| 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 | // TODO(intel): Improve error handling in this file; instead of CHECK failing |
| 17 | // all over the place, we should log an error and execute the original graph. |
| 18 | #ifdef INTEL_MKL |
| 19 | |
| 20 | #include "tensorflow/core/common_runtime/mkl_layout_pass.h" |
| 21 | |
| 22 | #include <algorithm> |
| 23 | #include <functional> |
| 24 | #include <memory> |
| 25 | #include <queue> |
| 26 | #include <set> |
| 27 | #include <stack> |
| 28 | #include <tuple> |
| 29 | #include <unordered_set> |
| 30 | #include <utility> |
| 31 | #include <vector> |
| 32 | |
| 33 | #include "absl/base/call_once.h" |
| 34 | #include "tensorflow/core/common_runtime/function.h" |
| 35 | #include "tensorflow/core/common_runtime/optimization_registry.h" |
| 36 | #include "tensorflow/core/framework/node_def_util.h" |
| 37 | #include "tensorflow/core/framework/tensor.pb.h" |
| 38 | #include "tensorflow/core/graph/algorithm.h" |
| 39 | #include "tensorflow/core/graph/graph.h" |
| 40 | #include "tensorflow/core/graph/mkl_graph_util.h" |
| 41 | #include "tensorflow/core/graph/node_builder.h" |
| 42 | #include "tensorflow/core/lib/core/status.h" |
| 43 | #include "tensorflow/core/lib/gtl/array_slice.h" |
| 44 | #include "tensorflow/core/lib/gtl/map_util.h" |
| 45 | #include "tensorflow/core/lib/hash/hash.h" |
| 46 | #include "tensorflow/core/platform/logging.h" |
| 47 | #include "tensorflow/core/util/tensor_format.h" |
| 48 | #include "tensorflow/core/util/util.h" |
| 49 | |
| 50 | namespace tensorflow { |
| 51 | |
| 52 | // This pass implements rewriting of graph to support following scenarios: |
| 53 | // (A) Merging nodes in the graph |
| 54 | // (B) Rewriting a node in the graph to a new node |
| 55 | // Rewrite happens under following scenario: |
| 56 | // - Propagating Mkl layout as an additional output tensor |
| 57 | // (we will loosely call a tensor that carries Mkl layout as Mkl tensor |
| 58 | // henceforth.) from every Mkl supported NN layer. |
| 59 | // |
| 60 | // Example of A : Merging nodes in the graph |
| 61 | // ----------------------------------------- |
| 62 | // Currently, we merge Conv2D+AddBias together. Consider Conv2D and BiasAdd as: |
| 63 | // |
| 64 | // O = Conv2D(A, B) |
| 65 | // P = BiasAdd(O, C) |
| 66 | // |
| 67 | // We merge them into Conv2DWithBias as: |
| 68 | // P = _MklConv2DWithBias(A, A_m, B, B_m, C, C_m) |
| 69 | // |
| 70 | // The meaning of A_m, B_m and C_m is explained in B.1. |
| 71 | // |
| 72 | // Merge rules: |
| 73 | // - The merge for Conv2D and BiasAdd happens when the output of Conv2D _only_ |
| 74 | // goes to BiasAdd. |
| 75 | // - Also, the intersection of attributes of both the nodes must have same |
| 76 | // values. |
| 77 | // - Both the nodes must have been assigned to same device (if any). |
| 78 | // |
| 79 | // Example of B.1 : Rewriting nodes to Mkl nodes |
| 80 | // --------------------------------------------- |
| 81 | // Consider a Relu node. Current definition of Relu node looks like: |
| 82 | // |
| 83 | // O = Relu(A) |
| 84 | // |
| 85 | // Relu has 1 input (A), and 1 output (O). |
| 86 | // |
| 87 | // This rewrite pass will generate a new graph node for Relu (new node is |
| 88 | // called MklRelu) as: |
| 89 | // |
| 90 | // O, O_m = MklRelu(A, A_m) |
| 91 | // |
| 92 | // MklRelu has 2 inputs (A and A_m) and 2 outputs (O and O_m). Here input A is |
| 93 | // same as input A of Relu; output O is same as output O of Relu. O_m is the |
| 94 | // additional output tensor that will be set by MklRelu, and it represents |
| 95 | // Mkl tensor corresponding to O -- in other words, O_m is some kind of |
| 96 | // metadata for O. A_m is additional input of Relu, and it represents metadata |
| 97 | // for A - as O_m is metadata for O, A_m is metadata for A. MklRelu receives |
| 98 | // this metadata from previous node in the graph. |
| 99 | // |
| 100 | // When a previous node in the graph is an Mkl node, A_m will represent a valid |
| 101 | // Mkl tensor. But when a previous node is not an Mkl node, A_m will represent |
| 102 | // a dummy Mkl tensor. |
| 103 | // |
| 104 | // Rewriting rules: |
| 105 | // - Selection of a node for rewriting happens by registering the op type of |
| 106 | // the node with the rewriting pass. If the op type is not registered, then |
| 107 | // all nodes of this op type will not be rewritten. |
| 108 | // - Number of inputs after rewriting: |
| 109 | // Since for every input Tensorflow tensor, the rewritten node gets Mkl |
| 110 | // tensor(s), rewritten node gets 2*N inputs, where N is the number of |
| 111 | // inputs for the original node. |
| 112 | // - Number of outputs after rewriting: |
| 113 | // Since for every output Tensorflow tensor, the rewritten node generates |
| 114 | // Mkl tensor(s), the rewritten node generates 2*N outputs, where N is the |
| 115 | // number of outputs of the original node. |
| 116 | // - Ordering of Tensorflow tensors and Mkl tensors: |
| 117 | // Since every rewritten node generates twice the number of inputs and |
| 118 | // outputs, one could imagine various orderings among Tensorflow tensors |
| 119 | // and Mkl tensors. E.g., assume an op 'Conv2D' that takes (A, B) as |
| 120 | // inputs, then the new op '_MklConv2D' can take inputs A, B, A_m and B_m |
| 121 | // in A, A_m, B, B_m order or it can also take them in A, B, A_m, B_m |
| 122 | // order. Among N inputs one can get N! permutations. |
| 123 | // |
| 124 | // So the question is: which order do we follow? We support 2 types of |
| 125 | // orderings: (1) interleaved, and (2) contiguous. Interleaved ordering |
| 126 | // follows an intuitive order where an Mkl tensor follows the |
| 127 | // corresponding Tensorflow tensor immediately. In the context of the |
| 128 | // above example, it will be: A, A_m, B, B_m. Note that the ordering rule |
| 129 | // applies to both the inputs and outputs. Contiguous ordering means |
| 130 | // all the Tensorflow tensors are contiguous followed by all the Mkl |
| 131 | // tensors. We use contiguous ordering as default. |
| 132 | // |
| 133 | // Graph rewrite algorithm: |
| 134 | // Algorithm: Graph Rewrite |
| 135 | // Input: Graph G, Names of the nodes to rewrite and their new names |
| 136 | // Output: Modified Graph G' if the nodes are modified, G otherwise. |
| 137 | // Start: |
| 138 | // N = Topological_Sort(G) // N is a set of nodes in toposort order. |
| 139 | // foreach node n in N |
| 140 | // do |
| 141 | // if (Is_MKL_Op(n)) // Can this node accept an Mkl layout as input. |
| 142 | // then |
| 143 | // E = set of <incoming edge and its src_output slot> of n |
| 144 | // E' = {} // a new set of edges for rewritten node |
| 145 | // foreach <e,s> in E |
| 146 | // do |
| 147 | // E' U {<e,s>} // First copy edge which generates Tensorflow |
| 148 | // // tensor as it is |
| 149 | // m = Source node of edge e |
| 150 | // if Is_Rewritten(m) // Did we rewrite this node in this pass? |
| 151 | // then |
| 152 | // E' U {<m,s+1>} // If yes, then m will generate an Mkl |
| 153 | // // tensor as an additional output. |
| 154 | // else |
| 155 | // d = Generate_Dummy_Mkl_Tensor() // If not, generate a dummy |
| 156 | // // Mkl tensor. |
| 157 | // E' U {<d,0>} // The dummy Mkl tensor has only 1 output slot. |
| 158 | // fi |
| 159 | // done |
| 160 | // n' = Build_New_Node(G,new_name,E') |
| 161 | // Mark_Rewritten(n') // Mark the new node as being rewritten. |
| 162 | // fi |
| 163 | // done |
| 164 | // |
| 165 | // Explanation: |
| 166 | // For graph rewrite, we visit nodes of the input graph in the |
| 167 | // topological sort order. With this ordering, we visit nodes in the |
| 168 | // top-to-bottom fashion. We need this order because while visiting a |
| 169 | // node we want that all of its input nodes are visited and rewritten if |
| 170 | // applicable. This is because if we need to rewrite a given node |
| 171 | // then all of its input nodes need to be fixed (in other words they |
| 172 | // cannot be deleted later.) |
| 173 | // |
| 174 | // While visiting a node, we first check if the op type of the node is |
| 175 | // an Mkl op. If it is, then we rewrite that node after constructing |
| 176 | // new inputs to the node. If the op type of the node is not Mkl op, |
| 177 | // then we do not rewrite that node. |
| 178 | // |
| 179 | // Handling workspace propagation for certain ops: |
| 180 | // |
| 181 | // Certain backward ops in MKL (MaxPool, LRN and BatchNorm) require |
| 182 | // passing of a workspace from their respective forward ops. Workspace |
| 183 | // tensors provide memory for storing results of intermediate operations |
| 184 | // which are helpful in backward propagation. TensorFlow does not have |
| 185 | // a notion of a workspace and as a result does not allow producing |
| 186 | // additional outputs from these forward ops. For these ops, we need |
| 187 | // to add 2 extra edges between forward ops and their corresponding |
| 188 | // backward ops - the first extra edge carries a workspace tensor and |
| 189 | // the second one carries an Mkl tensor for the workspace tensor. |
| 190 | // |
| 191 | // Example: |
| 192 | // |
| 193 | // Typical graph for MaxPool and its gradient looks like: |
| 194 | // |
| 195 | // A = MaxPool(T) |
| 196 | // B = MaxPoolGrad(X, A, Y) |
| 197 | // |
| 198 | // We will transform this graph to propagate the workspace as: |
| 199 | // (with the contiguous ordering) |
| 200 | // |
| 201 | // A, W, A_m, W_m = MklMaxPool(T, T_m) |
| 202 | // B, B_m = MklMaxPoolGrad(X, A, Y, W, X_m, A_m, Y_m, W_m) |
| 203 | // |
| 204 | // Here W is the workspace tensor. Transformed tensor names with the |
| 205 | // suffix _m are Mkl tensors, and this transformation has been done |
| 206 | // using the algorithm discussed earlier. The transformation for |
| 207 | // workspace propagation only adds extra outputs (W, W_m) for a forward |
| 208 | // op and connects them to the corresponding backward ops. |
| 209 | // |
| 210 | // Terms: |
| 211 | // |
| 212 | // Forward op name = name of the op in the forward pass |
| 213 | // where a workspace tensor originates (MaxPool in this example) |
| 214 | // Backward op name = name of the op in the backward pass that receives |
| 215 | // a workspace tensor from the forward op (MaxPoolGrad in the example) |
| 216 | // Slot = Position of the output or input slot that will be |
| 217 | // used by the workspace tensor (1 for MklMaxPool as W is the 2nd |
| 218 | // output of MaxPool (0 is 1st); 3 for MklMaxPoolGrad) |
| 219 | // |
| 220 | // Question: |
| 221 | // |
| 222 | // How do we associate a backward op to a forward op? There can be more |
| 223 | // than one op with the exact same name. |
| 224 | // |
| 225 | // In this example, we associate MaxPoolGrad with MaxPool. But there |
| 226 | // could be more than one MaxPool ops. To solve this problem, we look |
| 227 | // for _direct_ edge between a forward op and a backward op (tensor A is |
| 228 | // flowing along this edge in the example). |
| 229 | // |
| 230 | // How do we transform forward and backward ops when there is no direct |
| 231 | // edge between them? In such a case, we generate dummy tensors for |
| 232 | // workspace tensors. For the example, transformation of MaxPool will |
| 233 | // be exactly same as it would be when there is a direct edge between |
| 234 | // the forward and the backward op --- it is just that MaxPool won't |
| 235 | // generate any workspace tensor. For MaxPoolGrad, the transformation |
| 236 | // will also be same, but instead of connecting W and W_m with the |
| 237 | // outputs of MaxPool, we will produce dummy tensors for them, and we |
| 238 | // will set workspace_enabled attribute to false. |
| 239 | // |
| 240 | class MklLayoutRewritePass : public GraphOptimizationPass { |
| 241 | public: |
| 242 | MklLayoutRewritePass() { |
| 243 | // NOTE: names are alphabetically sorted. |
| 244 | csinfo_.addn = "AddN"; |
| 245 | csinfo_.avg_pool = "AvgPool"; |
| 246 | csinfo_.avg_pool_grad = "AvgPoolGrad"; |
| 247 | csinfo_.avg_pool3d = "AvgPool3D"; |
| 248 | csinfo_.avg_pool3d_grad = "AvgPool3DGrad"; |
| 249 | csinfo_.batch_matmul = "BatchMatMul"; |
| 250 | csinfo_.batch_matmul_v2 = "BatchMatMulV2"; |
| 251 | csinfo_.bias_add = "BiasAdd"; |
| 252 | csinfo_.bias_add_grad = "BiasAddGrad"; |
| 253 | csinfo_.concat = "Concat"; |
| 254 | csinfo_.concatv2 = "ConcatV2"; |
| 255 | csinfo_.conjugate_transpose = "ConjugateTranspose"; |
| 256 | csinfo_.conv2d = "Conv2D"; |
| 257 | csinfo_.conv2d_with_bias = "__MklDummyConv2DWithBias"; |
| 258 | csinfo_.conv2d_grad_input = "Conv2DBackpropInput"; |
| 259 | csinfo_.conv2d_grad_filter = "Conv2DBackpropFilter"; |
| 260 | csinfo_.conv2d_grad_filter_with_bias = |
| 261 | "__MklDummyConv2DBackpropFilterWithBias"; |
| 262 | csinfo_.conv3d = "Conv3D"; |
| 263 | csinfo_.conv3d_grad_input = "Conv3DBackpropInputV2"; |
| 264 | csinfo_.conv3d_grad_filter = "Conv3DBackpropFilterV2"; |
| 265 | csinfo_.depthwise_conv2d = "DepthwiseConv2dNative"; |
| 266 | csinfo_.depthwise_conv2d_grad_input = "DepthwiseConv2dNativeBackpropInput"; |
| 267 | csinfo_.depthwise_conv2d_grad_filter = |
| 268 | "DepthwiseConv2dNativeBackpropFilter"; |
| 269 | csinfo_.dequantize = "Dequantize"; |
| 270 | csinfo_.einsum = "Einsum"; |
| 271 | csinfo_.fused_batch_norm = "FusedBatchNorm"; |
| 272 | csinfo_.fused_batch_norm_grad = "FusedBatchNormGrad"; |
| 273 | csinfo_.fused_batch_norm_ex = "_FusedBatchNormEx"; |
| 274 | csinfo_.fused_batch_norm_v2 = "FusedBatchNormV2"; |
| 275 | csinfo_.fused_batch_norm_grad_v2 = "FusedBatchNormGradV2"; |
| 276 | csinfo_.fused_batch_norm_v3 = "FusedBatchNormV3"; |
| 277 | csinfo_.fused_batch_norm_grad_v3 = "FusedBatchNormGradV3"; |
| 278 | csinfo_.fused_conv2d = "_FusedConv2D"; |
| 279 | csinfo_.fused_conv3d = "_FusedConv3D"; |
| 280 | csinfo_.fused_depthwise_conv2d = "_FusedDepthwiseConv2dNative"; |
| 281 | csinfo_.fused_matmul = "_FusedMatMul"; |
| 282 | csinfo_.identity = "Identity"; |
| 283 | csinfo_.leakyrelu = "LeakyRelu"; |
| 284 | csinfo_.leakyrelu_grad = "LeakyReluGrad"; |
| 285 | csinfo_.lrn = "LRN"; |
| 286 | csinfo_.lrn_grad = "LRNGrad"; |
| 287 | csinfo_.matmul = "MatMul"; |
| 288 | csinfo_.max_pool = "MaxPool"; |
| 289 | csinfo_.max_pool_grad = "MaxPoolGrad"; |
| 290 | csinfo_.max_pool3d = "MaxPool3D"; |
| 291 | csinfo_.max_pool3d_grad = "MaxPool3DGrad"; |
| 292 | csinfo_.mkl_conv2d = "_MklConv2D"; |
| 293 | csinfo_.mkl_conv2d_grad_input = "_MklConv2DBackpropInput"; |
| 294 | csinfo_.mkl_conv2d_grad_filter = "_MklConv2DBackpropFilter"; |
| 295 | csinfo_.mkl_conv2d_with_bias = "_MklConv2DWithBias"; |
| 296 | csinfo_.mkl_conv2d_grad_filter_with_bias = |
| 297 | "_MklConv2DBackpropFilterWithBias"; |
| 298 | csinfo_.mkl_depthwise_conv2d_grad_input = |
| 299 | "_MklDepthwiseConv2dNativeBackpropInput"; |
| 300 | csinfo_.mkl_depthwise_conv2d_grad_filter = |
| 301 | "_MklDepthwiseConv2dNativeBackpropFilter"; |
| 302 | csinfo_.mkl_fused_batch_norm_ex = "_MklFusedBatchNormEx"; |
| 303 | csinfo_.mkl_fused_conv2d = "_MklFusedConv2D"; |
| 304 | csinfo_.mkl_fused_depthwise_conv2d = "_MklFusedDepthwiseConv2dNative"; |
| 305 | csinfo_.mkl_fused_matmul = "_MklFusedMatMul"; |
| 306 | csinfo_.mkl_native_conv2d_with_bias = "_MklNativeConv2DWithBias"; |
| 307 | csinfo_.mkl_native_conv2d_grad_filter_with_bias = |
| 308 | "_MklNativeConv2DBackpropFilterWithBias"; |
| 309 | csinfo_.mkl_native_fused_batch_norm_ex = "_MklNativeFusedBatchNormEx"; |
| 310 | csinfo_.mkl_native_fused_conv2d = "_MklNativeFusedConv2D"; |
| 311 | csinfo_.mkl_native_fused_conv3d = "_MklNativeFusedConv3D"; |
| 312 | csinfo_.mkl_native_fused_depthwise_conv2d = |
| 313 | "_MklNativeFusedDepthwiseConv2dNative"; |
| 314 | csinfo_.mkl_native_fused_matmul = "_MklNativeFusedMatMul"; |
| 315 | csinfo_.mkl_native_pad_with_conv2d = "_MklNativePadWithConv2D"; |
| 316 | csinfo_.mkl_native_pad_with_fused_conv2d = "_MklNativePadWithFusedConv2D"; |
| 317 | csinfo_.mkl_pad_with_conv2d = "_MklPadWithConv2D"; |
| 318 | csinfo_.mkl_pad_with_fused_conv2d = "_MklPadWithFusedConv2D"; |
| 319 | csinfo_.pad = "Pad"; |
| 320 | csinfo_.pad_with_conv2d = "__MklDummyPadWithConv2D"; |
| 321 | csinfo_.pad_with_fused_conv2d = "__MklDummyPadWithFusedConv2D"; |
| 322 | csinfo_.quantized_avg_pool = "QuantizedAvgPool"; |
| 323 | csinfo_.quantized_concatv2 = "QuantizedConcatV2"; |
| 324 | csinfo_.quantized_conv2d = "QuantizedConv2D"; |
| 325 | csinfo_.quantized_conv2d_per_channel = "QuantizedConv2DPerChannel"; |
