• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 // This file implements logic for lowering MHLO general dot to a regular dot.
17 
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/StringSwitch.h"
20 #include "mlir-hlo/Dialect/mhlo/IR/hlo_ops.h"
21 #include "mlir-hlo/Dialect/mhlo/transforms/PassDetail.h"
22 #include "mlir-hlo/Dialect/mhlo/transforms/passes.h"
23 #include "mlir-hlo/Dialect/mhlo/transforms/rewriters.h"
24 #include "mlir/Dialect/StandardOps/IR/Ops.h"
25 #include "mlir/IR/Attributes.h"
26 #include "mlir/IR/BuiltinOps.h"
27 #include "mlir/IR/BuiltinTypes.h"
28 #include "mlir/IR/Location.h"
29 #include "mlir/IR/Operation.h"
30 #include "mlir/IR/TypeUtilities.h"
31 #include "mlir/Pass/Pass.h"
32 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
33 
34 namespace mlir {
35 namespace mhlo {
36 namespace {
37 
TransposeReshape(Value arg,Location loc,llvm::ArrayRef<int64_t> left_dims,llvm::ArrayRef<int64_t> right_dims,llvm::ArrayRef<int64_t> arg_shape,PatternRewriter * rewriter)38 Value TransposeReshape(Value arg, Location loc,
39                        llvm::ArrayRef<int64_t> left_dims,
40                        llvm::ArrayRef<int64_t> right_dims,
41                        llvm::ArrayRef<int64_t> arg_shape,
42                        PatternRewriter *rewriter) {
43   auto element_type = getElementTypeOrSelf(arg.getType());
44 
45   int64_t left_size = 1;
46   for (auto dim : left_dims) {
47     left_size *= arg_shape[dim];
48   }
49 
50   int64_t right_size = 1;
51   for (auto dim : right_dims) {
52     right_size *= arg_shape[dim];
53   }
54 
55   // Generate the transpose permutation attribute.
56   llvm::SmallVector<int64_t, 5> transpose_permutation(left_dims.begin(),
57                                                       left_dims.end());
58   transpose_permutation.append(right_dims.begin(), right_dims.end());
59 
60   TensorType transpose_permutation_type = RankedTensorType::get(
61       {static_cast<int64_t>(transpose_permutation.size())},
62       rewriter->getIntegerType(64));
63 
64   auto transpose_permutation_attr =
65       DenseIntElementsAttr::get(transpose_permutation_type,
66                                 llvm::makeArrayRef(transpose_permutation))
67           .cast<DenseIntElementsAttr>();
68 
69   // Compute the resulting shape.
70   llvm::SmallVector<int64_t, 5> transposed_shape;
71   for (auto val : transpose_permutation) {
72     transposed_shape.push_back(arg_shape[val]);
73   }
74   auto transpose_type = RankedTensorType::get(transposed_shape, element_type);
75   auto transpose_result = rewriter->create<TransposeOp>(
76       loc, transpose_type, arg, transpose_permutation_attr);
77 
78   // Return the final result.
79   auto reshaped_type =
80       RankedTensorType::get({left_size, right_size}, element_type);
81   return rewriter->create<ReshapeOp>(loc, reshaped_type, transpose_result);
82 }
83 
ProcessDotArg(Value arg,Location loc,ElementsAttr contract_dims_attr,bool outer_dims_first,PatternRewriter * rewriter)84 Value ProcessDotArg(Value arg, Location loc, ElementsAttr contract_dims_attr,
85                     bool outer_dims_first, PatternRewriter *rewriter) {
86   auto shape = arg.getType().cast<ShapedType>().getShape();
87 
88   llvm::SmallVector<bool, 5> is_outer_dim;
89   is_outer_dim.resize(shape.size(), true);
90 
91   // Compute the contract dimension ordering.
92   llvm::SmallVector<int64_t, 5> contract_dims;
93   for (auto dim : contract_dims_attr.getValues<int64_t>()) {
94     contract_dims.push_back(dim);
95     is_outer_dim[dim] = false;
96   }
97 
98   // Compute the outer dimension orderings.
