xref: /llvm-project/mlir/lib/Conversion/SCFToGPU/SCFToGPU.cpp (revision 0e944a30954e666cba2bf17497fafe835e4b3519)
1 //===- SCFToGPU.cpp - Convert an affine loop nest to a GPU kernel -------===//
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 // This implements a straightforward conversion of an loop nest into a GPU
10 // kernel.  The caller is expected to guarantee that the conversion is correct
11 // or to further transform the kernel to ensure correctness.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "mlir/Conversion/SCFToGPU/SCFToGPU.h"
16 
17 #include "mlir/Conversion/AffineToStandard/AffineToStandard.h"
18 #include "mlir/Dialect/Affine/IR/AffineOps.h"
19 #include "mlir/Dialect/Arith/IR/Arith.h"
20 #include "mlir/Dialect/GPU/IR/GPUDialect.h"
21 #include "mlir/Dialect/GPU/Transforms/ParallelLoopMapper.h"
22 #include "mlir/Dialect/MemRef/IR/MemRef.h"
23 #include "mlir/Dialect/SCF/IR/SCF.h"
24 #include "mlir/IR/AffineExpr.h"
25 #include "mlir/IR/Builders.h"
26 #include "mlir/IR/IRMapping.h"
27 #include "mlir/Interfaces/SideEffectInterfaces.h"
28 #include "mlir/Pass/Pass.h"
29 #include "mlir/Transforms/DialectConversion.h"
30 #include "mlir/Transforms/Passes.h"
31 #include "mlir/Transforms/RegionUtils.h"
32 #include "llvm/ADT/Sequence.h"
33 #include "llvm/Support/Debug.h"
34 #include <optional>
35 
36 #define DEBUG_TYPE "loops-to-gpu"
37 
38 using namespace mlir;
39 using namespace mlir::affine;
40 using namespace mlir::scf;
41 
42 // Name of internal attribute to mark visited operations during conversion.
43 //
44 // NOTE: The conversion originally used the following legality criteria:
45 //   `!parallelOp->hasAttr(gpu::getMappingAttrName())`
46 // But the provided pattern might reject some cases based on more detailed
47 // analysis of the `mapping` attribute.
48 // To avoid dialect conversion failure due to non-converted illegal operation
49 // we use this extra Unit attribute as a marker, that the operation was checked
50 // by the pattern and is should be considered as legal in the following legality
51 // checks. The `finalizeParallelLoopToGPUConversion` function performs clean up
52 // of this extra attributes ans is supposed to be called after the dialect
53 // conversion.
54 //
55 // TODO: Implement a cleaner solution, factoring out the "matching" logic
56 // from the pattern and its callees into a separate function that can be called
57 // from both the pattern and the op legality check.
58 static constexpr StringLiteral kVisitedAttrName = "SCFToGPU_visited";
59 
60 // Extract an indexed value from KernelDim3.
61 static Value getDim3Value(const gpu::KernelDim3 &dim3, unsigned pos) {
62   switch (pos) {
63   case 0:
64     return dim3.x;
65   case 1:
66     return dim3.y;
67   case 2:
68     return dim3.z;
69   default:
70     llvm_unreachable("dim3 position out of bounds");
71   }
72   return nullptr;
73 }
74 
75 // Get the lower bound-related operands of a loop operation.
76 static Operation::operand_range getLowerBoundOperands(AffineForOp forOp) {
77   return forOp.getLowerBoundOperands();
78 }
79 
80 // Get the upper bound-related operands of a loop operation.
81 static Operation::operand_range getUpperBoundOperands(AffineForOp forOp) {
82   return forOp.getUpperBoundOperands();
83 }
84 
85 // Get a Value that corresponds to the loop step.  If the step is an attribute,
86 // materialize a corresponding constant using builder.
87 static Value getOrCreateStep(AffineForOp forOp, OpBuilder &builder) {
88   return builder.create<arith::ConstantIndexOp>(forOp.getLoc(),
89                                                 forOp.getStepAsInt());
90 }
91 
92 // Get a Value for the loop lower bound.  If the value requires computation,
93 // materialize the instructions using builder.
94 static Value getOrEmitLowerBound(AffineForOp forOp, OpBuilder &builder) {
95   return lowerAffineLowerBound(forOp, builder);
96 }
97 
98 // Get a Value for the loop upper bound.  If the value requires computation,
99 // materialize the instructions using builder.
