1 /* Copyright 2017 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 // XLA-specific Tile Op. 17 18 #include <vector> 19 #include "absl/algorithm/container.h" 20 #include "absl/types/span.h" 21 #include "tensorflow/compiler/tf2xla/lib/broadcast.h" 22 #include "tensorflow/compiler/tf2xla/type_util.h" 23 #include "tensorflow/compiler/tf2xla/xla_helpers.h" 24 #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" 25 #include "tensorflow/compiler/tf2xla/xla_op_registry.h" 26 #include "tensorflow/compiler/xla/client/xla_builder.h" 27 #include "tensorflow/core/framework/numeric_op.h" 28 #include "tensorflow/core/framework/op_kernel.h" 29 #include "tensorflow/core/framework/tensor.h" 30 #include "tensorflow/core/framework/tensor_shape.h" 31 #include "tensorflow/core/framework/type_index.h" 32 #include "tensorflow/core/lib/core/errors.h" 33 #include "tensorflow/core/platform/macros.h" 34 35 namespace tensorflow { 36 namespace { 37 38 // -------------------------------------------------------------------------- 39 class TileOp : public XlaOpKernel { 40 public: TileOp(OpKernelConstruction * ctx)41 explicit TileOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} 42 Compile(XlaOpKernelContext * ctx)43 void Compile(XlaOpKernelContext* ctx) override { 44 const TensorShape input_shape = ctx->InputShape("input"); 45 const TensorShape multiples_shape = ctx->InputShape("multiples"); 46 47 OP_REQUIRES( 48 ctx, TensorShapeUtils::IsVector(multiples_shape), 49 errors::InvalidArgument("Expected multiples to be 1-D, but got shape ", 50 multiples_shape.DebugString())); 51 OP_REQUIRES(ctx, input_shape.dims() == multiples_shape.num_elements(), 52 errors::InvalidArgument( 53 "Expected multiples argument to be a vector of length ", 54 input_shape.dims(), " but got length ", 55 multiples_shape.dim_size(0))); 56 const int input_dims = input_shape.dims(); 57 auto input = ctx->Input(0); 58 // If input is a scalar then multiples has 0 elements and this is 59 // a NoOp. 60 if (input_dims == 0) { 61 ctx->SetOutput(0, input); 62 return; 63 } 64 65 std::vector<int64> multiples; 66 OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntVector("multiples", &multiples)); 67 std::vector<int64> output_dims(input_shape.dims()); 68 for (int64 i = 0; i < input_shape.dims(); ++i) { 69 OP_REQUIRES(ctx, multiples[i] >= 0, 70 errors::InvalidArgument("Expected multiples[", i, 71 "] >= 0, but got ", output_dims[i])); 72 output_dims[i] = input_shape.dim_size(i) * multiples[i]; 73 } 74 75 // If all multiples are 1, than the input is the same as the output. 76 if (absl::c_all_of(multiples, 77 [](int64 multiple) { return multiple == 1; })) { 78 ctx->SetOutput(0, input); 79 return; 80 } 81 82 bool can_tile_with_implicit_broadcast = true; 83 for (int i = 0; i < input_dims; ++i) { 84 int64 multiple = multiples[i]; 85 // If the multiple and input dimension are not 1, then tile cannot be 86 // implemented with a single hlo broadcast. 87 if (multiple != 1 && input_shape.dim_size(i) != 1) { 88 can_tile_with_implicit_broadcast = false; 89 } 90 } 91 92 if (can_tile_with_implicit_broadcast) { 93 // Create a constant Zero the size of the output shape to leverage binary 94 // operation broadcast semantics. 95 auto broadcasted_zero = xla::Broadcast( 96 XlaHelpers::Zero(ctx->builder(), ctx->input_type(0)), output_dims); 97 if (ctx->input_type(0) == DT_BOOL) { 98 ctx->SetOutput(0, xla::Or(broadcasted_zero, input)); 99 } else { 100 ctx->SetOutput(0, xla::Add(broadcasted_zero, input)); 101 } 102 return; 103 } 104 105 auto result = BroadcastTo(ctx->Input("input"), output_dims); 106 OP_REQUIRES_OK(ctx, result.status()); 107 ctx->SetOutput(0, result.ValueOrDie()); 108 } 109 110 private: 111 TF_DISALLOW_COPY_AND_ASSIGN(TileOp); 112 }; 113 114 REGISTER_XLA_OP(Name("Tile").CompileTimeConstantInput("multiples"), TileOp); 115 116 } // namespace 117 } // namespace tensorflow 118