xref: /llvm-project/mlir/lib/Conversion/SCFToGPU/SCFToGPU.cpp (revision 5550c821897ab77e664977121a0e90ad5be1ff59)
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.getStep());
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 // Return true if the value is obviously a constant "one".
175 static bool isConstantOne(Value value) {
176   if (auto def = value.getDefiningOp<arith::ConstantIndexOp>())
177     return def.value() == 1;
178   return false;
179 }
180 
181 // Collect ranges, bounds, steps and induction variables in preparation for
182 // mapping a loop nest of depth "numLoops" rooted at "forOp" to a GPU kernel.
183 // This may fail if the IR for computing loop bounds cannot be constructed, for
184 // example if an affine loop uses semi-affine maps. Return the last loop to be
185 // mapped on success, std::nullopt on failure.
186 std::optional<AffineForOp>
187 AffineLoopToGpuConverter::collectBounds(AffineForOp forOp, unsigned numLoops) {
188   OpBuilder builder(forOp.getOperation());
189   dims.reserve(numLoops);
190   lbs.reserve(numLoops);
191   ivs.reserve(numLoops);
192   steps.reserve(numLoops);
193   AffineForOp currentLoop = forOp;
194   for (unsigned i = 0; i < numLoops; ++i) {
195     Value lowerBound = getOrEmitLowerBound(currentLoop, builder);
196     Value upperBound = getOrEmitUpperBound(currentLoop, builder);
197     if (!lowerBound || !upperBound) {
198       return std::nullopt;
199     }
200 
201     Value range = builder.create<arith::SubIOp>(currentLoop.getLoc(),
202                                                 upperBound, lowerBound);
203     Value step = getOrCreateStep(currentLoop, builder);
204     if (!isConstantOne(step))
205       range = builder.create<arith::DivSIOp>(currentLoop.getLoc(), range, step);
206     dims.push_back(range);
207 
208     lbs.push_back(lowerBound);
209     ivs.push_back(currentLoop.getInductionVar());
210     steps.push_back(step);
211 
212     if (i != numLoops - 1)
213       currentLoop = cast<AffineForOp>(&currentLoop.getBody()->front());
214   }
215   return currentLoop;
216 }
217 
218 // Replace the rooted at "rootForOp" with a GPU launch operation.  This expects
219 // "innermostForOp" to point to the last loop to be transformed to the kernel,
220 // and to have (numBlockDims + numThreadDims) perfectly nested loops between
221 // "rootForOp" and "innermostForOp".
222 void AffineLoopToGpuConverter::createLaunch(AffineForOp rootForOp,
223                                             AffineForOp innermostForOp,
224                                             unsigned numBlockDims,
225                                             unsigned numThreadDims) {
226   OpBuilder builder(rootForOp.getOperation());
227   // Prepare the grid and block sizes for the launch operation.  If there is
228   // no loop mapped to a specific dimension, use constant "1" as its size.
229   Value constOne =
230       (numBlockDims < 3 || numThreadDims < 3)
231           ? builder.create<arith::ConstantIndexOp>(rootForOp.getLoc(), 1)
232           : nullptr;
233   Value gridSizeX = numBlockDims > 0 ? dims[0] : constOne;
234   Value gridSizeY = numBlockDims > 1 ? dims[1] : constOne;
235   Value gridSizeZ = numBlockDims > 2 ? dims[2] : constOne;
236   Value blockSizeX = numThreadDims > 0 ? dims[numBlockDims] : constOne;
237   Value blockSizeY = numThreadDims > 1 ? dims[numBlockDims + 1] : constOne;
238   Value blockSizeZ = numThreadDims > 2 ? dims[numBlockDims + 2] : constOne;
239 
240   // Create a launch op and move the body region of the innermost loop to the
241   // launch op.
242   auto launchOp = builder.create<gpu::LaunchOp>(
243       rootForOp.getLoc(), gridSizeX, gridSizeY, gridSizeZ, blockSizeX,
244       blockSizeY, blockSizeZ);
245 
246   // Replace the loop terminator (loops contain only a single block) with the
247   // gpu terminator and move the operations from the loop body block to the gpu
248   // launch body block.  Do not move the entire block because of the difference
249   // in block arguments.