100 static Value getOrEmitUpperBound(AffineForOp forOp, OpBuilder &builder) {
101   return lowerAffineUpperBound(forOp, builder);
102 }
103 
104 // Check the structure of the loop nest:
105 //   - there are enough loops to map to numDims;
106 //   - the loops are perfectly nested;
107 //   - the loop bounds can be computed above the outermost loop.
108 // This roughly corresponds to the "matcher" part of the pattern-based
109 // rewriting infrastructure.
110 static LogicalResult checkAffineLoopNestMappableImpl(AffineForOp forOp,
111                                                      unsigned numDims) {
112   Region &limit = forOp.getRegion();
113   for (unsigned i = 0, e = numDims; i < e; ++i) {
114     Operation *nested = &forOp.getBody()->front();
115     if (!areValuesDefinedAbove(getLowerBoundOperands(forOp), limit) ||
116         !areValuesDefinedAbove(getUpperBoundOperands(forOp), limit))
117       return forOp.emitError(
118           "loops with bounds depending on other mapped loops "
119           "are not supported");
120 
121     // The innermost loop can have an arbitrary body, skip the perfect nesting
122     // check for it.
123     if (i == e - 1)
124       break;
125 
126     auto begin = forOp.getBody()->begin(), end = forOp.getBody()->end();
127     if (forOp.getBody()->empty() || std::next(begin, 2) != end)
128       return forOp.emitError("expected perfectly nested loops in the body");
129 
130     if (!(forOp = dyn_cast<AffineForOp>(nested)))
131       return nested->emitError("expected a nested loop");
132   }
133   return success();
134 }
135 
136 static LogicalResult checkAffineLoopNestMappable(AffineForOp forOp,
137                                                  unsigned numBlockDims,
138                                                  unsigned numThreadDims) {
139   if (numBlockDims < 1 || numThreadDims < 1) {
140     LLVM_DEBUG(llvm::dbgs() << "nothing to map");
141     return success();
142   }
143 
144   if (numBlockDims > 3) {
145     return forOp.emitError("cannot map to more than 3 block dimensions");
146   }
147   if (numThreadDims > 3) {
148     return forOp.emitError("cannot map to more than 3 thread dimensions");
149   }
150   return checkAffineLoopNestMappableImpl(forOp, numBlockDims + numThreadDims);
151 }
152 
153 namespace {
154 // Helper structure that holds common state of the loop to GPU kernel
155 // conversion.
156 struct AffineLoopToGpuConverter {
157   std::optional<AffineForOp> collectBounds(AffineForOp forOp,
158                                            unsigned numLoops);
159 
160   void createLaunch(AffineForOp rootForOp, AffineForOp innermostForOp,
161                     unsigned numBlockDims, unsigned numThreadDims);
162 
163   // Ranges of the loops mapped to blocks or threads.
164   SmallVector<Value, 6> dims;
165   // Lower bounds of the loops mapped to blocks or threads.
166   SmallVector<Value, 6> lbs;
167   // Induction variables of the loops mapped to blocks or threads.
168   SmallVector<Value, 6> ivs;
169   // Steps of the loops mapped to blocks or threads.
170   SmallVector<Value, 6> steps;
171 };
172 } // namespace
173 
174 // Collect ranges, bounds, steps and induction variables in preparation for
175 // mapping a loop nest of depth "numLoops" rooted at "forOp" to a GPU kernel.
176 // This may fail if the IR for computing loop bounds cannot be constructed, for
177 // example if an affine loop uses semi-affine maps. Return the last loop to be
178 // mapped on success, std::nullopt on failure.