| 326 | csinfo_.quantized_conv2d_with_requantize = "QuantizedConv2DAndRequantize"; |
| 327 | csinfo_.quantized_conv2d_with_bias = "QuantizedConv2DWithBias"; |
| 328 | csinfo_.quantized_conv2d_with_bias_and_requantize = |
| 329 | "QuantizedConv2DWithBiasAndRequantize"; |
| 330 | csinfo_.quantized_conv2d_and_relu = "QuantizedConv2DAndRelu"; |
| 331 | csinfo_.quantized_conv2d_and_relu_and_requantize = |
| 332 | "QuantizedConv2DAndReluAndRequantize"; |
| 333 | csinfo_.quantized_conv2d_with_bias_and_relu = |
| 334 | "QuantizedConv2DWithBiasAndRelu"; |
| 335 | csinfo_.quantized_conv2d_with_bias_and_relu_and_requantize = |
| 336 | "QuantizedConv2DWithBiasAndReluAndRequantize"; |
| 337 | csinfo_.quantized_max_pool = "QuantizedMaxPool"; |
| 338 | csinfo_.quantized_conv2d_with_bias_sum_and_relu = |
| 339 | "QuantizedConv2DWithBiasSumAndRelu"; |
| 340 | csinfo_.quantized_conv2d_with_bias_sum_and_relu_and_requantize = |
| 341 | "QuantizedConv2DWithBiasSumAndReluAndRequantize"; |
| 342 | csinfo_.quant_conv2d_with_bias_signed_sum_and_relu_and_requantize = |
| 343 | "QuantizedConv2DWithBiasSignedSumAndReluAndRequantize"; |
| 344 | csinfo_.quantized_matmul_with_bias = "QuantizedMatMulWithBias"; |
| 345 | csinfo_.quantized_matmul_with_bias_and_relu = |
| 346 | "QuantizedMatMulWithBiasAndRelu"; |
| 347 | csinfo_.quantized_matmul_with_bias_and_relu_and_requantize = |
| 348 | "QuantizedMatMulWithBiasAndReluAndRequantize"; |
| 349 | csinfo_.quantized_matmul_with_bias_and_dequantize = |
| 350 | "QuantizedMatMulWithBiasAndDequantize"; |
| 351 | csinfo_.quantized_matmul_with_bias_and_requantize = |
| 352 | "QuantizedMatMulWithBiasAndRequantize"; |
| 353 | csinfo_.quantized_depthwise_conv2d = "QuantizedDepthwiseConv2D"; |
| 354 | csinfo_.quantized_depthwise_conv2d_with_bias = |
| 355 | "QuantizedDepthwiseConv2DWithBias"; |
| 356 | csinfo_.quantized_depthwise_conv2d_with_bias_and_relu = |
| 357 | "QuantizedDepthwiseConv2DWithBiasAndRelu"; |
| 358 | csinfo_.quantized_depthwise_conv2d_with_bias_and_relu_and_requantize = |
| 359 | "QuantizedDepthwiseConv2DWithBiasAndReluAndRequantize"; |
| 360 | csinfo_.quantize_v2 = "QuantizeV2"; |
| 361 | csinfo_.relu = "Relu"; |
| 362 | csinfo_.relu_grad = "ReluGrad"; |
| 363 | csinfo_.relu6 = "Relu6"; |
| 364 | csinfo_.relu6_grad = "Relu6Grad"; |
| 365 | csinfo_.requantize = "Requantize"; |
| 366 | csinfo_.tanh = "Tanh"; |
| 367 | csinfo_.tanh_grad = "TanhGrad"; |
| 368 | csinfo_.reshape = "Reshape"; |
| 369 | csinfo_.slice = "Slice"; |
| 370 | csinfo_.softmax = "Softmax"; |
| 371 | csinfo_.split = "Split"; |
| 372 | csinfo_.transpose = "Transpose"; |
| 373 | // Element-wise ops. Ensure you also add any new ops to IsOpElementWise |
| 374 | // in the MklUtil.h (IsMklElementWiseOp method) to ensure that the |
| 375 | // MklInputConversion op is added before it. |
| 376 | csinfo_.add = "Add"; |
| 377 | csinfo_.add_v2 = "AddV2"; |
| 378 | csinfo_.maximum = "Maximum"; |
| 379 | csinfo_.mul = "Mul"; |
| 380 | csinfo_.squared_difference = "SquaredDifference"; |
| 381 | csinfo_.sub = "Sub"; |
| 382 | // End - element-wise ops. See note above. |
| 383 | |
| 384 | const bool native_fmt = NativeFormatEnabled(); |
| 385 | // NOTE: names are alphabetically sorted. |
| 386 | rinfo_.push_back({csinfo_.addn, mkl_op_registry::GetMklOpName(csinfo_.addn), |
| 387 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 388 | rinfo_.push_back({csinfo_.add, mkl_op_registry::GetMklOpName(csinfo_.add), |
| 389 | CopyAttrsAll, RewriteIfAtleastOneMklInput, |
| 390 | GetRewriteCause()}); |
| 391 | rinfo_.push_back( |
| 392 | {csinfo_.add_v2, mkl_op_registry::GetMklOpName(csinfo_.add_v2), |
| 393 | CopyAttrsAll, RewriteIfAtleastOneMklInput, GetRewriteCause()}); |
| 394 | rinfo_.push_back({csinfo_.avg_pool, |
| 395 | mkl_op_registry::GetMklOpName(csinfo_.avg_pool), |
| 396 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 397 | rinfo_.push_back({csinfo_.avg_pool_grad, |
| 398 | mkl_op_registry::GetMklOpName(csinfo_.avg_pool_grad), |
| 399 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 400 | rinfo_.push_back({csinfo_.avg_pool3d, |
| 401 | mkl_op_registry::GetMklOpName(csinfo_.avg_pool3d), |
| 402 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 403 | rinfo_.push_back({csinfo_.avg_pool3d_grad, |
| 404 | mkl_op_registry::GetMklOpName(csinfo_.avg_pool3d_grad), |
| 405 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 406 | rinfo_.push_back({csinfo_.batch_matmul, |
| 407 | mkl_op_registry::GetMklOpName(csinfo_.batch_matmul), |
| 408 | CopyAttrsAll, MatMulRewrite, kRewriteForOpNameChange}); |
| 409 | rinfo_.push_back({csinfo_.einsum, |
| 410 | mkl_op_registry::GetMklOpName(csinfo_.einsum), |
| 411 | CopyAttrsAll, MatMulRewrite, kRewriteForOpNameChange}); |
| 412 | rinfo_.push_back({csinfo_.batch_matmul_v2, |
| 413 | mkl_op_registry::GetMklOpName(csinfo_.batch_matmul_v2), |
| 414 | CopyAttrsAll, MatMulRewrite, kRewriteForOpNameChange}); |
| 415 | rinfo_.push_back({csinfo_.concat, |
| 416 | mkl_op_registry::GetMklOpName(csinfo_.concat), |
| 417 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 418 | rinfo_.push_back({csinfo_.concatv2, |
| 419 | mkl_op_registry::GetMklOpName(csinfo_.concatv2), |
| 420 | CopyAttrsAll, ConcatV2Rewrite, GetRewriteCause()}); |
| 421 | rinfo_.push_back( |
| 422 | {csinfo_.conjugate_transpose, |
| 423 | mkl_op_registry::GetMklOpName(csinfo_.conjugate_transpose), |
| 424 | CopyAttrsAll, AlwaysRewrite, kRewriteForOpNameChange}); |
| 425 | rinfo_.push_back( |
| 426 | {csinfo_.conv2d, mkl_op_registry::GetMklOpName(csinfo_.conv2d), |
| 427 | CopyAttrsConvCheckConstFilter, AlwaysRewrite, GetRewriteCause()}); |
| 428 | rinfo_.push_back({csinfo_.conv2d_with_bias, |
| 429 | native_fmt ? csinfo_.mkl_native_conv2d_with_bias |
| 430 | : csinfo_.mkl_conv2d_with_bias, |
| 431 | CopyAttrsConvCheckConstFilter, AlwaysRewrite, |
| 432 | GetRewriteCause()}); |
| 433 | rinfo_.push_back({csinfo_.conv2d_grad_filter, |
| 434 | mkl_op_registry::GetMklOpName(csinfo_.conv2d_grad_filter), |
| 435 | CopyAttrsConv, AlwaysRewrite, GetRewriteCause()}); |
| 436 | rinfo_.push_back({csinfo_.conv2d_grad_filter_with_bias, |
| 437 | native_fmt |
| 438 | ? csinfo_.mkl_native_conv2d_grad_filter_with_bias |
| 439 | : csinfo_.mkl_conv2d_grad_filter_with_bias, |
| 440 | CopyAttrsConv, AlwaysRewrite, GetRewriteCause()}); |
| 441 | rinfo_.push_back({csinfo_.conv2d_grad_input, |
| 442 | mkl_op_registry::GetMklOpName(csinfo_.conv2d_grad_input), |
| 443 | CopyAttrsConv, AlwaysRewrite, GetRewriteCause()}); |
| 444 | rinfo_.push_back( |
| 445 | {csinfo_.conv3d, mkl_op_registry::GetMklOpName(csinfo_.conv3d), |
| 446 | CopyAttrsConvCheckConstFilter, AlwaysRewrite, GetRewriteCause()}); |
| 447 | rinfo_.push_back({csinfo_.conv3d_grad_filter, |
| 448 | mkl_op_registry::GetMklOpName(csinfo_.conv3d_grad_filter), |
| 449 | CopyAttrsConv, AlwaysRewrite, GetRewriteCause()}); |
| 450 | rinfo_.push_back({csinfo_.conv3d_grad_input, |
| 451 | mkl_op_registry::GetMklOpName(csinfo_.conv3d_grad_input), |
| 452 | CopyAttrsConv, AlwaysRewrite, GetRewriteCause()}); |
| 453 | rinfo_.push_back({csinfo_.depthwise_conv2d, |
| 454 | mkl_op_registry::GetMklOpName(csinfo_.depthwise_conv2d), |
| 455 | CopyAttrsConvCheckConstFilter, AlwaysRewrite, |
| 456 | GetRewriteCause()}); |
| 457 | rinfo_.push_back( |
| 458 | {csinfo_.depthwise_conv2d_grad_input, |
| 459 | mkl_op_registry::GetMklOpName(csinfo_.depthwise_conv2d_grad_input), |
| 460 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 461 | rinfo_.push_back( |
| 462 | {csinfo_.depthwise_conv2d_grad_filter, |
| 463 | mkl_op_registry::GetMklOpName(csinfo_.depthwise_conv2d_grad_filter), |
| 464 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 465 | rinfo_.push_back( |
| 466 | {csinfo_.dequantize, mkl_op_registry::GetMklOpName(csinfo_.dequantize), |
| 467 | CopyAttrsAll, DequantizeRewrite, kRewriteForOpNameChange}); |
| 468 | rinfo_.push_back({csinfo_.fused_batch_norm, |
| 469 | mkl_op_registry::GetMklOpName(csinfo_.fused_batch_norm), |
| 470 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 471 | rinfo_.push_back( |
| 472 | {csinfo_.fused_batch_norm_grad, |
| 473 | mkl_op_registry::GetMklOpName(csinfo_.fused_batch_norm_grad), |
| 474 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 475 | rinfo_.push_back( |
| 476 | {csinfo_.fused_batch_norm_v2, |
| 477 | mkl_op_registry::GetMklOpName(csinfo_.fused_batch_norm_v2), |
| 478 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 479 | rinfo_.push_back( |
| 480 | {csinfo_.fused_batch_norm_grad_v2, |
| 481 | mkl_op_registry::GetMklOpName(csinfo_.fused_batch_norm_grad_v2), |
| 482 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 483 | |
| 484 | // Using CopyAttrsAll for V3 on CPU, as there are no additional |
| 485 | // attributes. |
| 486 | rinfo_.push_back( |
| 487 | {csinfo_.fused_batch_norm_v3, |
| 488 | mkl_op_registry::GetMklOpName(csinfo_.fused_batch_norm_v3), |
| 489 | CopyAttrsAll, FusedBatchNormV3Rewrite, GetRewriteCause()}); |
| 490 | rinfo_.push_back( |
| 491 | {csinfo_.fused_batch_norm_grad_v3, |
| 492 | mkl_op_registry::GetMklOpName(csinfo_.fused_batch_norm_grad_v3), |
| 493 | CopyAttrsAll, FusedBatchNormV3Rewrite, GetRewriteCause()}); |
| 494 | rinfo_.push_back({csinfo_.fused_batch_norm_ex, |
| 495 | native_fmt ? csinfo_.mkl_native_fused_batch_norm_ex |
| 496 | : csinfo_.mkl_fused_batch_norm_ex, |
| 497 | CopyAttrsAll, FusedBatchNormExRewrite, |
| 498 | GetRewriteCause()}); |
| 499 | rinfo_.push_back({csinfo_.fused_conv2d, |
| 500 | native_fmt ? csinfo_.mkl_native_fused_conv2d |
| 501 | : csinfo_.mkl_fused_conv2d, |
| 502 | CopyAttrsFusedConv2DCheckConstFilter, FusedConv2DRewrite, |
| 503 | GetRewriteCause()}); |
| 504 | rinfo_.push_back({csinfo_.fused_conv3d, csinfo_.mkl_native_fused_conv3d, |
| 505 | CopyAttrsAllCheckConstFilter, AlwaysRewrite, |
| 506 | kRewriteForOpNameChange}); |
| 507 | rinfo_.push_back({csinfo_.fused_depthwise_conv2d, |
| 508 | native_fmt ? csinfo_.mkl_native_fused_depthwise_conv2d |
| 509 | : csinfo_.mkl_fused_depthwise_conv2d, |
| 510 | CopyAttrsAllCheckConstFilter, FusedDepthwiseConv2DRewrite, |
| 511 | GetRewriteCause()}); |
| 512 | rinfo_.push_back({csinfo_.fused_matmul, |
| 513 | native_fmt ? csinfo_.mkl_native_fused_matmul |
| 514 | : csinfo_.mkl_fused_matmul, |
| 515 | CopyAttrsAllCheckConstFilter, FusedMatMulRewrite, |
| 516 | GetRewriteCause()}); |
| 517 | rinfo_.push_back( |
| 518 | {csinfo_.identity, mkl_op_registry::GetMklOpName(csinfo_.identity), |
| 519 | CopyAttrsAll, RewriteIfAtleastOneMklInput, GetRewriteCause()}); |
| 520 | rinfo_.push_back({csinfo_.lrn, mkl_op_registry::GetMklOpName(csinfo_.lrn), |
| 521 | CopyAttrsAll, LrnRewrite, GetRewriteCause()}); |
| 522 | rinfo_.push_back({csinfo_.lrn_grad, |
| 523 | mkl_op_registry::GetMklOpName(csinfo_.lrn_grad), |
| 524 | CopyAttrsAll, LrnGradRewrite, GetRewriteCause()}); |
| 525 | rinfo_.push_back({csinfo_.matmul, |
| 526 | mkl_op_registry::GetMklOpName(csinfo_.matmul), |
| 527 | CopyAttrsAll, MatMulRewrite, kRewriteForOpNameChange}); |
| 528 | rinfo_.push_back({csinfo_.leakyrelu, |
| 529 | mkl_op_registry::GetMklOpName(csinfo_.leakyrelu), |
| 530 | CopyAttrsAll, LeakyReluRewrite, GetRewriteCause()}); |
| 531 | rinfo_.push_back({csinfo_.leakyrelu_grad, |
| 532 | mkl_op_registry::GetMklOpName(csinfo_.leakyrelu_grad), |
| 533 | CopyAttrsAll, LeakyReluRewrite, GetRewriteCause()}); |
| 534 | rinfo_.push_back( |
| 535 | {csinfo_.max_pool, mkl_op_registry::GetMklOpName(csinfo_.max_pool), |
| 536 | CopyAttrsAll, NonDepthBatchWisePoolRewrite, GetRewriteCause()}); |
| 537 | rinfo_.push_back({csinfo_.max_pool_grad, |
| 538 | mkl_op_registry::GetMklOpName(csinfo_.max_pool_grad), |
| 539 | CopyAttrsAll, MaxpoolGradRewrite, GetRewriteCause()}); |
| 540 | rinfo_.push_back( |
| 541 | {csinfo_.max_pool3d, mkl_op_registry::GetMklOpName(csinfo_.max_pool3d), |
| 542 | CopyAttrsAll, NonDepthBatchWisePoolRewrite, GetRewriteCause()}); |
| 543 | rinfo_.push_back({csinfo_.max_pool3d_grad, |
| 544 | mkl_op_registry::GetMklOpName(csinfo_.max_pool3d_grad), |
| 545 | CopyAttrsAll, Maxpool3DGradRewrite, GetRewriteCause()}); |
| 546 | rinfo_.push_back( |
| 547 | {csinfo_.maximum, mkl_op_registry::GetMklOpName(csinfo_.maximum), |
| 548 | CopyAttrsAll, RewriteIfAtleastOneMklInput, GetRewriteCause()}); |
| 549 | rinfo_.push_back({csinfo_.mul, mkl_op_registry::GetMklOpName(csinfo_.mul), |
| 550 | CopyAttrsAll, RewriteIfAtleastOneMklInput, |
| 551 | GetRewriteCause()}); |
| 552 | rinfo_.push_back({csinfo_.pad_with_conv2d, |
| 553 | native_fmt ? csinfo_.mkl_native_pad_with_conv2d |
| 554 | : csinfo_.mkl_pad_with_conv2d, |
| 555 | CopyAttrsAllCheckConstFilter, AlwaysRewrite, |
| 556 | GetRewriteCause()}); |
| 557 | rinfo_.push_back({csinfo_.pad_with_fused_conv2d, |
| 558 | native_fmt ? csinfo_.mkl_native_pad_with_fused_conv2d |
| 559 | : csinfo_.mkl_pad_with_fused_conv2d, |
| 560 | CopyAttrsAllCheckConstFilter, AlwaysRewrite, |
| 561 | GetRewriteCause()}); |
| 562 | rinfo_.push_back({csinfo_.quantized_avg_pool, |
| 563 | mkl_op_registry::GetMklOpName(csinfo_.quantized_avg_pool), |
| 564 | CopyAttrsAll, AlwaysRewrite, kRewriteForOpNameChange}); |
| 565 | rinfo_.push_back({csinfo_.quantized_concatv2, |
| 566 | mkl_op_registry::GetMklOpName(csinfo_.quantized_concatv2), |
| 567 | CopyAttrsAll, ConcatV2Rewrite, kRewriteForOpNameChange}); |
| 568 | rinfo_.push_back({csinfo_.quantized_conv2d, |
| 569 | mkl_op_registry::GetMklOpName(csinfo_.quantized_conv2d), |
| 570 | CopyAttrsQuantizedConv2D, AlwaysRewrite, |
| 571 | kRewriteForOpNameChange}); |
| 572 | rinfo_.push_back( |
| 573 | {csinfo_.quantized_conv2d_per_channel, |
| 574 | mkl_op_registry::GetMklOpName(csinfo_.quantized_conv2d_per_channel), |
| 575 | CopyAttrsQuantizedConv2D, AlwaysRewrite, kRewriteForOpNameChange}); |
| 576 | rinfo_.push_back({csinfo_.quantized_conv2d_with_requantize, |
| 577 | mkl_op_registry::GetMklOpName( |
| 578 | csinfo_.quantized_conv2d_with_requantize), |
| 579 | CopyAttrsQuantizedConv2D, AlwaysRewrite, |
| 580 | kRewriteForOpNameChange}); |
| 581 | rinfo_.push_back( |
| 582 | {csinfo_.quantized_conv2d_with_bias, |
| 583 | mkl_op_registry::GetMklOpName(csinfo_.quantized_conv2d_with_bias), |
| 584 | CopyAttrsQuantizedConv2D, AlwaysRewrite, kRewriteForOpNameChange}); |
| 585 | rinfo_.push_back({csinfo_.quantized_conv2d_with_bias_and_requantize, |
| 586 | mkl_op_registry::GetMklOpName( |
| 587 | csinfo_.quantized_conv2d_with_bias_and_requantize), |
| 588 | CopyAttrsQuantizedConv2D, AlwaysRewrite, |
| 589 | kRewriteForOpNameChange}); |
| 590 | rinfo_.push_back( |
| 591 | {csinfo_.quantized_conv2d_and_relu, |
| 592 | mkl_op_registry::GetMklOpName(csinfo_.quantized_conv2d_and_relu), |
| 593 | CopyAttrsQuantizedConv2D, AlwaysRewrite, kRewriteForOpNameChange}); |
| 594 | rinfo_.push_back({csinfo_.quantized_conv2d_and_relu_and_requantize, |
| 595 | mkl_op_registry::GetMklOpName( |
| 596 | csinfo_.quantized_conv2d_and_relu_and_requantize), |
| 597 | CopyAttrsQuantizedConv2D, AlwaysRewrite, |
| 598 | kRewriteForOpNameChange}); |
| 599 | rinfo_.push_back({csinfo_.quantized_conv2d_with_bias_and_relu, |
| 600 | mkl_op_registry::GetMklOpName( |
| 601 | csinfo_.quantized_conv2d_with_bias_and_relu), |
| 602 | CopyAttrsQuantizedConv2D, AlwaysRewrite, |
| 603 | kRewriteForOpNameChange}); |
| 604 | rinfo_.push_back( |
| 605 | {csinfo_.quantized_conv2d_with_bias_and_relu_and_requantize, |
| 606 | mkl_op_registry::GetMklOpName( |
| 607 | csinfo_.quantized_conv2d_with_bias_and_relu_and_requantize), |
| 608 | CopyAttrsQuantizedConv2D, AlwaysRewrite, kRewriteForOpNameChange}); |
| 609 | rinfo_.push_back({csinfo_.quantized_max_pool, |
| 610 | mkl_op_registry::GetMklOpName(csinfo_.quantized_max_pool), |
| 611 | CopyAttrsAll, AlwaysRewrite, kRewriteForOpNameChange}); |
| 612 | rinfo_.push_back({csinfo_.quantized_conv2d_with_bias_sum_and_relu, |
| 613 | mkl_op_registry::GetMklOpName( |
| 614 | csinfo_.quantized_conv2d_with_bias_sum_and_relu), |
| 615 | CopyAttrsQuantizedConv2D, AlwaysRewrite, |
| 616 | kRewriteForOpNameChange}); |
| 617 | rinfo_.push_back( |
| 618 | {csinfo_.quantized_conv2d_with_bias_sum_and_relu_and_requantize, |
| 619 | mkl_op_registry::GetMklOpName( |
| 620 | csinfo_.quantized_conv2d_with_bias_sum_and_relu_and_requantize), |
| 621 | CopyAttrsQuantizedConv2D, AlwaysRewrite, kRewriteForOpNameChange}); |
| 622 | rinfo_.push_back( |
| 623 | {csinfo_.quant_conv2d_with_bias_signed_sum_and_relu_and_requantize, |
| 624 | mkl_op_registry::GetMklOpName( |
| 625 | csinfo_.quant_conv2d_with_bias_signed_sum_and_relu_and_requantize), |
