xref: /llvm-project/mlir/lib/Conversion/SCFToGPU/SCFToGPU.cpp (revision fc367dfa6770fda4adc4e5de79846f22e7e4e215)
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/BlockAndValueMapping.h"
26 #include "mlir/IR/Builders.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 
35 #define DEBUG_TYPE "loops-to-gpu"
36 
37 using namespace mlir;
38 using namespace mlir::scf;
39 
40 // Name of internal attribute to mark visited operations during conversion.
41 //
42 // NOTE: The conversion originally used the following legality criteria:
43 //   `!parallelOp->hasAttr(gpu::getMappingAttrName())`
44 // But the provided pattern might reject some cases based on more detailed
45 // analysis of the `mapping` attribute.
46 // To avoid dialect conversion failure due to non-converted illegal operation
47 // we use this extra Unit attribute as a marker, that the operation was checked
48 // by the pattern and is should be considered as legal in the following legality
49 // checks. The `finalizeParallelLoopToGPUConversion` function performs clean up
50 // of this extra attributes ans is supposed to be called after the dialect
51 // conversion.
52 //
53 // TODO: Implement a cleaner solution, factoring out the "matching" logic
54 // from the pattern and its callees into a separate function that can be called
55 // from both the pattern and the op legality check.
56 static constexpr StringLiteral kVisitedAttrName = "SCFToGPU_visited";
57 
58 // Extract an indexed value from KernelDim3.
59 static Value getDim3Value(const gpu::KernelDim3 &dim3, unsigned pos) {
60   switch (pos) {
61   case 0:
62     return dim3.x;
63   case 1:
64     return dim3.y;
65   case 2:
66     return dim3.z;
67   default:
68     llvm_unreachable("dim3 position out of bounds");
69   }
70   return nullptr;
71 }
72 
73 // Get the lower bound-related operands of a loop operation.
74 static Operation::operand_range getLowerBoundOperands(AffineForOp forOp) {
75   return forOp.getLowerBoundOperands();
76 }
77 
78 // Get the upper bound-related operands of a loop operation.
79 static Operation::operand_range getUpperBoundOperands(AffineForOp forOp) {
80   return forOp.getUpperBoundOperands();
81 }
82 
83 // Get a Value that corresponds to the loop step.  If the step is an attribute,
84 // materialize a corresponding constant using builder.
85 static Value getOrCreateStep(AffineForOp forOp, OpBuilder &builder) {
86   return builder.create<arith::ConstantIndexOp>(forOp.getLoc(),
87                                                 forOp.getStep());
88 }
89 
90 // Get a Value for the loop lower bound.  If the value requires computation,
91 // materialize the instructions using builder.
92 static Value getOrEmitLowerBound(AffineForOp forOp, OpBuilder &builder) {
93   return lowerAffineLowerBound(forOp, builder);
94 }
95 
96 // Get a Value for the loop upper bound.  If the value requires computation,
97 // materialize the instructions using builder.
98 static Value getOrEmitUpperBound(AffineForOp forOp, OpBuilder &builder) {
99   return lowerAffineUpperBound(forOp, builder);
100 }
101 
102 // Check the structure of the loop nest:
103 //   - there are enough loops to map to numDims;
104 //   - the loops are perfectly nested;
105 //   - the loop bounds can be computed above the outermost loop.
106 // This roughly corresponds to the "matcher" part of the pattern-based
107 // rewriting infrastructure.
108 static LogicalResult checkAffineLoopNestMappableImpl(AffineForOp forOp,
109                                                      unsigned numDims) {
110   Region &limit = forOp.getRegion();
111   for (unsigned i = 0, e = numDims; i < e; ++i) {
112     Operation *nested = &forOp.getBody()->front();
113     if (!areValuesDefinedAbove(getLowerBoundOperands(forOp), limit) ||
114         !areValuesDefinedAbove(getUpperBoundOperands(forOp), limit))
115       return forOp.emitError(
116           "loops with bounds depending on other mapped loops "
117           "are not supported");
118 
119     // The innermost loop can have an arbitrary body, skip the perfect nesting
120     // check for it.