179 std::optional<AffineForOp>
180 AffineLoopToGpuConverter::collectBounds(AffineForOp forOp, unsigned numLoops) {
181   OpBuilder builder(forOp.getOperation());
182   dims.reserve(numLoops);
183   lbs.reserve(numLoops);
184   ivs.reserve(numLoops);
185   steps.reserve(numLoops);
186   AffineForOp currentLoop = forOp;
187   for (unsigned i = 0; i < numLoops; ++i) {
188     Value lowerBound = getOrEmitLowerBound(currentLoop, builder);
189     Value upperBound = getOrEmitUpperBound(currentLoop, builder);
190     if (!lowerBound || !upperBound) {
191       return std::nullopt;
192     }
193 
194     Value range = builder.create<arith::SubIOp>(currentLoop.getLoc(),
195                                                 upperBound, lowerBound);
196     Value step = getOrCreateStep(currentLoop, builder);
197     if (getConstantIntValue(step) != static_cast<int64_t>(1))
198       range =
199           builder.create<arith::CeilDivSIOp>(currentLoop.getLoc(), range, step);
200     dims.push_back(range);
201 
202     lbs.push_back(lowerBound);
203     ivs.push_back(currentLoop.getInductionVar());
204     steps.push_back(step);
205 
206     if (i != numLoops - 1)
207       currentLoop = cast<AffineForOp>(&currentLoop.getBody()->front());
208   }
209   return currentLoop;
210 }
211 
212 // Replace the rooted at "rootForOp" with a GPU launch operation.  This expects
213 // "innermostForOp" to point to the last loop to be transformed to the kernel,
214 // and to have (numBlockDims + numThreadDims) perfectly nested loops between
215 // "rootForOp" and "innermostForOp".
216 void AffineLoopToGpuConverter::createLaunch(AffineForOp rootForOp,
217                                             AffineForOp innermostForOp,
218                                             unsigned numBlockDims,
219                                             unsigned numThreadDims) {
220   OpBuilder builder(rootForOp.getOperation());
221   // Prepare the grid and block sizes for the launch operation.  If there is
222   // no loop mapped to a specific dimension, use constant "1" as its size.
223   Value constOne =
224       (numBlockDims < 3 || numThreadDims < 3)
225           ? builder.create<arith::ConstantIndexOp>(rootForOp.getLoc(), 1)
226           : nullptr;
227   Value gridSizeX = numBlockDims > 0 ? dims[0] : constOne;
228   Value gridSizeY = numBlockDims > 1 ? dims[1] : constOne;
229   Value gridSizeZ = numBlockDims > 2 ? dims[2] : constOne;
230   Value blockSizeX = numThreadDims > 0 ? dims[numBlockDims] : constOne;
231   Value blockSizeY = numThreadDims > 1 ? dims[numBlockDims + 1] : constOne;
232   Value blockSizeZ = numThreadDims > 2 ? dims[numBlockDims + 2] : constOne;
233 
234   // Create a launch op and move the body region of the innermost loop to the
235   // launch op.
236   auto launchOp = builder.create<gpu::LaunchOp>(
237       rootForOp.getLoc(), gridSizeX, gridSizeY, gridSizeZ, blockSizeX,
238       blockSizeY, blockSizeZ);
239 
240   // Replace the loop terminator (loops contain only a single block) with the
241   // gpu terminator and move the operations from the loop body block to the gpu
242   // launch body block.  Do not move the entire block because of the difference
243   // in block arguments.
244   Operation &terminator = innermostForOp.getBody()->back();
245   Location terminatorLoc = terminator.getLoc();
246   terminator.erase();
247   builder.setInsertionPointToEnd(innermostForOp.getBody());
248   builder.create<gpu::TerminatorOp>(terminatorLoc, std::nullopt);
249   launchOp.getBody().front().getOperations().splice(
250       launchOp.getBody().front().begin(),
251       innermostForOp.getBody()->getOperations());
252 
253   // Remap the loop iterators to use block/thread identifiers instead.  Loops
254   // may iterate from LB with step S whereas GPU thread/block ids always iterate
255   // from 0 to N with step 1.  Therefore, loop induction variables are replaced
256   // with (gpu-thread/block-id * S) + LB.
257   builder.setInsertionPointToStart(&launchOp.getBody().front());
258   auto *lbArgumentIt = lbs.begin();
259   auto *stepArgumentIt = steps.begin();
260   for (const auto &en : llvm::enumerate(ivs)) {
261     Value id =
262         en.index() < numBlockDims
263             ? getDim3Value(launchOp.getBlockIds(), en.index())
264             : getDim3Value(launchOp.getThreadIds(), en.index() - numBlockDims);
265     Value step = steps[en.index()];
266     if (getConstantIntValue(step) != static_cast<int64_t>(1))
267       id = builder.create<arith::MulIOp>(rootForOp.getLoc(), step, id);
268 
269     Value ivReplacement =
270         builder.create<arith::AddIOp>(rootForOp.getLoc(), *lbArgumentIt, id);
271     en.value().replaceAllUsesWith(ivReplacement);
272     std::advance(lbArgumentIt, 1);
273     std::advance(stepArgumentIt, 1);
274   }
275 
276   // We are done and can erase the original outermost loop.