| 626 | CopyAttrsQuantizedConv2D, AlwaysRewrite, kRewriteForOpNameChange}); |
| 627 | rinfo_.push_back( |
| 628 | {csinfo_.quantized_matmul_with_bias, |
| 629 | mkl_op_registry::GetMklOpName(csinfo_.quantized_matmul_with_bias), |
| 630 | CopyAttrsQuantizedMatMulWithBias, AlwaysRewrite, |
| 631 | kRewriteForOpNameChange}); |
| 632 | rinfo_.push_back({csinfo_.quantized_matmul_with_bias_and_relu, |
| 633 | mkl_op_registry::GetMklOpName( |
| 634 | csinfo_.quantized_matmul_with_bias_and_relu), |
| 635 | CopyAttrsQuantizedMatMulWithBias, AlwaysRewrite, |
| 636 | kRewriteForOpNameChange}); |
| 637 | rinfo_.push_back( |
| 638 | {csinfo_.quantized_matmul_with_bias_and_relu_and_requantize, |
| 639 | mkl_op_registry::GetMklOpName( |
| 640 | csinfo_.quantized_matmul_with_bias_and_relu_and_requantize), |
| 641 | CopyAttrsQuantizedMatMulWithBias, AlwaysRewrite, |
| 642 | kRewriteForOpNameChange}); |
| 643 | rinfo_.push_back({csinfo_.quantized_matmul_with_bias_and_requantize, |
| 644 | mkl_op_registry::GetMklOpName( |
| 645 | csinfo_.quantized_matmul_with_bias_and_requantize), |
| 646 | CopyAttrsQuantizedMatMulWithBias, AlwaysRewrite, |
| 647 | kRewriteForOpNameChange}); |
| 648 | rinfo_.push_back({csinfo_.quantized_matmul_with_bias_and_dequantize, |
| 649 | mkl_op_registry::GetMklOpName( |
| 650 | csinfo_.quantized_matmul_with_bias_and_dequantize), |
| 651 | CopyAttrsQuantizedMatMulWithBiasAndDequantize, |
| 652 | AlwaysRewrite, kRewriteForOpNameChange}); |
| 653 | rinfo_.push_back( |
| 654 | {csinfo_.quantized_depthwise_conv2d, |
| 655 | mkl_op_registry::GetMklOpName(csinfo_.quantized_depthwise_conv2d), |
| 656 | CopyAttrsQuantizedConv2D, AlwaysRewrite, kRewriteForOpNameChange}); |
| 657 | rinfo_.push_back({csinfo_.quantized_depthwise_conv2d_with_bias, |
| 658 | mkl_op_registry::GetMklOpName( |
| 659 | csinfo_.quantized_depthwise_conv2d_with_bias), |
| 660 | CopyAttrsQuantizedConv2D, AlwaysRewrite, |
| 661 | kRewriteForOpNameChange}); |
| 662 | rinfo_.push_back( |
| 663 | {csinfo_.quantized_depthwise_conv2d_with_bias_and_relu, |
| 664 | mkl_op_registry::GetMklOpName( |
| 665 | csinfo_.quantized_depthwise_conv2d_with_bias_and_relu), |
| 666 | CopyAttrsQuantizedConv2D, AlwaysRewrite, kRewriteForOpNameChange}); |
| 667 | rinfo_.push_back( |
| 668 | {csinfo_.quantized_depthwise_conv2d_with_bias_and_relu_and_requantize, |
| 669 | mkl_op_registry::GetMklOpName( |
| 670 | csinfo_ |
| 671 | .quantized_depthwise_conv2d_with_bias_and_relu_and_requantize), |
| 672 | CopyAttrsQuantizedConv2D, AlwaysRewrite, kRewriteForOpNameChange}); |
| 673 | rinfo_.push_back({csinfo_.quantize_v2, |
| 674 | mkl_op_registry::GetMklOpName(csinfo_.quantize_v2), |
| 675 | CopyAttrsAll, QuantizeOpRewrite, |
| 676 | kRewriteForOpNameChange}); |
| 677 | rinfo_.push_back({csinfo_.relu, mkl_op_registry::GetMklOpName(csinfo_.relu), |
| 678 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 679 | rinfo_.push_back({csinfo_.relu_grad, |
| 680 | mkl_op_registry::GetMklOpName(csinfo_.relu_grad), |
| 681 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 682 | rinfo_.push_back({csinfo_.relu6, |
| 683 | mkl_op_registry::GetMklOpName(csinfo_.relu6), |
| 684 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 685 | rinfo_.push_back({csinfo_.relu6_grad, |
| 686 | mkl_op_registry::GetMklOpName(csinfo_.relu6_grad), |
| 687 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 688 | rinfo_.push_back({csinfo_.requantize, |
| 689 | mkl_op_registry::GetMklOpName(csinfo_.requantize), |
| 690 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 691 | // Optimized TanhGrad support exists only in DNNL 1.x. |
| 692 | rinfo_.push_back({csinfo_.tanh, mkl_op_registry::GetMklOpName(csinfo_.tanh), |
| 693 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 694 | rinfo_.push_back({csinfo_.tanh_grad, |
| 695 | mkl_op_registry::GetMklOpName(csinfo_.tanh_grad), |
| 696 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 697 | rinfo_.push_back({csinfo_.reshape, |
| 698 | mkl_op_registry::GetMklOpName(csinfo_.reshape), |
| 699 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 700 | rinfo_.push_back( |
| 701 | {csinfo_.slice, mkl_op_registry::GetMklOpName(csinfo_.slice), |
| 702 | CopyAttrsAll, RewriteIfAtleastOneMklInput, GetRewriteCause()}); |
| 703 | rinfo_.push_back({csinfo_.softmax, |
| 704 | mkl_op_registry::GetMklOpName(csinfo_.softmax), |
| 705 | CopyAttrsAll, AlwaysRewrite, GetRewriteCause()}); |
| 706 | |
| 707 | rinfo_.push_back({csinfo_.squared_difference, |
| 708 | mkl_op_registry::GetMklOpName(csinfo_.squared_difference), |
| 709 | CopyAttrsAll, RewriteIfAtleastOneMklInput, |
| 710 | GetRewriteCause()}); |
| 711 | rinfo_.push_back({csinfo_.sub, mkl_op_registry::GetMklOpName(csinfo_.sub), |
| 712 | CopyAttrsAll, RewriteIfAtleastOneMklInput, |
| 713 | GetRewriteCause()}); |
| 714 | rinfo_.push_back({csinfo_.transpose, |
| 715 | mkl_op_registry::GetMklOpName(csinfo_.transpose), |
| 716 | CopyAttrsAll, AlwaysRewrite, kRewriteForOpNameChange}); |
| 717 | |
| 718 | // Add info about which ops to add workspace edge to and the slots. |
| 719 | wsinfo_.push_back({csinfo_.lrn, csinfo_.lrn_grad, 0, 2, 1, 3}); |
| 720 | wsinfo_.push_back({csinfo_.max_pool, csinfo_.max_pool_grad, 0, 1, 1, 3}); |
| 721 | wsinfo_.push_back( |
| 722 | {csinfo_.max_pool3d, csinfo_.max_pool3d_grad, 0, 1, 1, 3}); |
| 723 | |
| 724 | // Add a rule for merging nodes |
| 725 | minfo_.push_back({csinfo_.conv2d, csinfo_.bias_add, |
| 726 | csinfo_.conv2d_with_bias, GetConv2DOrBiasAdd}); |
| 727 | |
| 728 | // Merge Pad and Conv2d, only if the pad op is "Pad" |
| 729 | // Doesn't merge if pad op is "PadV2" or "MirrorPad" |
| 730 | minfo_.push_back( |
| 731 | {csinfo_.pad, csinfo_.conv2d, csinfo_.pad_with_conv2d, GetPadOrConv2D}); |
| 732 | |
| 733 | minfo_.push_back({csinfo_.pad, csinfo_.fused_conv2d, |
| 734 | csinfo_.pad_with_fused_conv2d, GetPadOrFusedConv2D}); |
| 735 | |
| 736 | minfo_.push_back({csinfo_.conv2d_grad_filter, csinfo_.bias_add_grad, |
| 737 | csinfo_.conv2d_grad_filter_with_bias, |
| 738 | GetConv2DBackpropFilterOrBiasAddGrad}); |
| 739 | |
| 740 | // The fusion patterns in "finfo_" that show up first will get applied |
| 741 | // first, for example, graph "A->B->C-D" and finfo_ is {A->B->C to ABC, |
| 742 | // A->B->C->D to ABCD}, since the first gets applied first, the final |
| 743 | // graph will be ABC->D. |
| 744 | } |
| 745 | |
| 746 | // Standard interface to run pass |
| 747 | Status Run(const GraphOptimizationPassOptions& options); |
| 748 | |
| 749 | // Helper function which does most of heavy lifting for rewriting |
| 750 | // Mkl nodes to propagate Mkl tensor as additional output |
| 751 | // |
| 752 | // Extracts common functionality between Run public interface and |
| 753 | // test interface. |
| 754 | // |
| 755 | // @return true, if and only if graph is mutated; false otherwise. |
| 756 | bool RunPass(std::unique_ptr<Graph>* g); |
| 757 | |
| 758 | /// Cause for rewrite |
| 759 | /// Currently, we only support 2 causes - either for Mkl layout propagation |
| 760 | /// which is the most common case, or for just a name change (used in case |
| 761 | /// of ops like MatMul, Transpose, which do not support Mkl layout) |
| 762 | enum RewriteCause { kRewriteForLayoutPropagation, kRewriteForOpNameChange }; |
| 763 | |
| 764 | // Get the op rewrite cause depending on whether native format mode |
| 765 | // is enabled or not. |
| 766 | RewriteCause GetRewriteCause() { |
| 767 | if (NativeFormatEnabled()) { |
| 768 | return kRewriteForOpNameChange; |
| 769 | } else { |
| 770 | return kRewriteForLayoutPropagation; |
| 771 | } |
| 772 | } |
| 773 | |
| 774 | /// Structure to specify the name of an original node, its new name after |
| 775 | /// rewrite, the number of inputs to the original node, the function to |
| 776 | /// be used to copy attributes for the op, and the rule (if any) which |
| 777 | /// must hold for rewriting the node |
| 778 | typedef struct { |
| 779 | string name; // Original name of op of the node in the graph |
| 780 | string new_name; // New name of the op of the node in the graph |
| 781 | // A function handler to copy attributes from an old node to a new node. |
| 782 | std::function<void(const Node*, NodeBuilder*, bool)> copy_attrs; |
| 783 | // A rule under which to rewrite this node |
| 784 | std::function<bool(const Node*)> rewrite_rule; |
| 785 | // Why are we rewriting? |
| 786 | RewriteCause rewrite_cause; |
| 787 | } RewriteInfo; |
| 788 | |
| 789 | /// Structure to specify a forward op, a backward op, and the slot numbers |
| 790 | /// in the forward and backward ops where we will add a workspace edge. |
| 791 | typedef struct { |
| 792 | string fwd_op; // Name of a forward op in the graph |
| 793 | string bwd_op; // Name of a backward op in the graph |
| 794 | int fwd_slot; // Output slot in the forward op node where actual |
| 795 | // output tensor resides |
| 796 | int bwd_slot; // Input slot in the backward op node where actual |
| 797 | // input tensor resides |
| 798 | int ws_fwd_slot; // Output slot in the forward op node where workspace |
| 799 | // edge is added |
| 800 | int ws_bwd_slot; // Input slot in the backward op node where workspace |
| 801 | // edge is added |
| 802 | } WorkSpaceInfo; |
| 803 | |
| 804 | /// Structure to specify information used in node merge of 2 operators |
| 805 | typedef struct { |
| 806 | string op1; // Node string for one operator. |
| 807 | string op2; // Node string for second operator. |
| 808 | string new_node; // Name of the node after merge |
| 809 | // Function that enables user of the node merger to specify how to find |
| 810 | // second operator given the first operator. |
| 811 | std::function<Node*(const Node*)> get_node_to_be_merged; |
| 812 | } MergeInfo; |
| 813 | |
| 814 | // Structure to specify information used in node fusion of 3+ operators |
| 815 | typedef struct { |
| 816 | std::string pattern_name; // Name to describe this pattern, such as |
| 817 | // "Transpose_Mklop_Transpose". |
| 818 | std::vector<std::function<bool(const Node*)> > |
| 819 | node_checkers; // Extra restriction checker for these ops |
| 820 | std::function<Status( |
| 821 | std::unique_ptr<Graph>*, std::vector<Node*>&, |
| 822 | std::function<void(const Node*, NodeBuilder* nb, bool)>)> |
| 823 | fuse_func; |
| 824 | std::function<void(const Node*, NodeBuilder* nb, bool)> copy_attrs; |
| 825 | } FusionInfo; |
| 826 | |
| 827 | // |
| 828 | // Dimension indices for 2D tensor. |
| 829 | // |
| 830 | struct NCHW { |
| 831 | enum dim { N = 0, C = 1, H = 2, W = 3 }; |
| 832 | }; |
| 833 | |
| 834 | struct NHWC { |
| 835 | enum dim { N = 0, H = 1, W = 2, C = 3 }; |
| 836 | }; |
| 837 | |
| 838 | // |
| 839 | // dimension indices for 3D tensor. |
| 840 | // |
| 841 | struct NCDHW { |
| 842 | enum dim { N = 0, C = 1, D = 2, H = 3, W = 4 }; |
| 843 | }; |
| 844 | |
| 845 | struct NDHWC { |
| 846 | enum dim { N = 0, D = 1, H = 2, W = 3, C = 4 }; |
| 847 | }; |
| 848 | |
| 849 | /// Structure to store all constant strings |
| 850 | /// NOTE: names are alphabetically sorted. |
| 851 | typedef struct { |
| 852 | string addn; |
| 853 | string add; |
| 854 | string add_v2; |
| 855 | string avg_pool; |
| 856 | string avg_pool_grad; |
| 857 | string avg_pool3d; |
| 858 | string avg_pool3d_grad; |
| 859 | string batch_matmul; |
| 860 | string batch_matmul_v2; |
| 861 | string bias_add; |
| 862 | string bias_add_grad; |
| 863 | string concat; |
| 864 | string concatv2; |
| 865 | string conjugate_transpose; |
| 866 | string conv2d; |
| 867 | string conv2d_with_bias; |
| 868 | string conv2d_grad_input; |
| 869 | string conv2d_grad_filter; |
| 870 | string conv2d_grad_filter_with_bias; |
| 871 | string conv3d; |
| 872 | string conv3d_grad_input; |
| 873 | string conv3d_grad_filter; |
| 874 | string depthwise_conv2d; |
| 875 | string depthwise_conv2d_grad_input; |
| 876 | string depthwise_conv2d_grad_filter; |
| 877 | string dequantize; |
| 878 | string einsum; |
| 879 | string fused_batch_norm; |
| 880 | string fused_batch_norm_grad; |
| 881 | string fused_batch_norm_ex; |
| 882 | string fused_batch_norm_v2; |
| 883 | string fused_batch_norm_grad_v2; |
| 884 | string fused_batch_norm_v3; |
| 885 | string fused_batch_norm_grad_v3; |
| 886 | string fused_conv2d; |
| 887 | string fused_conv3d; |
| 888 | string fused_depthwise_conv2d; |
| 889 | string fused_matmul; |
| 890 | string identity; |
| 891 | string leakyrelu; |
| 892 | string leakyrelu_grad; |
| 893 | string lrn; |
| 894 | string lrn_grad; |
| 895 | string matmul; |
| 896 | string max_pool; |
| 897 | string max_pool_grad; |
| 898 | string max_pool3d; |
| 899 | string max_pool3d_grad; |
| 900 | string maximum; |
| 901 | string mkl_conv2d; |
| 902 | string mkl_conv2d_grad_input; |
| 903 | string mkl_conv2d_grad_filter; |
| 904 | string mkl_conv2d_grad_filter_with_bias; |
| 905 | string mkl_conv2d_with_bias; |
| 906 | string mkl_depthwise_conv2d_grad_input; |
| 907 | string mkl_depthwise_conv2d_grad_filter; |
| 908 | string mkl_fused_batch_norm_ex; |
| 909 | string mkl_fused_conv2d; |
| 910 | string mkl_fused_depthwise_conv2d; |
| 911 | string mkl_fused_matmul; |
| 912 | string mkl_native_conv2d_with_bias; |
| 913 | string mkl_native_conv2d_grad_filter_with_bias; |
| 914 | string mkl_native_fused_batch_norm_ex; |
| 915 | string mkl_native_fused_conv2d; |
| 916 | string mkl_native_fused_conv3d; |
| 917 | string mkl_native_fused_depthwise_conv2d; |
| 918 | string mkl_native_fused_matmul; |
| 919 | string mkl_native_pad_with_conv2d; |
| 920 | string mkl_native_pad_with_fused_conv2d; |
| 921 | string mkl_pad_with_conv2d; |
| 922 | string mkl_pad_with_fused_conv2d; |
| 923 | string mul; |
| 924 | string pad; |
| 925 | string pad_with_conv2d; |
| 926 | string pad_with_fused_conv2d; |
| 927 | string quantized_avg_pool; |
| 928 | string quantized_conv2d; |
| 929 | string quantized_conv2d_per_channel; |
| 930 | string quantized_conv2d_with_requantize; |
| 931 | string quantized_conv2d_with_bias; |
| 932 | string quantized_conv2d_with_bias_and_requantize; |
| 933 | string quantized_conv2d_and_relu; |
| 934 | string quantized_conv2d_and_relu_and_requantize; |
| 935 | string quantized_conv2d_with_bias_and_relu; |
| 936 | string quantized_conv2d_with_bias_and_relu_and_requantize; |
| 937 | string quantized_concatv2; |
| 938 | string quantized_max_pool; |
| 939 | string quantized_conv2d_with_bias_sum_and_relu; |
| 940 | string quantized_conv2d_with_bias_sum_and_relu_and_requantize; |
| 941 | string quant_conv2d_with_bias_signed_sum_and_relu_and_requantize; |
| 942 | string quantized_matmul_with_bias; |
| 943 | string quantized_matmul_with_bias_and_relu; |
| 944 | string quantized_matmul_with_bias_and_relu_and_requantize; |
| 945 | string quantized_matmul_with_bias_and_requantize; |
| 946 | string quantized_matmul_with_bias_and_dequantize; |
| 947 | string quantized_depthwise_conv2d; |
| 948 | string quantized_depthwise_conv2d_with_bias; |
| 949 | string quantized_depthwise_conv2d_with_bias_and_relu; |
| 950 | string quantized_depthwise_conv2d_with_bias_and_relu_and_requantize; |
| 951 | string quantize_v2; |
| 952 | string relu; |
| 953 | string relu_grad; |
| 954 | string relu6; |
| 955 | string relu6_grad; |
| 956 | string requantize; |
| 957 | string tanh; |
| 958 | string tanh_grad; |
| 959 | string transpose; |
| 960 | string reshape; |
| 961 | string slice; |
| 962 | string softmax; |
| 963 | string split; |
| 964 | string squared_difference; |
| 965 | string sub; |
| 966 | } ConstStringsInfo; |
| 967 | |
| 968 | private: |
| 969 | /// Maintain info about nodes to rewrite |
| 970 | std::vector<RewriteInfo> rinfo_; |
| 971 | |
| 972 | /// Maintain info about nodes to add workspace edge |
| 973 | std::vector<WorkSpaceInfo> wsinfo_; |
| 974 | |
| 975 | /// Maintain info about nodes to be merged |
| 976 | std::vector<MergeInfo> minfo_; |
| 977 | |
| 978 | /// Maintain info about nodes to be fused |
| 979 | std::vector<FusionInfo> finfo_; |
| 980 | |
| 981 | /// Maintain structure of constant strings |
| 982 | static ConstStringsInfo csinfo_; |
| 983 | |
| 984 | private: |
| 985 | // Is OpDef::ArgDef a list type? It could be N * T or list(type). |
| 986 | // Refer to opdef.proto for details of list type. |
| 987 | inline bool ArgIsList(const OpDef::ArgDef& arg) const { |
| 988 | return !arg.type_list_attr().empty() || !arg.number_attr().empty(); |
| 989 | } |
| 990 | |
| 991 | // Get length of a list in 'n' if 'arg' is of list type. Refer to |
| 992 | // description of ArgIsList for definition of list type. |