121     if (i == e - 1)
122       break;
123 
124     auto begin = forOp.getBody()->begin(), end = forOp.getBody()->end();
125     if (forOp.getBody()->empty() || std::next(begin, 2) != end)
126       return forOp.emitError("expected perfectly nested loops in the body");
127 
128     if (!(forOp = dyn_cast<AffineForOp>(nested)))
129       return nested->emitError("expected a nested loop");
130   }
131   return success();
132 }
133 
134 static LogicalResult checkAffineLoopNestMappable(AffineForOp forOp,
135                                                  unsigned numBlockDims,
136                                                  unsigned numThreadDims) {
137   if (numBlockDims < 1 || numThreadDims < 1) {
138     LLVM_DEBUG(llvm::dbgs() << "nothing to map");
139     return success();
140   }
141 
142   if (numBlockDims > 3) {
143     return forOp.emitError("cannot map to more than 3 block dimensions");
144   }
145   if (numThreadDims > 3) {
146     return forOp.emitError("cannot map to more than 3 thread dimensions");
147   }
148   return checkAffineLoopNestMappableImpl(forOp, numBlockDims + numThreadDims);
149 }
150 
151 namespace {
152 // Helper structure that holds common state of the loop to GPU kernel
153 // conversion.
154 struct AffineLoopToGpuConverter {
155   Optional<AffineForOp> collectBounds(AffineForOp forOp, unsigned numLoops);
156 
157   void createLaunch(AffineForOp rootForOp, AffineForOp innermostForOp,
158                     unsigned numBlockDims, unsigned numThreadDims);
159 
160   // Ranges of the loops mapped to blocks or threads.
161   SmallVector<Value, 6> dims;
162   // Lower bounds of the loops mapped to blocks or threads.
163   SmallVector<Value, 6> lbs;
164   // Induction variables of the loops mapped to blocks or threads.
165   SmallVector<Value, 6> ivs;
166   // Steps of the loops mapped to blocks or threads.
167   SmallVector<Value, 6> steps;
168 };
169 } // namespace
170 
171 // Return true if the value is obviously a constant "one".
172 static bool isConstantOne(Value value) {
173   if (auto def = value.getDefiningOp<arith::ConstantIndexOp>())
174     return def.value() == 1;
175   return false;
176 }
177 
178 // Collect ranges, bounds, steps and induction variables in preparation for
179 // mapping a loop nest of depth "numLoops" rooted at "forOp" to a GPU kernel.
180 // This may fail if the IR for computing loop bounds cannot be constructed, for
181 // example if an affine loop uses semi-affine maps. Return the last loop to be
182 // mapped on success, llvm::None on failure.
183 Optional<AffineForOp>
184 AffineLoopToGpuConverter::collectBounds(AffineForOp forOp, unsigned numLoops) {
185   OpBuilder builder(forOp.getOperation());
186   dims.reserve(numLoops);
187   lbs.reserve(numLoops);
188   ivs.reserve(numLoops);
189   steps.reserve(numLoops);
190   AffineForOp currentLoop = forOp;
191   for (unsigned i = 0; i < numLoops; ++i) {
192     Value lowerBound = getOrEmitLowerBound(currentLoop, builder);
193     Value upperBound = getOrEmitUpperBound(currentLoop, builder);
194     if (!lowerBound || !upperBound) {
195       return llvm::None;
196     }
197 
198     Value range = builder.create<arith::SubIOp>(currentLoop.getLoc(),
199                                                 upperBound, lowerBound);
200     Value step = getOrCreateStep(currentLoop, builder);
201     if (!isConstantOne(step))
202       range = builder.create<arith::DivSIOp>(currentLoop.getLoc(), range, step);
203     dims.push_back(range);
204 
205     lbs.push_back(lowerBound);
206     ivs.push_back(currentLoop.getInductionVar());
207     steps.push_back(step);
208 
209     if (i != numLoops - 1)
210       currentLoop = cast<AffineForOp>(&currentLoop.getBody()->front());
211   }
212   return currentLoop;
213 }
214 
215 // Replace the rooted at "rootForOp" with a GPU launch operation.  This expects
216 // "innermostForOp" to point to the last loop to be transformed to the kernel,
217 // and to have (numBlockDims + numThreadDims) perfectly nested loops between
218 // "rootForOp" and "innermostForOp".
