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 #ifndef TENSORFLOW_CORE_KERNELS_MKL_TFCONV_OP_H_ 17 #define TENSORFLOW_CORE_KERNELS_MKL_TFCONV_OP_H_ 18 19 #ifdef INTEL_MKL 20 21 #include <algorithm> 22 #include <vector> 23 #include "tensorflow/core/framework/numeric_op.h" 24 #include "tensorflow/core/framework/op.h" 25 #include "tensorflow/core/framework/op_kernel.h" 26 #include "tensorflow/core/framework/register_types.h" 27 #include "tensorflow/core/framework/tensor.h" 28 #include "tensorflow/core/framework/tensor_shape.h" 29 #include "tensorflow/core/kernels/ops_util.h" 30 #include "tensorflow/core/platform/cpu_info.h" 31 #include "tensorflow/core/platform/macros.h" 32 #include "tensorflow/core/util/tensor_format.h" 33 34 #include "mkl_dnn.h" 35 #include "mkl_dnn_types.h" 36 #include "tensorflow/core/util/mkl_util.h" 37 38 #ifndef INTEL_MKL_ML 39 using mkldnn::stream; 40 #endif 41 42 namespace tensorflow { 43 typedef Eigen::ThreadPoolDevice CPUDevice; 44 45 /////////////////////////////////////////////////////////// 46 // Op kernel 47 /////////////////////////////////////////////////////////// 48 49 template <typename Device, typename T> 50 class MklToTfOp : public OpKernel { 51 public: 52 explicit MklToTfOp(OpKernelConstruction* context) : OpKernel(context) { 53 OP_REQUIRES_OK(context, context->GetAttr("data_format", &data_format_str)); 54 OP_REQUIRES_OK(context, context->GetAttr("T", &op_data_type)); 55 has_avx512f_ = port::TestCPUFeature(port::CPUFeature::AVX512F); 56 } 57 58 void Compute(OpKernelContext* context) override { 59 ConvertMklToTf(this, context, data_format_str, op_data_type, has_avx512f_, 60 0); 61 VLOG(1) << "MKLToTFConversion complete successfully."; 62 } 63 64 #ifndef INTEL_MKL_ML 65 static void ConvertMklToTf(OpKernel* op_kernel, OpKernelContext* context, 66 string data_format_str, DataType op_data_type, 67 bool has_avx512f, uint input_number) { 68 try { 69 // Check that input tensor is in MKL format. 70 const Tensor& input_tensor = MklGetInput(context, input_number); 71 MklDnnShape input_shape; 72 GetMklShape(context, input_number, &input_shape); 73 74 // if input is already in Tf format, then copy input tensor to output. 75 if (!input_shape.IsMklTensor()) { 76 context->set_output(input_number, input_tensor); 77 VLOG(1) << "MKLToTFConversion: No conversion needed, " 78 << "copying input to output"; 79 return; 80 } 81 82 // Check that input data type is same as operator data type and that it 83 // is same as output data type. 84 DataType input_data_type = op_kernel->input_type(input_number); 85 DataType output_data_type = op_kernel->output_type(input_number); 86 CHECK_EQ(op_data_type, input_data_type); 87 CHECK_EQ(op_data_type, output_data_type); 88 89 auto cpu_engine = engine(engine::cpu, 0); 90 MklDnnData<T> input(&cpu_engine); 91 92 // Get Mkl layout of input tensor. 93 auto input_mkl_md = input_shape.GetMklLayout(); 94 // Get TensorFlow layout of input tensor. Expected output of conversion 95 // has same layout as Tensorflow layout of input tensor. 96 auto output_tf_md = input_shape.GetTfLayout(); 97 auto output_tf_pd = memory::primitive_desc(output_tf_md, cpu_engine); 98 // Set input Mkl layout as the user layout. 99 input.SetUsrMem(input_mkl_md, &input_tensor); 100 101 // Allocate output tensor. 102 TensorShape output_shape = input_shape.GetTfShape(); 103 Tensor* output_tensor = NULL; 104 OP_REQUIRES_OK(context, context->allocate_output( 105 input_number, output_shape, &output_tensor)); 106 CHECK_NOTNULL(output_tensor); 107 108 // Do we need to reorder Mkl layout into TensorFlow layout? 