250   Operation &terminator = innermostForOp.getBody()->back();
251   Location terminatorLoc = terminator.getLoc();
252   terminator.erase();
253   builder.setInsertionPointToEnd(innermostForOp.getBody());
254   builder.create<gpu::TerminatorOp>(terminatorLoc, std::nullopt);
255   launchOp.getBody().front().getOperations().splice(
256       launchOp.getBody().front().begin(),
257       innermostForOp.getBody()->getOperations());
258 
259   // Remap the loop iterators to use block/thread identifiers instead.  Loops
260   // may iterate from LB with step S whereas GPU thread/block ids always iterate
261   // from 0 to N with step 1.  Therefore, loop induction variables are replaced
262   // with (gpu-thread/block-id * S) + LB.
263   builder.setInsertionPointToStart(&launchOp.getBody().front());
264   auto *lbArgumentIt = lbs.begin();
265   auto *stepArgumentIt = steps.begin();
266   for (const auto &en : llvm::enumerate(ivs)) {
267     Value id =
268         en.index() < numBlockDims
269             ? getDim3Value(launchOp.getBlockIds(), en.index())
270             : getDim3Value(launchOp.getThreadIds(), en.index() - numBlockDims);
271     Value step = steps[en.index()];
272     if (!isConstantOne(step))
273       id = builder.create<arith::MulIOp>(rootForOp.getLoc(), step, id);
274 
275     Value ivReplacement =
276         builder.create<arith::AddIOp>(rootForOp.getLoc(), *lbArgumentIt, id);
277     en.value().replaceAllUsesWith(ivReplacement);
278     std::advance(lbArgumentIt, 1);
279     std::advance(stepArgumentIt, 1);
280   }
281 
282   // We are done and can erase the original outermost loop.
283   rootForOp.erase();
284 }
285 
286 // Generic loop to GPU kernel conversion function.
287 static LogicalResult convertAffineLoopNestToGPULaunch(AffineForOp forOp,
288                                                       unsigned numBlockDims,
289                                                       unsigned numThreadDims) {
290   if (failed(checkAffineLoopNestMappable(forOp, numBlockDims, numThreadDims)))
291     return failure();
292 
293   AffineLoopToGpuConverter converter;
294   auto maybeInnerLoop =
295       converter.collectBounds(forOp, numBlockDims + numThreadDims);
296   if (!maybeInnerLoop)
297     return failure();
298   converter.createLaunch(forOp, *maybeInnerLoop, numBlockDims, numThreadDims);
299 
300   return success();
301 }
302 
303 LogicalResult mlir::convertAffineLoopNestToGPULaunch(AffineForOp forOp,
304                                                      unsigned numBlockDims,
305                                                      unsigned numThreadDims) {
306   return ::convertAffineLoopNestToGPULaunch(forOp, numBlockDims, numThreadDims);
307 }
308 
309 namespace {
310 struct ParallelToGpuLaunchLowering : public OpRewritePattern<ParallelOp> {
311   using OpRewritePattern<ParallelOp>::OpRewritePattern;
312 
313   LogicalResult matchAndRewrite(ParallelOp parallelOp,
314                                 PatternRewriter &rewriter) const override;
315 };
316 } // namespace
317 
318 /// Tries to derive a static upper bound from the defining operation of
319 /// `upperBound`.
320 static Value deriveStaticUpperBound(Value upperBound,
321                                     PatternRewriter &rewriter) {
322   if (auto op = upperBound.getDefiningOp<arith::ConstantIndexOp>()) {
323     return op;
324   }
325 
326   if (auto minOp = upperBound.getDefiningOp<AffineMinOp>()) {
327     for (const AffineExpr &result : minOp.getMap().getResults()) {
328       if (auto constExpr = result.dyn_cast<AffineConstantExpr>()) {
329         return rewriter.create<arith::ConstantIndexOp>(minOp.getLoc(),
330                                                        constExpr.getValue());
331       }
332     }
333   }
334 
335   if (auto minOp = upperBound.getDefiningOp<arith::MinSIOp>()) {
336     for (Value operand : {minOp.getLhs(), minOp.getRhs()}) {
337       if (auto staticBound = deriveStaticUpperBound(operand, rewriter))
338         return staticBound;
339     }
340   }
341 
342   if (auto multiplyOp = upperBound.getDefiningOp<arith::MulIOp>()) {
343     if (auto lhs = dyn_cast_or_null<arith::ConstantIndexOp>(
344             deriveStaticUpperBound(multiplyOp.getOperand(0), rewriter)
345                 .getDefiningOp()))
346       if (auto rhs = dyn_cast_or_null<arith::ConstantIndexOp>(
347               deriveStaticUpperBound(multiplyOp.getOperand(1), rewriter)
348                   .getDefiningOp())) {
349         // Assumptions about the upper bound of minimum computations no longer
350         // work if multiplied by mixed signs, so abort in this case.