219 void AffineLoopToGpuConverter::createLaunch(AffineForOp rootForOp,
220                                             AffineForOp innermostForOp,
221                                             unsigned numBlockDims,
222                                             unsigned numThreadDims) {
223   OpBuilder builder(rootForOp.getOperation());
224   // Prepare the grid and block sizes for the launch operation.  If there is
225   // no loop mapped to a specific dimension, use constant "1" as its size.
226   Value constOne =
227       (numBlockDims < 3 || numThreadDims < 3)
228           ? builder.create<arith::ConstantIndexOp>(rootForOp.getLoc(), 1)
229           : nullptr;
230   Value gridSizeX = numBlockDims > 0 ? dims[0] : constOne;
231   Value gridSizeY = numBlockDims > 1 ? dims[1] : constOne;
232   Value gridSizeZ = numBlockDims > 2 ? dims[2] : constOne;
233   Value blockSizeX = numThreadDims > 0 ? dims[numBlockDims] : constOne;
234   Value blockSizeY = numThreadDims > 1 ? dims[numBlockDims + 1] : constOne;
235   Value blockSizeZ = numThreadDims > 2 ? dims[numBlockDims + 2] : constOne;
236 
237   // Create a launch op and move the body region of the innermost loop to the
238   // launch op.
239   auto launchOp = builder.create<gpu::LaunchOp>(
240       rootForOp.getLoc(), gridSizeX, gridSizeY, gridSizeZ, blockSizeX,
241       blockSizeY, blockSizeZ);
242 
243   // Replace the loop terminator (loops contain only a single block) with the
244   // gpu terminator and move the operations from the loop body block to the gpu
245   // launch body block.  Do not move the entire block because of the difference
246   // in block arguments.
247   Operation &terminator = innermostForOp.getBody()->back();
248   Location terminatorLoc = terminator.getLoc();
249   terminator.erase();
250   builder.setInsertionPointToEnd(innermostForOp.getBody());
251   builder.create<gpu::TerminatorOp>(terminatorLoc, llvm::None);
252   launchOp.getBody().front().getOperations().splice(
253       launchOp.getBody().front().begin(),
254       innermostForOp.getBody()->getOperations());
255 
256   // Remap the loop iterators to use block/thread identifiers instead.  Loops
257   // may iterate from LB with step S whereas GPU thread/block ids always iterate
258   // from 0 to N with step 1.  Therefore, loop induction variables are replaced
259   // with (gpu-thread/block-id * S) + LB.
260   builder.setInsertionPointToStart(&launchOp.getBody().front());
261   auto *lbArgumentIt = lbs.begin();
262   auto *stepArgumentIt = steps.begin();
263   for (const auto &en : llvm::enumerate(ivs)) {
264     Value id =
265         en.index() < numBlockDims
266             ? getDim3Value(launchOp.getBlockIds(), en.index())
267             : getDim3Value(launchOp.getThreadIds(), en.index() - numBlockDims);
268     Value step = steps[en.index()];
269     if (!isConstantOne(step))
270       id = builder.create<arith::MulIOp>(rootForOp.getLoc(), step, id);
271 
272     Value ivReplacement =
273         builder.create<arith::AddIOp>(rootForOp.getLoc(), *lbArgumentIt, id);
274     en.value().replaceAllUsesWith(ivReplacement);
275     std::advance(lbArgumentIt, 1);
276     std::advance(stepArgumentIt, 1);
277   }
278 
279   // We are done and can erase the original outermost loop.
280   rootForOp.erase();
281 }
282 
283 // Generic loop to GPU kernel conversion function.
