1 //===- SCFToSPIRV.cpp - SCF to SPIR-V Patterns ----------------------------===//
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 file implements patterns to convert SCF dialect to SPIR-V dialect.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "mlir/Conversion/SCFToSPIRV/SCFToSPIRV.h"
14 #include "mlir/Dialect/SCF/SCF.h"
15 #include "mlir/Dialect/SPIRV/IR/SPIRVDialect.h"
16 #include "mlir/Dialect/SPIRV/IR/SPIRVOps.h"
17 #include "mlir/Dialect/SPIRV/Transforms/SPIRVConversion.h"
18 #include "mlir/IR/BuiltinOps.h"
19 #include "mlir/Transforms/DialectConversion.h"
20 
21 using namespace mlir;
22 
23 //===----------------------------------------------------------------------===//
24 // Context
25 //===----------------------------------------------------------------------===//
26 
27 namespace mlir {
28 struct ScfToSPIRVContextImpl {
29   // Map between the spirv region control flow operation (spv.mlir.loop or
30   // spv.mlir.selection) to the VariableOp created to store the region results.
31   // The order of the VariableOp matches the order of the results.
32   DenseMap<Operation *, SmallVector<spirv::VariableOp, 8>> outputVars;
33 };
34 } // namespace mlir
35 
36 /// We use ScfToSPIRVContext to store information about the lowering of the scf
37 /// region that need to be used later on. When we lower scf.for/scf.if we create
38 /// VariableOp to store the results. We need to keep track of the VariableOp
39 /// created as we need to insert stores into them when lowering Yield. Those
40 /// StoreOp cannot be created earlier as they may use a different type than
41 /// yield operands.
42 ScfToSPIRVContext::ScfToSPIRVContext() {
43   impl = std::make_unique<ScfToSPIRVContextImpl>();
44 }
45 
46 ScfToSPIRVContext::~ScfToSPIRVContext() = default;
47 
48 //===----------------------------------------------------------------------===//
49 // Pattern Declarations
50 //===----------------------------------------------------------------------===//
51 
52 namespace {
53 /// Common class for all vector to GPU patterns.
54 template <typename OpTy>
55 class SCFToSPIRVPattern : public OpConversionPattern<OpTy> {
56 public:
57   SCFToSPIRVPattern<OpTy>(MLIRContext *context, SPIRVTypeConverter &converter,
58                           ScfToSPIRVContextImpl *scfToSPIRVContext)
59       : OpConversionPattern<OpTy>::OpConversionPattern(converter, context),
60         scfToSPIRVContext(scfToSPIRVContext), typeConverter(converter) {}
61 
62 protected:
63   ScfToSPIRVContextImpl *scfToSPIRVContext;
64   // FIXME: We explicitly keep a reference of the type converter here instead of
65   // passing it to OpConversionPattern during construction. This effectively
66   // bypasses the conversion framework's automation on type conversion. This is
67   // needed right now because the conversion framework will unconditionally
68   // legalize all types used by SCF ops upon discovering them, for example, the
69   // types of loop carried values. We use SPIR-V variables for those loop
70   // carried values. Depending on the available capabilities, the SPIR-V
71   // variable can be different, for example, cooperative matrix or normal
72   // variable. We'd like to detach the conversion of the loop carried values
73   // from the SCF ops (which is mainly a region). So we need to "mark" types
74   // used by SCF ops as legal, if to use the conversion framework for type
75   // conversion. There isn't a straightforward way to do that yet, as when
76   // converting types, ops aren't taken into consideration. Therefore, we just
77   // bypass the framework's type conversion for now.
78   SPIRVTypeConverter &typeConverter;
79 };
80 
81 /// Pattern to convert a scf::ForOp within kernel functions into spirv::LoopOp.
82 class ForOpConversion final : public SCFToSPIRVPattern<scf::ForOp> {
83 public:
84   using SCFToSPIRVPattern<scf::ForOp>::SCFToSPIRVPattern;
85 
86   LogicalResult
87   matchAndRewrite(scf::ForOp forOp, OpAdaptor adaptor,
88                   ConversionPatternRewriter &rewriter) const override;
89 };
90 
91 /// Pattern to convert a scf::IfOp within kernel functions into
92 /// spirv::SelectionOp.
