Implement a cost model to drive the lowering of scf.parallel.

[bakhtiyarneyman]

Summary

Depends On D110680

[mehdi_amini]

This cost model component is pretty interesting! It looks like something that really can be a generic pass that would drive a traversal and query opinterfaces?

mlir/lib/Dialect/Async/Transforms/CostModel.cpp

2

(license header)

90–92

96

Seems like all these kind of "estimateCostXXX" are the kind of things that should be behind an op interface instead, shouldn't it?

mlir/lib/Dialect/Async/Transforms/CostModel.h

2

You're missing the license header, and the guard does not match the usual format I believe.

7

Please avoid using namespace directive in a header.

15

Can you make it a class and increase the doc?
Right now the description says " Approximate cost of executing an op on a modern CPU" but the first member is a Builder which is confusing to me.

Edit: reading the implementation, I get a better feel for what is going on, but the comment stands :)

[nicolasvasilache]

This seems important and cross-cutting enough to warrant an RFC on discourse, I am surprised @mehdi_amini has not asked for this already :)
Putting a blocker until a discussion has happened.

[bakhtiyarneyman]

Address comments.

[bakhtiyarneyman]

mlir/lib/Dialect/Async/Transforms/CostModel.cpp

96

Yes, the code is written with that potential development in mind. But it is not clear that something as intrusive is necessary now. The concept needs to be validated first. The biggest question is whether the estimation precision of what we have now is sufficient. If it turns to be insufficient, we might need to do better analysis. For instance, it's conceivable that the whole approach would need to be scrapped, it might be imperative to do the memory component of the analysis earlier and/or incrementally, before the tiling and fusion have been lowered to ops which make it difficult to guess the cache performance. Furthermore, it might be beneficial to make a first-class cost-model which would help drive optimizations in other parts of the pipeline, most importantly the codegen.

Furthermore, I suspect that once we are trying to use op interfaces, the cost model needs to be reimagined as this first-class entity, which might mean that we would need to design it in a way, that is useful for other architectures, e.g. GPUs. That in itself will take a quarter. So my preference is to do small and local to the AsyncParallelFor pass.

mlir/lib/Dialect/Async/Transforms/CostModel.h

2

It seems to be matching the PassDetail.h from the same directory, no?

[ezhulenev]

It would be great to have the cost model to be a part of core MLIR, defined with op interface, but for it to be usable for all the potential targets it will require a lot of discussions (e.g. how cost model can capture GPU memory coalescing, should it be the concern of the cost model at all, or it should just track compute?). Let's start an RFC to figure out how it should look like. In the meantime we can hide the current crude implementation behind the AsyncBlockSizeComputationFunction API, and keep implementation inside the Tensorflow compiler. We can tune it for the compilation pipeline we have right now, without need to worry about the general case.

mlir/include/mlir/Dialect/Async/Passes.h

25

How about giving more control to the caller (similar to the TileSizeComputationFunction in linalg).

Something like this:

struct AsyncBlockSize {
  Value targetBlockSize; // must be a Value of index type
   ...                   // some other values?
}

using AsyncBlockSizeComputationFunction =
    std::function<AsyncBlockSize(OpBuilder &, scf::ParallelOp parallelLoop)>;

createAsyncParallelForPass(bool asyncDispatch, int32_t numWorkerThreads, 
                                                AsyncBlockSizeComputationFunction cost);

Constant int32_t targetBlockSize will be just a trivial cost compute function. Also Async dialect can provide the default cost function based on the ops properties.

mlir/include/mlir/Dialect/Async/Passes.td

34

and this will become a property of the caller pipeline

mlir/include/mlir/Dialect/Async/Passes.h

 #define MLIR_DIALECT_ASYNC_PASSES_H_
 
 #include "mlir/Pass/Pass.h"
 
 namespace mlir {
 
 std::unique_ptr<Pass> createAsyncParallelForPass();
 