| 993 | inline int GetTensorListLength(const OpDef::ArgDef& arg, const Node* n) { |
| 994 | CHECK_EQ(ArgIsList(arg), true); |
| 995 | int N = 0; |
| 996 | const string attr_name = !arg.type_list_attr().empty() |
| 997 | ? arg.type_list_attr() |
| 998 | : arg.number_attr(); |
| 999 | if (!arg.type_list_attr().empty()) { |
| 1000 | std::vector<DataType> value; |
| 1001 | TF_CHECK_OK(GetNodeAttr(n->def(), attr_name, &value)); |
| 1002 | N = value.size(); |
| 1003 | } else { |
| 1004 | TF_CHECK_OK(GetNodeAttr(n->def(), attr_name, &N)); |
| 1005 | } |
| 1006 | return N; |
| 1007 | } |
| 1008 | |
| 1009 | // Can op represented by node 'n' run on DEVICE_CPU? |
| 1010 | // Op can run on CPU with MKL if the runtime assigned device or the |
| 1011 | // user requested device contains device CPU, or both are empty. |
| 1012 | bool CanOpRunOnCPUDevice(const Node* n) { |
| 1013 | bool result = true; |
| 1014 | string reason; |
| 1015 | |
| 1016 | // Substring that should be checked for in device name for CPU device. |
| 1017 | const char* const kCPUDeviceSubStr = "CPU"; |
| 1018 | const char* const kXLACPUDeviceSubStr = "XLA_CPU"; |
| 1019 | |
| 1020 | // If Op has been specifically assigned to a non-CPU or XLA_CPU device, then |
| 1021 | // No. |
| 1022 | if (!n->assigned_device_name().empty() && |
| 1023 | (!absl::StrContains(n->assigned_device_name(), kCPUDeviceSubStr) || |
| 1024 | absl::StrContains(n->assigned_device_name(), kXLACPUDeviceSubStr))) { |
| 1025 | result = false; |
| 1026 | reason = "Op has been assigned a runtime device that is not CPU."; |
| 1027 | } |
| 1028 | |
| 1029 | // If user has specifically assigned this op to a non-CPU or XLA_CPU device, |
| 1030 | // then No. |
| 1031 | if (!n->def().device().empty() && |
| 1032 | (!absl::StrContains(n->def().device(), kCPUDeviceSubStr) || |
| 1033 | absl::StrContains(n->def().device(), kXLACPUDeviceSubStr))) { |
| 1034 | result = false; |
| 1035 | reason = "User has assigned a device that is not CPU."; |
| 1036 | } |
| 1037 | |
| 1038 | if (result == false) { |
| 1039 | VLOG(1) << "MklLayoutRewritePass: Skipping rewriting of the node " |
| 1040 | << n->type_string() << ", reason: "<< reason; |
| 1041 | } |
| 1042 | |
| 1043 | // Otherwise Yes. |
| 1044 | return result; |
| 1045 | } |
| 1046 | |
| 1047 | // Return a node that can be merged with input node 'n' |
| 1048 | // |
| 1049 | // @return pointer to the node if we can find such a |
| 1050 | // node. Otherwise, it returns nullptr. |
| 1051 | Node* CheckForNodeMerge(const Node* n) const; |
| 1052 | |
| 1053 | // Merge node 'm' with node 'n'. |
| 1054 | // Currently, we merge (1) Conv2D with BiasAdd, and (2) BiasAddGrad with |
| 1055 | // Conv2DBackpropFilter. |
| 1056 | // |
| 1057 | // Input nodes m and n may be deleted if the call to |
| 1058 | // this function is successful. Attempt to use the pointers |
| 1059 | // after the call to function may result in undefined behaviors. |
| 1060 | // |
| 1061 | // @input g - input graph, m - graph node, n - graph node to be merged with m |
| 1062 | // @return OkStatus(), if merging is successful and supported. |
| 1063 | // Returns appropriate Status error code otherwise. |
| 1064 | // Graph is updated in case nodes are merged. Otherwise, it is |
| 1065 | // not updated. |
| 1066 | Status MergeNode(std::unique_ptr<Graph>* g, Node* m, Node* n); |
| 1067 | |
| 1068 | // Helper function to merge different nodes |
| 1069 | Status MergeConv2DWithBiasAdd(std::unique_ptr<Graph>* g, Node* m, Node* n); |
| 1070 | Status MergePadWithConv2D(std::unique_ptr<Graph>* g, Node* m, Node* n); |
| 1071 | Status MergeConv2DBackpropFilterWithBiasAddGrad(std::unique_ptr<Graph>* g, |
| 1072 | Node* m, Node* n); |
| 1073 | |
| 1074 | // Find BiasAdd or Conv2D node that can be merged with input node 'm'. |
| 1075 | // If input 'm' is BiasAdd, then check if there exists Conv2D node that can be |
| 1076 | // merged with 'm'. If input 'm' is Conv2D, then check if there exists BiasAdd |
| 1077 | // node that can be merged with 'm'. |
| 1078 | static Node* GetConv2DOrBiasAdd(const Node* m) { |
| 1079 | DCHECK(m); |
| 1080 | Node* n = nullptr; |
| 1081 | |
| 1082 | DataType T_m; |
| 1083 | TF_CHECK_OK(GetNodeAttr(m->def(), "T", &T_m)); |
| 1084 | |
| 1085 | // Don't try to merge if datatype is not DT_FLOAT or DT_BFLOAT16 |
| 1086 | if (T_m != DT_FLOAT && T_m != DT_BFLOAT16) return n; |
| 1087 | |
| 1088 | if (m->type_string() == csinfo_.bias_add) { |
| 1089 | // If a is BiasAdd, then Conv2D is 0th input of BiasAdd. |
| 1090 | TF_CHECK_OK(m->input_node(0, &n)); |
| 1091 | } else { |
| 1092 | CHECK_EQ(m->type_string(), csinfo_.conv2d); |
| 1093 | // Go over all output edges and search for BiasAdd Node. |
| 1094 | // 0th input of BiasAdd is Conv2D. |
| 1095 | for (const Edge* e : m->out_edges()) { |
| 1096 | if (!e->IsControlEdge() && |
| 1097 | e->dst()->type_string() == csinfo_.bias_add && |
| 1098 | e->dst_input() == 0) { |
| 1099 | n = e->dst(); |
| 1100 | break; |
| 1101 | } |
| 1102 | } |
| 1103 | } |
| 1104 | |
| 1105 | if (n == nullptr) { |
| 1106 | VLOG(1) << "MklLayoutRewritePass: Could not find matching " |
| 1107 | << "Conv2D and BiasAdd node for merging. Input node: " |
| 1108 | << m->DebugString(); |
| 1109 | } |
| 1110 | |
| 1111 | return n; |
| 1112 | } |
| 1113 | |
| 1114 | // Find Pad or Conv2D node that can be merged with input node 'm'. |
| 1115 | // If input 'm' is Pad, then check if there exists Conv2D node that can be |
| 1116 | // merged with 'm'. If input 'm' is Conv2D, then check if there exists Pad |
| 1117 | // node that can be merged with 'm'. |
| 1118 | static Node* GetPadOrConv2D(const Node* m) { |
| 1119 | DCHECK(m); |
| 1120 | Node* n = nullptr; |
| 1121 | |
| 1122 | DataType T_m; |
| 1123 | TF_CHECK_OK(GetNodeAttr(m->def(), "T", &T_m)); |
| 1124 | |
| 1125 | // Don't try to merge if datatype is not DT_FLOAT or DT_BFLOAT16 |
| 1126 | if (T_m != DT_FLOAT && T_m != DT_BFLOAT16) return n; |
| 1127 | |
| 1128 | const Node* conv_node; |
| 1129 | if (m->type_string() == csinfo_.pad) { |
| 1130 | // If m is Pad, then Conv2D is the output of Pad. |
| 1131 | for (const Edge* e : m->out_edges()) { |
| 1132 | if (!e->IsControlEdge() && e->dst()->type_string() == csinfo_.conv2d) { |
| 1133 | n = e->dst(); |
| 1134 | conv_node = n; |
| 1135 | break; |
| 1136 | } |
| 1137 | } |
| 1138 | } else { |
| 1139 | DCHECK_EQ(m->type_string(), csinfo_.conv2d); |
| 1140 | // If m is conv2D, Go over all input edges |
| 1141 | // and search for Pad Node. |
| 1142 | for (const Edge* e : m->in_edges()) { |
| 1143 | if (!e->IsControlEdge() && e->src()->type_string() == csinfo_.pad) { |
| 1144 | n = e->src(); |
| 1145 | conv_node = m; |
| 1146 | break; |
| 1147 | } |
| 1148 | } |
| 1149 | } |
| 1150 | // Check if only VALID type of padding is used |
| 1151 | // or not. |
| 1152 | if (n != nullptr) { |
| 1153 | string padding; |
| 1154 | TF_CHECK_OK(GetNodeAttr(conv_node->def(), "padding", &padding)); |
| 1155 | if (padding != "VALID") |
| 1156 | // Then do not merge. |
| 1157 | // Only VALID type of padding in conv op can be |
| 1158 | // merged with Pad op. |
| 1159 | n = nullptr; |
| 1160 | } else { |
| 1161 | VLOG(1) << "MklLayoutRewritePass: Could not find matching " |
| 1162 | << "Pad and Conv2D node for merging. Input node: " |
| 1163 | << m->DebugString(); |
| 1164 | } |
| 1165 | |
| 1166 | return n; |
| 1167 | } |
| 1168 | |
| 1169 | // Find Pad or _FusedConv2D node that can be merged with input node 'm'. |
| 1170 | // If input 'm' is Pad, then check if there exists _FusedConv2D node that can |
| 1171 | // be merged with 'm'. If input 'm' is _FusedConv2D, then check if there |
| 1172 | // exists Pad node that can be merged with 'm'. |
| 1173 | static Node* GetPadOrFusedConv2D(const Node* m) { |
| 1174 | DCHECK(m); |
| 1175 | Node* n = nullptr; |
| 1176 | |
| 1177 | const Node* conv_node; |
| 1178 | if (m->type_string() == csinfo_.pad) { |
| 1179 | // If m is Pad, then _FusedConv2D is the output of Pad. |
| 1180 | for (const Edge* e : m->out_edges()) { |
| 1181 | if (!e->IsControlEdge() && |
| 1182 | e->dst()->type_string() == csinfo_.fused_conv2d) { |
| 1183 | n = e->dst(); |
| 1184 | conv_node = n; |
| 1185 | break; |
| 1186 | } |
| 1187 | } |
| 1188 | } else { |
| 1189 | DCHECK_EQ(m->type_string(), csinfo_.fused_conv2d); |
| 1190 | // If m is _FusedConv2D, Go over all input edges |
| 1191 | // and search for Pad node. |
| 1192 | for (const Edge* e : m->in_edges()) { |
| 1193 | if (!e->IsControlEdge() && e->src()->type_string() == csinfo_.pad) { |
| 1194 | n = e->src(); |
| 1195 | conv_node = m; |
| 1196 | break; |
| 1197 | } |
| 1198 | } |
| 1199 | } |
| 1200 | // Check if only VALID type of padding is used or not. |
| 1201 | if (n != nullptr) { |
| 1202 | string padding; |
| 1203 | string data_format; |
| 1204 | string filter_format; |
| 1205 | int num_host_args = 0; |
| 1206 | TF_CHECK_OK(GetNodeAttr(conv_node->def(), "padding", &padding)); |
| 1207 | TF_CHECK_OK(GetNodeAttr(conv_node->def(), "data_format", &data_format)); |
| 1208 | TF_CHECK_OK( |
| 1209 | GetNodeAttr(conv_node->def(), "filter_format", &filter_format)); |
| 1210 | TF_CHECK_OK( |
| 1211 | GetNodeAttr(conv_node->def(), "num_host_args", &num_host_args)); |
| 1212 | |
| 1213 | if ((data_format != "NCHW"&& data_format != "NHWC") || |
| 1214 | (filter_format != "HWIO"&& filter_format != "OIHW")) { |
| 1215 | n = nullptr; |
| 1216 | VLOG(1) << "MklLayoutRewritePass: Could match Pad and _FusedConv2D " |
| 1217 | << "nodes but cannot merge them. Only conv ops with NCHW or " |
| 1218 | << "NHWC data format and HWIO or OIHW filter format can be " |
| 1219 | << "merged with Pad op Input node: "<< m->DebugString(); |
| 1220 | } else if (padding != "VALID") { |
| 1221 | // Then do not merge. |
| 1222 | n = nullptr; |
| 1223 | VLOG(1) << "MklLayoutRewritePass: Could match Pad and _FusedConv2D " |
| 1224 | << "nodes but cannot merge them. Only conv ops with padding " |
| 1225 | << "type VALID can be merged with Pad op Input node: " |
| 1226 | << m->DebugString(); |
| 1227 | } else if (num_host_args != 0) { |
| 1228 | n = nullptr; |
| 1229 | VLOG(1) << "MklLayoutRewritePass: Could match Pad and _FusedConv2D " |
| 1230 | << "nodes but cannot merge them. Only conv ops without host " |
| 1231 | << "args can be merged with Pad op Input node: " |
| 1232 | << m->DebugString(); |
| 1233 | } |
| 1234 | } else { |
| 1235 | VLOG(1) << "MklLayoutRewritePass: Could not find matching " |
| 1236 | << "Pad and _FusedConv2D node for merging. Input node: " |
| 1237 | << m->DebugString(); |
| 1238 | } |
| 1239 | |
| 1240 | return n; |
| 1241 | } |
| 1242 | |
| 1243 | // Find Conv2DBackpropFilter or BiasAddGrad node that can be merged with input |
| 1244 | // node 'm'. If input 'm' is Conv2DBackpropFilter, then check if there exists |
| 1245 | // BiasAddGrad node that can be merged with 'm'. If input 'm' is BiasAddGrad, |
| 1246 | // then check if there exists Conv2DBackpropFilter node that can be merged |
| 1247 | // with 'm'. |
| 1248 | // |
| 1249 | // Graph that will allow us to connect Conv2DBackpropFilter with BiasAddGrad |
| 1250 | // would look like: |
| 1251 | // |
| 1252 | // _ = Conv2DBackpropFilter(F, _, G) |
| 1253 | // _ = BiasAddGrad(G) |
| 1254 | // |
| 1255 | // So 1st input of BiasAddGrad connects with 3rd input of |
| 1256 | // Conv2DBackpropFilter and vice versa. |
| 1257 | static Node* GetConv2DBackpropFilterOrBiasAddGrad(const Node* m) { |
| 1258 | DCHECK(m); |
| 1259 | Node* n = nullptr; |
| 1260 | const Node* conv2d_backprop_filter = nullptr; |
| 1261 | |
| 1262 | DataType T_m; |
| 1263 | TF_CHECK_OK(GetNodeAttr(m->def(), "T", &T_m)); |
| 1264 | |
| 1265 | // Don't try to merge if datatype is not DT_FLOAT or DT_BFLOAT16 |
| 1266 | if (T_m != DT_FLOAT && T_m != DT_BFLOAT16) return n; |
| 1267 | |
| 1268 | if (m->type_string() == csinfo_.bias_add_grad) { |
| 1269 | // Get 1st input 'g' of BiasAddGrad. |
| 1270 | Node* g = nullptr; |
| 1271 | TF_CHECK_OK(m->input_node(0, &g)); |
| 1272 | // Now traverse all outgoing edges from g that have destination node as |
| 1273 | // Conv2DBackpropFilter. |
| 1274 | for (const Edge* e : g->out_edges()) { |
| 1275 | if (!e->IsControlEdge() && |
| 1276 | e->dst()->type_string() == csinfo_.conv2d_grad_filter && |
| 1277 | e->dst_input() == 2 /* 3rd input of BackpropFilter */) { |
| 1278 | n = e->dst(); |
| 1279 | conv2d_backprop_filter = n; |
| 1280 | break; |
| 1281 | } |
| 1282 | } |
| 1283 | } else { |
| 1284 | conv2d_backprop_filter = m; |
| 1285 | CHECK_EQ(m->type_string(), csinfo_.conv2d_grad_filter); |
| 1286 | // Get 3rd input 'g' of Conv2DBackpropFilter. |
| 1287 | Node* g = nullptr; |
| 1288 | TF_CHECK_OK(m->input_node(2, &g)); |
| 1289 | // Now traverse all outgoing edges from g that have destination node as |
| 1290 | // BiasAddGrad. |
| 1291 | for (const Edge* e : g->out_edges()) { |
| 1292 | if (!e->IsControlEdge() && |
| 1293 | e->dst()->type_string() == csinfo_.bias_add_grad && |
| 1294 | e->dst_input() == 0 /* 1st input of BiasAddGrad */) { |
| 1295 | n = e->dst(); |
| 1296 | break; |
| 1297 | } |
| 1298 | } |
| 1299 | } |
| 1300 | |
| 1301 | // Do not merge if padding type is EXPLICIT. |
| 1302 | // TODO(intel): Support `EXPLICIT` padding for MklConv2DBackpropFilter. |
| 1303 | if (conv2d_backprop_filter != nullptr) { |
| 1304 | string padding; |
| 1305 | TF_CHECK_OK( |
| 1306 | GetNodeAttr(conv2d_backprop_filter->def(), "padding", &padding)); |
| 1307 | if (padding == "EXPLICIT") { |
| 1308 | // Then do not merge. |
| 1309 | VLOG(1) << "MklLayoutRewritePass: Could match Conv2DBackpropFilter " |
| 1310 | << "and BiasAddGrad nodes but cannot merge them. " |
| 1311 | << "EXPLICIT padding is not supported now. " |
| 1312 | << conv2d_backprop_filter->DebugString(); |
| 1313 | return nullptr; |
| 1314 | } |
| 1315 | } |
| 1316 | |
| 1317 | if (n == nullptr) { |
| 1318 | VLOG(1) << "MklLayoutRewritePass: Could not find matching " |
| 1319 | << "Conv2DBackpropFilter and BiasAddGrad node for merging. " |
| 1320 | << "Input node: "<< m->DebugString(); |
| 1321 | } |
| 1322 | return n; |
| 1323 | } |
| 1324 | |
| 1325 | // Return a node that can be fused with input node 'n' |
| 1326 | // |
| 1327 | // @return tuple. If we can find such nodes, the first |
| 1328 | // element of the tuple is a true. Otherwise, it's false. |
| 1329 | std::tuple<bool, std::vector<Node*>, const MklLayoutRewritePass::FusionInfo> |
| 1330 | CheckForNodeFusion(Node* n) const; |
| 1331 | |
| 1332 | // Fuse nodes in the vector "nodes" |
| 1333 | Status FuseNode(std::unique_ptr<Graph>* g, std::vector<Node*>& nodes, |
| 1334 | const MklLayoutRewritePass::FusionInfo fi); |
| 1335 | |
| 1336 | // Fuse transpose(to "NHWC") + mklop("NHWC") + transpose(to "NCHW") into |
| 1337 | // mklop("NCHW"). |
| 1338 | // Here "mklop" can be any MKL-DNN supported op, such as Conv2D. |
| 1339 | static Status FuseTransposeMklOpTranspose( |
| 1340 | std::unique_ptr<Graph>* g, std::vector<Node*>& nodes, |
| 1341 | std::function<void(const Node*, NodeBuilder* nb, bool)> copy_attrs, |
| 1342 | string data_format); |
| 1343 | |
| 1344 | static bool CheckForTranspose(const Node* node, std::vector<int> perm) { |
| 1345 | // Check if node's type is "Transpose" |
| 1346 | if (node->type_string() != "Transpose") return false; |
| 1347 | |
| 1348 | // If "Transpose" has multiple output data edges, also don't fuse it. |
| 1349 | if (node->num_outputs() > 1 || node->out_edges().size() > 1) return false; |
| 1350 | |
| 1351 | // Check if has out control edge. If true, this is a training graph. |
| 1352 | // Currently we focus on inference and do no fusion in training. |
| 1353 | // Note: this constraint will eventually be removed, if we enabled this |
| 1354 | // fusion for training |
| 1355 | // in the future. |
| 1356 | for (const Edge* e : node->out_edges()) { |
| 1357 | if (e->IsControlEdge()) { |
| 1358 | return false; |
| 1359 | } |
| 1360 | } |
| 1361 | |
| 1362 | // If "Transpose" has input control edges, don't fuse on it. |
| 1363 | for (const Edge* e : node->in_edges()) { |
| 1364 | if (e->IsControlEdge()) { |
| 1365 | return false; |
| 1366 | } |
| 1367 | } |
| 1368 | |
| 1369 | // We compared the tensor containing the permutation order ("perm_node") |
| 1370 | // with our desired order ("perm"). If they're exactly match, this check |
| 1371 | // succeed and returns true. |
| 1372 | for (const Edge* e : node->in_edges()) { |
| 1373 | if (!e->IsControlEdge()) { |
| 1374 | const Node* perm_node = e->src(); |
| 1375 | |
