xref: /llvm-project/mlir/lib/Dialect/GPU/Transforms/KernelOutlining.cpp (revision d566a5cd22b4a653f10698f90c691a1452dad5ce)
1 //===- KernelOutlining.cpp - Implementation of GPU kernel outlining -------===//
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 the GPU dialect kernel outlining pass.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "mlir/Dialect/GPU/Transforms/Passes.h"
14 
15 #include "mlir/AsmParser/AsmParser.h"
16 #include "mlir/Dialect/Arith/IR/Arith.h"
17 #include "mlir/Dialect/ControlFlow/IR/ControlFlowOps.h"
18 #include "mlir/Dialect/DLTI/DLTI.h"
19 #include "mlir/Dialect/Func/IR/FuncOps.h"
20 #include "mlir/Dialect/GPU/IR/GPUDialect.h"
21 #include "mlir/Dialect/GPU/Transforms/Utils.h"
22 #include "mlir/Dialect/MemRef/IR/MemRef.h"
23 #include "mlir/IR/Builders.h"
24 #include "mlir/IR/BuiltinAttributes.h"
25 #include "mlir/IR/IRMapping.h"
26 #include "mlir/IR/Matchers.h"
27 #include "mlir/IR/SymbolTable.h"
28 #include "mlir/Support/LLVM.h"
29 #include "mlir/Transforms/RegionUtils.h"
30 #include <limits>
31 
32 namespace mlir {
33 #define GEN_PASS_DEF_GPULAUNCHSINKINDEXCOMPUTATIONS
34 #define GEN_PASS_DEF_GPUKERNELOUTLINING
35 #include "mlir/Dialect/GPU/Transforms/Passes.h.inc"
36 } // namespace mlir
37 
38 using namespace mlir;
39 
40 template <typename OpTy>
41 static void createForAllDimensions(OpBuilder &builder, Location loc,
42                                    SmallVectorImpl<Value> &values) {
43   for (auto dim : {gpu::Dimension::x, gpu::Dimension::y, gpu::Dimension::z})
44     values.push_back(builder.create<OpTy>(loc, builder.getIndexType(), dim));
45 }
46 
47 /// Adds operations generating block/thread ids and grid/block dimensions at the
48 /// beginning of the `launchFuncOpBody` region. Add mapping from argument in
49 /// entry block of `launchOpBody`, to the corresponding result value of the
50 /// added operations.
51 static void injectGpuIndexOperations(Location loc, Region &launchFuncOpBody,
52                                      Region &launchOpBody, IRMapping &map,
53                                      bool hasCluster = false) {
54   OpBuilder builder(loc->getContext());
55   Block &firstBlock = launchOpBody.front();
56   builder.setInsertionPointToStart(&launchFuncOpBody.front());
57   SmallVector<Value> indexOps;
58   // The order is important here, as it must match the order of the arguments
59   createForAllDimensions<gpu::BlockIdOp>(builder, loc, indexOps);
60   createForAllDimensions<gpu::ThreadIdOp>(builder, loc, indexOps);
61   createForAllDimensions<gpu::GridDimOp>(builder, loc, indexOps);
62   createForAllDimensions<gpu::BlockDimOp>(builder, loc, indexOps);
63   if (hasCluster) {
64     createForAllDimensions<gpu::ClusterIdOp>(builder, loc, indexOps);
65     createForAllDimensions<gpu::ClusterDimOp>(builder, loc, indexOps);
66   }
67   // Replace the leading 12 function args with the respective thread/block index
68   // operations. Iterate backwards since args are erased and indices change.
69   for (const auto &indexOp : enumerate(indexOps))
70     map.map(firstBlock.getArgument(indexOp.index()), indexOp.value());
71 }
72 
73 /// Identifies operations that are beneficial to sink into kernels. These
74 /// operations may not have side-effects, as otherwise sinking (and hence
75 /// duplicating them) is not legal.
76 static bool isLikelyAnIndexComputation(Operation *op) {
77   return matchPattern(op, m_Constant()) ||
78          isa<memref::DimOp, arith::SelectOp, arith::CmpIOp>(op);
79 }
80 
81 /// For a given operation `op`, computes whether it is beneficial to sink the
82 /// operation into the kernel. An operation can be sunk if doing so does not
83 /// introduce new kernel arguments. Whether a value is already available in the
84 /// kernel (and hence does not introduce new arguments) is checked by
85 /// querying `existingDependencies` and `availableValues`.