284 static LogicalResult convertAffineLoopNestToGPULaunch(AffineForOp forOp,
285                                                       unsigned numBlockDims,
286                                                       unsigned numThreadDims) {
287   if (failed(checkAffineLoopNestMappable(forOp, numBlockDims, numThreadDims)))
288     return failure();
289 
290   AffineLoopToGpuConverter converter;
291   auto maybeInnerLoop =
292       converter.collectBounds(forOp, numBlockDims + numThreadDims);
293   if (!maybeInnerLoop)
294     return failure();
295   converter.createLaunch(forOp, *maybeInnerLoop, numBlockDims, numThreadDims);
296 
297   return success();
298 }
299 
300 LogicalResult mlir::convertAffineLoopNestToGPULaunch(AffineForOp forOp,
301                                                      unsigned numBlockDims,
302                                                      unsigned numThreadDims) {
303   return ::convertAffineLoopNestToGPULaunch(forOp, numBlockDims, numThreadDims);
304 }
305 
306 namespace {
307 struct ParallelToGpuLaunchLowering : public OpRewritePattern<ParallelOp> {
308   using OpRewritePattern<ParallelOp>::OpRewritePattern;
309 
310   LogicalResult matchAndRewrite(ParallelOp parallelOp,
311                                 PatternRewriter &rewriter) const override;
312 };
313 } // namespace
314 
315 /// Tries to derive a static upper bound from the defining operation of
316 /// `upperBound`.
317 static Value deriveStaticUpperBound(Value upperBound,
318                                     PatternRewriter &rewriter) {
319   if (auto op = upperBound.getDefiningOp<arith::ConstantIndexOp>()) {
320     return op;
321   }
322 
323   if (auto minOp = upperBound.getDefiningOp<AffineMinOp>()) {
324     for (const AffineExpr &result : minOp.getMap().getResults()) {
325       if (auto constExpr = result.dyn_cast<AffineConstantExpr>()) {
326         return rewriter.create<arith::ConstantIndexOp>(minOp.getLoc(),
327                                                        constExpr.getValue());
328       }
329     }
330   }
331 
332   if (auto minOp = upperBound.getDefiningOp<arith::MinSIOp>()) {
333     for (Value operand : {minOp.getLhs(), minOp.getRhs()}) {
334       if (auto staticBound = deriveStaticUpperBound(operand, rewriter))
335         return staticBound;
336     }
337   }
338 
339   if (auto multiplyOp = upperBound.getDefiningOp<arith::MulIOp>()) {
340     if (auto lhs = dyn_cast_or_null<arith::ConstantIndexOp>(
341             deriveStaticUpperBound(multiplyOp.getOperand(0), rewriter)
342                 .getDefiningOp()))
343       if (auto rhs = dyn_cast_or_null<arith::ConstantIndexOp>(
344               deriveStaticUpperBound(multiplyOp.getOperand(1), rewriter)
345                   .getDefiningOp())) {
346         // Assumptions about the upper bound of minimum computations no longer
347         // work if multiplied by mixed signs, so abort in this case.
348         if ((lhs.value() < 0) != (rhs.value() < 0))
349           return {};
350 
351         return rewriter.create<arith::ConstantIndexOp>(
352             multiplyOp.getLoc(), lhs.value() * rhs.value());
353       }
354   }
355 
356   return {};
357 }
358 
359 static bool isMappedToProcessor(gpu::Processor processor) {
360   return processor != gpu::Processor::Sequential;
361 }
362 
363 static unsigned getLaunchOpArgumentNum(gpu::Processor processor) {
364   switch (processor) {
365   case gpu::Processor::BlockX:
366     return 0;
367   case gpu::Processor::BlockY:
368     return 1;
369   case gpu::Processor::BlockZ:
370     return 2;
371   case gpu::Processor::ThreadX:
372     return 3;
373   case gpu::Processor::ThreadY:
374     return 4;
375   case gpu::Processor::ThreadZ:
376     return 5;
377   default:;
378   }
379   llvm_unreachable(
380       "invalid processor type while retrieving launch op argument number");
381 }
382 
383 /// Modifies the current transformation state to capture the effect of the given
384 /// `scf.parallel` operation on index substitutions and the operations to be
385 /// inserted.