| 1376 | const int kPermTensorIndex = 1; |
| 1377 | if (perm_node->type_string() == "Const"&& |
| 1378 | e->dst_input() == kPermTensorIndex) { |
| 1379 | // we find the "perm" node, now try to retrieve its value. |
| 1380 | const TensorProto* proto = nullptr; |
| 1381 | TF_CHECK_OK(GetNodeAttr(perm_node->def(), "value", &proto)); |
| 1382 | |
| 1383 | DataType type; |
| 1384 | TF_CHECK_OK(GetNodeAttr(perm_node->def(), "dtype", &type)); |
| 1385 | |
| 1386 | Tensor tensor; |
| 1387 | if (!tensor.FromProto(*proto)) { |
| 1388 | TF_CHECK_OK(errors::InvalidArgument( |
| 1389 | "Could not construct Tensor from TensorProto in node: ", |
| 1390 | node->name())); |
| 1391 | return false; |
| 1392 | } |
| 1393 | // Current fusion only supports 4D or 5D tensors according to `perm` |
| 1394 | // vector, return false otherwise. |
| 1395 | if (tensor.dim_size(0) != perm.size()) return false; |
| 1396 | DCHECK_EQ(tensor.dims(), 1); |
| 1397 | if (type == DT_INT32) { |
| 1398 | const auto tensor_content = tensor.flat<int>().data(); |
| 1399 | for (int i = 0; i < perm.size(); ++i) |
| 1400 | if (tensor_content[i] != perm[i]) return false; |
| 1401 | return true; |
| 1402 | } else if (type == DT_INT64) { |
| 1403 | const auto tensor_content = tensor.flat<int64_t>().data(); |
| 1404 | for (int i = 0; i < perm.size(); ++i) |
| 1405 | if (tensor_content[i] != perm[i]) return false; |
| 1406 | return true; |
| 1407 | } |
| 1408 | return false; |
| 1409 | } |
| 1410 | } |
| 1411 | } |
| 1412 | return false; |
| 1413 | } |
| 1414 | |
| 1415 | static bool CheckForMklOp(const Node* node, string name = "") { |
| 1416 | if (node == nullptr) return false; |
| 1417 | |
| 1418 | if (!name.empty() && node->type_string() != name) { |
| 1419 | return false; |
| 1420 | } |
| 1421 | |
| 1422 | // if mklop has multiple outputs, don't fuse it. |
| 1423 | if (node->num_outputs() > 1) return false; |
| 1424 | |
| 1425 | if (node->out_edges().size() > 1) return false; |
| 1426 | |
| 1427 | DataType T; |
| 1428 | TF_CHECK_OK(GetNodeAttr(node->def(), "T", &T)); |
| 1429 | return mkl_op_registry::IsMklOp( |
| 1430 | mkl_op_registry::GetMklOpName(node->type_string()), T); |
| 1431 | } |
| 1432 | |
| 1433 | // Check if the node 'n' has any applicable rewrite rule |
| 1434 | // We check for 2 scenarios for rewrite. |
| 1435 | // |
| 1436 | // @return RewriteInfo* for the applicable rewrite rule |
| 1437 | const RewriteInfo* CheckForNodeRewrite(const Node* n) const; |
| 1438 | const RewriteInfo* CheckForQuantizedNodeRewrite(const Node* n) const; |
| 1439 | |
| 1440 | // Default rewrite rule to be used in scenario 1 for rewrite. |
| 1441 | // @return - true (since we want to always rewrite) |
| 1442 | static bool AlwaysRewrite(const Node* n) { return true; } |
| 1443 | |
| 1444 | // Rewrite rule which considers "context" of the current node to decide if we |
| 1445 | // should rewrite. By "context" we currently mean all the inputs of current |
| 1446 | // node. The idea is if none of the inputs of current node are not MKL nodes, |
| 1447 | // then rewriting current node to MKL node _may not_ offer any performance |
| 1448 | // improvement. |
| 1449 | // |
| 1450 | // One such case is element-wise ops. For such ops, we reuse the Eigen |
| 1451 | // implementation and pass the MKL metadata tensor through so we can avoid |
| 1452 | // conversions. However, if all incoming edges are in TF format, we don't |
| 1453 | // need all this overhead, so replace the elementwise node only if at least |
| 1454 | // one of its parents is a MKL node. |
| 1455 | // |
| 1456 | // More generally, all memory- or IO-bound ops (such as Identity) may fall |
| 1457 | // under this category. |
| 1458 | // |
| 1459 | // @input - Input graph node to be rewritten |
| 1460 | // @return - true if node is to be rewritten as MKL node; false otherwise. |
| 1461 | static bool RewriteIfAtleastOneMklInput(const Node* n) { |
| 1462 | DataType T; |
| 1463 | if (GetNodeAttr(n->def(), "T", &T).ok() && |
| 1464 | mkl_op_registry::IsMklOp( |
| 1465 | mkl_op_registry::GetMklOpName(n->type_string()), T)) { |
| 1466 | for (auto e : n->in_edges()) { |
| 1467 | if (e->IsControlEdge()) continue; |
| 1468 | if (mkl_op_registry::IsMklOp(e->src())) { |
| 1469 | return true; |
| 1470 | } |
| 1471 | } |
| 1472 | } |
| 1473 | return false; |
| 1474 | } |
| 1475 | |
| 1476 | static bool MatMulRewrite(const Node* n) { |
| 1477 | DataType T; |
| 1478 | TF_CHECK_OK(GetNodeAttr(n->def(), "T", &T)); |
| 1479 | if ((T == DT_FLOAT) || (T == DT_BFLOAT16)) { |
| 1480 | VLOG(2) << "Rewriting MatMul to _MklMatMul"; |
| 1481 | return true; |
| 1482 | } |
| 1483 | return false; |
| 1484 | } |
| 1485 | // For oneDNN, only int32 is supported for axis data type |
| 1486 | static bool ConcatV2Rewrite(const Node* n) { |
| 1487 | DataType T; |
| 1488 | TF_CHECK_OK(GetNodeAttr(n->def(), "Tidx", &T)); |
| 1489 | return (T == DT_INT32); |
| 1490 | } |
| 1491 | |
| 1492 | static bool DequantizeRewrite(const Node* n) { |
| 1493 | DCHECK(n); |
| 1494 | Node* input = nullptr; |
| 1495 | TF_CHECK_OK(n->input_node(0, &input)); |
| 1496 | string mode_string; |
| 1497 | int axis = -1; |
| 1498 | TF_CHECK_OK(GetNodeAttr(n->def(), "mode", &mode_string)); |
| 1499 | TF_CHECK_OK(GetNodeAttr(n->def(), "axis", &axis)); |
| 1500 | if (mode_string != "SCALED") { |
| 1501 | VLOG(1) << "DequantizeRewrite: Mode is not SCALED. " |
| 1502 | << "This case is not optimized by Intel MKL kernel, thus using " |
| 1503 | "Eigen op for Dequantize op."; |
| 1504 | return false; |
| 1505 | } |
| 1506 | if (input->IsConstant()) { |
| 1507 | VLOG(1) << "DequantizeRewrite: Trying to dequantize a Const node which " |
| 1508 | << "could possibly be a filter. " |
| 1509 | << "This case is not supported by Intel MKL kernel, thus using " |
| 1510 | "Eigen op for Dequantize op."; |
| 1511 | return false; |
| 1512 | } |
| 1513 | |
| 1514 | if (axis != -1) { |
| 1515 | VLOG(1) << "DequantizeRewrite: Using Eigen op for Dequantize op because " |
| 1516 | << "dimension is specified for per slice dequantization. "; |
| 1517 | return false; |
| 1518 | } |
| 1519 | return true; |
| 1520 | } |
| 1521 | |
| 1522 | // Rewrite rule for _FusedMatMul. |
| 1523 | // @return - true (no transpose attribute for input 1); |
| 1524 | // false otherwise. |
| 1525 | static bool FusedMatMulRewrite(const Node* n) { |
| 1526 | bool trans_a; |
| 1527 | |
| 1528 | // Do not rewrite with transpose attribute because reorder has performance |
| 1529 | // impact. |
| 1530 | TF_CHECK_OK(GetNodeAttr(n->def(), "transpose_a", &trans_a)); |
| 1531 | |
| 1532 | return !trans_a; |
| 1533 | } |
| 1534 | |
| 1535 | // Check if we are performing pooling on depth or batch. If it is, then we |
| 1536 | // do not rewrite MaxPool node to Mkl version. |
| 1537 | // @return - true (if it is not a depth/batch wise pooling case); |
| 1538 | // false otherwise. |
| 1539 | static bool NonDepthBatchWisePoolRewrite(const Node* n) { |
| 1540 | DCHECK(n); |
| 1541 | |
| 1542 | string data_format_str; |
| 1543 | TensorFormat data_format; |
| 1544 | std::vector<int32> ksize, strides; |
| 1545 | TF_CHECK_OK(GetNodeAttr(n->def(), "ksize", &ksize)); |
| 1546 | TF_CHECK_OK(GetNodeAttr(n->def(), "strides", &strides)); |
| 1547 | TF_CHECK_OK(GetNodeAttr(n->def(), "data_format", &data_format_str)); |
| 1548 | bool result = FormatFromString(data_format_str, &data_format); |
| 1549 | DCHECK(result); |
| 1550 | |
| 1551 | // Condition that specifies non-batch-wise and non-depth-wise pooling. |
| 1552 | if (GetTensorDim(ksize, data_format, 'N') == 1 && |
| 1553 | GetTensorDim(strides, data_format, 'N') == 1 && |
| 1554 | GetTensorDim(ksize, data_format, 'C') == 1 && |
| 1555 | GetTensorDim(strides, data_format, 'C') == 1) { |
| 1556 | return true; |
| 1557 | } |
| 1558 | |
| 1559 | return false; |
| 1560 | } |
| 1561 | |
| 1562 | // If the depth_radius of LRN is not 2, then MKL DNN takes unoptimized |
| 1563 | // path. The unoptimized path is slow. Thus we don't rewrite the node |
| 1564 | // and use default Eigen. But for depth_radius=2, MKL DNN optimized |
| 1565 | // path is taken, i.e., eigen node is rewritten by MKl DNN node. |
| 1566 | static bool LrnRewrite(const Node* n) { |
| 1567 | DCHECK(n); |
| 1568 | |
| 1569 | int depth_radius; |
| 1570 | TF_CHECK_OK(GetNodeAttr(n->def(), "depth_radius", &depth_radius)); |
| 1571 | |
| 1572 | // if the depth_radius of LRN is not 2, don't rewrite the node by MKL DNN |
| 1573 | // and use eigen node instead |
| 1574 | if (depth_radius == 2) { |
| 1575 | return true; |
| 1576 | } |
| 1577 | VLOG(1) << "LrnRewrite: The model sets depth_radius as not 2 which" |
| 1578 | << "case is not optimized by Intel MKL, thus using Eigen op" |
| 1579 | << "for LRN "; |
| 1580 | |
| 1581 | return false; |
| 1582 | } |
| 1583 | |
| 1584 | static bool LrnGradRewrite(const Node* n) { |
| 1585 | DCHECK(n); |
| 1586 | bool do_rewrite = false; |
| 1587 | |
| 1588 | for (const Edge* e : n->in_edges()) { |
| 1589 | // Rewrite only if there is corresponding LRN, i.e workspace is available |
| 1590 | if (e->dst()->type_string() == csinfo_.lrn_grad && e->dst_input() == 2 && |
| 1591 | e->src()->type_string() == |
| 1592 | mkl_op_registry::GetMklOpName(csinfo_.lrn) && |
| 1593 | e->src_output() == 0) { |
| 1594 | do_rewrite = true; |
| 1595 | break; |
| 1596 | } |
| 1597 | } |
| 1598 | return do_rewrite; |
| 1599 | } |
| 1600 | |
| 1601 | // MKL-DNN's LeakyRelu(feature) = feature (if feature > 0), or |
| 1602 | // feature * alpha (otherwise), |
| 1603 | // while TensorFlow's LeakyRelu(feature) = max(feature, feature * alpha). |
| 1604 | // These two algorithms are not consistent when alpha > 1, |
| 1605 | // so we only rewrite LeakyRelu to MKL OP when alpha <= 1. |
| 1606 | static bool LeakyReluRewrite(const Node* n) { |
| 1607 | DCHECK(n); |
| 1608 | |
| 1609 | float alpha; |
| 1610 | bool has_attr = TryGetNodeAttr(n->def(), "alpha", &alpha); |
| 1611 | DCHECK(has_attr); |
| 1612 | |
| 1613 | // If the alpha of LeakyRelu is less than 1, rewrite the node. |
| 1614 | // Otherwise eigen node is used instead. |
| 1615 | if (alpha <= 1) { |
| 1616 | return true; |
| 1617 | } |
| 1618 | VLOG(1) << "LeakyReluRewrite: The model sets alpha is greater than 1 " |
| 1619 | << "which case is not optimized by Intel MKL, thus using Eigen op" |
| 1620 | << "for LeakyRelu "; |
| 1621 | |
| 1622 | return false; |
| 1623 | } |
| 1624 | |
| 1625 | static bool QuantizeOpRewrite(const Node* n) { |
| 1626 | DCHECK(n); |
| 1627 | Node* filter_node = nullptr; |
| 1628 | TF_CHECK_OK(n->input_node(0, &filter_node)); |
| 1629 | bool narrow_range = false; |
| 1630 | int axis = -1; |
| 1631 | string mode_string; |
| 1632 | string round_mode_string; |
| 1633 | DataType type; |
| 1634 | TryGetNodeAttr(n->def(), "narrow_range", &narrow_range); |
| 1635 | TryGetNodeAttr(n->def(), "axis", &axis); |
| 1636 | TF_CHECK_OK(GetNodeAttr(n->def(), "mode", &mode_string)); |
| 1637 | TF_CHECK_OK(GetNodeAttr(n->def(), "round_mode", &round_mode_string)); |
| 1638 | TF_CHECK_OK(GetNodeAttr(n->def(), "T", &type)); |
| 1639 | |
| 1640 | if (narrow_range) { |
| 1641 | VLOG(1) << "QuantizeOpRewrite: narrow range is enabled for quantization." |
| 1642 | << "This case is not optimized by Intel MKL, " |
| 1643 | << "thus using Eigen op for Quantize op "; |
| 1644 | return false; |
| 1645 | } |
| 1646 | if (axis != -1) { |
| 1647 | VLOG(1) << "QuantizeOpRewrite: dimension is specified for " |
| 1648 | << "per slice quantization." |
| 1649 | << "This case is not optimized by Intel MKL, " |
| 1650 | << "thus using Eigen op for Quantize op "; |
| 1651 | return false; |
| 1652 | } |
| 1653 | if (!((mode_string == "SCALED"&& round_mode_string == "HALF_TO_EVEN") || |
| 1654 | (mode_string == "MIN_FIRST"))) { |
| 1655 | VLOG(1) << "QuantizeOpRewrite: Mode is not SCALED or MIN_FIRST and/or" |
| 1656 | << "rounding mode is not HALF_TO_EVEN. " |
| 1657 | << "This case is not optimized by Intel MKL, thus using Eigen op" |
| 1658 | << "for Quantize op "; |
| 1659 | return false; |
| 1660 | } |
| 1661 | if (filter_node->IsConstant()) { |
| 1662 | VLOG(1) << "QuantizeOpRewrite: Trying to quantize a node which " |
| 1663 | << "is a constant. " |
| 1664 | << "This case is not supported by the kernel, thus using Eigen op" |
| 1665 | << "for Quantize op "; |
| 1666 | |
| 1667 | return false; |
| 1668 | } |
| 1669 | if (mode_string == "MIN_FIRST") { |
| 1670 | if (type != DT_QUINT8) { |
| 1671 | VLOG(1) << "QuantizeOpRewrite: For MIN_FIRST mode the data type is " |
| 1672 | << "not DT_UINT8. This case is not optimized by Intel MKL, " |
| 1673 | << "thus using Eigen op for Quantize op "; |
| 1674 | return false; |
| 1675 | } |
| 1676 | } |
| 1677 | return true; |
| 1678 | } |
| 1679 | |
| 1680 | static bool MaxpoolGradRewrite(const Node* n) { |
| 1681 | DCHECK(n); |
| 1682 | bool do_rewrite = false; |
| 1683 | for (const Edge* e : n->in_edges()) { |
| 1684 | // Rewrite only if there is corresponding Maxpool, i.e workspace is |
| 1685 | // available |
| 1686 | if (e->dst()->type_string() == csinfo_.max_pool_grad && |
| 1687 | e->dst_input() == 1 && |
| 1688 | e->src()->type_string() == |
| 1689 | mkl_op_registry::GetMklOpName(csinfo_.max_pool) && |
| 1690 | e->src_output() == 0) { |
| 1691 | do_rewrite = true; |
| 1692 | break; |
| 1693 | } |
| 1694 | } |
| 1695 | return do_rewrite; |
| 1696 | } |
| 1697 | |
| 1698 | static bool Maxpool3DGradRewrite(const Node* n) { |
| 1699 | DCHECK(n); |
| 1700 | for (const Edge* e : n->in_edges()) { |
| 1701 | // Rewrite only if there is corresponding Maxpool3D, i.e., workspace is |
| 1702 | // available |
| 1703 | if (e->dst()->type_string() == csinfo_.max_pool3d_grad && |
| 1704 | e->dst_input() == 1 && |
| 1705 | e->src()->type_string() == |
| 1706 | mkl_op_registry::GetMklOpName(csinfo_.max_pool3d) && |
| 1707 | e->src_output() == 0) { |
| 1708 | return true; |
| 1709 | } |
| 1710 | } |
| 1711 | return false; |
| 1712 | } |
| 1713 | |
| 1714 | static bool FusedBatchNormV3Rewrite(const Node* n) { |
| 1715 | DCHECK(n); |
| 1716 | if (Check5DFormat(n->def())) { |
| 1717 | VLOG(1) << "Graph Rewrite: FusedBatchNorm(Grad)V3 op currently does not " |
| 1718 | << "support 5D tensors."; |
| 1719 | return false; |
| 1720 | } |
| 1721 | return true; |
| 1722 | } |
| 1723 | |
| 1724 | static bool FusedBatchNormExRewrite(const Node* n) { |
| 1725 | DCHECK(n); |
| 1726 | |
| 1727 | int num_side_inputs; |
| 1728 | TF_CHECK_OK(GetNodeAttr(n->def(), "num_side_inputs", &num_side_inputs)); |
| 1729 | string activation_mode; |
| 1730 | TF_CHECK_OK(GetNodeAttr(n->def(), "activation_mode", &activation_mode)); |
| 1731 | |
| 1732 | // if the num_side_inputs is not 0, don't rewrite the node. |
| 1733 | if (num_side_inputs != 0) { |
| 1734 | VLOG(1) << "FusedBatchNormExRewrite: The model sets num_side_inputs" |
| 1735 | << "larger than 0 is not optimized by Intel MKL."; |
| 1736 | return false; |
| 1737 | } |
| 1738 | |
| 1739 | // if the activation_mode is not 'Relu', don't rewrite the node. |
| 1740 | if (activation_mode != "Relu") { |
| 1741 | VLOG(1) << "FusedBatchNormExRewrite: Only Relu activation mode is" |
| 1742 | << "supported by Intel MKL."; |
| 1743 | return false; |
| 1744 | } |
| 1745 | |
| 1746 | return true; |
| 1747 | } |
| 1748 | |
| 1749 | static bool FusedConv2DRewrite(const Node* n) { |
| 1750 | // MKL DNN currently doesn't support all fusions that grappler fuses |
| 1751 | // together with Conv2D (ex. batchnorm). We rewrite _FusedConv2D only if |
| 1752 | // it includes those we support. |
| 1753 | DataType T; |
| 1754 | if (!TryGetNodeAttr(n->def(), "T", &T) || |
| 1755 | !mkl_op_registry::IsMklOp(NativeFormatEnabled() |
| 1756 | ? csinfo_.mkl_native_fused_conv2d |
| 1757 | : csinfo_.mkl_fused_conv2d, |
| 1758 | T)) { |
| 1759 | return false; |
| 1760 | } |
| 1761 | |
| 1762 | string data_format; |
| 1763 | string filter_format; |
| 1764 | int num_host_args = 0; |
| 1765 | TF_CHECK_OK(GetNodeAttr(n->def(), "data_format", &data_format)); |
| 1766 | TF_CHECK_OK(GetNodeAttr(n->def(), "filter_format", &filter_format)); |
| 1767 | TF_CHECK_OK(GetNodeAttr(n->def(), "num_host_args", &num_host_args)); |
| 1768 | if ((data_format != "NCHW"&& data_format != "NHWC") || |
| 1769 | (filter_format != "HWIO"&& filter_format != "OIHW") || |
| 1770 | (num_host_args != 0)) { |
| 1771 | return false; |
| 1772 | } |
| 1773 | |
| 1774 | std::vector<string> fused_ops; |
| 1775 | TF_CHECK_OK(GetNodeAttr(n->def(), "fused_ops", &fused_ops)); |
| 1776 | return (fused_ops == std::vector<string>{"BiasAdd"} || |