86 /// If an operand is not yet available, we recursively check whether it can be
87 /// made available by siking its defining op.
88 /// Operations that are indentified for sinking are added to `beneficiaryOps` in
89 /// the order they should appear in the kernel. Furthermore, `availableValues`
90 /// is updated with results that will be available after sinking the identified
91 /// ops.
92 static bool extractBeneficiaryOps(
93     Operation *op, const SetVector<Value> &existingDependencies,
94     SetVector<Operation *> &beneficiaryOps,
95     llvm::SmallPtrSetImpl<Value> &availableValues,
96     llvm::function_ref<bool(Operation *)> isSinkingBeneficiary) {
97   if (beneficiaryOps.count(op))
98     return true;
99 
100   if (!isSinkingBeneficiary(op))
101     return false;
102 
103   for (Value operand : op->getOperands()) {
104     // It is already visible in the kernel, keep going.
105     if (availableValues.count(operand))
106       continue;
107     // Else check whether it can be made available via sinking or already is a
108     // dependency.
109     Operation *definingOp = operand.getDefiningOp();
110     if ((!definingOp || !extractBeneficiaryOps(definingOp, existingDependencies,
111                                                beneficiaryOps, availableValues,
112                                                isSinkingBeneficiary)) &&
113         !existingDependencies.count(operand))
114       return false;
115   }
116   // We will sink the operation, mark its results as now available.
117   beneficiaryOps.insert(op);
118   for (Value result : op->getResults())
119     availableValues.insert(result);
120   return true;
121 }
122 
123 LogicalResult mlir::sinkOperationsIntoLaunchOp(
124     gpu::LaunchOp launchOp,
125     llvm::function_ref<bool(Operation *)> isSinkingBeneficiary) {
126   assert(isSinkingBeneficiary);
127   Region &launchOpBody = launchOp.getBody();
128 
129   // Identify uses from values defined outside of the scope of the launch
130   // operation.
131   SetVector<Value> sinkCandidates;
132   getUsedValuesDefinedAbove(launchOpBody, sinkCandidates);
133 
134   SetVector<Operation *> toBeSunk;
135   llvm::SmallPtrSet<Value, 4> availableValues;
136   for (Value operand : sinkCandidates) {
137     Operation *operandOp = operand.getDefiningOp();
138     if (!operandOp)
139       continue;
140     extractBeneficiaryOps(operandOp, sinkCandidates, toBeSunk, availableValues,
141                           isSinkingBeneficiary);
142   }
143 
144   // Insert operations so that the defs get cloned before uses.
145   IRMapping map;
146   OpBuilder builder(launchOpBody);
147   for (Operation *op : toBeSunk) {
148     Operation *clonedOp = builder.clone(*op, map);
149     // Only replace uses within the launch op.
150     for (auto pair : llvm::zip(op->getResults(), clonedOp->getResults()))
151       replaceAllUsesInRegionWith(std::get<0>(pair), std::get<1>(pair),
152                                  launchOp.getBody());
153   }
154   return success();
155 }
156 
157 /// Return the provided KernelDim3 as an array of i32 constants if possible.
158 static DenseI32ArrayAttr maybeConstantDimsAttr(gpu::KernelDim3 dims) {
159   SmallVector<int32_t, 3> constants;
160   MLIRContext *ctx = dims.x.getContext();
161   for (Value v : {dims.x, dims.y, dims.z}) {
162     APInt constValue;
163     if (!matchPattern(v, m_ConstantInt(&constValue)))
164       return nullptr;
165     // In the event someone called for a too-large block or grid dimension,
166     // don't set bounds as it is likely to cause more confusing behavior.
167     if (constValue.ugt(std::numeric_limits<uint32_t>::max()))
168       return nullptr;
169     constants.push_back(
170         constValue.getLimitedValue(std::numeric_limits<uint32_t>::max()));
171   }
172   return DenseI32ArrayAttr::get(ctx, constants);
173 }
174 
175 /// Outline the `gpu.launch` operation body into a kernel function. Replace
176 /// `gpu.terminator` operations by `gpu.return` in the generated function.
177 /// Set block and grid size bounds if known.