-std::unique_ptr<Pass> createAsyncParallelForPass(bool asyncDispatch,
-                                                 int32_t numWorkerThreads,
-                                                 int32_t targetBlockSize);
+std::unique_ptr<Pass>
+createAsyncParallelForPass(bool asyncDispatch, int32_t numWorkerThreads,
+                           bool useCostModel, int32_t minTaskSize = 512 * 1024);
 
 std::unique_ptr<OperationPass<ModuleOp>> createAsyncToAsyncRuntimePass();
 
 std::unique_ptr<Pass> createAsyncRuntimeRefCountingPass();
 
 std::unique_ptr<Pass> createAsyncRuntimeRefCountingOptPass();
 
 std::unique_ptr<Pass> createAsyncRuntimePolicyBasedRefCountingPass();
 

mlir/include/mlir/Dialect/Async/Passes.td

Option<"asyncDispatch", "async-dispatch",
       "caller thread.">,
 
     Option<"numWorkerThreads", "num-workers",
       "int32_t", /*default=*/"8",
       "The number of available workers to execute async operations.">,
 
     Option<"minTaskSize", "min-task-size",
       "int32_t", /*default=*/"1000",
-      "The minimum task size for sharding parallel operation.">
+      "The minimum task size for sharding parallel operation.">,
+
+    Option<"useCostModel", "use-cost-model",
+      "bool", /*default=*/"false",
+      "Use experimental cost model to determine the parallelism granularity.">
+
   ];
 
   let dependentDialects = [
     "arith::ArithmeticDialect",
     "async::AsyncDialect",
     "scf::SCFDialect"
   ];
 }

mlir/lib/Dialect/Async/Transforms/AsyncParallelFor.cpp

 //===- AsyncParallelFor.cpp - Implementation of Async Parallel For --------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements scf.parallel to scf.for + async.execute conversion pass.
 //
 //===----------------------------------------------------------------------===//
 
+#include "CostModel.h"
 #include "PassDetail.h"
 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
 #include "mlir/Dialect/Async/IR/Async.h"
 #include "mlir/Dialect/Async/Passes.h"
 #include "mlir/Dialect/SCF/SCF.h"
 #include "mlir/Dialect/StandardOps/IR/Ops.h"
 #include "mlir/IR/BlockAndValueMapping.h"
 #include "mlir/IR/ImplicitLocOpBuilder.h"

 //      call @parallel_compute_fn(%block_index, %block_size, ...)
 //   }
 //
 struct AsyncParallelForPass
     : public AsyncParallelForBase<AsyncParallelForPass> {
   AsyncParallelForPass() = default;
 
   AsyncParallelForPass(bool asyncDispatch, int32_t numWorkerThreads,
-                       int32_t minTaskSize) {
+                       bool useCostModel, int32_t minTaskSize) {
     this->asyncDispatch = asyncDispatch;
     this->numWorkerThreads = numWorkerThreads;
+    this->useCostModel = useCostModel;
     this->minTaskSize = minTaskSize;
   }
 
   void runOnOperation() override;
 };
 
 struct AsyncParallelForRewrite : public OpRewritePattern<scf::ParallelOp> {
 public:
   AsyncParallelForRewrite(MLIRContext *ctx, bool asyncDispatch,
                           int32_t numWorkerThreads, int32_t minTaskSize)
       : OpRewritePattern(ctx), asyncDispatch(asyncDispatch),
         numWorkerThreads(numWorkerThreads), minTaskSize(minTaskSize) {}
 
   LogicalResult matchAndRewrite(scf::ParallelOp op,
                                 PatternRewriter &rewriter) const override;
 
 private:
   bool asyncDispatch;
   int32_t numWorkerThreads;
+  bool useCostModel;
   int32_t minTaskSize;
 };
 
 struct ParallelComputeFunctionType {
   FunctionType type;
   llvm::SmallVector<Value> captures;
 };
 
auto noOp = [&](OpBuilder &nestedBuilder, Location loc) {
     nestedBuilder.create<scf::YieldOp>(loc);
   };
 
   // Compute the parallel block size and dispatch concurrent tasks computing
   // results for each block.
   auto dispatch = [&](OpBuilder &nestedBuilder, Location loc) {
     ImplicitLocOpBuilder nb(loc, nestedBuilder);
 
