-
Notifications
You must be signed in to change notification settings - Fork 160
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
add shape propagate pass on mhlo (#1298)
add shape propagation pass on mhlo
- Loading branch information
1 parent
f9de05a
commit 037e6ba
Showing
7 changed files
with
352 additions
and
2 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
296 changes: 296 additions & 0 deletions
296
tao_compiler/mlir/disc/transforms/disc_shape_propagate.cc
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,296 @@ | ||
| /* Copyright 2022 The BladeDISC Authors. All Rights Reserved. | ||
| Licensed under the Apache License, Version 2.0 (the "License"); | ||
| you may not use this file except in compliance with the License. | ||
| You may obtain a copy of the License at | ||
| http://www.apache.org/licenses/LICENSE-2.0 | ||
| Unless required by applicable law or agreed to in writing, software | ||
| distributed under the License is distributed on an "AS IS" BASIS, | ||
| WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| See the License for the specific language governing permissions and | ||
| limitations under the License. | ||
| ==============================================================================*/ | ||
|
|
||
| // This file implements the logic to do some shape optimizations on tensor | ||
| // level. | ||
| #include <chrono> | ||
| #include <unordered_set> | ||
| #include <utility> | ||
|
|
||
| #include "absl/strings/str_split.h" | ||
| #include "llvm/ADT/DenseMap.h" | ||
| #include "llvm/ADT/StringRef.h" | ||
| #include "llvm/Support/Debug.h" | ||
| #include "mhlo/IR/hlo_ops.h" | ||
| #include "mlir/Dialect/Arith/IR/Arith.h" | ||
| #include "mlir/Dialect/Func/IR/FuncOps.h" | ||
| #include "mlir/Dialect/Shape/IR/Shape.h" | ||
| #include "mlir/Dialect/Tensor/IR/Tensor.h" // TF:llvm-project | ||
| #include "mlir/IR/Dominance.h" | ||
| #include "mlir/IR/MLIRContext.h" // TF:llvm-project | ||
| #include "mlir/IR/Matchers.h" | ||
| #include "mlir/IR/OpDefinition.h" | ||
| #include "mlir/Pass/Pass.h" // TF:local_config_mlir | ||
| #include "mlir/Pass/PassManager.h" | ||
| #include "mlir/Transforms/GreedyPatternRewriteDriver.h" | ||
| #include "mlir/Transforms/Passes.h" | ||
| #include "mlir/Transforms/Passes.h" // TF:llvm-project | ||
| #include "mlir/disc/IR/disc_shape_ops.h" | ||
| #include "mlir/disc/IR/hlo_disc_ops.h" | ||
| #include "mlir/disc/disc_util.h" | ||
| #include "mlir/disc/transforms/PassDetail.h" | ||
| #include "mlir/disc/transforms/shape_utils.h" | ||
|
|
||
| namespace mlir { | ||
| namespace disc_ral { | ||
|
|
||
| using ::mlir::func::FuncOp; | ||
|
|
||
| namespace { | ||
| std::string kDynamicDimsAttr = "input_dynamic_dims"; | ||
| struct ShapeContext { | ||
| ShapeContext() = default; | ||
| ShapeContext(Value value, SmallVector<int64_t> shape) | ||
| : value(value), shape(shape){}; | ||
|
|
||
| Value value; | ||
| SmallVector<int64_t> shape; | ||
| }; | ||
| struct DiscShapePropagatePass | ||
| : public DiscShapePropagatePassBase<DiscShapePropagatePass> { | ||
| DiscShapePropagatePass() | ||
| : DiscShapePropagatePassBase< | ||
| DiscShapePropagatePass>::DiscShapePropagatePassBase() {} | ||
| void getDependentDialects(DialectRegistry& registry) const override { | ||
| DiscShapePropagatePassBase<DiscShapePropagatePass>::getDependentDialects( | ||
| registry); | ||
| registry.insert<shape::ShapeDialect>(); | ||
| } | ||
| void runOnOperation() override; | ||
| }; | ||
| bool isBinaryOp(Operation* op) { | ||
| return isa<mhlo::AddOp>(*op) || isa<mhlo::CompareOp>(*op) || | ||
| isa<mhlo::SelectOp>(*op) || isa<mhlo::ConvertOp>(*op); | ||
| } | ||
|
|
||
| bool isUnaryOp(Operation* op) { return isa<mhlo::ConvertOp>(op); } | ||
| bool isConcreteShape(ShapeContext& ctx) { | ||
| for (auto dim : ctx.shape) { | ||
| if (dim == ShapedType::kDynamic) return false; | ||
| } | ||
| return true; | ||
| } | ||
|
|
||
| std::optional<Value> getConstTensor(OpBuilder& b, Operation* op, | ||
| ArrayRef<int> vec, | ||
