 | } |
| 678 | }; |
| 679 | |
| 680 | struct SparseDisassembleDemapper |
| 681 | : public DemapInsRewriter<SparseDisassembleDemapper, DisassembleOp> { |
| 682 | using DemapInsRewriter::DemapInsRewriter; |
| 683 | LogicalResult rewriteOp(DisassembleOp op, OpAdaptor adaptor, |
| 684 | PatternRewriter &rewriter) const { |
| 685 | if (!hasAnyNonIdentityOperandsOrResults(op)) |
| 686 | return failure(); |
| 687 | |
| 688 | assert(hasAnySparseOperandOrResult(op)); |
| 689 | rewriter.modifyOpInPlace(root: op, callable: [&op, &adaptor]() { |
| 690 | op.getTensorMutable().assign(value: adaptor.getTensor()); |
| 691 | }); |
| 692 | return success(); |
| 693 | } |
| 694 | }; |
| 695 | |
| 696 | struct ForeachOpDemapper |
| 697 | : public DemapInsRewriter<ForeachOpDemapper, ForeachOp> { |
| 698 | using DemapInsRewriter::DemapInsRewriter; |
| 699 | LogicalResult rewriteOp(ForeachOp op, OpAdaptor adaptor, |
| 700 | PatternRewriter &rewriter) const { |
| 701 | // Only handle operations with sparse input/output with non-identity dim2lvl |
| 702 | // maps. |
| 703 | if (!hasAnyNonIdentityOperandsOrResults(op)) |
| 704 | return failure(); |
| 705 | |
| 706 | // TODO: demap constant as well. |
| 707 | if (auto constOp = op.getTensor().getDefiningOp<arith::ConstantOp>()) |
| 708 | if (auto attr = dyn_cast<SparseElementsAttr>(Val: constOp.getValue())) |
| 709 | return failure(); |
| 710 | |
| 711 | Location loc = op.getLoc(); |
| 712 | // Cache the type information since we update the foreach op in-place. |
| 713 | auto srcStt = getSparseTensorType(val: op.getTensor()); |
| 714 | SmallVector<Type> prevRetTps(op.getResultTypes()); |
| 715 | |
| 716 | rewriter.startOpModification(op); |
| 717 | op.getTensorMutable().assign(value: adaptor.getTensor()); |
| 718 | op.getInitArgsMutable().assign(values: adaptor.getInitArgs()); |
| 719 | // Update results' types. |
| 720 | for (auto r : op.getResults()) |
| 721 | if (auto stt = tryGetSparseTensorType(val: r); stt && !stt->isIdentity()) |
| 722 | r.setType(stt->getDemappedType()); |
| 723 | |
| 724 | Level lvlRank = getSparseTensorType(val: adaptor.getTensor()).getLvlRank(); |
| 725 | // Update the foreach body. |
| 726 | SmallVector<Type> blockArgTps(lvlRank, rewriter.getIndexType()); |
| 727 | blockArgTps.push_back(Elt: srcStt.getElementType()); |
| 728 | blockArgTps.append(in_start: adaptor.getInitArgs().getTypes().begin(), |
| 729 | in_end: adaptor.getInitArgs().getTypes().end()); |
| 730 | Block *body = op.getBody(); |
| 731 | // Block Args: [dimCrd, val, initArgs] |
| 732 | unsigned preArgNum = body->getNumArguments(); |
| 733 | for (Type t : blockArgTps) |
| 734 | body->addArgument(type: t, loc); |
| 735 | |
| 736 | // Block Args: [dimCrd, val, initArgs, lvlCrds, val, DemappedArgs] |
| 737 | rewriter.setInsertionPointToStart(body); |
| 738 | ValueRange lvlCrds = body->getArguments().slice(N: preArgNum, M: lvlRank); |
| 739 | |
| 740 | ValueRange dimCrds = srcStt.translateCrds(builder&: rewriter, loc, crds: lvlCrds, |
| 741 | dir: CrdTransDirectionKind::lvl2dim); |