386 /// Specifically, if a dimension of a parallel loop is mapped to a hardware id,
387 /// this function will
388 /// - compute the loop index based on the hardware id and affine map from the
389 ///   mapping and update `cloningMap` to substitute all uses.
390 /// - derive a new upper bound for the hardware id and augment the provided
391 ///   `gpu.launch operation` accordingly.
392 /// - if the upper bound is imprecise, insert a conditional in the `gpu.launch`
393 ///   and update the rewriter to insert into the conditional's body.
394 /// If the dimension is mapped to sequential,
395 /// - insert a for loop into the body and update the rewriter to insert into
396 ///   the for loop's body.
397 /// - update the `cloningMap` to replace uses of the index with the index of
398 ///   the new for loop.
399 /// In either case,
400 /// - append the instructions from the loops body to worklist, in reverse order.
401 /// To note the end of the current scope in case a loop or conditional was
402 /// inserted, a sentinel (the `gpu.launch` operation) is inserted into the
403 /// worklist. This signals the processor of the worklist to pop the rewriter
404 /// one scope-level up.
405 static LogicalResult processParallelLoop(
406     ParallelOp parallelOp, gpu::LaunchOp launchOp,
407     BlockAndValueMapping &cloningMap, SmallVectorImpl<Operation *> &worklist,
408     DenseMap<gpu::Processor, Value> &bounds, PatternRewriter &rewriter) {
409   // TODO: Verify that this is a valid GPU mapping.
410   // processor ids: 0-2 block [x/y/z], 3-5 -> thread [x/y/z], 6-> sequential
411   ArrayAttr mapping =
412       parallelOp->getAttrOfType<ArrayAttr>(gpu::getMappingAttrName());
413 
414   // TODO: Support reductions.
415   if (!mapping || parallelOp.getNumResults() != 0)
416     return failure();
417 
418   Location loc = parallelOp.getLoc();
419 
420   auto launchIndependent = [&launchOp](Value val) {
421     return val.getParentRegion()->isAncestor(launchOp->getParentRegion());
422   };
423 
424   auto ensureLaunchIndependent = [&rewriter,
425                                   launchIndependent](Value val) -> Value {
426     if (launchIndependent(val))
427       return val;
428     if (auto constOp = val.getDefiningOp<arith::ConstantOp>())
429       return rewriter.create<arith::ConstantOp>(constOp.getLoc(),
430                                                 constOp.getValue());
431     return {};
432   };
433 
434   for (auto config : llvm::zip(
435            mapping, parallelOp.getInductionVars(), parallelOp.getLowerBound(),
436            parallelOp.getUpperBound(), parallelOp.getStep())) {
437     Attribute mappingAttribute;
438     Value iv, lowerBound, upperBound, step;
439     std::tie(mappingAttribute, iv, lowerBound, upperBound, step) = config;
440     auto annotation =
441         mappingAttribute.dyn_cast<gpu::ParallelLoopDimMappingAttr>();
442     if (!annotation)
443       return parallelOp.emitOpError()
444              << "expected mapping attribute for lowering to GPU";
445     Value newIndex;
446     gpu::Processor processor = annotation.getProcessor();
447 
448     if (isMappedToProcessor(processor)) {
449       // Use the corresponding thread/grid index as replacement for the loop iv.
450       Value operand =
451           launchOp.getBody().getArgument(getLaunchOpArgumentNum(processor));
452       // Take the indexmap and add the lower bound and step computations in.
453       // This computes operand * step + lowerBound.
454       // Use an affine map here so that it composes nicely with the provided
455       // annotation.