277   rootForOp.erase();
278 }
279 
280 // Generic loop to GPU kernel conversion function.
281 static LogicalResult convertAffineLoopNestToGPULaunch(AffineForOp forOp,
282                                                       unsigned numBlockDims,
283                                                       unsigned numThreadDims) {
284   if (failed(checkAffineLoopNestMappable(forOp, numBlockDims, numThreadDims)))
285     return failure();
286 
287   AffineLoopToGpuConverter converter;
288   auto maybeInnerLoop =
289       converter.collectBounds(forOp, numBlockDims + numThreadDims);
290   if (!maybeInnerLoop)
291     return failure();
292   converter.createLaunch(forOp, *maybeInnerLoop, numBlockDims, numThreadDims);
293 
294   return success();
295 }
296 
297 LogicalResult mlir::convertAffineLoopNestToGPULaunch(AffineForOp forOp,
298                                                      unsigned numBlockDims,
299                                                      unsigned numThreadDims) {
300   return ::convertAffineLoopNestToGPULaunch(forOp, numBlockDims, numThreadDims);
301 }
302 
303 namespace {
304 struct ParallelToGpuLaunchLowering : public OpRewritePattern<ParallelOp> {
305   using OpRewritePattern<ParallelOp>::OpRewritePattern;
306 
307   LogicalResult matchAndRewrite(ParallelOp parallelOp,
308                                 PatternRewriter &rewriter) const override;
309 };
310 } // namespace
311 
312 /// Tries to derive a static upper bound from the defining operation of
313 /// `upperBound`.
314 static Value deriveStaticUpperBound(Value upperBound,
315                                     PatternRewriter &rewriter) {
316   if (auto op = upperBound.getDefiningOp<arith::ConstantIndexOp>()) {
317     return op;
318   }
319 
320   if (auto minOp = upperBound.getDefiningOp<AffineMinOp>()) {
321     for (const AffineExpr &result : minOp.getMap().getResults()) {
322       if (auto constExpr = dyn_cast<AffineConstantExpr>(result)) {
323         return rewriter.create<arith::ConstantIndexOp>(minOp.getLoc(),
324                                                        constExpr.getValue());
325       }
326     }
327   }
328 
329   if (auto minOp = upperBound.getDefiningOp<arith::MinSIOp>()) {
330     for (Value operand : {minOp.getLhs(), minOp.getRhs()}) {
331       if (auto staticBound = deriveStaticUpperBound(operand, rewriter))
332         return staticBound;
333     }
334   }
335 
336   if (auto multiplyOp = upperBound.getDefiningOp<arith::MulIOp>()) {
337     if (auto lhs = dyn_cast_or_null<arith::ConstantIndexOp>(
338             deriveStaticUpperBound(multiplyOp.getOperand(0), rewriter)
339                 .getDefiningOp()))
340       if (auto rhs = dyn_cast_or_null<arith::ConstantIndexOp>(
341               deriveStaticUpperBound(multiplyOp.getOperand(1), rewriter)
342                   .getDefiningOp())) {
343         // Assumptions about the upper bound of minimum computations no longer
344         // work if multiplied by mixed signs, so abort in this case.
345         if ((lhs.value() < 0) != (rhs.value() < 0))
346           return {};
347 
348         return rewriter.create<arith::ConstantIndexOp>(
349             multiplyOp.getLoc(), lhs.value() * rhs.value());
350       }
351   }
352 
353   return {};
354 }
355 
356 static bool isMappedToProcessor(gpu::Processor processor) {
357   return processor != gpu::Processor::Sequential;
358 }
359 
360 static unsigned getLaunchOpArgumentNum(gpu::Processor processor) {
361   switch (processor) {
362   case gpu::Processor::BlockX:
363     return 0;
364   case gpu::Processor::BlockY:
365     return 1;
366   case gpu::Processor::BlockZ:
367     return 2;
368   case gpu::Processor::ThreadX:
369     return 3;
370   case gpu::Processor::ThreadY:
371     return 4;
372   case gpu::Processor::ThreadZ:
373     return 5;
374   default:;
375   }
376   llvm_unreachable(
377       "invalid processor type while retrieving launch op argument number");
378 }
379 
380 /// Modifies the current transformation state to capture the effect of the given
381 /// `scf.parallel` operation on index substitutions and the operations to be
382 /// inserted.
