[mlir][gpu] Move gpu.wait ops from async.execute regions to its dependencies.

This can prevent unnecessary host synchronization.

Reviewed By: herhut

Differential Revision: https://reviews.llvm.org/D90346

mlir/lib/Dialect/GPU/Transforms/AsyncRegionRewriter.cpp

@@ -12,6 +12,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "PassDetail.h"
+#include "mlir/Dialect/Async/IR/Async.h"
 #include "mlir/Dialect/GPU/GPUDialect.h"
 #include "mlir/Dialect/GPU/Passes.h"
 #include "mlir/Dialect/GPU/Utils.h"
@@ -22,24 +23,35 @@
 #include "mlir/IR/SymbolTable.h"
 #include "mlir/Support/LLVM.h"
 #include "mlir/Transforms/RegionUtils.h"
+#include "llvm/ADT/TypeSwitch.h"
 
 using namespace mlir;
 namespace {
 class GpuAsyncRegionPass : public GpuAsyncRegionPassBase<GpuAsyncRegionPass> {
-  struct Callback;
+  struct ThreadTokenCallback;
+  struct DeferWaitCallback;
   void runOnFunction() override;
 };
 } // namespace
 
+static bool isTerminator(Operation *op) { return !op->isKnownNonTerminator(); }
+static bool hasSideEffects(Operation *op) {
+  return !MemoryEffectOpInterface::hasNoEffect(op);
+}
+
 // Region walk callback which makes GPU ops implementing the AsyncOpInterface
 // execute asynchronously.
-struct GpuAsyncRegionPass::Callback {
+struct GpuAsyncRegionPass::ThreadTokenCallback {
+  ThreadTokenCallback(MLIRContext &context) : builder(&context) {}
+
   // If `op` implements the AsyncOpInterface, insert a `gpu.wait async` to
   // create a current token (unless it already exists), and 'thread' that token
   // through the `op` so that it executes asynchronously.
   //
   // If `op` is a terminator or an op with side-effects, insert a `gpu.wait` to
-  // host-synchronize execution.
+  // host-synchronize execution. A `!gpu.async.token` will therefore only be
+  // used inside of its block and GPU execution will always synchronize with
+  // the host at block boundaries.
   WalkResult operator()(Operation *op) {
     if (isa<gpu::LaunchOp>(op))
       return op->emitOpError("replace with gpu.launch_func first");
@@ -50,14 +62,13 @@ struct GpuAsyncRegionPass::Callback {
       return rewriteAsyncOp(asyncOp); // Replace GPU op with async version.
     if (!currentToken)
       return success();
-    if (!op->hasTrait<OpTrait::IsTerminator>() &&
-        MemoryEffectOpInterface::hasNoEffect(op))
-      return success();
     // Insert host synchronization before terminator or op with side effects.
-    currentToken = createWaitOp(op->getLoc(), Type(), {currentToken});
+    if (isTerminator(op) || hasSideEffects(op))
+      currentToken = createWaitOp(op->getLoc(), Type(), {currentToken});
     return success();
   }
 
+private:
   // Replaces asyncOp with a clone that returns a token.
   LogicalResult rewriteAsyncOp(gpu::AsyncOpInterface asyncOp) {
     auto *op = asyncOp.getOperation();
@@ -104,13 +115,159 @@ struct GpuAsyncRegionPass::Callback {
   Value currentToken = {};
 };
 