178 static gpu::GPUFuncOp outlineKernelFuncImpl(gpu::LaunchOp launchOp,
179                                             StringRef kernelFnName,
180                                             SetVector<Value> &operands) {
181   Location loc = launchOp.getLoc();
182   // Create a builder with no insertion point, insertion will happen separately
183   // due to symbol table manipulation.
184   OpBuilder builder(launchOp.getContext());
185   Region &launchOpBody = launchOp.getBody();
186 
187   // Identify uses from values defined outside of the scope of the launch
188   // operation.
189   getUsedValuesDefinedAbove(launchOpBody, operands);
190 
191   // Create the gpu.func operation.
192   SmallVector<Type, 4> kernelOperandTypes;
193   kernelOperandTypes.reserve(operands.size());
194   for (Value operand : operands) {
195     kernelOperandTypes.push_back(operand.getType());
196   }
197   FunctionType type =
198       FunctionType::get(launchOp.getContext(), kernelOperandTypes, {});
199   auto outlinedFunc = builder.create<gpu::GPUFuncOp>(
200       loc, kernelFnName, type,
201       TypeRange(ValueRange(launchOp.getWorkgroupAttributions())),
202       TypeRange(ValueRange(launchOp.getPrivateAttributions())));
203   outlinedFunc->setAttr(gpu::GPUDialect::getKernelFuncAttrName(),
204                         builder.getUnitAttr());
205 
206   // If we can infer bounds on the grid and/or block sizes from the arguments
207   // to the launch op, propagate them to the generated kernel. This is safe
208   // because multiple launches with the same body are not deduplicated.
209   if (auto blockBounds =
210           maybeConstantDimsAttr(launchOp.getBlockSizeOperandValues()))
211     outlinedFunc->setAttr(gpu::GPUFuncOp::getKnownBlockSizeAttrName(),
212                           blockBounds);
213   if (auto gridBounds =
214           maybeConstantDimsAttr(launchOp.getGridSizeOperandValues()))
215     outlinedFunc->setAttr(gpu::GPUFuncOp::getKnownGridSizeAttrName(),
216                           gridBounds);
217 
218   IRMapping map;
219 
220   // Map the arguments corresponding to the launch parameters like blockIdx,
221   // threadIdx, etc. If cluster is present, then we also generate clusterIdx and
222   // clusterDim.
223   Region &outlinedFuncBody = outlinedFunc.getBody();
224   injectGpuIndexOperations(loc, outlinedFuncBody, launchOpBody, map,
225                            launchOp.hasClusterSize());
226 
227   // Map memory attributions from the LaunOp op to the GPUFuncOp attributions.
228   for (const auto &[launchArg, funcArg] :
229        llvm::zip(launchOp.getWorkgroupAttributions(),
230                  outlinedFunc.getWorkgroupAttributions()))
231     map.map(launchArg, funcArg);
232   for (const auto &[launchArg, funcArg] :
233        llvm::zip(launchOp.getPrivateAttributions(),
234                  outlinedFunc.getPrivateAttributions()))
235     map.map(launchArg, funcArg);
236 
237   // Map arguments from gpu.launch region to the arguments of the gpu.func
238   // operation.
239   Block &entryBlock = outlinedFuncBody.front();
240   for (const auto &operand : enumerate(operands))
241     map.map(operand.value(), entryBlock.getArgument(operand.index()));
242 
243   // Clone the region of the gpu.launch operation into the gpu.func operation.
244   launchOpBody.cloneInto(&outlinedFuncBody, map);
245 
246   // Replace the terminator op with returns.
247   for (Block &block : launchOpBody) {
248     Block *clonedBlock = map.lookup(&block);
249     auto terminator = dyn_cast<gpu::TerminatorOp>(clonedBlock->getTerminator());
250     if (!terminator)
251       continue;
252     OpBuilder replacer(terminator);
253     replacer.create<gpu::ReturnOp>(terminator->getLoc());
254     terminator->erase();
255   }
256 
257   // Splice now the entry block of the gpu.launch operation at the end of the
258   // gpu.func entry block and erase the redundant block.