383 /// Specifically, if a dimension of a parallel loop is mapped to a hardware id,
384 /// this function will
385 /// - compute the loop index based on the hardware id and affine map from the
386 ///   mapping and update `cloningMap` to substitute all uses.
387 /// - derive a new upper bound for the hardware id and augment the provided
388 ///   `gpu.launch operation` accordingly.
389 /// - if the upper bound is imprecise, insert a conditional in the `gpu.launch`
390 ///   and update the rewriter to insert into the conditional's body.
391 /// If the dimension is mapped to sequential,
392 /// - insert a for loop into the body and update the rewriter to insert into
393 ///   the for loop's body.
394 /// - update the `cloningMap` to replace uses of the index with the index of
395 ///   the new for loop.
396 /// In either case,
397 /// - append the instructions from the loops body to worklist, in reverse order.
398 /// To note the end of the current scope in case a loop or conditional was
399 /// inserted, a sentinel (the `gpu.launch` operation) is inserted into the
400 /// worklist. This signals the processor of the worklist to pop the rewriter
401 /// one scope-level up.
402 static LogicalResult processParallelLoop(
403     ParallelOp parallelOp, gpu::LaunchOp launchOp, IRMapping &cloningMap,
404     SmallVectorImpl<Operation *> &worklist,
405     DenseMap<gpu::Processor, Value> &bounds, PatternRewriter &rewriter) {
406   // TODO: Verify that this is a valid GPU mapping.
407   // processor ids: 0-2 block [x/y/z], 3-5 -> thread [x/y/z], 6-> sequential
408   ArrayAttr mapping =
409       parallelOp->getAttrOfType<ArrayAttr>(gpu::getMappingAttrName());
410 
411   // TODO: Support multiple reductions.
412   if (!mapping || parallelOp.getNumResults() > 1)
413     return failure();
414 
415   Location loc = parallelOp.getLoc();
416 
417   auto launchIndependent = [&launchOp](Value val) {
418     return val.getParentRegion()->isAncestor(launchOp->getParentRegion());
419   };
420 
421   auto ensureLaunchIndependent = [&rewriter,
422                                   launchIndependent](Value val) -> Value {
423     if (launchIndependent(val))
424       return val;
425     if (auto constOp = val.getDefiningOp<arith::ConstantOp>())
426       return rewriter.create<arith::ConstantOp>(constOp.getLoc(),
427                                                 constOp.getValue());
428     return {};
429   };
430 
431   for (auto config : llvm::zip(
432            mapping, parallelOp.getInductionVars(), parallelOp.getLowerBound(),
433            parallelOp.getUpperBound(), parallelOp.getStep())) {
434     Attribute mappingAttribute;
435     Value iv, lowerBound, upperBound, step;
436     std::tie(mappingAttribute, iv, lowerBound, upperBound, step) = config;
437     auto annotation =
438         dyn_cast<gpu::ParallelLoopDimMappingAttr>(mappingAttribute);
439     if (!annotation)
440       return parallelOp.emitOpError()
441              << "expected mapping attribute for lowering to GPU";
442     Value newIndex;
443     gpu::Processor processor = annotation.getProcessor();
444 
445     if (isMappedToProcessor(processor)) {
446       // Use the corresponding thread/grid index as replacement for the loop iv.
447       Value operand =
448           launchOp.getBody().getArgument(getLaunchOpArgumentNum(processor));
449       // Take the indexmap and add the lower bound and step computations in.
450       // This computes operand * step + lowerBound.
451       // Use an affine map here so that it composes nicely with the provided
452       // annotation.
453       AffineMap lowerAndStep = AffineMap::get(
454           1, 2,
455           rewriter.getAffineDimExpr(0) * rewriter.getAffineSymbolExpr(0) +
456               rewriter.getAffineSymbolExpr(1));
457       newIndex = rewriter.create<AffineApplyOp>(
458           loc, annotation.getMap().compose(lowerAndStep),
459           ValueRange{operand, ensureLaunchIndependent(step),
460                      ensureLaunchIndependent(lowerBound)});
461       // If there was also a bound, insert that, too.
462       // TODO: Check that we do not assign bounds twice.
463       if (annotation.getBound()) {
464         // We pass as the single operand to the bound-map the number of
465         // iterations, which is (upperBound - lowerBound) ceilDiv step. To
466         // support inner loops with dynamic upper bounds (as generated by e.g.