351         if ((lhs.value() < 0) != (rhs.value() < 0))
352           return {};
353 
354         return rewriter.create<arith::ConstantIndexOp>(
355             multiplyOp.getLoc(), lhs.value() * rhs.value());
356       }
357   }
358 
359   return {};
360 }
361 
362 static bool isMappedToProcessor(gpu::Processor processor) {
363   return processor != gpu::Processor::Sequential;
364 }
365 
366 static unsigned getLaunchOpArgumentNum(gpu::Processor processor) {
367   switch (processor) {
368   case gpu::Processor::BlockX:
369     return 0;
370   case gpu::Processor::BlockY:
371     return 1;
372   case gpu::Processor::BlockZ:
373     return 2;
374   case gpu::Processor::ThreadX:
375     return 3;
376   case gpu::Processor::ThreadY:
377     return 4;
378   case gpu::Processor::ThreadZ:
379     return 5;
380   default:;
381   }
382   llvm_unreachable(
383       "invalid processor type while retrieving launch op argument number");
384 }
385 
386 /// Modifies the current transformation state to capture the effect of the given
387 /// `scf.parallel` operation on index substitutions and the operations to be
388 /// inserted.
389 /// Specifically, if a dimension of a parallel loop is mapped to a hardware id,
390 /// this function will
391 /// - compute the loop index based on the hardware id and affine map from the
392 ///   mapping and update `cloningMap` to substitute all uses.
393 /// - derive a new upper bound for the hardware id and augment the provided
394 ///   `gpu.launch operation` accordingly.
395 /// - if the upper bound is imprecise, insert a conditional in the `gpu.launch`
396 ///   and update the rewriter to insert into the conditional's body.
397 /// If the dimension is mapped to sequential,
398 /// - insert a for loop into the body and update the rewriter to insert into
399 ///   the for loop's body.
400 /// - update the `cloningMap` to replace uses of the index with the index of
401 ///   the new for loop.
402 /// In either case,
403 /// - append the instructions from the loops body to worklist, in reverse order.
404 /// To note the end of the current scope in case a loop or conditional was
405 /// inserted, a sentinel (the `gpu.launch` operation) is inserted into the
406 /// worklist. This signals the processor of the worklist to pop the rewriter
407 /// one scope-level up.
408 static LogicalResult processParallelLoop(
409     ParallelOp parallelOp, gpu::LaunchOp launchOp, IRMapping &cloningMap,
410     SmallVectorImpl<Operation *> &worklist,
411     DenseMap<gpu::Processor, Value> &bounds, PatternRewriter &rewriter) {
412   // TODO: Verify that this is a valid GPU mapping.
413   // processor ids: 0-2 block [x/y/z], 3-5 -> thread [x/y/z], 6-> sequential
414   ArrayAttr mapping =
415       parallelOp->getAttrOfType<ArrayAttr>(gpu::getMappingAttrName());
416 
417   // TODO: Support reductions.
418   if (!mapping || parallelOp.getNumResults() != 0)
419     return failure();
420 
421   Location loc = parallelOp.getLoc();
422 
423   auto launchIndependent = [&launchOp](Value val) {
424     return val.getParentRegion()->isAncestor(launchOp->getParentRegion());
425   };
426 
427   auto ensureLaunchIndependent = [&rewriter,
428                                   launchIndependent](Value val) -> Value {
429     if (launchIndependent(val))
430       return val;
431     if (auto constOp = val.getDefiningOp<arith::ConstantOp>())
432       return rewriter.create<arith::ConstantOp>(constOp.getLoc(),
433                                                 constOp.getValue());
434     return {};
435   };
436 
437   for (auto config : llvm::zip(
438            mapping, parallelOp.getInductionVars(), parallelOp.getLowerBound(),
439            parallelOp.getUpperBound(), parallelOp.getStep())) {
440     Attribute mappingAttribute;
441     Value iv, lowerBound, upperBound, step;
442     std::tie(mappingAttribute, iv, lowerBound, upperBound, step) = config;
443     auto annotation =
444         dyn_cast<gpu::ParallelLoopDimMappingAttr>(mappingAttribute);
445     if (!annotation)
446       return parallelOp.emitOpError()
447              << "expected mapping attribute for lowering to GPU";
448     Value newIndex;
449     gpu::Processor processor = annotation.getProcessor();
450 
451     if (isMappedToProcessor(processor)) {
452       // Use the corresponding thread/grid index as replacement for the loop iv.