+    // Create a parallel compute function that takes a block id and computes the
+    // parallel operation body for a subset of iteration space.
+    ParallelComputeFunction parallelComputeFunction =
+        createParallelComputeFunction(op, rewriter);
+
     // With large number of threads the value of creating many compute blocks
     // is reduced because the problem typically becomes memory bound. For small
     // number of threads it helps with stragglers.
     float overshardingFactor = numWorkerThreads <= 4    ? 8.0
                                : numWorkerThreads <= 8  ? 4.0
                                : numWorkerThreads <= 16 ? 2.0
                                : numWorkerThreads <= 32 ? 1.0
                                : numWorkerThreads <= 64 ? 0.8
                                                         : 0.6;
 
     // Do not overload worker threads with too many compute blocks.
     Value maxComputeBlocks = b.create<arith::ConstantIndexOp>(
         std::max(1, static_cast<int>(numWorkerThreads * overshardingFactor)));
 
     // Target block size from the pass parameters.
     Value minTaskSizeCst = b.create<arith::ConstantIndexOp>(minTaskSize);
 
     // Compute parallel block size from the parallel problem size:
     //   blockSize = min(tripCount,
     //                   max(ceil_div(tripCount, maxComputeBlocks),
     //                       ceil_div(minTaskSize, bodySize)))
+    Value bodySize;
+    if (useCostModel) {
+      CostModel helper(b, *op);
+      Cost cost = helper.estimateCost(op.getLoopBody());
+      bodySize = helper.costToNanoseconds(cost);
+    } else {
+      bodySize = b.create<arith::ConstantIndexOp>(32);
+    }
     Value bs0 = b.create<arith::CeilDivSIOp>(tripCount, maxComputeBlocks);
-    Value bs1 = b.create<arith::MaxSIOp>(bs0, minTaskSizeCst);
-    Value blockSize = b.create<arith::MinSIOp>(tripCount, bs1);
+    Value bs1 = b.create<arith::CeilDivSIOp>(minTaskSizeCst, bodySize);
+    Value bs2 = b.create<arith::MaxSIOp>(bs0, bs1);
+    Value blockSize = b.create<arith::MinSIOp>(tripCount, bs2);
     Value blockCount = b.create<arith::CeilDivSIOp>(tripCount, blockSize);
 
-    // Create a parallel compute function that takes a block id and computes the
-    // parallel operation body for a subset of iteration space.
-    ParallelComputeFunction parallelComputeFunction =
-        createParallelComputeFunction(op, rewriter);
-
     // Dispatch parallel compute function using async recursive work splitting,
     // or by submitting compute task sequentially from a caller thread.
     if (asyncDispatch) {
       doAsyncDispatch(b, rewriter, parallelComputeFunction, op, blockSize,
                       blockCount, tripCounts);
     } else {
       doSequentialDispatch(b, rewriter, parallelComputeFunction, op, blockSize,
                            blockCount, tripCounts);

 }
 
 std::unique_ptr<Pass> mlir::createAsyncParallelForPass() {
   return std::make_unique<AsyncParallelForPass>();
 }
 
 std::unique_ptr<Pass> mlir::createAsyncParallelForPass(bool asyncDispatch,
                                                        int32_t numWorkerThreads,
+                                                       bool useCostModel,
                                                        int32_t minTaskSize) {
   return std::make_unique<AsyncParallelForPass>(asyncDispatch, numWorkerThreads,
-                                                minTaskSize);
+                                                useCostModel, minTaskSize);
 }

mlir/lib/Dialect/Async/Transforms/CMakeLists.txt

 add_mlir_dialect_library(MLIRAsyncTransforms
   AsyncParallelFor.cpp
   AsyncRuntimeRefCounting.cpp
   AsyncRuntimeRefCountingOpt.cpp
   AsyncToAsyncRuntime.cpp
+  CostModel.cpp
   PassDetail.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/Async
 