| 1777 | fused_ops == std::vector<string>{"Relu"} || |
| 1778 | fused_ops == std::vector<string>{"Relu6"} || |
| 1779 | fused_ops == std::vector<string>{"Elu"} || |
| 1780 | fused_ops == std::vector<string>{"BiasAdd", "Relu"} || |
| 1781 | fused_ops == std::vector<string>{"BiasAdd", "Relu6"} || |
| 1782 | fused_ops == std::vector<string>{"BiasAdd", "Elu"} || |
| 1783 | fused_ops == std::vector<string>{"BiasAdd", "Add"} || |
| 1784 | fused_ops == std::vector<string>{"BiasAdd", "Add", "Relu"} || |
| 1785 | fused_ops == std::vector<string>{"BiasAdd", "Add", "Relu6"} || |
| 1786 | fused_ops == std::vector<string>{"BiasAdd", "Add", "Elu"} || |
| 1787 | fused_ops == std::vector<string>{"LeakyRelu"} || |
| 1788 | fused_ops == std::vector<string>{"BiasAdd", "LeakyRelu"} || |
| 1789 | fused_ops == std::vector<string>{"BiasAdd", "Add", "LeakyRelu"} || |
| 1790 | fused_ops == std::vector<string>{"FusedBatchNorm"} || |
| 1791 | fused_ops == std::vector<string>{"FusedBatchNorm", "Relu"} || |
| 1792 | fused_ops == std::vector<string>{"FusedBatchNorm", "Relu6"} || |
| 1793 | fused_ops == std::vector<string>{"FusedBatchNorm", "Elu"} || |
| 1794 | fused_ops == std::vector<string>{"FusedBatchNorm", "LeakyRelu"}); |
| 1795 | } |
| 1796 | |
| 1797 | static bool FusedDepthwiseConv2DRewrite(const Node* n) { |
| 1798 | // MKL DNN currently doesn't support all fusions that grappler fuses |
| 1799 | // together with DepthwiseConv2D (ex. batchnorm). We rewrite |
| 1800 | // _FusedDepthwiseConv2DNative only if it includes those we support. |
| 1801 | DataType T; |
| 1802 | if (!TryGetNodeAttr(n->def(), "T", &T) || |
| 1803 | !mkl_op_registry::IsMklOp( |
| 1804 | NativeFormatEnabled() ? csinfo_.mkl_native_fused_depthwise_conv2d |
| 1805 | : csinfo_.mkl_fused_depthwise_conv2d, |
| 1806 | T)) { |
| 1807 | return false; |
| 1808 | } |
| 1809 | |
| 1810 | std::vector<string> fused_ops; |
| 1811 | TF_CHECK_OK(GetNodeAttr(n->def(), "fused_ops", &fused_ops)); |
| 1812 | return (fused_ops == std::vector<string>{"BiasAdd"} || |
| 1813 | fused_ops == std::vector<string>{"BiasAdd", "Relu"} || |
| 1814 | fused_ops == std::vector<string>{"BiasAdd", "Relu6"} || |
| 1815 | fused_ops == std::vector<string>{"BiasAdd", "Elu"}); |
| 1816 | } |
| 1817 | |
| 1818 | // Rewrites input node to a new node specified by its matching rewrite info. |
| 1819 | // |
| 1820 | // Method first searches matching rewrite info for input node and then |
| 1821 | // uses that info to rewrite. |
| 1822 | // |
| 1823 | // Input node may be deleted in case of rewrite. Attempt to use the node |
| 1824 | // after the call can result in undefined behaviors. |
| 1825 | // |
| 1826 | // @input g - input graph, n - Node to be rewritten, |
| 1827 | // ri - matching rewriteinfo |
| 1828 | // @return OkStatus(), if the input node is rewritten; |
| 1829 | // Returns appropriate Status error code otherwise. |
| 1830 | // Graph is updated in case the input node is rewritten. |
| 1831 | // Otherwise, it is not updated. |
| 1832 | Status RewriteNode(std::unique_ptr<Graph>* g, Node* n, const RewriteInfo* ri); |
| 1833 | |
| 1834 | // Rewrites input node to just change its operator name. The number of |
| 1835 | // inputs to the node and the number of outputs remain the same. Attributes |
| 1836 | // of the new node could be copied from attributes of the old node or |
| 1837 | // modified. copy_attrs field of RewriteInfo controls this. |
| 1838 | // |
| 1839 | // Conceptually, it allows us to rewrite: |
| 1840 | // |
| 1841 | // f[a=v1,b=v2](x,y) -> g[a'=v3,b'=v4](x,y) |
| 1842 | // |
| 1843 | // Attributes can be altered without any restrictions --- they could be |
| 1844 | // copied, modified, or deleted completely. |
| 1845 | // |
| 1846 | // @input g - input graph, orig_node - Node to be rewritten, |
| 1847 | // ri - matching rewriteinfo |
| 1848 | // @output new_node - points to newly created node |
| 1849 | // @return OkStatus(), if the input node is rewritten; |
| 1850 | // Returns appropriate Status error code otherwise. |
| 1851 | // Graph is only updated when the input node is rewritten. |
| 1852 | Status RewriteNodeForJustOpNameChange(std::unique_ptr<Graph>* g, |
| 1853 | const Node* orig_node, Node** new_node, |
| 1854 | const RewriteInfo* ri); |
| 1855 | |
| 1856 | // Rewrites input node to enable MKL layout propagation. Please also refer to |
| 1857 | // documentation for the function RewriteNodeForJustOpNameChange() to |
| 1858 | // understand what it means. |
| 1859 | // |
| 1860 | // @input g - input graph, orig_node - Node to be rewritten, |
| 1861 | // ri - matching rewriteinfo |
| 1862 | // @output new_node - points to newly created node |
| 1863 | // @return OkStatus(), if the input node is rewritten; |
| 1864 | // Returns appropriate Status error code otherwise. |
| 1865 | // Graph is updated in case the input node is rewritten. |
| 1866 | // Otherwise, it is not updated. |
| 1867 | Status RewriteNodeForLayoutPropagation(std::unique_ptr<Graph>* g, |
| 1868 | const Node* orig_node, Node** new_node, |
| 1869 | const RewriteInfo* ri); |
| 1870 | |
| 1871 | // Get nodes that will feed a list of TF tensors to the new |
| 1872 | // node that we are constructing. |
| 1873 | // |
| 1874 | // @input g - input graph, |
| 1875 | // @input inputs - inputs to old node that we are using for constructing |
| 1876 | // new inputs, |
| 1877 | // @input input_idx - the index in the 'inputs' vector pointing to the |
| 1878 | // current input that we have processed so far |
| 1879 | // @output input_idx - index will be incremented by the number of nodes |
| 1880 | // from 'inputs' that are processed |
| 1881 | // @input list_length - The expected length of list of TF tensors |
| 1882 | // @output output_nodes - the list of new nodes creating TF tensors |
| 1883 | // |
| 1884 | // @return None |
| 1885 | void GetNodesProducingTFTensorList( |
| 1886 | const gtl::InlinedVector<std::pair<Node*, int>, 4>& inputs, |
| 1887 | int* input_idx, int list_length, |
| 1888 | std::vector<NodeBuilder::NodeOut>* output_nodes); |
| 1889 | |
| 1890 | // Get nodes that will feed a list of Mkl tensors to the new |
| 1891 | // node that we are constructing. |
| 1892 | // |
| 1893 | // @input g - input graph, |
| 1894 | // @input orig_node - Original node that we are rewriting |
| 1895 | // @input inputs - inputs to old node that we are using for constructing |
| 1896 | // new inputs, |
| 1897 | // @input input_idx - the index in the 'inputs' vector pointing to the |
| 1898 | // current input that we have processed so far |
| 1899 | // @output input_idx - index will be incremented by the number of nodes |
| 1900 | // from 'inputs' that are processed |
| 1901 | // @input list_length - The expected length of list of Mkl tensors |
| 1902 | // @output output_nodes - the list of new nodes creating Mkl tensors |
| 1903 | // |
| 1904 | // @return None |
| 1905 | void GetNodesProducingMklTensorList( |
| 1906 | std::unique_ptr<Graph>* g, const Node* orig_node, |
| 1907 | const gtl::InlinedVector<std::pair<Node*, int>, 4>& inputs, |
| 1908 | int* input_idx, int list_length, |
| 1909 | std::vector<NodeBuilder::NodeOut>* output_nodes); |
| 1910 | |
| 1911 | // Get a node that will feed an Mkl tensor to the new |
| 1912 | // node that we are constructing. The output node could be (1) 'n' |
| 1913 | // if it is Mkl layer, or (2) a dummy node producing dummy Mkl tensor |
| 1914 | // if 'n' is not an Mkl layer. |
| 1915 | // |
| 1916 | // @input g - input graph, |
| 1917 | // @input orig_node - Original node that we are rewriting, |
| 1918 | // @input n - Node based on which we are creating Mkl node, |
| 1919 | // @input n_output_slot - the output slot of node 'n' |
| 1920 | // which is feeding to the node that we are constructing |
| 1921 | // @output mkl_node - the new node that will feed Mkl tensor |
| 1922 | // @output mkl_node_output_slot - the slot number of mkl_node that |
| 1923 | // will feed the tensor |
| 1924 | // @return None |
| 1925 | void GetNodeProducingMklTensor(std::unique_ptr<Graph>* g, |
| 1926 | const Node* orig_node, Node* n, |
| 1927 | int n_output_slot, Node** mkl_node, |
| 1928 | int* mkl_node_output_slot); |
| 1929 | |
| 1930 | // Setup new inputs using old inputs 'inputs' for the rewritten node in 'nb' |
| 1931 | // in graph 'g'. Original node is input in 'old_node'. Inputs to 'nb' are |
| 1932 | // set up in contiguous fashion. 'workspace_tensors' carry graph nodes |
| 1933 | // producing workspace edges if 'are_workspace_tensors_available' is true. |
| 1934 | // Otherwise, 'workspace_tensors' is empty vector. |
| 1935 | // |
| 1936 | // For details, refer to 'Ordering of inputs after rewriting' section in the |
| 1937 | // documentation above. |
| 1938 | // |
| 1939 | // Returns OkStatus() if setting up inputs is successful, otherwise |
| 1940 | // returns appropriate status code. |
| 1941 | int SetUpContiguousInputs( |
| 1942 | std::unique_ptr<Graph>* g, |
| 1943 | const gtl::InlinedVector<std::pair<Node*, int>, 4>& old_node_inputs, |
| 1944 | NodeBuilder* nb, const Node* old_node, |
| 1945 | std::vector<NodeBuilder::NodeOut>* workspace_tensors, |
| 1946 | bool are_workspace_tensors_available); |
| 1947 | |
| 1948 | // Setup new inputs using old inputs 'inputs' for the rewritten node in 'nb' |
| 1949 | // in graph 'g'. Original node is input in 'orig_node'. |
| 1950 | // |
| 1951 | // For details, refer to 'Ordering of Tensorflow tensors and Mkl tensors' |
| 1952 | // section in the documentation above. |
| 1953 | // |
| 1954 | // Returns OkStatus() if setting up inputs is successful, otherwise |
| 1955 | // returns appropriate status code. |
| 1956 | Status SetUpInputs(std::unique_ptr<Graph>* g, |
| 1957 | const gtl::InlinedVector<std::pair<Node*, int>, 4>& inputs, |
| 1958 | NodeBuilder* nb, const Node* orig_node); |
| 1959 | |
| 1960 | // Create new inputs by copying old inputs 'inputs' for the rewritten node |
| 1961 | // in 'nb' in graph 'g'. Original node is input in 'orig_node'. This is mostly |
| 1962 | // used in the context of rewrite for just operator name change in which |
| 1963 | // inputs of old operator and new operator are same. |
| 1964 | // |
| 1965 | // Returns OkStatus() if setting up inputs is successful, otherwise |
| 1966 | // returns appropriate status code. |
| 1967 | Status CopyInputs(const Node* orig_node, |
| 1968 | const gtl::InlinedVector<std::pair<Node*, int>, 4>& inputs, |
| 1969 | NodeBuilder* nb); |
| 1970 | |
| 1971 | // Add workspace edge on the input or output side of Node 'orig_node' by using |
| 1972 | // NodeBuilder 'nb' for the new node provided. If 'orig_node' does not dictate |
| 1973 | // adding workspace edge then do not add it. Workspace Tensorflow and Mkl |
| 1974 | // tensors, if they need to be added, will be set into these tensors. |
| 1975 | // If we set workspace tensors, then are_ws_tensors_added should be true. |
| 1976 | void AddWorkSpaceEdgeIfNeeded(std::unique_ptr<Graph>* g, |
| 1977 | const Node* orig_node, NodeBuilder* nb, |
| 1978 | std::vector<NodeBuilder::NodeOut>* ws_tensors, |
| 1979 | bool* are_ws_tensors_added); |
| 1980 | |
| 1981 | // Helper function used by FixMklMetaDataEdges. Fixes the metadata edge |
| 1982 | // pointed by 'e_metadata' corresponding to the data edge 'e_data' in graph |
| 1983 | // 'g'. Returns true if fixup was done; otherwise, it returns false. |
| 1984 | bool FixMklMetaDataEdgeIfNeeded(std::unique_ptr<Graph>* g, const Edge* e_data, |
| 1985 | const Edge* e_metadata); |
| 1986 | |
| 1987 | // Are the input Mkl metadata edges for node 'n' in graph 'g' correctly |
| 1988 | // connected? If not, then fix them. This is needed because a graph may have |
| 1989 | // some input Mkl metadata edges incorrectly setup after node merge and |
| 1990 | // rewrite passes. This could happen because GetReversePostOrder function may |
| 1991 | // not provide topologically sorted order if a graph contains cycles. The |
| 1992 | // function returns true if at least one Mkl metadata edge for node 'n' was |
| 1993 | // fixed. Otherwise, it returns false. |
| 1994 | // |
| 1995 | // Example: |
| 1996 | // |
| 1997 | // X = MklConv2D(_, _, _) |
| 1998 | // Y = MklConv2DWithBias(_, _, _, _, _, _) |
| 1999 | // Z = MklAdd(X, Y, DummyMklTensor, Y:1) |
| 2000 | // |
| 2001 | // For a graph such as shown above, note that 3rd argument of MklAdd contains |
| 2002 | // DummyMklTensor. Actually, it should be getting the Mkl metadata from |
| 2003 | // MklConv2D op (specifically, X:2). This incorrect plumbing could be possible |
| 2004 | // (although rare) if the Mkl NodeMerge + NodeRewrite passes visit Z before X |
| 2005 | // (possible if X, Y, Z are part of a loop.) This function fixes the Mkl |
| 2006 | // metadata edges only - it does not rewrite nodes nor does it modify the Mkl |
| 2007 | // data edges (1st and 2nd arguments of MklAdd). |
| 2008 | bool FixMklMetaDataEdges(std::unique_ptr<Graph>* g, Node* n); |
| 2009 | |
| 2010 | // Functions specific to operators to copy attributes |
| 2011 | // We need operator-specific function to copy attributes because the framework |
| 2012 | // does not provide any generic function for it. |
| 2013 | // NOTE: names are alphabetically sorted. |
| 2014 | static void CopyAttrsAll(const Node* orig_node, NodeBuilder* nb, |
| 2015 | bool change_format = false); |
| 2016 | static void CopyAttrsAllCheckConstFilter(const Node* orig_node, |
| 2017 | NodeBuilder* nb, |
| 2018 | bool change_format = false); |
| 2019 | |
| 2020 | static void CopyAttrsConv(const Node* orig_node, NodeBuilder* nb, |
| 2021 | bool change_format = false); |
| 2022 | static void CopyAttrsConvCheckConstFilter(const Node* orig_node, |
| 2023 | NodeBuilder* nb, |
| 2024 | bool change_format = false); |
| 2025 | static void CopyAttrsFusedConv2DCheckConstFilter(const Node* orig_node, |
| 2026 | NodeBuilder* nb, |
| 2027 | bool change_format = false); |
| 2028 | static void CopyAttrsFromPadAndConv2D(const Node* orig_node1, |
| 2029 | const Node* orig_node2, NodeBuilder* nb, |
| 2030 | bool change_format = false); |
| 2031 | static void CopyAttrsFromPadAndFusedConv2D(const Node* orig_node1, |
| 2032 | const Node* orig_node2, |
| 2033 | NodeBuilder* nb, |
| 2034 | bool change_format = false); |
| 2035 | static void CopyAttrsQuantizedConv2D(const Node* orig_node, NodeBuilder* nb, |
| 2036 | bool change_format = false); |
| 2037 | static void CopyFormatAttrsConv(const Node* orig_node, NodeBuilder* nb, |
| 2038 | const std::vector<int32>& strides, |
| 2039 | const std::vector<int32>& dilations, |
| 2040 | bool change_format = false); |
| 2041 | |
| 2042 | static void CopyAttrsQuantizedMatMulWithBias(const Node* orig_node, |
| 2043 | NodeBuilder* nb, |
| 2044 | bool change_format = false); |
| 2045 | static void CopyAttrsQuantizedMatMulWithBiasAndDequantize( |
| 2046 | const Node* orig_node, NodeBuilder* nb, bool change_format = false); |
| 2047 | static void CopyAttrsPooling(const Node* orig_node, NodeBuilder* nb, |
| 2048 | bool change_format = false); |
| 2049 | |
| 2050 | // Generate a graph node in graph 'g' representing a dummy Mkl tensor node, |
| 2051 | // using node for original node 'orig_node' and return it in '*out'. |
| 2052 | // TODO(nhasabni) We should move this to mkl_util.h |
| 2053 | void GetDummyMklTensorNode(std::unique_ptr<Graph>* g, Node** out, |
| 2054 | const Node* orig_node); |
| 2055 | void GetDummyWorkspaceTensorNode(std::unique_ptr<Graph>* g, Node** out, |
| 2056 | const Node* orig_node); |
| 2057 | }; |
| 2058 | |
| 2059 | MklLayoutRewritePass::ConstStringsInfo MklLayoutRewritePass::csinfo_; |
| 2060 | |
| 2061 | // We register Mkl rewrite pass for phase 1 in post partitioning group. |
| 2062 | // We register it here so that we get a complete picture of all users of Mkl |
| 2063 | // nodes. Do not change the ordering of the Mkl passes. |
| 2064 | const OptimizationPassRegistry::Grouping kMklLayoutRewritePassGroup = |
| 2065 | OptimizationPassRegistry::POST_PARTITIONING; |
| 2066 | REGISTER_OPTIMIZATION(kMklLayoutRewritePassGroup, 1, MklLayoutRewritePass); |
| 2067 | |
| 2068 | ////////////////////////////////////////////////////////////////////////// |
| 2069 | // Helper functions for creating new node |
| 2070 | ////////////////////////////////////////////////////////////////////////// |
| 2071 | |
| 2072 | static void FillInputs(const Node* n, |
| 2073 | gtl::InlinedVector<Node*, 4>* control_edges, |
| 2074 | gtl::InlinedVector<std::pair<Node*, int>, 4>* in) { |
| 2075 | control_edges->clear(); |