109 if (input.IsReorderNeeded(output_tf_pd)) { 110 // Insert reorder between Mkl layout and TensorFlow layout. 111 std::vector<primitive> net; 112 CHECK_EQ(input.CheckReorderToOpMem(output_tf_pd, output_tensor, &net), 113 true); 114 stream(stream::kind::eager).submit(net).wait(); 115 } else { 116 // If not, just forward input tensor to output tensor. 117 CHECK(output_tensor->CopyFrom(input_tensor, output_shape)); 118 } 119 } catch (mkldnn::error& e) { 120 string error_msg = "Status: " + std::to_string(e.status) + 121 ", message: " + std::string(e.message) + ", in file " + 122 std::string(__FILE__) + ":" + std::to_string(__LINE__); 123 OP_REQUIRES_OK( 124 context, 125 errors::Aborted("Operation received an exception:", error_msg)); 126 } 127 } 128 #else 129 static void ConvertMklToTf(OpKernel* op_kernel, OpKernelContext* context, 130 string data_format_str, DataType op_data_type, 131 bool has_avx512f, uint32 input_number) { 132 // Check that input tensor is in MKL format. 133 const Tensor& input_tensor = MklGetInput(context, input_number); 134 MklShape input_shape; 135 GetMklShape(context, input_number, &input_shape); 136 137 // if input is already in Tf format, then just copy input tensor to output. 138 if (!input_shape.IsMklTensor()) { 139 context->set_output(input_number, input_tensor); 140 VLOG(1) << "MKLToTFConversion: No conversion needed, " 141 << "copying input to output"; 142 return; 143 } 144 145 // Check that input data type is same as operator data type and that it is 146 // same as output data type. 147 DataType input_data_type = op_kernel->input_type(input_number); 148 DataType output_data_type = op_kernel->output_type(input_number); 149 CHECK_EQ(op_data_type, input_data_type); 150 CHECK_EQ(op_data_type, output_data_type); 151 152 TensorShape output_shape; 153 size_t ndims = input_shape.GetDimension(); 154 size_t* in_sizes = new size_t[ndims]; 155 for (size_t i = 0; i < ndims; i++) { 156 // Outermost to innermost dimension 157 output_shape.AddDim(input_shape.GetSizes()[input_shape.tf_dim_idx(i)]); 158 in_sizes[i] = input_shape.GetSizes()[i]; 159 } 160 161 // Allocate output tensor. 162 Tensor* output_tensor = NULL; 163 OP_REQUIRES_OK(context, context->allocate_output(input_number, output_shape, 164 &output_tensor)); 165 166 dnnLayout_t output_layout = 167 static_cast<dnnLayout_t>(input_shape.GetTfLayout()); 168 // Execute DNNConversion. 169 void* input_buffer = 170 static_cast<void*>(const_cast<T*>(input_tensor.flat<T>().data())); 171 delete[] in_sizes; 172 void* output_buffer = 173 static_cast<void*>(const_cast<T*>(output_tensor->flat<T>().data())); 174 input_shape.GetConvertedFlatData(output_layout, input_buffer, 175 output_buffer); 176 VLOG(1) << "MKLToTFConversion complete successfully."; 177 } 178 #endif 179 180 private: 181 /// Data format of the operation 182 string data_format_str; 183 184 /// Data type of the operation 185 DataType op_data_type; 186 187 /// CPUIDInfo 188 bool has_avx512f_ = false; 189 }; 190 191 /////////////////////////////////////////////////////////// 192 // Register kernel 193 /////////////////////////////////////////////////////////// 194 195 #define REGISTER_CPU(T) \ 196 REGISTER_KERNEL_BUILDER(Name("_MklToTf") \ 197 .Device(DEVICE_CPU) \ 198 .TypeConstraint<T>("T") \ 199 .Label(mkl_op_registry::kMklOpLabel), \ 200 MklToTfOp<CPUDevice, T>); 201 202 TF_CALL_NUMBER_TYPES(REGISTER_CPU); 203 #undef REGISTER_CPU 204 } // namespace tensorflow 205 #endif // INTEL_MKL 206 #endif // TENSORFLOW_CORE_KERNELS_MKL_TFCONV_OP_H_ 207