   DEPENDS
   MLIRAsyncPassIncGen
 
   LINK_LIBS PUBLIC
   MLIRArithmetic
   MLIRAsync
   MLIRIR
+  MLIRMath
+  MLIRMemref
   MLIRPass
   MLIRSCF
   MLIRSCFToStandard
   MLIRStandard
   MLIRTransforms
   MLIRTransformUtils
+  MLIRVector
 )

mlir/lib/Dialect/Async/Transforms/CostModel.h

+//===- CostModel.h - Declaration of the cost model.           --------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===////
+//
+// This file declares a cost model that drives deciding the parallelization
+// granularity when lowering parallel operations into async tasks.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef DIALECT_ASYNC_TRANSFORMS_COSTMODEL_H_
+#define DIALECT_ASYNC_TRANSFORMS_COSTMODEL_H_
+
+#include "mlir/IR/ImplicitLocOpBuilder.h"
+
+namespace mlir {
+
+// Cost measuring unit.
+//
+// For CPU roughly corresponds to cycles. For RAM roughly corresponds to bytes.
+using InverseThroughput = Value;
+
+// Approximate cost of executing an op on a modern CPU.
+//
+// Encapsulates the leaf nodes of the IR that is being emitted as the cost
+// estimation progresses. The leaf nodes correspond to CPU and RAM
+// inverse-throughputs.
+class Cost {
+private:
+  InverseThroughput ram;
+  InverseThroughput cpu;
+  ImplicitLocOpBuilder builder;
+
+public:
+  Cost(ImplicitLocOpBuilder &builder, size_t ramCost = 0, size_t cpuCost = 0);
+
+  Cost(ImplicitLocOpBuilder &builder, InverseThroughput &ramCost,
+       InverseThroughput &cpuCost);
+
+  InverseThroughput &getRAM();
+  InverseThroughput &getCPU();
+
+  Cost &operator*=(const size_t &rhs);
+  Cost operator*(const size_t &rhs);
+  Cost &operator+=(const Cost &rhs);
+  Cost operator+(const Cost &rhs);
+};
+
+// Estimates execution time for an op on a modern CPU.
+//
+// Errs on the lower end, but is not strictly a lower bound estimate. Targeting
+// being within an order of magnitude of the correct value.
+class CostModel {
+private:
+  ImplicitLocOpBuilder builder;
+  Operation &rootOp;
+
+public:
+  CostModel(ImplicitLocOpBuilder &builder, Operation &rootOp)
+      : builder(builder), rootOp(rootOp) {}
+
+  ImplicitLocOpBuilder &getBuilder();
+  bool isBeforeRoot(Operation &op);
+  bool isBeforeRoot(Value &value);
+  Optional<Value> getIterations(Value lowerBound, Value upperBound, Value step);
+  Cost newCost(size_t ramCost, size_t cpuCost);
+  Cost newCost(Value ramCost, Value cpuCost);
+  Cost zeroCost();
+  Cost estimateCost(Operation &op);
+  Cost estimateCost(Region &region);
+  Value costToNanoseconds(Cost cost);
+};
+
+} // namespace mlir
+#endif // DIALECT_ASYNC_TRANSFORMS_COSTMODEL_H_

