1 //===- AsyncRegionRewriter.cpp - Implementation of GPU async rewriters ----===//
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 pattern rewriters that make GPU op
10 // within a region execute asynchronously.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "PassDetail.h"
15 #include "mlir/Dialect/Async/IR/Async.h"
16 #include "mlir/Dialect/GPU/GPUDialect.h"
17 #include "mlir/Dialect/GPU/Passes.h"
18 #include "mlir/Dialect/GPU/Utils.h"
19 #include "mlir/Dialect/StandardOps/IR/Ops.h"
20 #include "mlir/IR/BlockAndValueMapping.h"
21 #include "mlir/IR/Builders.h"
22 #include "mlir/IR/PatternMatch.h"
23 #include "mlir/IR/SymbolTable.h"
24 #include "mlir/Support/LLVM.h"
25 #include "mlir/Transforms/RegionUtils.h"
26 #include "llvm/ADT/TypeSwitch.h"
27 
28 using namespace mlir;
29 namespace {
30 class GpuAsyncRegionPass : public GpuAsyncRegionPassBase<GpuAsyncRegionPass> {
31   struct ThreadTokenCallback;
32   struct DeferWaitCallback;
33   void runOnFunction() override;
34 };
35 } // namespace
36 
37 static bool isTerminator(Operation *op) { return !op->isKnownNonTerminator(); }
38 static bool hasSideEffects(Operation *op) {
39   return !MemoryEffectOpInterface::hasNoEffect(op);
40 }
41 
42 // Region walk callback which makes GPU ops implementing the AsyncOpInterface
43 // execute asynchronously.
44 struct GpuAsyncRegionPass::ThreadTokenCallback {
45   ThreadTokenCallback(MLIRContext &context) : builder(&context) {}
46 
47   // If `op` implements the AsyncOpInterface, insert a `gpu.wait async` to
48   // create a current token (unless it already exists), and 'thread' that token
49   // through the `op` so that it executes asynchronously.
50   //
51   // If `op` is a terminator or an op with side-effects, insert a `gpu.wait` to
52   // host-synchronize execution. A `!gpu.async.token` will therefore only be
53   // used inside of its block and GPU execution will always synchronize with
54   // the host at block boundaries.
55   WalkResult operator()(Operation *op) {
56     if (isa<gpu::LaunchOp>(op))
57       return op->emitOpError("replace with gpu.launch_func first");
58     if (isa<gpu::WaitOp>(op))
59       return op->emitOpError("unexpected pre-existing gpu.wait");
60     builder.setInsertionPoint(op);
61     if (auto asyncOp = dyn_cast<gpu::AsyncOpInterface>(op))
62       return rewriteAsyncOp(asyncOp); // Replace GPU op with async version.
63     if (!currentToken)
64       return success();
65     // Insert host synchronization before terminator or op with side effects.
66     if (isTerminator(op) || hasSideEffects(op))
67       currentToken = createWaitOp(op->getLoc(), Type(), {currentToken});
68     return success();
69   }
70 
71 private:
72   // Replaces asyncOp with a clone that returns a token.
73   LogicalResult rewriteAsyncOp(gpu::AsyncOpInterface asyncOp) {
74     auto *op = asyncOp.getOperation();
75     if (asyncOp.getAsyncToken())
76       // TODO: Support ops that are already async.
77       return op->emitOpError("is already async");
78     if (op->getNumRegions() > 0)
79       return op->emitOpError("regions are not supported");
80 
81     // If there is no current token, insert a `gpu.wait async` without
82     // dependencies to create one.
83     if (!currentToken)
84       currentToken = createWaitOp(op->getLoc(), tokenType, {});
85     asyncOp.addAsyncDependency(currentToken);
86 
87     // Clone the op to return a token in addition to the other results.
88     SmallVector<Type, 1> resultTypes = {tokenType};
89     resultTypes.reserve(1 + op->getNumResults());
90     copy(op->getResultTypes(), std::back_inserter(resultTypes));
91     auto *newOp = Operation::create(op->getLoc(), op->getName(), resultTypes,
92                                     op->getOperands(), op->getMutableAttrDict(),
93                                     op->getSuccessors());
94 
95     // Replace the op with the async clone.