+// Callback for `async.execute` ops which tries to push the contained
+// synchronous `gpu.wait` op to the dependencies of the `async.execute`.
+struct GpuAsyncRegionPass::DeferWaitCallback {
+  // If the `executeOp`s token is used only in `async.execute` or `async.await`
+  // ops, add the region's last `gpu.wait` op to the worklist if it is
+  // synchronous and is the last op with side effects.
+  void operator()(async::ExecuteOp executeOp) {
+    if (!areAllUsersExecuteOrAwait(executeOp.token()))
+      return;
+    // async.execute's region is currently restricted to one block.
+    for (auto &op : llvm::reverse(executeOp.getBody()->without_terminator())) {
+      if (auto waitOp = dyn_cast<gpu::WaitOp>(op)) {
+        if (!waitOp.asyncToken())
+          worklist.push_back(waitOp);
+        return;
+      }
+      if (hasSideEffects(&op))
+        return;
+    }
+  }
+
+  // The destructor performs the actual rewrite work.
+  ~DeferWaitCallback() {
+    for (size_t i = 0; i < worklist.size(); ++i) {
+      auto waitOp = worklist[i];
+      auto executeOp = waitOp.getParentOfType<async::ExecuteOp>();
+      auto numDependencies = waitOp.asyncDependencies().size();
+
+      // Erase `gpu.wait` and return async dependencies from region instead.
+      auto &yieldOp = executeOp.getBody()->getOperations().back();
+      yieldOp.insertOperands(yieldOp.getNumOperands(),
+                             waitOp.asyncDependencies());
+      waitOp.erase();
+      auto asyncTokens = addAsyncTokenResults(executeOp, numDependencies);
+
+      // Add the async dependency to each user of the `async.execute` token.
+      for (Operation *user : executeOp.token().getUsers())
+        addAsyncDependencyAfter(asyncTokens, user);
+    }
+  }
+
+private:
+  // Append `count` `!async.value<!gpu.async.token>` results to `executeOp`.
+  static ValueRange addAsyncTokenResults(async::ExecuteOp &executeOp,
+                                         unsigned count) {
+    auto numResults = executeOp.getNumResults() + count;
+
+    // Construct new result type list with `count` additional types.
+    SmallVector<Type, 2> resultTypes;
+    resultTypes.reserve(numResults);
+    copy(executeOp.getResultTypes(), std::back_inserter(resultTypes));
+    OpBuilder builder(executeOp);
+    auto tokenType = builder.getType<gpu::AsyncTokenType>();
+    resultTypes.resize(numResults, tokenType);
+
+    // Clone executeOp with the extra `!gpu.async.token` results.
+    auto newOp = builder.create<async::ExecuteOp>(
+        executeOp.getLoc(), TypeRange{resultTypes}.drop_front() /*drop token*/,
+        executeOp.dependencies(), executeOp.operands());
+    BlockAndValueMapping mapper;
+    newOp.getRegion().getBlocks().clear();
+    executeOp.getRegion().cloneInto(&newOp.getRegion(), mapper);
+
+    // Replace executeOp with cloned one.
+    executeOp.getOperation()->replaceAllUsesWith(
+        newOp.getResults().drop_back(count));
+    executeOp.erase();
+    executeOp = newOp;
+
+    // Return the new result values.
+    return executeOp.getResults().take_back(count);
+  }
+
+  // Returns whether all token users are either 'async.execute' or 'async.await'
+  // ops. This is used as a requirement for pushing 'gpu.wait' ops from a
+  // 'async.execute' body to it's users. Specifically, we do not allow
+  // terminator users, because it could mean that the `async.execute` is inside
+  // control flow code.
+  static bool areAllUsersExecuteOrAwait(Value token) {
+    return llvm::all_of(token.getUsers(), [](Operation *user) {
+      return isa<async::ExecuteOp, async::AwaitOp>(user);
+    });
+  }
+
+  // Add the `asyncToken` as dependency as needed after `op`.
+  void addAsyncDependencyAfter(ValueRange asyncTokens, Operation *op) {
+    OpBuilder builder(op->getContext());
+    auto loc = op->getLoc();
+
+    Block::iterator it;
+    SmallVector<Value, 1> tokens;
+    tokens.reserve(asyncTokens.size());
+    TypeSwitch<Operation *>(op)
+        .Case<async::AwaitOp>([&](auto awaitOp) {
+          // Add async.await ops to wait for the !gpu.async.tokens.
+          builder.setInsertionPointAfter(op);
+          for (auto asyncToken : asyncTokens)
+            tokens.push_back(
+                builder.create<async::AwaitOp>(loc, asyncToken).result());
+          // Set `it` after the inserted async.await ops.
+          it = builder.getInsertionPoint();
+        })
+        .Case<async::ExecuteOp>([&](auto executeOp) {
+          // Set `it` to the beginning of the region and add asyncTokens to the
+          // async.execute operands.
+          it = executeOp.getBody()->begin();
+          executeOp.operandsMutable().append(asyncTokens);
+          SmallVector<Type, 1> tokenTypes(
+              asyncTokens.size(), builder.getType<gpu::AsyncTokenType>());
+          copy(executeOp.getBody()->addArguments(tokenTypes),
+               std::back_inserter(tokens));
+        });
+
+    // Advance `it` to terminator or op with side-effects.
+    it = std::find_if(it, Block::iterator(), [](Operation &op) {
+      return isTerminator(&op) || hasSideEffects(&op);
+    });
+
+    // If `op` implements the AsyncOpInterface, add `token` to the list of async
+    // dependencies.
+    if (auto asyncOp = dyn_cast<gpu::AsyncOpInterface>(*it)) {
+      for (auto token : tokens)
+        asyncOp.addAsyncDependency(token);
+      return;
+    }
+
+    // Otherwise, insert a gpu.wait before 'it'.
+    builder.setInsertionPoint(it->getBlock(), it);
+    auto waitOp = builder.create<gpu::WaitOp>(loc, Type{}, tokens);
+
+    // If the new waitOp is at the end of an async.execute region, add it to the
+    // worklist. 'operator()(executeOp)' would do the same, but this is faster.
+    auto executeOp = dyn_cast<async::ExecuteOp>(it->getParentOp());
+    if (executeOp && areAllUsersExecuteOrAwait(executeOp.token()) &&
+        !it->getNextNode())
+      worklist.push_back(waitOp);
+  }
+
+  SmallVector<gpu::WaitOp, 8> worklist;
+};
+
 // Replaces synchronous GPU ops in the op's region with asynchronous ones and
 // inserts the necessary synchronization (as gpu.wait ops). Assumes sequential
 // execution semantics and that no GPU ops are asynchronous yet.
 void GpuAsyncRegionPass::runOnFunction() {
-  Callback callback{OpBuilder(&getContext())};
-  if (getFunction().getRegion().walk(callback).wasInterrupted())
+  if (getFunction()
+          .getRegion()
+          .walk(ThreadTokenCallback(getContext()))
+          .wasInterrupted())
     return signalPassFailure();
+
+  // Collect gpu.wait ops that we can move out of gpu.execute regions.
+  getFunction().getRegion().walk(DeferWaitCallback());
 }
 