| 2076 | for (const Edge* e : n->in_edges()) { |
| 2077 | if (e->IsControlEdge()) { |
| 2078 | control_edges->push_back(e->src()); |
| 2079 | } else { |
| 2080 | (*in)[e->dst_input()] = std::make_pair(e->src(), e->src_output()); |
| 2081 | } |
| 2082 | } |
| 2083 | std::sort(control_edges->begin(), control_edges->end()); |
| 2084 | } |
| 2085 | |
| 2086 | void MklLayoutRewritePass::GetNodesProducingTFTensorList( |
| 2087 | const gtl::InlinedVector<std::pair<Node*, int>, 4>& inputs, int* input_idx, |
| 2088 | int list_length, std::vector<NodeBuilder::NodeOut>* output_nodes) { |
| 2089 | CHECK_LT(*input_idx, inputs.size()); |
| 2090 | CHECK_GT(list_length, 0); |
| 2091 | DCHECK(output_nodes); |
| 2092 | output_nodes->reserve(list_length); |
| 2093 | |
| 2094 | while (list_length != 0) { |
| 2095 | CHECK_GT(list_length, 0); |
| 2096 | CHECK_LT(*input_idx, inputs.size()); |
| 2097 | Node* n = inputs[*input_idx].first; |
| 2098 | int slot = inputs[*input_idx].second; |
| 2099 | // If input node 'n' is just producing a single tensor at |
| 2100 | // output slot 'slot' then we just add that single node. |
| 2101 | output_nodes->push_back(NodeBuilder::NodeOut(n, slot)); |
| 2102 | (*input_idx)++; |
| 2103 | list_length--; |
| 2104 | } |
| 2105 | } |
| 2106 | |
| 2107 | // TODO(nhasabni) We should move this to mkl_util.h. |
| 2108 | void MklLayoutRewritePass::GetDummyMklTensorNode(std::unique_ptr<Graph>* g, |
| 2109 | Node** out, |
| 2110 | const Node* orig_node) { |
| 2111 | // We use a tensor of shape {8} and value 0,0,0,0,0,0,0,0 to represent |
| 2112 | // dummy Mkl tensor. 8 = 2*size_t. |
| 2113 | const DataType dt = DataTypeToEnum<uint8>::v(); |
| 2114 | TensorProto proto; |
| 2115 | proto.set_dtype(dt); |
| 2116 | uint8 zero[8] = {0, 0, 0, 0, 0, 0, 0, 0}; |
| 2117 | proto.set_tensor_content(string(reinterpret_cast<char*>(&zero), 8)); |
| 2118 | TensorShape dummy_shape({8}); |
| 2119 | dummy_shape.AsProto(proto.mutable_tensor_shape()); |
| 2120 | TF_CHECK_OK(NodeBuilder((*g)->NewName("DMT"), "Const") |
| 2121 | .Attr("value", proto) |
| 2122 | .Attr("dtype", dt) |
| 2123 | .Device(orig_node->def().device()) // We place this node on |
| 2124 | // the same device as the |
| 2125 | // device of the original |
| 2126 | // node. |
| 2127 | .Finalize(&**g, out)); |
| 2128 | DCHECK(*out); // Make sure we got a valid object before using it |
| 2129 | |
| 2130 | // If number of inputs to the original node is > 0, then we add |
| 2131 | // control dependency between 1st input (index 0) of the original node and |
| 2132 | // the dummy Mkl node. This is needed because control-flow ops such as Enter, |
| 2133 | // Merge, etc, require frame_name of the dummy Mkl node to be same as the |
| 2134 | // rewritten node. Adding control edge between 1st input of the original node |
| 2135 | // and the dummy Mkl node ensures that the dummy node is in the same frame |
| 2136 | // as the original node. Choosing 1st input is not necessary - any input of |
| 2137 | // the original node is fine because all the inputs of a node are always in |
| 2138 | // the same frame. |
| 2139 | if (orig_node->num_inputs() > 0) { |
| 2140 | Node* orig_input0 = nullptr; |
| 2141 | TF_CHECK_OK( |
| 2142 | orig_node->input_node(0, const_cast<const Node**>(&orig_input0))); |
| 2143 | auto edge = (*g)->AddControlEdge(orig_input0, *out, false); |
| 2144 | DCHECK(edge != nullptr || DoesControlEdgeExist(orig_input0, *out)); |
| 2145 | } |
| 2146 | |
| 2147 | (*out)->set_assigned_device_name(orig_node->assigned_device_name()); |
| 2148 | } |
| 2149 | |
| 2150 | void MklLayoutRewritePass::GetNodesProducingMklTensorList( |
| 2151 | std::unique_ptr<Graph>* g, const Node* orig_node, |
| 2152 | const gtl::InlinedVector<std::pair<Node*, int>, 4>& inputs, int* input_idx, |
| 2153 | int list_length, std::vector<NodeBuilder::NodeOut>* output_nodes) { |
| 2154 | CHECK_LT(*input_idx, inputs.size()); |
| 2155 | CHECK_GT(list_length, 0); |
| 2156 | DCHECK(output_nodes); |
| 2157 | output_nodes->reserve(list_length); |
| 2158 | |
| 2159 | while (list_length != 0) { |
| 2160 | CHECK_GT(list_length, 0); |
| 2161 | CHECK_LT(*input_idx, inputs.size()); |
| 2162 | Node* n = inputs[*input_idx].first; |
| 2163 | int slot = inputs[*input_idx].second; |
| 2164 | // If 'n' is producing a single tensor, then create a single Mkl tensor |
| 2165 | // node. |
| 2166 | Node* mkl_node = nullptr; |
| 2167 | int mkl_node_output_slot = 0; |
| 2168 | GetNodeProducingMklTensor(g, orig_node, n, slot, &mkl_node, |
| 2169 | &mkl_node_output_slot); |
| 2170 | output_nodes->push_back( |
| 2171 | NodeBuilder::NodeOut(mkl_node, mkl_node_output_slot)); |
| 2172 | (*input_idx)++; |
| 2173 | list_length--; |
| 2174 | } |
| 2175 | } |
| 2176 | |
| 2177 | // Get an input node that will feed Mkl tensor to the new |
| 2178 | // node that we are constructing. An input node could be (1) 'n' |
| 2179 | // if it is Mkl layer, or (2) a dummy node producing dummy Mkl tensor |
| 2180 | // if 'n' is not an Mkl layer. |
| 2181 | void MklLayoutRewritePass::GetNodeProducingMklTensor( |
| 2182 | std::unique_ptr<Graph>* g, const Node* orig_node, Node* n, |
| 2183 | int n_output_slot, Node** mkl_node, int* mkl_node_output_slot) { |
| 2184 | DCHECK(n); |
| 2185 | DCHECK(mkl_node); |
| 2186 | DCHECK(mkl_node_output_slot); |
| 2187 | |
| 2188 | // If this is an MKL op, then it will create extra output for MKL layout. |
| 2189 | DataType T; |
| 2190 | if (TryGetNodeAttr(n->def(), "T", &T) && |
| 2191 | mkl_op_registry::IsMklOp(n->type_string(), T, false)) { |
| 2192 | // If this is an MKL op, then it will generate an edge that will receive |
| 2193 | // Mkl tensor from a node. |
| 2194 | // output slot number for Mkl tensor would be N+slot number of TensorFlow |
| 2195 | // tensor, where N is total number of TensorFlow tensors. |
| 2196 | *mkl_node = n; |
| 2197 | *mkl_node_output_slot = |
| 2198 | GetTensorMetaDataIndex(n_output_slot, n->num_outputs()); |
| 2199 | } else { |
| 2200 | // If we have not visited the node and rewritten it, then we need |
| 2201 | // to create a dummy node that will feed a dummy Mkl tensor to this node. |
| 2202 | // DummyMklTensor node has no input and generates only 1 output |
| 2203 | // (dummy Mkl tensor) as output slot number 0. |
| 2204 | GetDummyMklTensorNode(g, mkl_node, orig_node); |
| 2205 | DCHECK(*mkl_node); |
| 2206 | *mkl_node_output_slot = 0; |
| 2207 | } |
| 2208 | } |
| 2209 | |
| 2210 | int MklLayoutRewritePass::SetUpContiguousInputs( |
| 2211 | std::unique_ptr<Graph>* g, |
| 2212 | const gtl::InlinedVector<std::pair<Node*, int>, 4>& old_node_inputs, |
| 2213 | NodeBuilder* nb, const Node* old_node, |
| 2214 | std::vector<NodeBuilder::NodeOut>* workspace_tensors, |
| 2215 | bool are_workspace_tensors_available) { |
| 2216 | DCHECK(workspace_tensors); |
| 2217 | CHECK_EQ(kTensorOrdering, MklTfTensorOrdering::TENSORS_CONTIGUOUS); |
| 2218 | |
| 2219 | // TODO(nhasabni): Temporary solution to connect filter input of |
| 2220 | // BackpropInput with the converted filter from Conv2D. |
| 2221 | bool do_connect_conv2d_backprop_input_filter = false; |
| 2222 | Node* conv2d_node = nullptr; |
| 2223 | // Filter node is 2nd input (slot index 1) of Conv2D. |
| 2224 | int kConv2DFilterInputSlotIdx = 1; |
| 2225 | int kConv2DBackpropInputFilterInputSlotIdx = 1; |
| 2226 | int kConv2DFilterOutputSlotIdx = 1; |
| 2227 | if (old_node->type_string() == csinfo_.conv2d_grad_input) { |
| 2228 | // We need to find Conv2D node from Conv2DBackpropInput. |
| 2229 | // For that let's first find filter node that is 2nd input (slot 1) |
| 2230 | // of BackpropInput. |
| 2231 | Node* filter_node = nullptr; |
| 2232 | TF_CHECK_OK(old_node->input_node(kConv2DBackpropInputFilterInputSlotIdx, |
| 2233 | &filter_node)); |
| 2234 | DCHECK(filter_node); |
| 2235 | |
| 2236 | // Now check which nodes receive from filter_node. Filter feeds as |
| 2237 | // 2nd input (slot 1) of _MklConv2D, _MklConv2DWithBias, and |
| 2238 | // _MklFusedConv2D. |
| 2239 | for (const Edge* e : filter_node->out_edges()) { |
| 2240 | if ((e->dst()->type_string() == csinfo_.mkl_conv2d || |
| 2241 | e->dst()->type_string() == csinfo_.mkl_pad_with_conv2d || |
| 2242 | e->dst()->type_string() == csinfo_.mkl_pad_with_fused_conv2d || |
| 2243 | e->dst()->type_string() == csinfo_.mkl_conv2d_with_bias || |
| 2244 | e->dst()->type_string() == csinfo_.mkl_fused_conv2d) && |
| 2245 | e->dst_input() == kConv2DFilterInputSlotIdx |
| 2246 | /* filter is 2nd input of Conv2D and _MklConv2D. */) { |
| 2247 | if (conv2d_node != nullptr) { |
| 2248 | VLOG(1) << "MklLayoutRewritePass: unusual case of same filter" |
| 2249 | << " feeding multiple Conv2D nodes: " |
| 2250 | << filter_node->DebugString(); |
| 2251 | // We will not connect filter input of Conv2DBackpropInput |
| 2252 | // to be safe here. |
| 2253 | do_connect_conv2d_backprop_input_filter = false; |
| 2254 | break; |
| 2255 | } else { |
| 2256 | conv2d_node = e->dst(); |
| 2257 | do_connect_conv2d_backprop_input_filter = true; |
| 2258 | } |
| 2259 | } |
| 2260 | } |
| 2261 | } |
| 2262 | |
| 2263 | // Number of input slots to original op |
| 2264 | // Input slots are represented by .Input() calls in REGISTER_OP. |
| 2265 | int old_node_input_slots = old_node->op_def().input_arg_size(); |
| 2266 | int nn_slot_idx = 0; // slot index for inputs of new node |
| 2267 | |
| 2268 | // Let's copy all inputs (TF tensors) of original node to new node. |
| 2269 | int iidx = 0; |
| 2270 | for (int on_slot_idx = 0; on_slot_idx < old_node_input_slots; on_slot_idx++) { |
| 2271 | // An input slot could be a single tensor or a list. We need |
| 2272 | // to handle this case accordingly. |
| 2273 | const OpDef::ArgDef& arg = old_node->op_def().input_arg(on_slot_idx); |
| 2274 | if (ArgIsList(arg)) { |
| 2275 | std::vector<NodeBuilder::NodeOut> new_node_inputs; |
| 2276 | int tensor_list_length = GetTensorListLength(arg, old_node); |
| 2277 | if (tensor_list_length != 0) { |
| 2278 | GetNodesProducingTFTensorList(old_node_inputs, &iidx, |
| 2279 | tensor_list_length, &new_node_inputs); |
| 2280 | } |
| 2281 | nb->Input(new_node_inputs); |
| 2282 | nn_slot_idx++; |
| 2283 | } else { |
| 2284 | // Special case for connecting filter input of Conv2DBackpropInput |
| 2285 | if (do_connect_conv2d_backprop_input_filter && |
| 2286 | iidx == kConv2DBackpropInputFilterInputSlotIdx) { |
| 2287 | nb->Input(conv2d_node, kConv2DFilterOutputSlotIdx); |
| 2288 | } else { |
| 2289 | nb->Input(old_node_inputs[iidx].first, old_node_inputs[iidx].second); |
| 2290 | } |
| 2291 | iidx++; |
| 2292 | nn_slot_idx++; |
| 2293 | } |
| 2294 | } |
| 2295 | |
| 2296 | // If workspace tensors are available for this op and we are using |
| 2297 | // contiguous ordering then we need to add Tensorflow tensor for |
| 2298 | // workspace here because Tensorflow tensor for workspace is the |
| 2299 | // last tensor in the list of Tensorflow tensors. |
| 2300 | if (are_workspace_tensors_available) { |
| 2301 | CHECK_EQ(workspace_tensors->size(), 2); |
| 2302 | // Tensorflow tensor |
| 2303 | nb->Input((*workspace_tensors)[0].node, (*workspace_tensors)[0].index); |
| 2304 | nn_slot_idx++; |
| 2305 | } |
| 2306 | |
| 2307 | // Let's now setup all Mkl inputs to a new node. |
| 2308 | // Number of Mkl inputs must be same as number of TF inputs. |
| 2309 | iidx = 0; |
| 2310 | for (int on_slot_idx = 0; on_slot_idx < old_node_input_slots; on_slot_idx++) { |
| 2311 | // An input slot could be a single tensor or a list. We need |
| 2312 | // to handle this case accordingly. |
| 2313 | const OpDef::ArgDef& arg = old_node->op_def().input_arg(on_slot_idx); |
| 2314 | if (ArgIsList(arg)) { |
| 2315 | std::vector<NodeBuilder::NodeOut> new_node_inputs; |
| 2316 | int tensor_list_length = GetTensorListLength(arg, old_node); |
| 2317 | if (tensor_list_length != 0) { |
| 2318 | GetNodesProducingMklTensorList(g, old_node, old_node_inputs, &iidx, |
| 2319 | tensor_list_length, &new_node_inputs); |
| 2320 | } |
| 2321 | nb->Input(new_node_inputs); |
| 2322 | nn_slot_idx++; |
| 2323 | } else { |
| 2324 | Node* mkl_node = nullptr; |
| 2325 | int mkl_node_output_slot = 0; |
| 2326 | // Special case for connecting filter input of Conv2DBackpropInput |
| 2327 | if (do_connect_conv2d_backprop_input_filter && |
| 2328 | iidx == kConv2DBackpropInputFilterInputSlotIdx) { |
| 2329 | GetNodeProducingMklTensor(g, old_node, conv2d_node, |
| 2330 | kConv2DFilterOutputSlotIdx, &mkl_node, |
| 2331 | &mkl_node_output_slot); |
| 2332 | } else { |
| 2333 | GetNodeProducingMklTensor(g, old_node, old_node_inputs[iidx].first, |
| 2334 | old_node_inputs[iidx].second, &mkl_node, |
| 2335 | &mkl_node_output_slot); |
| 2336 | } |
| 2337 | nb->Input(mkl_node, mkl_node_output_slot); |
| 2338 | iidx++; |
| 2339 | nn_slot_idx++; |
| 2340 | } |
| 2341 | } |
| 2342 | |
| 2343 | // If workspace tensors are available for this op and we are using |
| 2344 | // contiguous ordering then we need to add Mkl tensor for |
| 2345 | // workspace here because Mkl tensor for workspace is the |
| 2346 | // last tensor in the list of Mkl tensors. |
| 2347 | if (are_workspace_tensors_available) { |
| 2348 | CHECK_EQ(workspace_tensors->size(), 2); |
| 2349 | // Mkl tensor |
| 2350 | nb->Input((*workspace_tensors)[1].node, (*workspace_tensors)[1].index); |
| 2351 | nn_slot_idx++; |
| 2352 | } |
| 2353 | |
| 2354 | return nn_slot_idx; |
| 2355 | } |
| 2356 | |
| 2357 | // This method finds out if checking workspace is needed or not. Workspace is |
| 2358 | // not used in quantized ops, so checking that would fail as quantized ops |
| 2359 | // don't have attribute: "T". |
| 2360 | bool IsWorkspaceCheckNeeded(const Node* node) { |
| 2361 | std::vector<string> quant_ops{ |
| 2362 | "Dequantize", |
| 2363 | "QuantizeV2", |
| 2364 | "QuantizedConv2D", |
| 2365 | "QuantizedConv2DWithBias", |
| 2366 | "QuantizedConv2DAndRelu", |
| 2367 | "QuantizedConv2DWithBiasAndRelu", |
| 2368 | "QuantizedConv2DWithBiasSumAndRelu", |
| 2369 | "QuantizedConv2DPerChannel", |
| 2370 | "QuantizedConv2DAndRequantize", |
| 2371 | "QuantizedConv2DWithBiasAndRequantize", |
| 2372 | "QuantizedConv2DAndReluAndRequantize", |
| 2373 | "QuantizedConv2DWithBiasAndReluAndRequantize", |
| 2374 | "QuantizedConv2DWithBiasSumAndReluAndRequantize", |
| 2375 | "QuantizedConv2DWithBiasSignedSumAndReluAndRequantize", |
| 2376 | "QuantizedMatMulWithBias", |
| 2377 | "QuantizedMatMulWithBiasAndRequantize", |
| 2378 | "QuantizedMatMulWithBiasAndDequantize", |
| 2379 | "QuantizedMatMulWithBiasAndRelu", |
| 2380 | "QuantizedMatMulWithBiasAndReluAndRequantize", |
| 2381 | "QuantizedDepthwiseConv2D", |
| 2382 | "QuantizedDepthwiseConv2DWithBias", |
| 2383 | "QuantizedDepthwiseConv2DWithBiasAndRelu", |
| 2384 | "QuantizedDepthwiseConv2DWithBiasAndReluAndRequantize"}; |
| 2385 | return std::find(std::begin(quant_ops), std::end(quant_ops), |
| 2386 | node->type_string()) == std::end(quant_ops); |
| 2387 | } |
| 2388 | |
| 2389 | Status MklLayoutRewritePass::SetUpInputs( |
| 2390 | std::unique_ptr<Graph>* g, |
| 2391 | const gtl::InlinedVector<std::pair<Node*, int>, 4>& old_node_inputs, |
| 2392 | NodeBuilder* nb, const Node* old_node) { |
| 2393 | // Let's check if we need to add workspace tensors for this node. |
| 2394 | // We add workspace edge only for MaxPool, LRN and BatchNorm. |
| 2395 | std::vector<NodeBuilder::NodeOut> workspace_tensors; |
| 2396 | bool are_workspace_tensors_available = false; |
| 2397 | |
| 2398 | if (IsWorkspaceCheckNeeded(old_node)) { |
| 2399 | AddWorkSpaceEdgeIfNeeded(g, old_node, nb, &workspace_tensors, |
| 2400 | &are_workspace_tensors_available); |
| 2401 | } |
| 2402 | |
| 2403 | int new_node_input_slots = 0; |
| 2404 | if (kTensorOrdering == MklTfTensorOrdering::TENSORS_INTERLEAVED) { |
| 2405 | // TODO(nhasabni): implement this function just for same of completion. |
| 2406 | // We do not use interleaved ordering right now. |
| 2407 | return Status( |
| 2408 | error::Code::UNIMPLEMENTED, |
| 2409 | "Interleaved ordering of tensors is currently not supported."); |
| 2410 | } else { |
| 2411 | CHECK_EQ(kTensorOrdering, MklTfTensorOrdering::TENSORS_CONTIGUOUS); |
| 2412 | new_node_input_slots = SetUpContiguousInputs( |
| 2413 | g, old_node_inputs, nb, old_node, &workspace_tensors, |
| 2414 | are_workspace_tensors_available); |
| 2415 | } |
| 2416 | |
| 2417 | // Sanity check |
| 2418 | int old_node_input_slots = old_node->op_def().input_arg_size(); |
| 2419 | if (!are_workspace_tensors_available) { |