96     auto results = newOp->getResults();
97     currentToken = results.front();
98     builder.insert(newOp);
99     op->replaceAllUsesWith(results.drop_front());
100     op->erase();
101 
102     return success();
103   }
104 
105   Value createWaitOp(Location loc, Type resultType, ValueRange operands) {
106     return builder.create<gpu::WaitOp>(loc, resultType, operands).asyncToken();
107   }
108 
109   OpBuilder builder;
110   const Type tokenType = builder.getType<gpu::AsyncTokenType>();
111   // The token that represents the current asynchronous dependency. It's valid
112   // range starts with a `gpu.wait async` op, and ends with a `gpu.wait` op.
113   // In between, each gpu::AsyncOpInterface depends on the current token and
114   // produces the new one.
115   Value currentToken = {};
116 };
117 
118 // Callback for `async.execute` ops which tries to push the contained
119 // synchronous `gpu.wait` op to the dependencies of the `async.execute`.
120 struct GpuAsyncRegionPass::DeferWaitCallback {
121   // If the `executeOp`s token is used only in `async.execute` or `async.await`
122   // ops, add the region's last `gpu.wait` op to the worklist if it is
123   // synchronous and is the last op with side effects.
124   void operator()(async::ExecuteOp executeOp) {
125     if (!areAllUsersExecuteOrAwait(executeOp.token()))
126       return;
127     // async.execute's region is currently restricted to one block.
128     for (auto &op : llvm::reverse(executeOp.getBody()->without_terminator())) {
129       if (auto waitOp = dyn_cast<gpu::WaitOp>(op)) {
130         if (!waitOp.asyncToken())
131           worklist.push_back(waitOp);
132         return;
133       }
134       if (hasSideEffects(&op))
135         return;
136     }
137   }
138 
139   // The destructor performs the actual rewrite work.
140   ~DeferWaitCallback() {
141     for (size_t i = 0; i < worklist.size(); ++i) {
142       auto waitOp = worklist[i];
143       auto executeOp = waitOp.getParentOfType<async::ExecuteOp>();
144       auto numDependencies = waitOp.asyncDependencies().size();
145 
146       // Erase `gpu.wait` and return async dependencies from region instead.
147       auto &yieldOp = executeOp.getBody()->getOperations().back();
148       yieldOp.insertOperands(yieldOp.getNumOperands(),
149                              waitOp.asyncDependencies());
150       waitOp.erase();
151       auto asyncTokens = addAsyncTokenResults(executeOp, numDependencies);
152 
153       // Add the async dependency to each user of the `async.execute` token.
154       for (Operation *user : executeOp.token().getUsers())
155         addAsyncDependencyAfter(asyncTokens, user);
156     }
157   }
158 
159 private:
160   // Append `count` `!async.value<!gpu.async.token>` results to `executeOp`.
161   static ValueRange addAsyncTokenResults(async::ExecuteOp &executeOp,
162                                          unsigned count) {
163     auto numResults = executeOp.getNumResults() + count;
164 
165     // Construct new result type list with `count` additional types.
166     SmallVector<Type, 2> resultTypes;
167     resultTypes.reserve(numResults);
168     copy(executeOp.getResultTypes(), std::back_inserter(resultTypes));
169     OpBuilder builder(executeOp);
170     auto tokenType = builder.getType<gpu::AsyncTokenType>();
171     resultTypes.resize(numResults, tokenType);
172 
173     // Clone executeOp with the extra `!gpu.async.token` results.
174     auto newOp = builder.create<async::ExecuteOp>(
175         executeOp.getLoc(), TypeRange{resultTypes}.drop_front() /*drop token*/,
176         executeOp.dependencies(), executeOp.operands());
177     BlockAndValueMapping mapper;
178     newOp.getRegion().getBlocks().clear();
179     executeOp.getRegion().cloneInto(&newOp.getRegion(), mapper);
180 
181     // Replace executeOp with cloned one.
182     executeOp.getOperation()->replaceAllUsesWith(
183         newOp.getResults().drop_back(count));
184     executeOp.erase();
185     executeOp = newOp;
186 
187     // Return the new result values.