93 class IfOpConversion final : public SCFToSPIRVPattern<scf::IfOp> {
94 public:
95   using SCFToSPIRVPattern<scf::IfOp>::SCFToSPIRVPattern;
96 
97   LogicalResult
98   matchAndRewrite(scf::IfOp ifOp, OpAdaptor adaptor,
99                   ConversionPatternRewriter &rewriter) const override;
100 };
101 
102 class TerminatorOpConversion final : public SCFToSPIRVPattern<scf::YieldOp> {
103 public:
104   using SCFToSPIRVPattern<scf::YieldOp>::SCFToSPIRVPattern;
105 
106   LogicalResult
107   matchAndRewrite(scf::YieldOp terminatorOp, OpAdaptor adaptor,
108                   ConversionPatternRewriter &rewriter) const override;
109 };
110 
111 class WhileOpConversion final : public SCFToSPIRVPattern<scf::WhileOp> {
112 public:
113   using SCFToSPIRVPattern<scf::WhileOp>::SCFToSPIRVPattern;
114 
115   LogicalResult
116   matchAndRewrite(scf::WhileOp forOp, OpAdaptor adaptor,
117                   ConversionPatternRewriter &rewriter) const override;
118 };
119 } // namespace
120 
121 /// Helper function to replaces SCF op outputs with SPIR-V variable loads.
122 /// We create VariableOp to handle the results value of the control flow region.
123 /// spv.mlir.loop/spv.mlir.selection currently don't yield value. Right after
124 /// the loop we load the value from the allocation and use it as the SCF op
125 /// result.
126 template <typename ScfOp, typename OpTy>
127 static void replaceSCFOutputValue(ScfOp scfOp, OpTy newOp,
128                                   ConversionPatternRewriter &rewriter,
129                                   ScfToSPIRVContextImpl *scfToSPIRVContext,
130                                   ArrayRef<Type> returnTypes) {
131 
132   Location loc = scfOp.getLoc();
133   auto &allocas = scfToSPIRVContext->outputVars[newOp];
134   // Clearing the allocas is necessary in case a dialect conversion path failed
135   // previously, and this is the second attempt of this conversion.
136   allocas.clear();
137   SmallVector<Value, 8> resultValue;
138   for (Type convertedType : returnTypes) {
139     auto pointerType =
140         spirv::PointerType::get(convertedType, spirv::StorageClass::Function);
141     rewriter.setInsertionPoint(newOp);
142     auto alloc = rewriter.create<spirv::VariableOp>(
143         loc, pointerType, spirv::StorageClass::Function,
144         /*initializer=*/nullptr);
145     allocas.push_back(alloc);
146     rewriter.setInsertionPointAfter(newOp);
147     Value loadResult = rewriter.create<spirv::LoadOp>(loc, alloc);
148     resultValue.push_back(loadResult);
149   }
150   rewriter.replaceOp(scfOp, resultValue);
151 }
152 
153 static Region::iterator getBlockIt(Region &region, unsigned index) {
154   return std::next(region.begin(), index);
155 }
156 
157 //===----------------------------------------------------------------------===//
158 // scf::ForOp
159 //===----------------------------------------------------------------------===//
160 
161 LogicalResult
162 ForOpConversion::matchAndRewrite(scf::ForOp forOp, OpAdaptor adaptor,
163                                  ConversionPatternRewriter &rewriter) const {
164   // scf::ForOp can be lowered to the structured control flow represented by
165   // spirv::LoopOp by making the continue block of the spirv::LoopOp the loop
166   // latch and the merge block the exit block. The resulting spirv::LoopOp has a
167   // single back edge from the continue to header block, and a single exit from
168   // header to merge.
169   auto loc = forOp.getLoc();
170   auto loopOp = rewriter.create<spirv::LoopOp>(loc, spirv::LoopControl::None);
171   loopOp.addEntryAndMergeBlock();
172 
173   OpBuilder::InsertionGuard guard(rewriter);
174   // Create the block for the header.
175   auto *header = new Block();
176   // Insert the header.