| 742 | rewriter.replaceAllUsesWith( |
| 743 | from: body->getArguments().take_front(N: srcStt.getDimRank()), to: dimCrds); |
| 744 | body->eraseArguments(start: 0, num: srcStt.getDimRank()); |
| 745 | // Block Args: [val, initArgs, lvlCrds, val, DemappedArgs] |
| 746 | unsigned numInitArgs = op.getInitArgs().size(); |
| 747 | rewriter.replaceAllUsesWith(from: body->getArgument(i: 0), |
| 748 | to: body->getArgument(i: lvlRank + numInitArgs + 1)); |
| 749 | body->eraseArgument(index: 0); |
| 750 | // Block Args: [initArgs, lvlCrds, val, DemappedArgs] |
| 751 | ValueRange srcArgs = body->getArguments().take_front(N: numInitArgs); |
| 752 | ValueRange dstArgs = body->getArguments().take_back(N: numInitArgs); |
| 753 | // Remap back before replacement. |
| 754 | SmallVector<Value> reMappedArgs = |
| 755 | remapValueRange(rewriter, types: srcArgs.getTypes(), outs: dstArgs); |
| 756 | rewriter.replaceAllUsesWith(from: srcArgs, to: reMappedArgs); |
| 757 | body->eraseArguments(start: 0, num: numInitArgs); |
| 758 | // Block Args: [lvlCrds, DemappedArgs] and we are done. |
| 759 | |
| 760 | // Update yield operations. |
| 761 | if (numInitArgs != 0) { |
| 762 | rewriter.setInsertionPointToEnd(body); |
| 763 | auto yield = llvm::cast<YieldOp>(Val: body->getTerminator()); |
| 764 | if (auto stt = tryGetSparseTensorType(val: yield.getSingleResult()); |
| 765 | stt && !stt->isIdentity()) { |
| 766 | Value y = |
| 767 | genDemap(builder&: rewriter, enc: stt->getEncoding(), val: yield.getSingleResult()); |
| 768 | rewriter.create<YieldOp>(location: loc, args&: y); |
| 769 | rewriter.eraseOp(op: yield); |
| 770 | } |
| 771 | } |
| 772 | rewriter.finalizeOpModification(op); |
| 773 | |
| 774 | rewriter.setInsertionPointAfter(op); |
| 775 | SmallVector<Value> outs = |
| 776 | remapValueRange(rewriter, types: prevRetTps, outs: op.getResults()); |
| 777 | |
| 778 | // Replace all the uses of the foreach results, expect the use in |
| 779 | // reinterpret_map used to remap the output. |
| 780 | for (auto [from, to] : llvm::zip(t: op.getResults(), u&: outs)) |
| 781 | rewriter.replaceAllUsesExcept(from, to, exceptedUser: to.getDefiningOp()); |
| 782 | |
| 783 | return success(); |
| 784 | } |
| 785 | }; |
| 786 | |
| 787 | } // namespace |
| 788 | |
| 789 | void mlir::populateSparseReinterpretMap(RewritePatternSet &patterns, |
| 790 | ReinterpretMapScope scope) { |
| 791 | if (scope == ReinterpretMapScope::kAll || |
| 792 | scope == ReinterpretMapScope::kGenericOnly) { |
| 793 | patterns.add<GenericOpReinterpretMap, GenericOpScheduler>( |
| 794 | arg: patterns.getContext()); |
| 795 | } |
| 796 | if (scope == ReinterpretMapScope::kAll || |
| 797 | scope == ReinterpretMapScope::kExceptGeneric) { |
| 798 | patterns.add<TensorAllocDemapper<bufferization::AllocTensorOp>, |
| 799 | TensorAllocDemapper<tensor::EmptyOp>, SparseAssembleDemapper, |
| 800 | SparseDisassembleDemapper, TensorInsertDemapper, |
| 801 | ForeachOpDemapper>(arg: patterns.getContext()); |
| 802 | } |
| 803 | } |
| 804 | |