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 20 #include "absl/algorithm/container.h" 21 #include "absl/types/span.h" 22 #include "tensorflow/compiler/tf2xla/lib/broadcast.h" 23 #include "tensorflow/compiler/tf2xla/type_util.h" 24 #include "tensorflow/compiler/tf2xla/xla_helpers.h" 25 #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" 26 #include "tensorflow/compiler/tf2xla/xla_op_registry.h" 27 #include "tensorflow/compiler/xla/client/value_inference.h" 28 #include "tensorflow/compiler/xla/client/xla_builder.h" 29 #include "tensorflow/core/framework/numeric_op.h" 30 #include "tensorflow/core/framework/op_kernel.h" 31 #include "tensorflow/core/framework/tensor.h" 32 #include "tensorflow/core/framework/tensor_shape.h" 33 #include "tensorflow/core/framework/type_index.h" 34 #include "tensorflow/core/lib/core/errors.h" 35 #include "tensorflow/core/platform/macros.h" 36 37 namespace tensorflow { 38 namespace { 39 40 // -------------------------------------------------------------------------- 41 class TileOp : public XlaOpKernel { 42 public: TileOp(OpKernelConstruction * ctx)43 explicit TileOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} 44 Compile(XlaOpKernelContext * ctx)45 void Compile(XlaOpKernelContext* ctx) override { 46 const TensorShape input_shape = ctx->InputShape("input"); 47 const TensorShape multiples_shape = ctx->InputShape("multiples"); 48 49 OP_REQUIRES( 50 ctx, TensorShapeUtils::IsVector(multiples_shape), 51 errors::InvalidArgument("Expected multiples to be 1-D, but got shape ", 52 multiples_shape.DebugString())); 53 OP_REQUIRES(ctx, input_shape.dims() == multiples_shape.num_elements(), 54 errors::InvalidArgument( 55 "Expected multiples argument to be a vector of length ", 56 input_shape.dims(), " but got length ", 57 multiples_shape.dim_size(0))); 58 const int input_dims = input_shape.dims(); 59 auto input = ctx->Input(0); 60 // If input is a scalar then multiples has 0 elements and this is 61 // a NoOp. 62 if (input_dims == 0) { 63 ctx->SetOutput(0, input); 64 return; 65 } 66 67 std::vector<int64_t> multiples_bounds; 68 OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntVector( 69 "multiples", &multiples_bounds, 70 xla::ValueInferenceMode::kUpperBound)); 71 72 std::vector<int64_t> output_dims(input_shape.dims()); 73 for (int64_t i = 0; i < input_shape.dims(); ++i) { 74 OP_REQUIRES(ctx, multiples_bounds[i] >= 0, 75 errors::InvalidArgument("Expected multiples[", i, 76 "] >= 0, but got ", output_dims[i])); 77 output_dims[i] = input_shape.dim_size(i) * multiples_bounds[i]; 78 } 79 80 std::vector<bool> multiples_are_dynamic; 81 82 OP_REQUIRES_OK(ctx, ctx->ResolveInputDynamismIntoPredVector( 83 1, &multiples_are_dynamic)); 84 85 bool all_multiples_are_static = absl::c_all_of( 86 multiples_are_dynamic, [](bool dynamic) { return !dynamic; }); 87 // If a value is static, it means the upper bound is the value itself: 88 // constant_value = constant_upper_boudn = counstant_lower_bound 89 if (all_multiples_are_static) { 90 // If all multiples are 1, than the input is the same as the output. 91 if (absl::c_all_of(multiples_bounds, 92 [](int64_t multiple) { return multiple == 1; })) { 93 ctx->SetOutput(0, input); 94 return; 95 } 96 } 97 98 auto result_or = BroadcastTo(ctx->Input("input"), output_dims); 99 100 OP_REQUIRES_OK(ctx, result_or.status()); 101 auto result = result_or.ValueOrDie(); 102 if (!all_multiples_are_static) { 103 // Some values of multiples are unknown at compile time, this is a dynamic 104 // tile op. We need to call set dimension size. 105 for (int64_t i = 0; i < multiples_are_dynamic.size(); ++i) { 106 if (!multiples_are_dynamic[i]) { 107 continue; 108 } 109 // If a dimension is dynamic, call set-dimension-size on the output. 110 auto dynamic_dim_size = 111 xla::Slice(ctx->Input("multiples"), {i}, {i + 1}, {1}); 112 dynamic_dim_size = xla::Reshape(dynamic_dim_size, {}); 113 dynamic_dim_size = xla::ConvertElementType(dynamic_dim_size, xla::S32); 114 result = xla::SetDimensionSize(result, dynamic_dim_size, i); 115 } 116 } 117 118 ctx->SetOutput(0, result); 119 } 120 121 private: 122 TF_DISALLOW_COPY_AND_ASSIGN(TileOp); 123 }; 124 125 REGISTER_XLA_OP(Name("Tile").CompileTimeConstantInput("multiples"), TileOp); 126 127 } // namespace 128 } // namespace tensorflow 129