177   loopOp.body().getBlocks().insert(getBlockIt(loopOp.body(), 1), header);
178 
179   // Create the new induction variable to use.
180   BlockArgument newIndVar = header->addArgument(adaptor.lowerBound().getType());
181   for (Value arg : adaptor.initArgs())
182     header->addArgument(arg.getType());
183   Block *body = forOp.getBody();
184 
185   // Apply signature conversion to the body of the forOp. It has a single block,
186   // with argument which is the induction variable. That has to be replaced with
187   // the new induction variable.
188   TypeConverter::SignatureConversion signatureConverter(
189       body->getNumArguments());
190   signatureConverter.remapInput(0, newIndVar);
191   for (unsigned i = 1, e = body->getNumArguments(); i < e; i++)
192     signatureConverter.remapInput(i, header->getArgument(i));
193   body = rewriter.applySignatureConversion(&forOp.getLoopBody(),
194                                            signatureConverter);
195 
196   // Move the blocks from the forOp into the loopOp. This is the body of the
197   // loopOp.
198   rewriter.inlineRegionBefore(forOp->getRegion(0), loopOp.body(),
199                               getBlockIt(loopOp.body(), 2));
200 
201   SmallVector<Value, 8> args(1, adaptor.lowerBound());
202   args.append(adaptor.initArgs().begin(), adaptor.initArgs().end());
203   // Branch into it from the entry.
204   rewriter.setInsertionPointToEnd(&(loopOp.body().front()));
205   rewriter.create<spirv::BranchOp>(loc, header, args);
206 
207   // Generate the rest of the loop header.
208   rewriter.setInsertionPointToEnd(header);
209   auto *mergeBlock = loopOp.getMergeBlock();
210   auto cmpOp = rewriter.create<spirv::SLessThanOp>(
211       loc, rewriter.getI1Type(), newIndVar, adaptor.upperBound());
212 
213   rewriter.create<spirv::BranchConditionalOp>(
214       loc, cmpOp, body, ArrayRef<Value>(), mergeBlock, ArrayRef<Value>());
215 
216   // Generate instructions to increment the step of the induction variable and
217   // branch to the header.
218   Block *continueBlock = loopOp.getContinueBlock();
219   rewriter.setInsertionPointToEnd(continueBlock);
220 
221   // Add the step to the induction variable and branch to the header.
222   Value updatedIndVar = rewriter.create<spirv::IAddOp>(
223       loc, newIndVar.getType(), newIndVar, adaptor.step());
224   rewriter.create<spirv::BranchOp>(loc, header, updatedIndVar);
225 
226   // Infer the return types from the init operands. Vector type may get
227   // converted to CooperativeMatrix or to Vector type, to avoid having complex
228   // extra logic to figure out the right type we just infer it from the Init
229   // operands.
230   SmallVector<Type, 8> initTypes;
231   for (auto arg : adaptor.initArgs())
232     initTypes.push_back(arg.getType());
233   replaceSCFOutputValue(forOp, loopOp, rewriter, scfToSPIRVContext, initTypes);
234   return success();
235 }
236 
237 //===----------------------------------------------------------------------===//
238 // scf::IfOp
239 //===----------------------------------------------------------------------===//
240 
241 LogicalResult
242 IfOpConversion::matchAndRewrite(scf::IfOp ifOp, OpAdaptor adaptor,
243                                 ConversionPatternRewriter &rewriter) const {
244   // When lowering `scf::IfOp` we explicitly create a selection header block
245   // before the control flow diverges and a merge block where control flow
246   // subsequently converges.
247   auto loc = ifOp.getLoc();
248 
249   // Create `spv.selection` operation, selection header block and merge block.
250   auto selectionOp =
251       rewriter.create<spirv::SelectionOp>(loc, spirv::SelectionControl::None);
252   auto *mergeBlock =
253       rewriter.createBlock(&selectionOp.body(), selectionOp.body().end());
254   rewriter.create<spirv::MergeOp>(loc);
255 
256   OpBuilder::InsertionGuard guard(rewriter);
257   auto *selectionHeaderBlock =
258       rewriter.createBlock(&selectionOp.body().front());
259 
260   // Inline `then` region before the merge block and branch to it.