453       Value operand =
454           launchOp.getBody().getArgument(getLaunchOpArgumentNum(processor));
455       // Take the indexmap and add the lower bound and step computations in.
456       // This computes operand * step + lowerBound.
457       // Use an affine map here so that it composes nicely with the provided
458       // annotation.
459       AffineMap lowerAndStep = AffineMap::get(
460           1, 2,
461           rewriter.getAffineDimExpr(0) * rewriter.getAffineSymbolExpr(0) +
462               rewriter.getAffineSymbolExpr(1));
463       newIndex = rewriter.create<AffineApplyOp>(
464           loc, annotation.getMap().compose(lowerAndStep),
465           ValueRange{operand, step, lowerBound});
466       // If there was also a bound, insert that, too.
467       // TODO: Check that we do not assign bounds twice.
468       if (annotation.getBound()) {
469         // We pass as the single operand to the bound-map the number of
470         // iterations, which is (upperBound - lowerBound) ceilDiv step. To
471         // support inner loops with dynamic upper bounds (as generated by e.g.
472         // tiling), try to derive a max for the bounds. If the used bound for
473         // the hardware id is imprecise, wrap the contained code into a
474         // conditional. If the lower-bound is constant or defined before the
475         // launch, we can use it in the launch bounds. Otherwise fail.
476         if (!launchIndependent(lowerBound) &&
477             !isa_and_nonnull<arith::ConstantOp>(lowerBound.getDefiningOp()))
478           return failure();
479         // The step must also be constant or defined outside of the loop nest.
480         if (!launchIndependent(step) &&
481             !isa_and_nonnull<arith::ConstantOp>(step.getDefiningOp()))
482           return failure();
483         // If the upper-bound is constant or defined before the launch, we can
484         // use it in the launch bounds directly. Otherwise try derive a bound.
485         bool boundIsPrecise =
486             launchIndependent(upperBound) ||
487             isa_and_nonnull<arith::ConstantOp>(upperBound.getDefiningOp());
488         {
489           PatternRewriter::InsertionGuard guard(rewriter);
490           rewriter.setInsertionPoint(launchOp);
491           if (!boundIsPrecise) {
492             upperBound = deriveStaticUpperBound(upperBound, rewriter);
493             if (!upperBound) {
494               return rewriter.notifyMatchFailure(
495                   parallelOp,
496                   "cannot derive loop-invariant upper bound for number of"
497                   "iterations");
498             }
499           }
500           // Compute the number of iterations needed. We compute this as an
501           // affine expression ceilDiv (upperBound - lowerBound) step. We use
502           // affine.apply here so that it composes nicely with the provided map.
503           AffineMap stepMap = AffineMap::get(
504               1, 2,
505               ((rewriter.getAffineDimExpr(0) - rewriter.getAffineSymbolExpr(0))
506                    .ceilDiv(rewriter.getAffineSymbolExpr(1))));
507           Value launchBound = rewriter.create<AffineApplyOp>(
508               loc, annotation.getBound().compose(stepMap),
509               ValueRange{
510                   ensureLaunchIndependent(
511                       cloningMap.lookupOrDefault(upperBound)),
512                   ensureLaunchIndependent(
513                       cloningMap.lookupOrDefault(lowerBound)),
514                   ensureLaunchIndependent(cloningMap.lookupOrDefault(step))});
515           // todo(herhut,ravishankarm): Update the behavior of setMappingAttr
516           // when this condition is relaxed.
517           if (bounds.contains(processor)) {
518             return rewriter.notifyMatchFailure(
519                 parallelOp, "cannot redefine the bound for processor " +
520                                 Twine(static_cast<int64_t>(processor)));
521           }
522           bounds[processor] = launchBound;
523         }
524         if (!boundIsPrecise) {
525           // We are using an approximation, create a surrounding conditional.