456       AffineMap lowerAndStep = AffineMap::get(
457           1, 2,
458           rewriter.getAffineDimExpr(0) * rewriter.getAffineSymbolExpr(0) +
459               rewriter.getAffineSymbolExpr(1));
460       newIndex = rewriter.create<AffineApplyOp>(
461           loc, annotation.getMap().compose(lowerAndStep),
462           ValueRange{operand, step, lowerBound});
463       // If there was also a bound, insert that, too.
464       // TODO: Check that we do not assign bounds twice.
465       if (annotation.getBound()) {
466         // We pass as the single operand to the bound-map the number of
467         // iterations, which is (upperBound - lowerBound) ceilDiv step. To
468         // support inner loops with dynamic upper bounds (as generated by e.g.
469         // tiling), try to derive a max for the bounds. If the used bound for
470         // the hardware id is imprecise, wrap the contained code into a
471         // conditional. If the lower-bound is constant or defined before the
472         // launch, we can use it in the launch bounds. Otherwise fail.
473         if (!launchIndependent(lowerBound) &&
474             !isa_and_nonnull<arith::ConstantOp>(lowerBound.getDefiningOp()))
475           return failure();
476         // The step must also be constant or defined outside of the loop nest.
477         if (!launchIndependent(step) &&
478             !isa_and_nonnull<arith::ConstantOp>(step.getDefiningOp()))
479           return failure();
480         // If the upper-bound is constant or defined before the launch, we can
481         // use it in the launch bounds directly. Otherwise try derive a bound.
482         bool boundIsPrecise =
483             launchIndependent(upperBound) ||
484             isa_and_nonnull<arith::ConstantOp>(upperBound.getDefiningOp());
485         {
486           PatternRewriter::InsertionGuard guard(rewriter);
487           rewriter.setInsertionPoint(launchOp);
488           if (!boundIsPrecise) {
489             upperBound = deriveStaticUpperBound(upperBound, rewriter);
490             if (!upperBound) {
491               return rewriter.notifyMatchFailure(
492                   parallelOp,
493                   "cannot derive loop-invariant upper bound for number of"
494                   "iterations");
495             }
496           }
497           // Compute the number of iterations needed. We compute this as an
498           // affine expression ceilDiv (upperBound - lowerBound) step. We use
499           // affine.apply here so that it composes nicely with the provided map.
500           AffineMap stepMap = AffineMap::get(
501               1, 2,
502               ((rewriter.getAffineDimExpr(0) - rewriter.getAffineSymbolExpr(0))
503                    .ceilDiv(rewriter.getAffineSymbolExpr(1))));
504           Value launchBound = rewriter.create<AffineApplyOp>(
505               loc, annotation.getBound().compose(stepMap),
506               ValueRange{
507                   ensureLaunchIndependent(
508                       cloningMap.lookupOrDefault(upperBound)),
509                   ensureLaunchIndependent(
510                       cloningMap.lookupOrDefault(lowerBound)),
511                   ensureLaunchIndependent(cloningMap.lookupOrDefault(step))});
512           // todo(herhut,ravishankarm): Update the behavior of setMappingAttr
513           // when this condition is relaxed.
514           if (bounds.find(processor) != bounds.end()) {
515             return rewriter.notifyMatchFailure(
516                 parallelOp, "cannot redefine the bound for processor " +
517                                 Twine(static_cast<int64_t>(processor)));
518           }
519           bounds[processor] = launchBound;
520         }
521         if (!boundIsPrecise) {
522           // We are using an approximation, create a surrounding conditional.
523           Value originalBound = std::get<3>(config);
524           arith::CmpIOp pred = rewriter.create<arith::CmpIOp>(
525               loc, arith::CmpIPredicate::slt, newIndex,
526               cloningMap.lookupOrDefault(originalBound));
527           scf::IfOp ifOp = rewriter.create<scf::IfOp>(loc, pred, false);
528           rewriter.setInsertionPointToStart(&ifOp.getThenRegion().front());
529           // Put a sentinel into the worklist so we know when to pop out of the
530           // if body again. We use the launchOp here, as that cannot be part of
531           // the bodies instruction.
532           worklist.push_back(launchOp.getOperation());
533         }
534       }
535     } else {
536       // Create a sequential for loop.