261   auto &thenRegion = ifOp.thenRegion();
262   auto *thenBlock = &thenRegion.front();
263   rewriter.setInsertionPointToEnd(&thenRegion.back());
264   rewriter.create<spirv::BranchOp>(loc, mergeBlock);
265   rewriter.inlineRegionBefore(thenRegion, mergeBlock);
266 
267   auto *elseBlock = mergeBlock;
268   // If `else` region is not empty, inline that region before the merge block
269   // and branch to it.
270   if (!ifOp.elseRegion().empty()) {
271     auto &elseRegion = ifOp.elseRegion();
272     elseBlock = &elseRegion.front();
273     rewriter.setInsertionPointToEnd(&elseRegion.back());
274     rewriter.create<spirv::BranchOp>(loc, mergeBlock);
275     rewriter.inlineRegionBefore(elseRegion, mergeBlock);
276   }
277 
278   // Create a `spv.BranchConditional` operation for selection header block.
279   rewriter.setInsertionPointToEnd(selectionHeaderBlock);
280   rewriter.create<spirv::BranchConditionalOp>(loc, adaptor.condition(),
281                                               thenBlock, ArrayRef<Value>(),
282                                               elseBlock, ArrayRef<Value>());
283 
284   SmallVector<Type, 8> returnTypes;
285   for (auto result : ifOp.results()) {
286     auto convertedType = typeConverter.convertType(result.getType());
287     returnTypes.push_back(convertedType);
288   }
289   replaceSCFOutputValue(ifOp, selectionOp, rewriter, scfToSPIRVContext,
290                         returnTypes);
291   return success();
292 }
293 
294 //===----------------------------------------------------------------------===//
295 // scf::YieldOp
296 //===----------------------------------------------------------------------===//
297 
298 /// Yield is lowered to stores to the VariableOp created during lowering of the
299 /// parent region. For loops we also need to update the branch looping back to
300 /// the header with the loop carried values.
301 LogicalResult TerminatorOpConversion::matchAndRewrite(
302     scf::YieldOp terminatorOp, OpAdaptor adaptor,
303     ConversionPatternRewriter &rewriter) const {
304   ValueRange operands = adaptor.getOperands();
305 
306   // If the region is return values, store each value into the associated
307   // VariableOp created during lowering of the parent region.
308   if (!operands.empty()) {
309     auto &allocas = scfToSPIRVContext->outputVars[terminatorOp->getParentOp()];
310     if (allocas.size() != operands.size())
311       return failure();
312 
313     auto loc = terminatorOp.getLoc();
314     for (unsigned i = 0, e = operands.size(); i < e; i++)
315       rewriter.create<spirv::StoreOp>(loc, allocas[i], operands[i]);
316     if (isa<spirv::LoopOp>(terminatorOp->getParentOp())) {
317       // For loops we also need to update the branch jumping back to the header.
318       auto br =
319           cast<spirv::BranchOp>(rewriter.getInsertionBlock()->getTerminator());
320       SmallVector<Value, 8> args(br.getBlockArguments());
321       args.append(operands.begin(), operands.end());
322       rewriter.setInsertionPoint(br);
323       rewriter.create<spirv::BranchOp>(terminatorOp.getLoc(), br.getTarget(),
324                                        args);
325       rewriter.eraseOp(br);
326     }
327   }
328   rewriter.eraseOp(terminatorOp);
329   return success();
330 }
331 
332 //===----------------------------------------------------------------------===//
333 // scf::WhileOp
334 //===----------------------------------------------------------------------===//
335 
336 LogicalResult
337 WhileOpConversion::matchAndRewrite(scf::WhileOp whileOp, OpAdaptor adaptor,
338                                    ConversionPatternRewriter &rewriter) const {
339   auto loc = whileOp.getLoc();
340   auto loopOp = rewriter.create<spirv::LoopOp>(loc, spirv::LoopControl::None);
341   loopOp.addEntryAndMergeBlock();
342 
343   OpBuilder::InsertionGuard guard(rewriter);
344 
345   Region &beforeRegion = whileOp.before();
346   Region &afterRegion = whileOp.after();
347 
348   Block &entryBlock = *loopOp.getEntryBlock();
349   Block &beforeBlock = beforeRegion.front();
350   Block &afterBlock = afterRegion.front();
351   Block &mergeBlock = *loopOp.getMergeBlock();
352 
353   auto cond = cast<scf::ConditionOp>(beforeBlock.getTerminator());
354   SmallVector<Value> condArgs;
355   if (failed(rewriter.getRemappedValues(cond.args(), condArgs)))
356     return failure();
357 
358   Value conditionVal = rewriter.getRemappedValue(cond.condition());
359   if (!conditionVal)
360     return failure();
361 
362   auto yield = cast<scf::YieldOp>(afterBlock.getTerminator());
363   SmallVector<Value> yieldArgs;
364   if (failed(rewriter.getRemappedValues(yield.results(), yieldArgs)))
365     return failure();
366 
367   // Move the while before block as the initial loop header block.