| ArrayRef<int64_t> shape) { | ||
| uint64_t num_total_elements = 1; | ||
| for (int64_t a : shape) { | ||
| num_total_elements *= a; | ||
| } | ||
|
|
||
| if (vec.size() != num_total_elements) { | ||
| op->emitOpError("getConstTensor(): number of elements mismatch."); | ||
| return std::nullopt; | ||
| } | ||
| auto const_type = RankedTensorType::get(shape, b.getI64Type()); | ||
| auto const_attr = DenseElementsAttr::get(const_type, vec); | ||
| auto const_op = | ||
| b.create<mhlo::ConstantOp>(op->getLoc(), const_type, const_attr); | ||
| return const_op.getResult(); | ||
| } | ||
|
|
||
| std::optional<ShapeContext> HandleBinaryOp(OpBuilder& b, Operation* op, | ||
| ShapeContext& inputCtx) { | ||
| if (!isBinaryOp(op)) return std::nullopt; | ||
| if (op->getOperand(1).isa<BlockArgument>()) { | ||
| return ShapeContext(op->getResult(0), inputCtx.shape); | ||
| } | ||
| if (auto const_op = | ||
| dyn_cast<mhlo::ConstantOp>(op->getOperand(1).getDefiningOp())) { | ||
| auto elemTy = | ||
| op->getOperand(0).getType().cast<RankedTensorType>().getElementType(); | ||
| b.setInsertionPoint(op); | ||
| auto dense_attr = const_op.getValue().dyn_cast<mlir::DenseElementsAttr>(); | ||
| int64_t value = (*dense_attr.getValues<APInt>().begin()).getSExtValue(); | ||
| auto scalar_const_op = getConstTensor(b, op, {value}, {}); | ||
| Value inputShape = | ||
| b.create<shape::ShapeOfOp>(op->getLoc(), op->getOperand(0)); | ||
| auto rank = inputCtx.shape.size(); | ||
| SmallVector<int64_t> boradcast_dim; | ||
| boradcast_dim.push_back(static_cast<int64_t>(rank)); | ||
|
|
||
| auto bcast_op = b.create<mhlo::DynamicBroadcastInDimOp>( | ||
| op->getLoc(), RankedTensorType::get(inputCtx.shape, elemTy), | ||
| scalar_const_op.value(), inputShape, b.getI64TensorAttr({})); | ||
| const_op.getResult().replaceAllUsesWith(bcast_op.getResult()); | ||
| const_op.erase(); | ||
| } | ||
| return ShapeContext(op->getResult(0), inputCtx.shape); | ||
| } | ||
|
|
||
| template <typename OpTy> | ||
| std::optional<ShapeContext> propagateHelper(OpBuilder& b, Operation* op, | ||
| ShapeContext& inputCtx) { | ||
| return std::nullopt; | ||
| } | ||
| template <> | ||
| std::optional<ShapeContext> propagateHelper<mhlo::DotOp>( | ||
| OpBuilder& b, Operation* op, ShapeContext& inputCtx) { | ||
| auto dot_op = cast<mhlo::DotOp>(op); | ||
| auto lhs_shape = | ||
| dot_op.getOperand(0).getType().cast<RankedTensorType>().getShape(); | ||
| auto rhs_shape = | ||
| dot_op.getOperand(1).getType().cast<RankedTensorType>().getShape(); | ||
| auto result_shape = | ||
| dot_op.getResult().getType().cast<RankedTensorType>().getShape(); | ||
| SmallVector<int64_t> new_shape; | ||
| new_shape.push_back(lhs_shape[0]); | ||
| new_shape.push_back(rhs_shape[1]); | ||
| return ShapeContext(op->getResult(0), new_shape); | ||
| } | ||
|
|
||
| LogicalResult parseInputDynamicDims( | ||
| func::FuncOp main, | ||
| std::vector<std::pair<int, std::vector<int>>>& input_dynamic_dims) { | ||
| auto dict_attr = main->getAttrOfType<DictionaryAttr>("tf.entry_function"); | ||
| if (!dict_attr) { | ||
| return failure(); | ||
| } | ||
| if (!dict_attr.get(kDynamicDimsAttr)) { | ||
| return failure(); | ||
| } | ||
| StringRef param_str = | ||
| dict_attr.get(kDynamicDimsAttr).dyn_cast<mlir::StringAttr>(); | ||
|
|
||
| SmallVector<StringRef, 4> parsed_dynamic_dims; | ||
| param_str.split(parsed_dynamic_dims, "|"); | ||
| for (auto kv : parsed_dynamic_dims) { | ||
| SmallVector<StringRef, 4> pair; | ||
| kv.split(pair, ":"); | ||
| if (pair.size() != 2) { | ||
| return failure(); | ||
| } | ||
| int arg_index = std::stoi(pair[0].str()); | ||
| SmallVector<StringRef, 4> dims; | ||
| pair[1].split(dims, ","); | ||
| std::vector<int> dim_vec; | ||
| for (auto dim : dims) { | ||
| dim_vec.push_back(std::stoi(dim.str())); | ||
| } | ||
| input_dynamic_dims.push_back({arg_index, dim_vec}); | ||
| } | ||
| return success(); | ||
| } | ||
|
|
||
| void applyShapeContext(ShapeContext& ctx) { | ||