467         // tiling), try to derive a max for the bounds. If the used bound for
468         // the hardware id is imprecise, wrap the contained code into a
469         // conditional. If the lower-bound is constant or defined before the
470         // launch, we can use it in the launch bounds. Otherwise fail.
471         if (!launchIndependent(lowerBound) &&
472             !isa_and_nonnull<arith::ConstantOp>(lowerBound.getDefiningOp()))
473           return failure();
474         // The step must also be constant or defined outside of the loop nest.
475         if (!launchIndependent(step) &&
476             !isa_and_nonnull<arith::ConstantOp>(step.getDefiningOp()))
477           return failure();
478         // If the upper-bound is constant or defined before the launch, we can
479         // use it in the launch bounds directly. Otherwise try derive a bound.
480         bool boundIsPrecise =
481             launchIndependent(upperBound) ||
482             isa_and_nonnull<arith::ConstantOp>(upperBound.getDefiningOp());
483         {
484           PatternRewriter::InsertionGuard guard(rewriter);
485           rewriter.setInsertionPoint(launchOp);
486           if (!boundIsPrecise) {
487             upperBound = deriveStaticUpperBound(upperBound, rewriter);
488             if (!upperBound) {
489               return rewriter.notifyMatchFailure(
490                   parallelOp,
491                   "cannot derive loop-invariant upper bound for number of"
492                   "iterations");
493             }
494           }
495           // Compute the number of iterations needed. We compute this as an
496           // affine expression ceilDiv (upperBound - lowerBound) step. We use
497           // affine.apply here so that it composes nicely with the provided map.
498           AffineMap stepMap = AffineMap::get(
499               1, 2,
500               ((rewriter.getAffineDimExpr(0) - rewriter.getAffineSymbolExpr(0))
501                    .ceilDiv(rewriter.getAffineSymbolExpr(1))));
502           Value launchBound = rewriter.create<AffineApplyOp>(
503               loc, annotation.getBound().compose(stepMap),
504               ValueRange{
505                   ensureLaunchIndependent(
506                       cloningMap.lookupOrDefault(upperBound)),
507                   ensureLaunchIndependent(
508                       cloningMap.lookupOrDefault(lowerBound)),
509                   ensureLaunchIndependent(cloningMap.lookupOrDefault(step))});
510           // todo(herhut,ravishankarm): Update the behavior of setMappingAttr
511           // when this condition is relaxed.
512           if (!bounds.try_emplace(processor, launchBound).second) {
513             return rewriter.notifyMatchFailure(
514                 parallelOp, "cannot redefine the bound for processor " +
515                                 Twine(static_cast<int64_t>(processor)));
516           }
517         }
518         if (!boundIsPrecise) {
519           // We are using an approximation, create a surrounding conditional.
520           Value originalBound = std::get<3>(config);
521           arith::CmpIOp pred = rewriter.create<arith::CmpIOp>(
522               loc, arith::CmpIPredicate::slt, newIndex,
523               cloningMap.lookupOrDefault(originalBound));
524           scf::IfOp ifOp = rewriter.create<scf::IfOp>(loc, pred, false);
525           rewriter.setInsertionPointToStart(&ifOp.getThenRegion().front());
526           // Put a sentinel into the worklist so we know when to pop out of the
527           // if body again. We use the launchOp here, as that cannot be part of
528           // the bodies instruction.
529           worklist.push_back(launchOp.getOperation());
530         }
531       }
532     } else {
533       // Create a sequential for loop.
534       auto loopOp = rewriter.create<scf::ForOp>(
535           loc, cloningMap.lookupOrDefault(lowerBound),
536           cloningMap.lookupOrDefault(upperBound),
537           cloningMap.lookupOrDefault(step));
538       newIndex = loopOp.getInductionVar();
539       rewriter.setInsertionPointToStart(loopOp.getBody());
540       // Put a sentinel into the worklist so we know when to pop out of the loop
541       // body again. We use the launchOp here, as that cannot be part of the
542       // bodies instruction.
543       worklist.push_back(launchOp.getOperation());
544     }
545     cloningMap.map(iv, newIndex);
546   }
547 
548   // Propagate custom user defined optional attributes, that can be used at
549   // later stage, such as extension data for GPU kernel dispatch
550   for (const auto &namedAttr : parallelOp->getAttrs()) {
551     if (namedAttr.getName() == gpu::getMappingAttrName() ||
552         namedAttr.getName() == ParallelOp::getOperandSegmentSizeAttr())
553       continue;
554     launchOp->setAttr(namedAttr.getName(), namedAttr.getValue());
555   }
556 
557   Block *body = parallelOp.getBody();
558   worklist.reserve(worklist.size() + body->getOperations().size());
559   // Include scf.reduce terminator if exists and has an operand.