368   rewriter.inlineRegionBefore(beforeRegion, loopOp.body(),
369                               getBlockIt(loopOp.body(), 1));
370 
371   // Move the while after block as the initial loop body block.
372   rewriter.inlineRegionBefore(afterRegion, loopOp.body(),
373                               getBlockIt(loopOp.body(), 2));
374 
375   // Jump from the loop entry block to the loop header block.
376   rewriter.setInsertionPointToEnd(&entryBlock);
377   rewriter.create<spirv::BranchOp>(loc, &beforeBlock, adaptor.inits());
378 
379   auto condLoc = cond.getLoc();
380 
381   SmallVector<Value> resultValues(condArgs.size());
382 
383   // For other SCF ops, the scf.yield op yields the value for the whole SCF op.
384   // So we use the scf.yield op as the anchor to create/load/store SPIR-V local
385   // variables. But for the scf.while op, the scf.yield op yields a value for
386   // the before region, which may not matching the whole op's result. Instead,
387   // the scf.condition op returns values matching the whole op's results. So we
388   // need to create/load/store variables according to that.
389   for (auto it : llvm::enumerate(condArgs)) {
390     auto res = it.value();
391     auto i = it.index();
392     auto pointerType =
393         spirv::PointerType::get(res.getType(), spirv::StorageClass::Function);
394 
395     // Create local variables before the scf.while op.
396     rewriter.setInsertionPoint(loopOp);
397     auto alloc = rewriter.create<spirv::VariableOp>(
398         condLoc, pointerType, spirv::StorageClass::Function,
399         /*initializer=*/nullptr);
400 
401     // Load the final result values after the scf.while op.
402     rewriter.setInsertionPointAfter(loopOp);
403     auto loadResult = rewriter.create<spirv::LoadOp>(condLoc, alloc);
404     resultValues[i] = loadResult;
405 
406     // Store the current iteration's result value.
407     rewriter.setInsertionPointToEnd(&beforeBlock);
408     rewriter.create<spirv::StoreOp>(condLoc, alloc, res);
409   }
410 
411   rewriter.setInsertionPointToEnd(&beforeBlock);
412   rewriter.replaceOpWithNewOp<spirv::BranchConditionalOp>(
413       cond, conditionVal, &afterBlock, condArgs, &mergeBlock, llvm::None);
414 
415   // Convert the scf.yield op to a branch back to the header block.
416   rewriter.setInsertionPointToEnd(&afterBlock);
417   rewriter.replaceOpWithNewOp<spirv::BranchOp>(yield, &beforeBlock, yieldArgs);
418 
419   rewriter.replaceOp(whileOp, resultValues);
420   return success();
421 }
422 
423 //===----------------------------------------------------------------------===//
424 // Hooks
425 //===----------------------------------------------------------------------===//
426 
427 void mlir::populateSCFToSPIRVPatterns(SPIRVTypeConverter &typeConverter,
428                                       ScfToSPIRVContext &scfToSPIRVContext,
429                                       RewritePatternSet &patterns) {
430   patterns.add<ForOpConversion, IfOpConversion, TerminatorOpConversion,
431                WhileOpConversion>(patterns.getContext(), typeConverter,
432                                   scfToSPIRVContext.getImpl());
433 }
434