537       auto loopOp = rewriter.create<scf::ForOp>(
538           loc, cloningMap.lookupOrDefault(lowerBound),
539           cloningMap.lookupOrDefault(upperBound),
540           cloningMap.lookupOrDefault(step));
541       newIndex = loopOp.getInductionVar();
542       rewriter.setInsertionPointToStart(loopOp.getBody());
543       // Put a sentinel into the worklist so we know when to pop out of the loop
544       // body again. We use the launchOp here, as that cannot be part of the
545       // bodies instruction.
546       worklist.push_back(launchOp.getOperation());
547     }
548     cloningMap.map(iv, newIndex);
549   }
550 
551   // Propagate custom user defined optional attributes, that can be used at
552   // later stage, such as extension data for GPU kernel dispatch
553   for (const auto &namedAttr : parallelOp->getAttrs()) {
554     if (namedAttr.getName() == gpu::getMappingAttrName() ||
555         namedAttr.getName() == ParallelOp::getOperandSegmentSizeAttr())
556       continue;
557     launchOp->setAttr(namedAttr.getName(), namedAttr.getValue());
558   }
559 
560   Block *body = parallelOp.getBody();
561   worklist.reserve(worklist.size() + body->getOperations().size());
562   for (Operation &op : llvm::reverse(body->without_terminator()))
563     worklist.push_back(&op);
564   return success();
565 }
566 
567 /// Lower a `scf.parallel` operation into a corresponding `gpu.launch`
568 /// operation.
569 ///
570 /// This essentially transforms a loop nest into a corresponding SIMT function.
571 /// The conversion is driven by mapping annotations on the `scf.parallel`
572 /// operations. The mapping is provided via a `DictionaryAttribute` named
573 /// `mapping`, which has three entries:
574 ///  - processor: the hardware id to map to. 0-2 are block dimensions, 3-5 are
575 ///               thread dimensions and 6 is sequential.
576 ///  - map : An affine map that is used to pre-process hardware ids before
577 ///          substitution.
578 ///  - bound : An affine map that is used to compute the bound of the hardware
579 ///            id based on an upper bound of the number of iterations.
580 /// If the `scf.parallel` contains nested `scf.parallel` operations, those
581 /// need to be annotated, as well. Structurally, the transformation works by
582 /// splicing all operations from nested `scf.parallel` operations into a single
583 /// sequence. Indices mapped to hardware ids are substituted with those ids,
584 /// wheras sequential mappings result in a sequential for-loop. To have more
585 /// flexibility when mapping code to hardware ids, the transform supports two
586 /// affine maps. The first `map` is used to compute the actual index for
587 /// substitution from the hardware id. The second `bound` is used to compute the
588 /// launch dimension for the hardware id from the number of iterations the
589 /// mapped loop is performing. Note that the number of iterations might be
590 /// imprecise if the corresponding loop-bounds are loop-dependent. In such case,
591 /// the hardware id might iterate over additional indices. The transformation
592 /// caters for this by predicating the created sequence of instructions on
593 /// the actual loop bound. This only works if an static upper bound for the
594 /// dynamic loop bound can be derived, currently via analyzing `affine.min`
595 /// operations.
596 LogicalResult
597 ParallelToGpuLaunchLowering::matchAndRewrite(ParallelOp parallelOp,
598                                              PatternRewriter &rewriter) const {
599   // Mark the operation as visited for recursive legality check.
600   parallelOp->setAttr(kVisitedAttrName, rewriter.getUnitAttr());
601 
602   // We can only transform starting at the outer-most loop. Launches inside of
603   // parallel loops are not supported.
604   if (auto parentLoop = parallelOp->getParentOfType<ParallelOp>())
605     return failure();
606   // Create a launch operation. We start with bound one for all grid/block
607   // sizes. Those will be refined later as we discover them from mappings.