188     return executeOp.getResults().take_back(count);
189   }
190 
191   // Returns whether all token users are either 'async.execute' or 'async.await'
192   // ops. This is used as a requirement for pushing 'gpu.wait' ops from a
193   // 'async.execute' body to it's users. Specifically, we do not allow
194   // terminator users, because it could mean that the `async.execute` is inside
195   // control flow code.
196   static bool areAllUsersExecuteOrAwait(Value token) {
197     return llvm::all_of(token.getUsers(), [](Operation *user) {
198       return isa<async::ExecuteOp, async::AwaitOp>(user);
199     });
200   }
201 
202   // Add the `asyncToken` as dependency as needed after `op`.
203   void addAsyncDependencyAfter(ValueRange asyncTokens, Operation *op) {
204     OpBuilder builder(op->getContext());
205     auto loc = op->getLoc();
206 
207     Block::iterator it;
208     SmallVector<Value, 1> tokens;
209     tokens.reserve(asyncTokens.size());
210     TypeSwitch<Operation *>(op)
211         .Case<async::AwaitOp>([&](auto awaitOp) {
212           // Add async.await ops to wait for the !gpu.async.tokens.
213           builder.setInsertionPointAfter(op);
214           for (auto asyncToken : asyncTokens)
215             tokens.push_back(
216                 builder.create<async::AwaitOp>(loc, asyncToken).result());
217           // Set `it` after the inserted async.await ops.
218           it = builder.getInsertionPoint();
219         })
220         .Case<async::ExecuteOp>([&](auto executeOp) {
221           // Set `it` to the beginning of the region and add asyncTokens to the
222           // async.execute operands.
223           it = executeOp.getBody()->begin();
224           executeOp.operandsMutable().append(asyncTokens);
225           SmallVector<Type, 1> tokenTypes(
226               asyncTokens.size(), builder.getType<gpu::AsyncTokenType>());
227           copy(executeOp.getBody()->addArguments(tokenTypes),
228                std::back_inserter(tokens));
229         });
230 
231     // Advance `it` to terminator or op with side-effects.
232     it = std::find_if(it, Block::iterator(), [](Operation &op) {
233       return isTerminator(&op) || hasSideEffects(&op);
234     });
235 
236     // If `op` implements the AsyncOpInterface, add `token` to the list of async
237     // dependencies.
238     if (auto asyncOp = dyn_cast<gpu::AsyncOpInterface>(*it)) {
239       for (auto token : tokens)
240         asyncOp.addAsyncDependency(token);
241       return;
242     }
243 
244     // Otherwise, insert a gpu.wait before 'it'.
245     builder.setInsertionPoint(it->getBlock(), it);
246     auto waitOp = builder.create<gpu::WaitOp>(loc, Type{}, tokens);
247 
248     // If the new waitOp is at the end of an async.execute region, add it to the
249     // worklist. 'operator()(executeOp)' would do the same, but this is faster.
250     auto executeOp = dyn_cast<async::ExecuteOp>(it->getParentOp());
251     if (executeOp && areAllUsersExecuteOrAwait(executeOp.token()) &&
252         !it->getNextNode())
253       worklist.push_back(waitOp);
254   }
255 
256   SmallVector<gpu::WaitOp, 8> worklist;
257 };
258 
259 // Replaces synchronous GPU ops in the op's region with asynchronous ones and
260 // inserts the necessary synchronization (as gpu.wait ops). Assumes sequential
261 // execution semantics and that no GPU ops are asynchronous yet.
262 void GpuAsyncRegionPass::runOnFunction() {
263   if (getFunction()
264           .getRegion()
265           .walk(ThreadTokenCallback(getContext()))
266           .wasInterrupted())
267     return signalPassFailure();
268 
269   // Collect gpu.wait ops that we can move out of gpu.execute regions.
270   getFunction().getRegion().walk(DeferWaitCallback());
271 }
272 
273 std::unique_ptr<OperationPass<FuncOp>> mlir::createGpuAsyncRegionPass() {
274   return std::make_unique<GpuAsyncRegionPass>();
275 }
276