526           Value originalBound = std::get<3>(config);
527           arith::CmpIOp pred = rewriter.create<arith::CmpIOp>(
528               loc, arith::CmpIPredicate::slt, newIndex,
529               cloningMap.lookupOrDefault(originalBound));
530           scf::IfOp ifOp = rewriter.create<scf::IfOp>(loc, pred, false);
531           rewriter.setInsertionPointToStart(&ifOp.getThenRegion().front());
532           // Put a sentinel into the worklist so we know when to pop out of the
533           // if body again. We use the launchOp here, as that cannot be part of
534           // the bodies instruction.
535           worklist.push_back(launchOp.getOperation());
536         }
537       }
538     } else {
539       // Create a sequential for loop.
540       auto loopOp = rewriter.create<scf::ForOp>(
541           loc, cloningMap.lookupOrDefault(lowerBound),
542           cloningMap.lookupOrDefault(upperBound),
543           cloningMap.lookupOrDefault(step));
544       newIndex = loopOp.getInductionVar();
545       rewriter.setInsertionPointToStart(loopOp.getBody());
546       // Put a sentinel into the worklist so we know when to pop out of the loop
547       // body again. We use the launchOp here, as that cannot be part of the
548       // bodies instruction.
549       worklist.push_back(launchOp.getOperation());
550     }
551     cloningMap.map(iv, newIndex);
552   }
553 
554   // Propagate custom user defined optional attributes, that can be used at
555   // later stage, such as extension data for GPU kernel dispatch
556   for (const auto &namedAttr : parallelOp->getAttrs()) {
557     if (namedAttr.getName() == gpu::getMappingAttrName() ||
558         namedAttr.getName() == ParallelOp::getOperandSegmentSizeAttr())
559       continue;
560     launchOp->setAttr(namedAttr.getName(), namedAttr.getValue());
561   }
562 
563   Block *body = parallelOp.getBody();
564   worklist.reserve(worklist.size() + body->getOperations().size());
565   for (Operation &op : llvm::reverse(body->without_terminator()))
566     worklist.push_back(&op);
567   return success();
568 }
569 
570 /// Lower a `scf.parallel` operation into a corresponding `gpu.launch`
571 /// operation.
572 ///
573 /// This essentially transforms a loop nest into a corresponding SIMT function.
574 /// The conversion is driven by mapping annotations on the `scf.parallel`
575 /// operations. The mapping is provided via a `DictionaryAttribute` named
576 /// `mapping`, which has three entries:
577 ///  - processor: the hardware id to map to. 0-2 are block dimensions, 3-5 are
578 ///               thread dimensions and 6 is sequential.
579 ///  - map : An affine map that is used to pre-process hardware ids before
580 ///          substitution.
581 ///  - bound : An affine map that is used to compute the bound of the hardware
582 ///            id based on an upper bound of the number of iterations.
583 /// If the `scf.parallel` contains nested `scf.parallel` operations, those
584 /// need to be annotated, as well. Structurally, the transformation works by
585 /// splicing all operations from nested `scf.parallel` operations into a single
586 /// sequence. Indices mapped to hardware ids are substituted with those ids,
587 /// wheras sequential mappings result in a sequential for-loop. To have more
588 /// flexibility when mapping code to hardware ids, the transform supports two
589 /// affine maps. The first `map` is used to compute the actual index for
590 /// substitution from the hardware id. The second `bound` is used to compute the
591 /// launch dimension for the hardware id from the number of iterations the
592 /// mapped loop is performing. Note that the number of iterations might be
593 /// imprecise if the corresponding loop-bounds are loop-dependent. In such case,
594 /// the hardware id might iterate over additional indices. The transformation
595 /// caters for this by predicating the created sequence of instructions on
596 /// the actual loop bound. This only works if an static upper bound for the
597 /// dynamic loop bound can be derived, currently via analyzing `affine.min`
598 /// operations.
599 LogicalResult
600 ParallelToGpuLaunchLowering::matchAndRewrite(ParallelOp parallelOp,
601                                              PatternRewriter &rewriter) const {
602   // Mark the operation as visited for recursive legality check.
603   parallelOp->setAttr(kVisitedAttrName, rewriter.getUnitAttr());
604 
605   // We can only transform starting at the outer-most loop. Launches inside of
606   // parallel loops are not supported.