259   Block *clonedLaunchOpEntry = map.lookup(&launchOpBody.front());
260   entryBlock.getOperations().splice(entryBlock.getOperations().end(),
261                                     clonedLaunchOpEntry->getOperations());
262   clonedLaunchOpEntry->erase();
263 
264   return outlinedFunc;
265 }
266 
267 gpu::GPUFuncOp mlir::outlineKernelFunc(gpu::LaunchOp launchOp,
268                                        StringRef kernelFnName,
269                                        llvm::SmallVectorImpl<Value> &operands) {
270   DenseSet<Value> inputOperandSet;
271   inputOperandSet.insert(operands.begin(), operands.end());
272   SetVector<Value> operandSet(operands.begin(), operands.end());
273   auto funcOp = outlineKernelFuncImpl(launchOp, kernelFnName, operandSet);
274   for (auto operand : operandSet) {
275     if (!inputOperandSet.count(operand))
276       operands.push_back(operand);
277   }
278   return funcOp;
279 }
280 
281 /// Replace `gpu.launch` operations with an `gpu.launch_func` operation
282 /// launching `kernelFunc`. The kernel func contains the body of the
283 /// `gpu.launch` with constant region arguments inlined.
284 static void convertToLaunchFuncOp(gpu::LaunchOp launchOp,
285                                   gpu::GPUFuncOp kernelFunc,
286                                   ValueRange operands) {
287   OpBuilder builder(launchOp);
288   // The launch op has an optional dynamic shared memory size. If it doesn't
289   // exist, we use zero.
290   Value asyncToken = launchOp.getAsyncToken();
291   std::optional<gpu::KernelDim3> clusterSize =
292       launchOp.getClusterSizeOperandValues();
293   auto launchFunc = builder.create<gpu::LaunchFuncOp>(
294       launchOp.getLoc(), kernelFunc, launchOp.getGridSizeOperandValues(),
295       launchOp.getBlockSizeOperandValues(),
296       launchOp.getDynamicSharedMemorySize(), operands,
297       asyncToken ? asyncToken.getType() : nullptr,
298       launchOp.getAsyncDependencies(), clusterSize);
299   launchOp.replaceAllUsesWith(launchFunc);
300   launchOp.erase();
301 }
302 
303 namespace {
304 /// Pass that moves ops which are likely an index computation into gpu.launch
305 /// body.
306 class GpuLaunchSinkIndexComputationsPass
307     : public impl::GpuLaunchSinkIndexComputationsBase<
308           GpuLaunchSinkIndexComputationsPass> {
309 public:
310   void runOnOperation() override {
311     Operation *op = getOperation();
312     if (op->walk([](gpu::LaunchOp launch) {
313             // Pull in instructions that can be sunk
314             if (failed(sinkOperationsIntoLaunchOp(launch,
315                                                   isLikelyAnIndexComputation)))
316               return WalkResult::interrupt();
317 
318             return WalkResult::advance();
319           }).wasInterrupted())
320       signalPassFailure();
321   }
322 };
323 
324 /// Pass that moves the kernel of each LaunchOp into its separate nested module.
325 ///
326 /// This pass moves the kernel code of each LaunchOp into a function created
327 /// inside a nested module. It also creates an external function of the same
328 /// name in the parent module.
329 ///
330 /// The gpu.modules are intended to be compiled to a cubin blob independently in
331 /// a separate pass. The external functions can then be annotated with the
332 /// symbol of the cubin accessor function.
333 class GpuKernelOutliningPass
334     : public impl::GpuKernelOutliningBase<GpuKernelOutliningPass> {
335 public:
336   GpuKernelOutliningPass(StringRef dlStr) {
337     if (!dlStr.empty() && !dataLayoutStr.hasValue())
338       dataLayoutStr = dlStr.str();
339   }
340 
341   GpuKernelOutliningPass(const GpuKernelOutliningPass &other)
342       : GpuKernelOutliningBase(other), dataLayoutSpec(other.dataLayoutSpec) {
343     dataLayoutStr = other.dataLayoutStr.getValue();
344   }
345 
346   LogicalResult initialize(MLIRContext *context) override {
347     // Initialize the data layout specification from the data layout string.
348     if (!dataLayoutStr.empty()) {
349       Attribute resultAttr = mlir::parseAttribute(dataLayoutStr, context);
350       if (!resultAttr)
351         return failure();
352 
353       dataLayoutSpec = dyn_cast<DataLayoutSpecInterface>(resultAttr);
354       if (!dataLayoutSpec)
355         return failure();
356     }
357 
358     return success();
359   }
360 
361   void runOnOperation() override {
362     SymbolTable symbolTable(getOperation());
363     bool modified = false;
364     for (auto func : getOperation().getOps<SymbolOpInterface>()) {
365       // Insert just after the function.