560   if (auto terminator = body->getTerminator();
561       isa<scf::ReduceOp>(terminator) && terminator->getOperands().size() == 1) {
562     worklist.push_back(terminator);
563   }
564   for (Operation &op : llvm::reverse(body->without_terminator()))
565     worklist.push_back(&op);
566   return success();
567 }
568 
569 /// Lower a `scf.parallel` operation into a corresponding `gpu.launch`
570 /// operation.
571 ///
572 /// This essentially transforms a loop nest into a corresponding SIMT function.
573 /// The conversion is driven by mapping annotations on the `scf.parallel`
574 /// operations. The mapping is provided via a `DictionaryAttribute` named
575 /// `mapping`, which has three entries:
576 ///  - processor: the hardware id to map to. 0-2 are block dimensions, 3-5 are
577 ///               thread dimensions and 6 is sequential.
578 ///  - map : An affine map that is used to pre-process hardware ids before
579 ///          substitution.
580 ///  - bound : An affine map that is used to compute the bound of the hardware
581 ///            id based on an upper bound of the number of iterations.
582 /// If the `scf.parallel` contains nested `scf.parallel` operations, those
583 /// need to be annotated, as well. Structurally, the transformation works by
584 /// splicing all operations from nested `scf.parallel` operations into a single
585 /// sequence. Indices mapped to hardware ids are substituted with those ids,
586 /// wheras sequential mappings result in a sequential for-loop. To have more
587 /// flexibility when mapping code to hardware ids, the transform supports two
588 /// affine maps. The first `map` is used to compute the actual index for
589 /// substitution from the hardware id. The second `bound` is used to compute the
590 /// launch dimension for the hardware id from the number of iterations the
591 /// mapped loop is performing. Note that the number of iterations might be
592 /// imprecise if the corresponding loop-bounds are loop-dependent. In such case,
593 /// the hardware id might iterate over additional indices. The transformation
594 /// caters for this by predicating the created sequence of instructions on
595 /// the actual loop bound. This only works if an static upper bound for the
596 /// dynamic loop bound can be derived, currently via analyzing `affine.min`
597 /// operations.
598 LogicalResult
599 ParallelToGpuLaunchLowering::matchAndRewrite(ParallelOp parallelOp,
600                                              PatternRewriter &rewriter) const {
601   // Mark the operation as visited for recursive legality check.
602   parallelOp->setAttr(kVisitedAttrName, rewriter.getUnitAttr());
603 
604   // We can only transform starting at the outer-most loop. Launches inside of
605   // parallel loops are not supported.
606   if (auto parentLoop = parallelOp->getParentOfType<ParallelOp>())
607     return failure();
608   // Create a launch operation. We start with bound one for all grid/block
609   // sizes. Those will be refined later as we discover them from mappings.
610   Location loc = parallelOp.getLoc();
611   Value constantOne =
612       rewriter.create<arith::ConstantIndexOp>(parallelOp.getLoc(), 1);
613   gpu::LaunchOp launchOp = rewriter.create<gpu::LaunchOp>(
614       parallelOp.getLoc(), constantOne, constantOne, constantOne, constantOne,
615       constantOne, constantOne);
616   rewriter.setInsertionPointToEnd(&launchOp.getBody().front());
617   rewriter.create<gpu::TerminatorOp>(loc);
618   rewriter.setInsertionPointToStart(&launchOp.getBody().front());
619 
620   IRMapping cloningMap;
621   llvm::DenseMap<gpu::Processor, Value> launchBounds;
622   SmallVector<Operation *, 16> worklist;
623   if (failed(processParallelLoop(parallelOp, launchOp, cloningMap, worklist,
624                                  launchBounds, rewriter)))
625     return failure();
626 
627   // Whether we have seen any side-effects. Reset when leaving an inner scope.
628   bool seenSideeffects = false;
629   // Whether we have left a nesting scope (and hence are no longer innermost).
630   bool leftNestingScope = false;
631   while (!worklist.empty()) {
632     Operation *op = worklist.pop_back_val();
633     // Now walk over the body and clone it.