607   if (auto parentLoop = parallelOp->getParentOfType<ParallelOp>())
608     return failure();
609   // Create a launch operation. We start with bound one for all grid/block
610   // sizes. Those will be refined later as we discover them from mappings.
611   Location loc = parallelOp.getLoc();
612   Value constantOne =
613       rewriter.create<arith::ConstantIndexOp>(parallelOp.getLoc(), 1);
614   gpu::LaunchOp launchOp = rewriter.create<gpu::LaunchOp>(
615       parallelOp.getLoc(), constantOne, constantOne, constantOne, constantOne,
616       constantOne, constantOne);
617   rewriter.setInsertionPointToEnd(&launchOp.getBody().front());
618   rewriter.create<gpu::TerminatorOp>(loc);
619   rewriter.setInsertionPointToStart(&launchOp.getBody().front());
620 
621   IRMapping cloningMap;
622   llvm::DenseMap<gpu::Processor, Value> launchBounds;
623   SmallVector<Operation *, 16> worklist;
624   if (failed(processParallelLoop(parallelOp, launchOp, cloningMap, worklist,
625                                  launchBounds, rewriter)))
626     return failure();
627 
628   // Whether we have seen any side-effects. Reset when leaving an inner scope.
629   bool seenSideeffects = false;
630   // Whether we have left a nesting scope (and hence are no longer innermost).
631   bool leftNestingScope = false;
632   while (!worklist.empty()) {
633     Operation *op = worklist.pop_back_val();
634     // Now walk over the body and clone it.
635     // TODO: This is only correct if there either is no further scf.parallel
636     //       nested or this code is side-effect free. Otherwise we might need
637     //       predication. We are overly conservative for now and only allow
638     //       side-effects in the innermost scope.
639     if (auto nestedParallel = dyn_cast<ParallelOp>(op)) {
640       // Before entering a nested scope, make sure there have been no
641       // sideeffects until now.
642       if (seenSideeffects)
643         return failure();
644       // A nested scf.parallel needs insertion of code to compute indices.
645       // Insert that now. This will also update the worklist with the loops
646       // body.
647       if (failed(processParallelLoop(nestedParallel, launchOp, cloningMap,
648                                      worklist, launchBounds, rewriter)))
649         return failure();
650     } else if (op == launchOp.getOperation()) {
651       // Found our sentinel value. We have finished the operations from one
652       // nesting level, pop one level back up.
653       auto *parent = rewriter.getInsertionPoint()->getParentOp();
654       rewriter.setInsertionPointAfter(parent);
655       leftNestingScope = true;
656       seenSideeffects = false;
657     } else {
658       // Otherwise we copy it over.
659       Operation *clone = rewriter.clone(*op, cloningMap);
660       cloningMap.map(op->getResults(), clone->getResults());
661       // Check for side effects.
662       // TODO: Handle region side effects properly.
663       seenSideeffects |=
664           !isMemoryEffectFree(clone) || clone->getNumRegions() != 0;
665       // If we are no longer in the innermost scope, sideeffects are disallowed.
666       if (seenSideeffects && leftNestingScope)
667         return failure();
668     }
669   }
670 
671   // Now that we succeeded creating the launch operation, also update the
672   // bounds.
673   for (auto bound : launchBounds)
674     launchOp.setOperand(getLaunchOpArgumentNum(std::get<0>(bound)),
675                         std::get<1>(bound));
676 
677   rewriter.eraseOp(parallelOp);
678   return success();
679 }
680 
681 void mlir::populateParallelLoopToGPUPatterns(RewritePatternSet &patterns) {
682   patterns.add<ParallelToGpuLaunchLowering>(patterns.getContext());
683 }
684 
685 void mlir::configureParallelLoopToGPULegality(ConversionTarget &target) {
686   target.addLegalDialect<memref::MemRefDialect>();
687   target.addDynamicallyLegalOp<scf::ParallelOp>([](scf::ParallelOp parallelOp) {
688     return !parallelOp->hasAttr(gpu::getMappingAttrName()) ||
689            parallelOp->hasAttr(kVisitedAttrName);
690   });
691 }
692 
693 void mlir::finalizeParallelLoopToGPUConversion(Operation *op) {
694   op->walk([](scf::ParallelOp parallelOp) {
695     parallelOp->removeAttr(kVisitedAttrName);
696   });
697 }
698