608   Location loc = parallelOp.getLoc();
609   Value constantOne =
610       rewriter.create<arith::ConstantIndexOp>(parallelOp.getLoc(), 1);
611   gpu::LaunchOp launchOp = rewriter.create<gpu::LaunchOp>(
612       parallelOp.getLoc(), constantOne, constantOne, constantOne, constantOne,
613       constantOne, constantOne);
614   rewriter.setInsertionPointToEnd(&launchOp.getBody().front());
615   rewriter.create<gpu::TerminatorOp>(loc);
616   rewriter.setInsertionPointToStart(&launchOp.getBody().front());
617 
618   BlockAndValueMapping cloningMap;
619   llvm::DenseMap<gpu::Processor, Value> launchBounds;
620   SmallVector<Operation *, 16> worklist;
621   if (failed(processParallelLoop(parallelOp, launchOp, cloningMap, worklist,
622                                  launchBounds, rewriter)))
623     return failure();
624 
625   // Whether we have seen any side-effects. Reset when leaving an inner scope.
626   bool seenSideeffects = false;
627   // Whether we have left a nesting scope (and hence are no longer innermost).
628   bool leftNestingScope = false;
629   while (!worklist.empty()) {
630     Operation *op = worklist.pop_back_val();
631     // Now walk over the body and clone it.
632     // TODO: This is only correct if there either is no further scf.parallel
633     //       nested or this code is side-effect free. Otherwise we might need
634     //       predication. We are overly conservative for now and only allow
635     //       side-effects in the innermost scope.
636     if (auto nestedParallel = dyn_cast<ParallelOp>(op)) {
637       // Before entering a nested scope, make sure there have been no
638       // sideeffects until now.
639       if (seenSideeffects)
640         return failure();
641       // A nested scf.parallel needs insertion of code to compute indices.
642       // Insert that now. This will also update the worklist with the loops
643       // body.
644       if (failed(processParallelLoop(nestedParallel, launchOp, cloningMap,
645                                      worklist, launchBounds, rewriter)))
646         return failure();
647     } else if (op == launchOp.getOperation()) {
648       // Found our sentinel value. We have finished the operations from one
649       // nesting level, pop one level back up.
650       auto *parent = rewriter.getInsertionPoint()->getParentOp();
651       rewriter.setInsertionPointAfter(parent);
652       leftNestingScope = true;
653       seenSideeffects = false;
654     } else {
655       // Otherwise we copy it over.
656       Operation *clone = rewriter.clone(*op, cloningMap);
657       cloningMap.map(op->getResults(), clone->getResults());
658       // Check for side effects.
659       // TODO: Handle region side effects properly.
660       seenSideeffects |=
661           !isMemoryEffectFree(clone) || clone->getNumRegions() != 0;
662       // If we are no longer in the innermost scope, sideeffects are disallowed.
663       if (seenSideeffects && leftNestingScope)
664         return failure();
665     }
666   }
667 
668   // Now that we succeeded creating the launch operation, also update the
669   // bounds.
670   for (auto bound : launchBounds)
671     launchOp.setOperand(getLaunchOpArgumentNum(std::get<0>(bound)),
672                         std::get<1>(bound));
673 
674   rewriter.eraseOp(parallelOp);
675   return success();
676 }
677 
678 void mlir::populateParallelLoopToGPUPatterns(RewritePatternSet &patterns) {
679   patterns.add<ParallelToGpuLaunchLowering>(patterns.getContext());
680 }
681 
682 void mlir::configureParallelLoopToGPULegality(ConversionTarget &target) {
683   target.addLegalDialect<memref::MemRefDialect>();
684   target.addDynamicallyLegalOp<scf::ParallelOp>([](scf::ParallelOp parallelOp) {
685     return !parallelOp->hasAttr(gpu::getMappingAttrName()) ||
686            parallelOp->hasAttr(kVisitedAttrName);
687   });
688 }
689 
690 void mlir::finalizeParallelLoopToGPUConversion(Operation *op) {
691   op->walk([](scf::ParallelOp parallelOp) {
692     parallelOp->removeAttr(kVisitedAttrName);
693   });
694 }
695