634     // TODO: This is only correct if there either is no further scf.parallel
635     //       nested or this code is side-effect free. Otherwise we might need
636     //       predication. We are overly conservative for now and only allow
637     //       side-effects in the innermost scope.
638     if (auto nestedParallel = dyn_cast<ParallelOp>(op)) {
639       // Before entering a nested scope, make sure there have been no
640       // sideeffects until now.
641       if (seenSideeffects)
642         return failure();
643       // A nested scf.parallel needs insertion of code to compute indices.
644       // Insert that now. This will also update the worklist with the loops
645       // body.
646       if (failed(processParallelLoop(nestedParallel, launchOp, cloningMap,
647                                      worklist, launchBounds, rewriter)))
648         return failure();
649     } else if (op == launchOp.getOperation()) {
650       // Found our sentinel value. We have finished the operations from one
651       // nesting level, pop one level back up.
652       auto *parent = rewriter.getInsertionPoint()->getParentOp();
653       rewriter.setInsertionPointAfter(parent);
654       leftNestingScope = true;
655       seenSideeffects = false;
656     } else if (auto reduceOp = dyn_cast<scf::ReduceOp>(op)) {
657       // Convert scf.reduction op
658       auto parentLoop = op->getParentOfType<ParallelOp>();
659       if (!parentLoop || op->getOperands().size() != 1)
660         return failure();
661       auto operand = op->getOperands().front();
662       auto newValue = cloningMap.lookupOrNull(operand);
663       if (!newValue || !operand.getType().isSignlessIntOrFloat())
664         return failure();
665       // Ensure reduction region is isolated from above.
666       llvm::SetVector<Value> externalValues;
667       getUsedValuesDefinedAbove(reduceOp.getRegion(0), externalValues);
668       if (externalValues.size())
669         return failure();
670       // Replace by gpu.all_reduce.
671       auto gpuRedOp = rewriter.create<gpu::AllReduceOp>(loc, newValue);
672       cloningMap.map(parentLoop->getResult(0), gpuRedOp.getResult());
673       // Copy region.
674       rewriter.inlineRegionBefore(reduceOp.getRegion(0), gpuRedOp.getRegion(),
675                                   gpuRedOp.getRegion().begin());
676       // Replace src.reduce.return with gpu.yield.
677       auto scfReturn = gpuRedOp.getRegion().front().getTerminator();
678       auto ip = rewriter.saveInsertionPoint();
679       rewriter.setInsertionPointToEnd(&gpuRedOp.getRegion().front());
680       rewriter.replaceOpWithNewOp<gpu::YieldOp>(
681           scfReturn, scfReturn->getOperands().front());
682       rewriter.restoreInsertionPoint(ip);
683     } else {
684       // Otherwise we copy it over.
685       Operation *clone = rewriter.clone(*op, cloningMap);
686       cloningMap.map(op->getResults(), clone->getResults());
687       // Check for side effects.
688       // TODO: Handle region side effects properly.
689       seenSideeffects |=
690           !isMemoryEffectFree(clone) || clone->getNumRegions() != 0;
691       // If we are no longer in the innermost scope, sideeffects are disallowed.
692       if (seenSideeffects && leftNestingScope)
693         return failure();
694     }
695   }
696 
697   // Now that we succeeded creating the launch operation, also update the
698   // bounds.
699   for (auto bound : launchBounds)
700     launchOp.setOperand(getLaunchOpArgumentNum(std::get<0>(bound)),
701                         std::get<1>(bound));
702 
703   rewriter.eraseOp(parallelOp);
704   return success();
705 }
706 
707 void mlir::populateParallelLoopToGPUPatterns(RewritePatternSet &patterns) {
708   patterns.add<ParallelToGpuLaunchLowering>(patterns.getContext());
709 }
710 
711 void mlir::configureParallelLoopToGPULegality(ConversionTarget &target) {
712   target.addLegalDialect<memref::MemRefDialect>();
713   target.addDynamicallyLegalOp<scf::ParallelOp>([](scf::ParallelOp parallelOp) {
714     return !parallelOp->hasAttr(gpu::getMappingAttrName()) ||
715            parallelOp->hasAttr(kVisitedAttrName);
716   });
717 }
718 
719 void mlir::finalizeParallelLoopToGPUConversion(Operation *op) {
720   op->walk([](scf::ParallelOp parallelOp) {
721     parallelOp->removeAttr(kVisitedAttrName);
722   });
723 }
724