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 #include "tensorflow/compiler/tf2xla/xla_helpers.h" 17 #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" 18 #include "tensorflow/compiler/tf2xla/xla_op_registry.h" 19 #include "tensorflow/core/util/tensor_format.h" 20 21 namespace tensorflow { 22 namespace { 23 24 class SpaceToDepthOp : public XlaOpKernel { 25 public: 26 explicit SpaceToDepthOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { 27 string data_format_str; 28 OP_REQUIRES_OK(ctx, ctx->GetAttr("data_format", &data_format_str)); 29 OP_REQUIRES(ctx, FormatFromString(data_format_str, &data_format_), 30 errors::InvalidArgument("Invalid data format")); 31 32 OP_REQUIRES(ctx, data_format_ == FORMAT_NCHW || data_format_ == FORMAT_NHWC, 33 errors::InvalidArgument("Unsupported data format ", 34 ToString(data_format_), 35 "; expected formats NHWC or NCHW")); 36 37 OP_REQUIRES_OK(ctx, ctx->GetAttr("block_size", &block_size_)); 38 OP_REQUIRES( 39 ctx, block_size_ > 1, 40 errors::InvalidArgument("Block size should be > 1: ", block_size_)); 41 } 42 43 void Compile(XlaOpKernelContext* ctx) override { 44 const TensorShape input_tensor_shape = ctx->InputShape(0); 45 int input_rank = input_tensor_shape.dims(); 46 static const int kRequiredDims = 4; 47 OP_REQUIRES(ctx, kRequiredDims == input_rank, 48 errors::InvalidArgument("Input rank should be ", kRequiredDims, 49 "; got ", input_rank)); 50 const gtl::InlinedVector<int64, 4> input_shape = 51 input_tensor_shape.dim_sizes(); 52 53 xla::ComputationBuilder* b = ctx->builder(); 54 xla::ComputationDataHandle input = ctx->Input(0); 55 56 int feature_dim = GetTensorFeatureDimIndex(input_rank, data_format_); 57 int num_spatial_dims = GetTensorSpatialDims(input_rank, data_format_); 58 59 std::vector<int64> reshaped_shape; 60 std::vector<int64> transpose_order; 61 std::vector<int64> output_shape; 62 reshaped_shape.reserve(input_rank); 63 transpose_order.reserve(input_rank); 64 output_shape.reserve(input_rank); 65 if (data_format_ == FORMAT_NHWC) { 66 int64 block_elems = 1; 67 for (int i = 0; i < num_spatial_dims; ++i) { 68 OP_REQUIRES(ctx, input_shape[1 + i] % block_size_ == 0, 69 errors::InvalidArgument( 70 "input shape[", 1 + i, "]=", input_shape[1 + i], 71 " is not divisible by block_size=", block_size_)); 72 block_elems *= block_size_; 73 } 74 75 reshaped_shape.push_back(input_shape[0]); 76 for (int i = 0; i < num_spatial_dims; ++i) { 77 reshaped_shape.push_back(input_shape[1 + i] / block_size_); 78 reshaped_shape.push_back(block_size_); 79 } 80 reshaped_shape.push_back(input_shape[feature_dim]); 81 82 transpose_order.push_back(0); 83 for (int i = 0; i < num_spatial_dims; ++i) { 84 transpose_order.push_back(i * 2 + 1); 85 } 86 for (int i = 0; i < num_spatial_dims; ++i) { 87 transpose_order.push_back(i * 2 + 2); 88 } 89 transpose_order.push_back(feature_dim + num_spatial_dims); 90 91 output_shape.push_back(input_shape[0]); 92 for (int i = 0; i < num_spatial_dims; ++i) { 93 output_shape.push_back(input_shape[1 + i] / block_size_); 94 } 95 output_shape.push_back(input_shape[feature_dim] * block_elems); 96 } else { 97 // FORMAT_NCHW 98 int64 block_elems = 1; 99 for (int i = 0; i < num_spatial_dims; ++i) { 100 OP_REQUIRES(ctx, input_shape[2 + i] % block_size_ == 0, 101 errors::InvalidArgument( 102 "input shape[", 2 + i, "]=", input_shape[2 + i], 103 " is not divisible by block_size=", block_size_)); 104 block_elems *= block_size_; 105 } 106 107 reshaped_shape.push_back(input_shape[0]); 108 reshaped_shape.push_back(input_shape[feature_dim]); 109 for (int i = 0; i < num_spatial_dims; ++i) { 110 reshaped_shape.push_back(input_shape[2 + i] / block_size_); 111 reshaped_shape.push_back(block_size_); 112 } 113 114 transpose_order.push_back(0); 115 for (int i = 0; i < num_spatial_dims; ++i) { 116 transpose_order.push_back(i * 2 + 3); 117 } 118 transpose_order.push_back(feature_dim); 119 for (int i = 0; i < num_spatial_dims; ++i) { 120 transpose_order.push_back(i * 2 + 2); 121 } 122 123 output_shape.push_back(input_shape[0]); 124 output_shape.push_back(input_shape[feature_dim] * block_elems); 125 for (int i = 0; i < num_spatial_dims; ++i) { 126 output_shape.push_back(input_shape[2 + i] / block_size_); 127 } 128 } 129 130 // Note: comments are given in NHWC format; NCHW is similar with a different 131 // dimension order. 132 // 1. Reshape `input` to `reshaped` of shape: 133 // 134 // [batch, 135 // input_shape[1] / block_size_, block_size_, 136 // input_shape[2] / block_size_, block_size_, 137 // depth] 138 xla::ComputationDataHandle reshaped = b->Reshape(input, reshaped_shape); 139 140 // 2. Permute dimensions of `reshaped` to produce 141 // `permuted_reshaped` of shape: 142 // 143 // [batch, 144 // input_shape[1] / block_size_, 145 // input_shape[2] / block_size_, 146 // block_size_, block_size_, 147 // depth] 148 xla::ComputationDataHandle permuted_reshaped = 149 b->Transpose(reshaped, transpose_order); 150 151 // 3. Reshape `permuted_reshaped` to flatten `block_shape` into the 152 // batch dimension, producing an output tensor of shape: 153 // 154 // [batch, 155 // input_shape[1] / block_size_, 156 // input_shape[2] / block_size_, 157 // block_size_ * block_size_ * depth] 158 // 159 xla::ComputationDataHandle output = 160 b->Reshape(permuted_reshaped, output_shape); 161 162 ctx->SetOutput(0, output); 163 } 164 165 private: 166 TensorFormat data_format_; 167 int block_size_; 168 }; 169 REGISTER_XLA_OP(Name("SpaceToDepth"), SpaceToDepthOp); 170 171 } // namespace 172 } // namespace tensorflow 173