| auto res_ty = ctx.value.getType().dyn_cast<RankedTensorType>(); | ||
| if (!res_ty) return; | ||
| auto elemTy = res_ty.getElementType(); | ||
| ctx.value.setType(RankedTensorType::get(ctx.shape, elemTy)); | ||
| } | ||
|
|
||
| std::optional<ShapeContext> propagateOpShape(OpBuilder& rewriter, Operation* op, | ||
| ShapeContext& inputCtx) { | ||
| if (isUnaryOp(op)) { | ||
| return ShapeContext(op->getResult(0), inputCtx.shape); | ||
| } | ||
| if (auto ctx = HandleBinaryOp(rewriter, op, inputCtx)) { | ||
| return ctx; | ||
| } | ||
| using PropagationFunc = | ||
| std::optional<ShapeContext> (*)(OpBuilder&, Operation*, ShapeContext&); | ||
| const std::vector<PropagationFunc> propagationFunctions = { | ||
| propagateHelper<mhlo::DotOp>, | ||
| }; | ||
| // Iterate over the propagation functions and apply each one | ||
| for (const auto& propagate : propagationFunctions) { | ||
| if (auto ctx = propagate(rewriter, op, inputCtx)) { | ||
| return ctx; | ||
| } | ||
| } | ||
| return std::nullopt; | ||
| } | ||
|
|
||
| void visitOperator(ModuleOp& m, OpBuilder& rewriter, Operation* op, | ||
| ShapeContext& ctx) { | ||
| if (isConcreteShape(ctx)) return; | ||
| // later to process return operators | ||
| if (isa<func::ReturnOp>(op)) return; | ||
|
|
||
| auto resultShapeCtx = propagateOpShape(rewriter, op, ctx); | ||
| if (!resultShapeCtx) { | ||
| m.emitError("failed update shape context on op:" + | ||
| op->getName().stripDialect().str()); | ||
| return; | ||
| } | ||
| for (auto user : op->getResult(0).getUsers()) { | ||
| visitOperator(m, rewriter, user, resultShapeCtx.value()); | ||
| } | ||
| applyShapeContext(*resultShapeCtx); | ||
| } | ||
|
|
||
| void DiscShapePropagatePass::runOnOperation() { | ||
| ModuleOp m = getOperation(); | ||
| auto main = m.lookupSymbol<FuncOp>("main"); | ||
| MLIRContext* context = &getContext(); | ||
| mlir::OpBuilder rewriter(context); | ||
| OpBuilder b(main); | ||
| if (!main) { | ||
| m.emitError("entry func: main not found"); | ||
| signalPassFailure(); | ||
| return; | ||
| } | ||
| SmallVector<Type, 4> new_arg_types, new_return_types; | ||
| for (auto arg : main.getArguments()) { | ||
| new_arg_types.push_back(arg.getType()); | ||
| } | ||
| // stage1: parse attribute input_dynamic_dims to a map | ||
| std::vector<std::pair<int, std::vector<int>>> input_dynamic_dims; | ||
| if (failed(parseInputDynamicDims(main, input_dynamic_dims))) { | ||
| return; | ||
| } | ||
| // skip this pass if no dynamic dims attribute | ||
| if (input_dynamic_dims.size() == 0) return; | ||
| // stage2: visit all operators to propagate shape | ||
| for (auto pair : input_dynamic_dims) { | ||
| int argIdx = pair.first; | ||
| Value value = main.getArgument(argIdx); | ||
| auto ty = value.getType().cast<RankedTensorType>(); | ||
| SmallVector<int64_t> newShape; | ||
| std::copy(ty.getShape().begin(), ty.getShape().end(), | ||
| std::back_inserter(newShape)); | ||
| for (auto dim : pair.second) { | ||
| newShape[dim] = ShapedType::kDynamic; | ||
| } | ||
| ShapeContext ctx(value, newShape); | ||
| auto newType = RankedTensorType::get(newShape, ty.getElementType()); | ||
| for (auto user : main.getArgument(argIdx).getUsers()) { | ||
| visitOperator(m, rewriter, user, ctx); | ||
| } | ||
| new_arg_types[argIdx] = newType; | ||
| applyShapeContext(ctx); | ||
| } | ||
|
|
||
| // stage3: visit all return operators to update function signature | ||
| main.walk([&](Operation* op) { | ||
| if (isa<func::ReturnOp>(*op)) { | ||
| for (auto operand : op->getOperands()) { | ||
| new_return_types.push_back(operand.getType()); | ||
| } | ||
| } | ||
| }); | ||
| main.setType( | ||
| FunctionType::get(main.getContext(), new_arg_types, new_return_types)); | ||
| } | ||
|
|
||
| } // namespace | ||
|
|
||
| std::unique_ptr<OperationPass<ModuleOp>> createDiscShapePropagatePass() { | ||
| return std::make_unique<DiscShapePropagatePass>(); | ||
| } | ||
|
|
||
| } // namespace disc_ral | ||
| } // namespace mlir |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.