mlir/lib/Dialect/Async/Transforms/CostModel.cpp

+//===- CostModel.cpp - Implementation of the cost model.           --------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements a cost model that drives deciding the parallelization
+// granularity when lowering parallel operations into async tasks.
+//
+//===----------------------------------------------------------------------===//
+
+#include "CostModel.h"
+#include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
+#include "mlir/Dialect/Math/IR/Math.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/Dialect/SCF/SCF.h"
+#include "mlir/Dialect/Vector/VectorOps.h"
+
+namespace mlir {
+
+Cost::Cost(ImplicitLocOpBuilder &builder, size_t ramCost, size_t cpuCost)
+    : builder(builder), ram(builder.create<arith::ConstantIndexOp>(ramCost)),
+      cpu(builder.create<arith::ConstantIndexOp>(cpuCost)) {}
+
+Cost::Cost(ImplicitLocOpBuilder &builder, InverseThroughput &ramCost,
+           InverseThroughput &cpuCost)
+    : builder(builder), ram(ramCost), cpu(cpuCost) {}
+
+InverseThroughput &Cost::getRAM() { return ram; }
+InverseThroughput &Cost::getCPU() { return cpu; }
+
+Cost &Cost::operator*=(const size_t &rhs) {
+  auto factor = builder.create<arith::ConstantIndexOp>(rhs);
+  auto scale = [&](InverseThroughput resource) {
+    return builder.create<arith::MulIOp>(resource, factor);
+  };
+  ram = scale(ram);
+  cpu = scale(cpu);
+  return *this;
+}
+
+Cost Cost::operator*(const size_t &rhs) {
+  Cost ret = *this;
+  ret *= rhs;
+  return ret;
+}
+
+Cost &Cost::operator+=(const Cost &rhs) {
+  ram = builder.create<arith::AddIOp>(ram, rhs.ram);
+  cpu = builder.create<arith::AddIOp>(cpu, rhs.cpu);
+  return *this; // return the result by reference
+}
+
+Cost Cost::operator+(const Cost &rhs) {
+  Cost ret = *this;
+  ret += rhs;
+  return ret;
+}
+
+Cost estimateCostVector(CostModel &helper, Operation &op) {
+
+  auto newCost = [&](Value value) {
+    if (auto type = value.getType().dyn_cast<ShapedType>()) {
+      if (type.hasStaticShape()) {
+        return helper.newCost(/* ram */ type.getNumElements() *
+                                  type.getElementTypeBitWidth() / 8,
+                              /* cpu */ 0);
+      }
+    }
+    return helper.zeroCost();
+  };
+
+  if (auto loadOp = dyn_cast<vector::LoadOp>(op)) {
+    return newCost(loadOp.getResult());
+  }
+
+  if (auto storeOp = dyn_cast<vector::StoreOp>(op)) {
+    return newCost(storeOp.getOperand(0));
+  }
+
+  return helper.zeroCost();
+}
+
+Cost estimateCostMemref(CostModel &helper, Operation &op) {
+
+  auto newCost = [&](Value value) {
+    return helper.newCost(/* ram */ value.getType().getIntOrFloatBitWidth() / 8,
+                          /* cpu */ 0);
+  };
+
+  if (auto loadOp = dyn_cast<memref::LoadOp>(op)) {
+    return newCost(loadOp.getResult());
+  }
+
+  if (auto storeOp = dyn_cast<memref::StoreOp>(op)) {
+    return newCost(storeOp.getOperand(0));
+  }
+
+  return helper.zeroCost();
+}
+
+Cost estimateCostArith(CostModel &helper, Operation &op) {
+
+  if (isa<arith::DivUIOp, arith::DivSIOp, arith::CeilDivSIOp,
+          arith::FloorDivSIOp, arith::RemSIOp, arith::RemUIOp>(op)) {
+    return helper.newCost(/* ram */ 0, /* cpu */ 3);
+  }
+
+  return helper.newCost(/* ram */ 0, /* cpu */ 1);
+}
+
+Cost estimateCostMath(CostModel &helper, Operation &op) {
+
+  if (isa<math::AtanOp, math::Atan2Op, math::CosOp, math::SinOp, math::ExpOp,
+          math::Exp2Op, math::ExpM1Op, math::LogOp, math::Log10Op,
+          math::Log1pOp, math::Log2Op, math::PowFOp, math::RsqrtOp,
+          math::SqrtOp, math::TanhOp>(op)) {
+    return helper.newCost(/* ram */ 0, /* cpu */ 100);
+  }
+
+  return helper.newCost(/* ram */ 0, /* cpu */ 1);
+}
+
+Cost estimateCostSCF(CostModel &helper, Operation &op) {
+
+  auto scaleCost = [&](Value iterations, Cost cost) {
+    return helper.newCost(
+        helper.getBuilder().create<arith::MulIOp>(cost.getRAM(), iterations),