 std::unique_ptr<OperationPass<FuncOp>> mlir::createGpuAsyncRegionPass() {

mlir/test/Dialect/GPU/async-region.mlir

@@ -24,4 +24,78 @@ module attributes {gpu.container_module} {
     return
   }
 
+  // CHECK-LABEL:func @defer_wait(%{{.*}}: index)
+  func @defer_wait(%sz : index) {
+    // CHECK: %[[a0:.*]], %[[f0:.*]] = async.execute
+    %a0 = async.execute {
+      // CHECK: %[[t:.*]] = gpu.launch_func async
+      gpu.launch_func @kernels::@kernel
+          blocks in (%sz, %sz, %sz) threads in (%sz, %sz, %sz)
+      // CHECK-NOT: gpu.wait
+      // CHECK: async.yield %[[t]]
+      async.yield
+    }
+
+    // CHECK: %[[a1:.*]], %[[f1:.*]] = async.execute
+    // CHECK-SAME: %[[f0]]
+    %a1 = async.execute [%a0] {
+      // CHECK: %[[t:.*]] = gpu.launch_func async
+      gpu.launch_func @kernels::@kernel
+          blocks in (%sz, %sz, %sz) threads in (%sz, %sz, %sz)
+      // CHECK-NOT: gpu.wait
+      // CHECK: async.yield %[[t]]
+      async.yield
+    }
+
+    // CHECK: async.await %[[a1]]
+    // CHECK: %[[t:.*]] = async.await %[[f1]]
+    // CHECK: gpu.wait [%[[t]]]
+    async.await %a1 : !async.token
+    return
+  }
+
+  // CHECK-LABEL:func @defer_wait_blocked_by_side_effect(%{{.*}}: index)
+  func @defer_wait_blocked_by_side_effect(%sz : index) {
+    // CHECK: %[[a:.*]] = async.execute
+    %a = async.execute {
+      // CHECK: %[[t:.*]] = gpu.launch_func async
+      gpu.launch_func @kernels::@kernel
+          blocks in (%sz, %sz, %sz) threads in (%sz, %sz, %sz)
+      // CHECK: gpu.wait [%[[t]]]
+      call @foo() : () -> ()
+      async.yield
+    }
+
+    // CHECK: async.await %[[a]]
+    // CHECK-NOT: gpu.wait
+    async.await %a : !async.token
+    return
+  }
+
+  // CHECK-LABEL:func @defer_wait_pass_through(%{{.*}}: index)
+  func @defer_wait_pass_through(%sz : index) {
+    // CHECK: %[[a0:.*]], %[[f0:.*]] = async.execute
+    %a0 = async.execute {
+      // CHECK: %[[t:.*]] = gpu.launch_func async
+      gpu.launch_func @kernels::@kernel
+          blocks in (%sz, %sz, %sz) threads in (%sz, %sz, %sz)
+      // CHECK-NOT: gpu.wait
+      // CHECK: async.yield %[[t]]
+      async.yield
+    }
+
+    // CHECK: %[[a1:.*]], %[[f1:.*]] = async.execute
+    // CHECK-SAME: %[[f0]]
+    %a1 = async.execute [%a0] {
+      // CHECK-NOT: gpu.wait
+      // CHECK: async.yield %{{.*}}
+      async.yield
+    }
+
+    // CHECK: async.await %[[a1]]
+    // CHECK: %[[t:.*]] = async.await %[[f1]]
+    // CHECK: gpu.wait [%[[t]]]
+    async.await %a1 : !async.token
+    return
+  }
 }