366       Block::iterator insertPt(func->getNextNode());
367       auto funcWalkResult = func.walk([&](gpu::LaunchOp op) {
368         SetVector<Value> operands;
369         std::string kernelFnName =
370             Twine(op->getParentOfType<SymbolOpInterface>().getName(), "_kernel")
371                 .str();
372 
373         gpu::GPUFuncOp outlinedFunc =
374             outlineKernelFuncImpl(op, kernelFnName, operands);
375 
376         // Create nested module and insert outlinedFunc. The module will
377         // originally get the same name as the function, but may be renamed on
378         // insertion into the parent module.
379         auto kernelModule = createKernelModule(outlinedFunc, symbolTable);
380         symbolTable.insert(kernelModule, insertPt);
381 
382         // Potentially changes signature, pulling in constants.
383         convertToLaunchFuncOp(op, outlinedFunc, operands.getArrayRef());
384         modified = true;
385         return WalkResult::advance();
386       });
387       if (funcWalkResult.wasInterrupted())
388         return signalPassFailure();
389     }
390 
391     // If any new module was inserted in this module, annotate this module as
392     // a container module.
393     if (modified)
394       getOperation()->setAttr(gpu::GPUDialect::getContainerModuleAttrName(),
395                               UnitAttr::get(&getContext()));
396   }
397 
398 private:
399   /// Returns a gpu.module containing kernelFunc and all callees (recursive).
400   gpu::GPUModuleOp createKernelModule(gpu::GPUFuncOp kernelFunc,
401                                       const SymbolTable &parentSymbolTable) {
402     // TODO: This code cannot use an OpBuilder because it must be inserted into
403     // a SymbolTable by the caller. SymbolTable needs to be refactored to
404     // prevent manual building of Ops with symbols in code using SymbolTables
405     // and then this needs to use the OpBuilder.
406     auto *context = getOperation().getContext();
407     OpBuilder builder(context);
408     auto kernelModule = builder.create<gpu::GPUModuleOp>(kernelFunc.getLoc(),
409                                                          kernelFunc.getName());
410 
411     // If a valid data layout spec was provided, attach it to the kernel module.
412     // Otherwise, the default data layout will be used.
413     if (dataLayoutSpec)
414       kernelModule->setAttr(DLTIDialect::kDataLayoutAttrName, dataLayoutSpec);
415 
416     SymbolTable symbolTable(kernelModule);
417     symbolTable.insert(kernelFunc);
418 
419     SmallVector<Operation *, 8> symbolDefWorklist = {kernelFunc};
420     while (!symbolDefWorklist.empty()) {
421       if (std::optional<SymbolTable::UseRange> symbolUses =
422               SymbolTable::getSymbolUses(symbolDefWorklist.pop_back_val())) {
423         for (SymbolTable::SymbolUse symbolUse : *symbolUses) {
424           StringRef symbolName =
425               cast<FlatSymbolRefAttr>(symbolUse.getSymbolRef()).getValue();
426           if (symbolTable.lookup(symbolName))
427             continue;
428 
429           Operation *symbolDefClone =
430               parentSymbolTable.lookup(symbolName)->clone();
431           symbolDefWorklist.push_back(symbolDefClone);
432           symbolTable.insert(symbolDefClone);
433         }
434       }
435     }
436 
437     return kernelModule;
438   }
439 
440   Option<std::string> dataLayoutStr{
441       *this, "data-layout-str",
442       llvm::cl::desc("String containing the data layout specification to be "
443                      "attached to the GPU kernel module")};
444 
445   DataLayoutSpecInterface dataLayoutSpec;
446 };
447 
448 } // namespace
449 
450 std::unique_ptr<Pass> mlir::createGpuLauchSinkIndexComputationsPass() {
451   return std::make_unique<GpuLaunchSinkIndexComputationsPass>();
452 }
453 
454 std::unique_ptr<OperationPass<ModuleOp>>
455 mlir::createGpuKernelOutliningPass(StringRef dataLayoutStr) {
456   return std::make_unique<GpuKernelOutliningPass>(dataLayoutStr);
457 }
458