+        helper.getBuilder().create<arith::MulIOp>(cost.getCPU(), iterations));
+  };
+
+  if (auto forOp = dyn_cast<scf::ForOp>(op)) {
+    Cost cost = helper.estimateCost(forOp.getLoopBody());
+    if (auto iterations = helper.getIterations(
+            forOp.lowerBound(), forOp.upperBound(), forOp.step())) {
+      return scaleCost(*iterations, cost);
+    }
+    return cost;
+  }
+
+  if (auto parallelOp = dyn_cast<scf::ParallelOp>(op)) {
+    Cost cost = helper.estimateCost(parallelOp.getLoopBody());
+    for (auto &inductionVariableDomain : llvm::enumerate(
+             llvm::zip(parallelOp.lowerBound(), parallelOp.upperBound(),
+                       parallelOp.step()))) {
+      Value lb, ub, step;
+      std::tie(lb, ub, step) = inductionVariableDomain.value();
+
+      if (auto iterations = helper.getIterations(lb, ub, step)) {
+        cost = scaleCost(*iterations, cost);
+      }
+    }
+    return cost;
+  }
+
+  return helper.zeroCost();
+}
+
+ImplicitLocOpBuilder &CostModel::getBuilder() { return builder; }
+
+bool CostModel::isBeforeRoot(Operation &op) {
+  return op.getBlock() == rootOp.getBlock() && op.isBeforeInBlock(&rootOp);
+}
+
+bool CostModel::isBeforeRoot(Value &value) {
+  return isBeforeRoot(*value.getDefiningOp());
+}
+
+Optional<Value> CostModel::getIterations(Value lowerBound, Value upperBound,
+                                         Value step) {
+  if (isBeforeRoot(lowerBound) && isBeforeRoot(upperBound) &&
+      isBeforeRoot(step)) {
+    return (Value)builder.create<arith::CeilDivSIOp>(
+        builder.create<arith::SubIOp>(upperBound, lowerBound), step);
+  }
+  return llvm::None;
+}
+
+Cost CostModel::newCost(size_t ramCost, size_t cpuCost) {
+  return Cost(builder, ramCost, cpuCost);
+}
+
+Cost CostModel::newCost(Value ramCost, Value cpuCost) {
+  return Cost(builder, ramCost, cpuCost);
+}
+
+Cost CostModel::zeroCost() { return Cost(builder, 0, 0); }
+
+Cost CostModel::estimateCost(Operation &op) {
+  MLIRContext *ctx = op.getContext();
+  if (op.getDialect() == ctx->getLoadedDialect("scf")) {
+    return estimateCostSCF(*this, op);
+  }
+  if (op.getDialect() == ctx->getLoadedDialect("memref")) {
+    return estimateCostMemref(*this, op);
+  }
+  if (op.getDialect() == ctx->getLoadedDialect("vector")) {
+    return estimateCostVector(*this, op);
+  }
+  if (op.getDialect() == ctx->getLoadedDialect("arith")) {
+    return estimateCostArith(*this, op);
+  }
+  if (op.getDialect() == ctx->getLoadedDialect("math")) {
+    return estimateCostMath(*this, op);
+  }
+  return zeroCost();
+}
+
+Cost CostModel::estimateCost(Region &region) {
+  Cost cost = zeroCost();
+  for (auto &op : region.getOps()) {
+    cost += estimateCost(op);
+  }
+  return cost;
+}
+
+Value CostModel::costToNanoseconds(Cost cost) {
+  // Assume that RAM throughput is 16 GB/s and that CPU runs at 3 GHz.
+  auto ramRuntime = builder.create<arith::DivUIOp>(
+      cost.getRAM(), builder.create<arith::ConstantIndexOp>(16));
+  auto cpuRuntime = builder.create<arith::DivUIOp>(
+      cost.getCPU(), builder.create<arith::ConstantIndexOp>(3));
+  auto max = [&](Value x, Value y) {
+    return builder.create<arith::MaxSIOp>(x, y);
+  };
+  return max(max(ramRuntime, cpuRuntime),
+             builder.create<arith::ConstantIndexOp>(1));
+}
+
+} // namespace mlir

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

cc_library(
     hdrs = ["include/mlir/Dialect/Async/Passes.h"],
     includes = ["include"],
     deps = [
         ":Analysis",
         ":ArithmeticDialect",
         ":Async",
         ":AsyncPassIncGen",
         ":IR",
+        ":MathDialect",
+        ":MemRefDialect",
         ":Pass",
         ":SCFDialect",
         ":SCFToStandard",
         ":StandardOps",
         ":Support",
         ":TransformUtils",
         ":Transforms",
         ":TransformsPassIncGen",
+        ":VectorOps",
         "//llvm:Core",
         "//llvm:Support",
     ],
 )
 
 cc_library(
     name = "AffineUtils",
     srcs = glob(