99   llvm::SmallVector<int64_t, 5> outer_dims;
100   for (auto it : llvm::enumerate(is_outer_dim)) {
101     if (it.value()) {
102       outer_dims.push_back(it.index());
103     }
104   }
105 
106   if (outer_dims_first) {
107     return TransposeReshape(arg, loc, outer_dims, contract_dims, shape,
108                             rewriter);
109   }
110 
111   return TransposeReshape(arg, loc, contract_dims, outer_dims, shape, rewriter);
112 }
113 
114 struct GeneralDotConvert : public OpRewritePattern<DotGeneralOp> {
115   // Attempts to lower a General Dot operator to a standard Dot operator.
116   // General dots include batching dimensions and can have collapsing
117   // dimensions along any axis. Inserting correctly arrange transpose and
118   // reshape operators organizes the tensors and allows the General Dot to be
119   // replaced with the standard Dot operator.
120   //
121   // Note: This requires an empty list of batch dimensions.
122 
GeneralDotConvertmlir::mhlo::__anonf50d5d480111::GeneralDotConvert123   explicit GeneralDotConvert(MLIRContext *context)
124       : OpRewritePattern(context) {}
125 
matchAndRewritemlir::mhlo::__anonf50d5d480111::GeneralDotConvert126   LogicalResult matchAndRewrite(DotGeneralOp op,
127                                 PatternRewriter &rewriter) const override {
128     auto dot_element_type = getElementTypeOrSelf(op);
129 
130     auto dot_numbers = op.dot_dimension_numbers();
131     if (dot_numbers.lhs_batching_dimensions().getNumElements() != 0 ||
132         dot_numbers.rhs_batching_dimensions().getNumElements() != 0) {
133       return failure();
134     }
135 
136     auto lhs = ProcessDotArg(op.lhs(), op.getLoc(),
137                              dot_numbers.lhs_contracting_dimensions(),
138                              /*outer_dims_first=*/true, &rewriter);
139 
140     auto rhs = ProcessDotArg(op.rhs(), op.getLoc(),
141                              dot_numbers.rhs_contracting_dimensions(),
142                              /*outer_dims_first=*/false, &rewriter);
143 
144     // Accept only static shaped types.
145     auto lhs_shape_type = lhs.getType().dyn_cast_or_null<ShapedType>();
146     auto rhs_shape_type = rhs.getType().dyn_cast_or_null<ShapedType>();
147     if (!lhs_shape_type || !rhs_shape_type) return failure();
148     if (!lhs_shape_type.hasStaticShape() || !rhs_shape_type.hasStaticShape())
149       return failure();
150 
151     // Dot resulting shape.
152     auto lhs_shape = lhs_shape_type.getShape();
153     auto rhs_shape = rhs_shape_type.getShape();
154     auto new_dot_type =
155         RankedTensorType::get({lhs_shape[0], rhs_shape[1]}, dot_element_type);
156 
157     ArrayAttr precision_config;
158     if (op.precision_config()) precision_config = *op.precision_config();
159     auto new_dot_op = rewriter.create<DotOp>(op.getLoc(), new_dot_type, lhs,
160                                              rhs, precision_config);
161 
162     rewriter.replaceOpWithNewOp<ReshapeOp>(op, op.getType(), new_dot_op);
163     return success();
164   }
165 };
166 
167 struct LegalizeGeneralDotPass
168     : public LegalizeGeneralDotPassBase<LegalizeGeneralDotPass> {
169   /// Lower all general dots that can be represented as a non-batched matmul.
runOnFunctionmlir::mhlo::__anonf50d5d480111::LegalizeGeneralDotPass170   void runOnFunction() override {
171     OwningRewritePatternList patterns(&getContext());
172     PopulateGeneralDotOpLoweringPatterns(&patterns, &getContext());
173     (void)applyPatternsAndFoldGreedily(getFunction(), std::move(patterns));
174   }
175 };
176 
177 }  // namespace
178 }  // namespace mhlo
179 }  // namespace mlir
180 
PopulateGeneralDotOpLoweringPatterns(OwningRewritePatternList * patterns,MLIRContext * ctx)181 void mlir::mhlo::PopulateGeneralDotOpLoweringPatterns(
182     OwningRewritePatternList *patterns, MLIRContext *ctx) {
183   patterns->insert<GeneralDotConvert>(ctx);
184 }
185 
createLegalizeGeneralDotPass()186 std::unique_ptr<::mlir::Pass> mlir::mhlo::createLegalizeGeneralDotPass() {
187   return std::make_unique<LegalizeGeneralDotPass>();
188 }
189