| 2420 | // If we are not adding workspace tensors for this op, then the total |
| 2421 | // number of input slots to the new node _must_ be 2 times the number |
| 2422 | // of input slots to the original node: N original Tensorflow tensors and |
| 2423 | // N for Mkl tensors corresponding to each Tensorflow tensors. |
| 2424 | CHECK_EQ(new_node_input_slots, old_node_input_slots * 2); |
| 2425 | } else { |
| 2426 | // If we are adding workspace tensors for this op, then the total |
| 2427 | // The total number of input slots to new node _must_ be 2 times the number |
| 2428 | // of input slots to the original node: N original Tensorflow tensors and |
| 2429 | // N for Mkl tensors corresponding to each Tensorflow tensors plus 2 |
| 2430 | // (for workspace Tensorflow tensor and workspace Mkl tensor). |
| 2431 | CHECK_EQ(new_node_input_slots, old_node_input_slots * 2 + 2); |
| 2432 | } |
| 2433 | |
| 2434 | return OkStatus(); |
| 2435 | } |
| 2436 | |
| 2437 | Status MklLayoutRewritePass::CopyInputs( |
| 2438 | const Node* old_node, |
| 2439 | const gtl::InlinedVector<std::pair<Node*, int>, 4>& old_node_inputs, |
| 2440 | NodeBuilder* nb) { |
| 2441 | // Number of input slots to old node |
| 2442 | // Input slots are represented by .Input() calls in REGISTER_OP. |
| 2443 | int old_node_input_slots = old_node->op_def().input_arg_size(); |
| 2444 | // Actual number of inputs can be greater than or equal to number |
| 2445 | // of Input slots because inputs of type list could be unfolded. |
| 2446 | auto old_node_input_size = old_node_inputs.size(); |
| 2447 | |
| 2448 | if (old_node->type_string() == "_FusedConv2D") { |
| 2449 | // [TODO(intel-tf)] |
| 2450 | // commit 5be9a5 updates _FusedConv2D with additional host_args, |
| 2451 | // but mkl version currently doesn't have this input arg, needs to |
| 2452 | // remove this extra input when replace node with mkl node. |
| 2453 | old_node_input_slots--; |
| 2454 | } |
| 2455 | |
| 2456 | DCHECK_GE(old_node_input_size, old_node_input_slots); |
| 2457 | |
| 2458 | // Let's copy all inputs of old node to new node. |
| 2459 | int iidx = 0; |
| 2460 | for (int on_slot_idx = 0; on_slot_idx < old_node_input_slots; on_slot_idx++) { |
| 2461 | // An input slot could be a single tensor or a list. We need |
| 2462 | // to handle this case accordingly. |
| 2463 | DCHECK_LT(iidx, old_node_input_size); |
| 2464 | const OpDef::ArgDef& arg = old_node->op_def().input_arg(on_slot_idx); |
| 2465 | if (ArgIsList(arg)) { |
| 2466 | std::vector<NodeBuilder::NodeOut> new_node_inputs; |
| 2467 | int N = GetTensorListLength(arg, old_node); |
| 2468 | if (N != 0) { |
| 2469 | GetNodesProducingTFTensorList(old_node_inputs, &iidx, N, |
| 2470 | &new_node_inputs); |
| 2471 | } |
| 2472 | nb->Input(new_node_inputs); |
| 2473 | } else { |
| 2474 | nb->Input(old_node_inputs[iidx].first, old_node_inputs[iidx].second); |
| 2475 | iidx++; |
| 2476 | } |
| 2477 | } |
| 2478 | return OkStatus(); |
| 2479 | } |
| 2480 | |
| 2481 | ////////////////////////////////////////////////////////////////////////// |
| 2482 | // Helper functions related to workspace pass |
| 2483 | ////////////////////////////////////////////////////////////////////////// |
| 2484 | |
| 2485 | // TODO(nhasabni) We should move this to mkl_util.h. |
| 2486 | void MklLayoutRewritePass::GetDummyWorkspaceTensorNode( |
| 2487 | std::unique_ptr<Graph>* g, Node** out, const Node* orig_node) { |
| 2488 | // We use uint8 tensor of shape 8 with content {0,0,0,0,0,0,0,0} to represent |
| 2489 | // workspace tensor. |
| 2490 | GetDummyMklTensorNode(g, out, orig_node); |
| 2491 | } |
| 2492 | |
| 2493 | void MklLayoutRewritePass::AddWorkSpaceEdgeIfNeeded( |
| 2494 | std::unique_ptr<Graph>* g, const Node* orig_node, NodeBuilder* nb, |
| 2495 | std::vector<NodeBuilder::NodeOut>* ws_tensors, bool* are_ws_tensors_added) { |
| 2496 | bool workspace_edge_added = false; // Default initializer |
| 2497 | DCHECK(are_ws_tensors_added); |
| 2498 | *are_ws_tensors_added = false; // Default initializer |
| 2499 | |
| 2500 | DataType T; |
| 2501 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "T", &T)); |
| 2502 | for (auto ws : wsinfo_) { |
| 2503 | if (orig_node->type_string() == ws.fwd_op && |
| 2504 | mkl_op_registry::IsMklOp( |
| 2505 | mkl_op_registry::GetMklOpName(orig_node->type_string()), T)) { |
| 2506 | // If this op is a fwd op, then we need to check if there is an |
| 2507 | // edge from this node's fwd_slot to bwdop's bwd_slot. If there is |
| 2508 | // an edge, then we just add an attribute on this node for setting |
| 2509 | // workspace_passed to true. We don't add actual workspace edge |
| 2510 | // in this node. Actual workspace edge gets added in the backward |
| 2511 | // op for this node. |
| 2512 | for (const Edge* e : orig_node->out_edges()) { |
| 2513 | if (e->src_output() == ws.fwd_slot && |
| 2514 | e->dst()->type_string() == ws.bwd_op && |
| 2515 | e->dst_input() == ws.bwd_slot) { |
| 2516 | nb->Attr("workspace_enabled", true); |
| 2517 | VLOG(1) << "MklLayoutRewritePass: workspace_enabled for " |
| 2518 | << orig_node->type_string(); |
| 2519 | workspace_edge_added = true; |
| 2520 | // We found the edge that we were looking for, so break. |
| 2521 | break; |
| 2522 | } |
| 2523 | } |
| 2524 | |
| 2525 | if (!workspace_edge_added) { |
| 2526 | // If we are here, then we did not find backward operator for this |
| 2527 | // node. |
| 2528 | nb->Attr("workspace_enabled", false); |
| 2529 | } |
| 2530 | } else if (orig_node->type_string() == ws.bwd_op && |
| 2531 | mkl_op_registry::IsMklOp( |
| 2532 | mkl_op_registry::GetMklOpName(orig_node->type_string()), |
| 2533 | T)) { |
| 2534 | // If this op is a bwd op, then we need to add workspace edge and |
| 2535 | // it's Mkl tensor edge between its corresponding fwd op and this |
| 2536 | // op. Corresponding fwd op is specified in 'fwd_op' field of |
| 2537 | // workspace info. fwd_slot and bwd_slot in workspace info specify |
| 2538 | // an edge between which slots connect forward and backward op. |
| 2539 | // Once all these criteria match, we add a workspace edge between |
| 2540 | // ws_fwd_slot and ws_bwd_slot. Its corresponding Mkl tensor is |
| 2541 | // determined by interleaved/contiguous ordering. Function |
| 2542 | // DataIndexToMetaDataIndex tells us the location of Mkl tensor |
| 2543 | // from the location of the Tensorflow tensor. |
| 2544 | for (const Edge* e : orig_node->in_edges()) { |
| 2545 | if (e->src_output() == ws.fwd_slot && |
| 2546 | // We would have rewritten the forward op, so we need to use |
| 2547 | // GetMklOpName call to get its Mkl name. |
| 2548 | e->src()->type_string() == |
| 2549 | mkl_op_registry::GetMklOpName(ws.fwd_op) && |
| 2550 | e->dst_input() == ws.bwd_slot) { |
| 2551 | nb->Attr("workspace_enabled", true); |
| 2552 | DCHECK(ws_tensors); |
| 2553 | // Add workspace edge between fwd op and bwd op. |
| 2554 | ws_tensors->push_back(NodeBuilder::NodeOut(e->src(), ws.ws_fwd_slot)); |
| 2555 | // Check if we are running in native format mode. If so, |
| 2556 | // we don't need to have an Mkl metadata tensor for the workspace. |
| 2557 | if (!NativeFormatEnabled()) { |
| 2558 | // Add Mkl tensor edge for workspace edge between fwd op and bwd op. |
| 2559 | ws_tensors->push_back(NodeBuilder::NodeOut( |
| 2560 | e->src(), DataIndexToMetaDataIndex(ws.ws_fwd_slot, |
| 2561 | e->src()->num_outputs()))); |
| 2562 | } |
| 2563 | *are_ws_tensors_added = true; |
| 2564 | // In terms of input ordering, we add these calls to add Input |
| 2565 | // here because workspace edge (and its Mkl tensor) is the last |
| 2566 | // edge in the fwdop and bwdop. So all inputs before workspace |
| 2567 | // tensor have been added by SetUpInputs function. |
| 2568 | VLOG(1) << "MklLayoutRewritePass: workspace_enabled for " |
| 2569 | << orig_node->type_string(); |
| 2570 | workspace_edge_added = true; |
| 2571 | // We found the edge that we were looking for, so break. |
| 2572 | break; |
| 2573 | } |
| 2574 | } |
| 2575 | |
| 2576 | // If we are here means we did not find fwd op that feeds to this |
| 2577 | // bwd op. So in this case, we need to generate dummy tensors for |
| 2578 | // workspace input and Mkl tensor for workspace, and set |
| 2579 | // workspace_enabled to false. |
| 2580 | if (!workspace_edge_added) { |
| 2581 | nb->Attr("workspace_enabled", false); |
| 2582 | Node* dmt_ws = nullptr; // Dummy tensor for workspace |
| 2583 | Node* dmt_mkl_ws = nullptr; // Dummy Mkl tensor for workspace |
| 2584 | GetDummyWorkspaceTensorNode(g, &dmt_ws, orig_node); |
| 2585 | GetDummyMklTensorNode(g, &dmt_mkl_ws, orig_node); |
| 2586 | DCHECK(dmt_ws); |
| 2587 | DCHECK(dmt_mkl_ws); |
| 2588 | DCHECK(ws_tensors); |
| 2589 | // We add dummy tensor as workspace tensor. |
| 2590 | ws_tensors->push_back(NodeBuilder::NodeOut(dmt_ws, 0)); |
| 2591 | // We add dummy tensor as Mkl tensor for workspace tensor. |
| 2592 | ws_tensors->push_back(NodeBuilder::NodeOut(dmt_mkl_ws, 0)); |
| 2593 | *are_ws_tensors_added = true; |
| 2594 | VLOG(1) << "MklLayoutRewritePass: dummy workspace_enabled for " |
| 2595 | << orig_node->type_string(); |
| 2596 | } |
| 2597 | } else { |
| 2598 | // If this node does not match any workspace info, then we do not |
| 2599 | // do anything special for workspace propagation for it. |
| 2600 | } |
| 2601 | } |
| 2602 | } |
| 2603 | |
| 2604 | ////////////////////////////////////////////////////////////////////////// |
| 2605 | // Op-specific functions to copy attributes from old node to new node |
| 2606 | ////////////////////////////////////////////////////////////////////////// |
| 2607 | |
| 2608 | // Generic function to copy all attributes from original node to target. |
| 2609 | void MklLayoutRewritePass::CopyAttrsAll(const Node* orig_node, NodeBuilder* nb, |
| 2610 | bool change_format) { |
| 2611 | string name; |
| 2612 | AttrSlice attr_list(orig_node->def()); |
| 2613 | |
| 2614 | auto iter = attr_list.begin(); |
| 2615 | while (iter != attr_list.end()) { |
| 2616 | name = iter->first; |
| 2617 | auto attr = iter->second; |
| 2618 | nb->Attr(name, attr); |
| 2619 | ++iter; |
| 2620 | } |
| 2621 | } |
| 2622 | |
| 2623 | // Generic function to copy all attributes and check if filter is const. |
| 2624 | void MklLayoutRewritePass::CopyAttrsAllCheckConstFilter(const Node* orig_node, |
| 2625 | NodeBuilder* nb, |
| 2626 | bool change_format) { |
| 2627 | CopyAttrsAll(orig_node, nb, change_format); |
| 2628 | |
| 2629 | // Check and set filter attribute. |
| 2630 | Node* filter_node = nullptr; |
| 2631 | TF_CHECK_OK(orig_node->input_node(1, &filter_node)); |
| 2632 | |
| 2633 | bool is_filter_const = false; |
| 2634 | if (HasNodeAttr(orig_node->def(), "is_filter_const")) { |
| 2635 | (void)GetNodeAttr(orig_node->def(), "is_filter_const", &is_filter_const); |
| 2636 | } |
| 2637 | |
| 2638 | // In case that (1) orig_node does not have attribute 'is_filter_const', |
| 2639 | // or (2) it has the attribute but with the false value, then we set the |
| 2640 | // attribute for 'nb' with a value based on filter_node being const or not. |
| 2641 | // If is_filter_const == true, then there is no need to call nb->Attr() as |
| 2642 | // CopyAttrsAll() has already copied the attribute from orig_node to nb. |
| 2643 | if (!is_filter_const) { |
| 2644 | nb->Attr("is_filter_const", filter_node->IsConstant()); |
| 2645 | } |
| 2646 | } |
| 2647 | |
| 2648 | void MklLayoutRewritePass::CopyAttrsConvCheckConstFilter(const Node* orig_node, |
| 2649 | NodeBuilder* nb, |
| 2650 | bool change_format) { |
| 2651 | CopyAttrsConv(orig_node, nb, change_format); |
| 2652 | |
| 2653 | // Check and set filter attribute. |
| 2654 | Node* filter_node = nullptr; |
| 2655 | TF_CHECK_OK(orig_node->input_node(1, &filter_node)); |
| 2656 | nb->Attr("is_filter_const", filter_node->IsConstant()); |
| 2657 | } |
| 2658 | |
| 2659 | void MklLayoutRewritePass::CopyAttrsConv(const Node* orig_node, NodeBuilder* nb, |
| 2660 | bool change_format) { |
| 2661 | DataType T; |
| 2662 | string padding; |
| 2663 | std::vector<int32> strides; |
| 2664 | std::vector<int32> dilations; |
| 2665 | std::vector<int32> explicit_paddings; |
| 2666 | |
| 2667 | // Get all attributes from old node. |
| 2668 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "T", &T)); |
| 2669 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "strides", &strides)); |
| 2670 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "dilations", &dilations)); |
| 2671 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "padding", &padding)); |
| 2672 | |
| 2673 | // Check `explicit_paddings` first because some Conv ops don't have |
| 2674 | // this attribute. |
| 2675 | if (TryGetNodeAttr(orig_node->def(), "explicit_paddings", |
| 2676 | &explicit_paddings) && |
| 2677 | !explicit_paddings.empty()) { |
| 2678 | nb->Attr("explicit_paddings", explicit_paddings); |
| 2679 | } |
| 2680 | |
| 2681 | // Add attributes to new node. |
| 2682 | nb->Attr("T", T); |
| 2683 | nb->Attr("padding", padding); |
| 2684 | |
| 2685 | // Add attributes related to `data_format`. |
| 2686 | CopyFormatAttrsConv(orig_node, nb, strides, dilations, change_format); |
| 2687 | } |
| 2688 | |
| 2689 | // Used with MergePadWithConv2D |
| 2690 | void MklLayoutRewritePass::CopyAttrsFromPadAndConv2D(const Node* orig_node1, |
| 2691 | const Node* orig_node2, |
| 2692 | NodeBuilder* nb, |
| 2693 | bool change_format) { |
| 2694 | DataType Tpaddings; |
| 2695 | DataType T; |
| 2696 | string data_format; |
| 2697 | string padding; |
| 2698 | std::vector<int32> strides; |
| 2699 | std::vector<int32> dilations; |
| 2700 | bool use_cudnn_on_gpu; |
| 2701 | |
| 2702 | // Get all attributes from old node 1. |
| 2703 | TF_CHECK_OK(GetNodeAttr(orig_node1->def(), "T", &T)); |
| 2704 | TF_CHECK_OK(GetNodeAttr(orig_node1->def(), "strides", &strides)); |
| 2705 | TF_CHECK_OK(GetNodeAttr(orig_node1->def(), "dilations", &dilations)); |
| 2706 | TF_CHECK_OK(GetNodeAttr(orig_node1->def(), "padding", &padding)); |
| 2707 | TF_CHECK_OK(GetNodeAttr(orig_node1->def(), "data_format", &data_format)); |
| 2708 | TF_CHECK_OK( |
| 2709 | GetNodeAttr(orig_node1->def(), "use_cudnn_on_gpu", &use_cudnn_on_gpu)); |
| 2710 | // Get all attributes from old node 2. |
| 2711 | TF_CHECK_OK(GetNodeAttr(orig_node2->def(), "Tpaddings", &Tpaddings)); |
| 2712 | |
| 2713 | // Add attributes to new node. |
| 2714 | nb->Attr("T", T); |
| 2715 | nb->Attr("strides", strides); |
| 2716 | nb->Attr("dilations", dilations); |
| 2717 | nb->Attr("padding", padding); |
| 2718 | nb->Attr("data_format", data_format); |
| 2719 | nb->Attr("use_cudnn_on_gpu", use_cudnn_on_gpu); |
| 2720 | nb->Attr("Tpaddings", Tpaddings); |
| 2721 | } |
| 2722 | |
| 2723 | void MklLayoutRewritePass::CopyAttrsFusedConv2DCheckConstFilter( |
| 2724 | const Node* orig_node, NodeBuilder* nb, bool change_format) { |
| 2725 | DataType T; |
| 2726 | int num_args; |
| 2727 | string data_format; |
| 2728 | string padding; |
| 2729 | std::vector<int32> strides; |
| 2730 | std::vector<int32> dilations; |
| 2731 | std::vector<int32> explicit_paddings; |
| 2732 | float epsilon; |
| 2733 | std::vector<string> fused_ops; |
| 2734 | float leakyrelu_alpha; |
| 2735 | bool use_cudnn_on_gpu; |
| 2736 | |
| 2737 | // Get all attributes from old node. |
| 2738 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "T", &T)); |
| 2739 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "num_args", &num_args)); |
| 2740 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "strides", &strides)); |
| 2741 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "padding", &padding)); |
| 2742 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "data_format", &data_format)); |
| 2743 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "dilations", &dilations)); |
| 2744 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "fused_ops", &fused_ops)); |
| 2745 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "epsilon", &epsilon)); |
| 2746 | TF_CHECK_OK( |
| 2747 | GetNodeAttr(orig_node->def(), "leakyrelu_alpha", &leakyrelu_alpha)); |
| 2748 | TF_CHECK_OK( |
| 2749 | GetNodeAttr(orig_node->def(), "use_cudnn_on_gpu", &use_cudnn_on_gpu)); |
| 2750 | |
| 2751 | // Add attributes to new node. |
| 2752 | nb->Attr("T", T); |
| 2753 | nb->Attr("num_args", num_args); |
| 2754 | nb->Attr("strides", strides); |
| 2755 | nb->Attr("padding", padding); |
| 2756 | nb->Attr("data_format", data_format); |
| 2757 | nb->Attr("dilations", dilations); |
| 2758 | nb->Attr("epsilon", epsilon); |
| 2759 | nb->Attr("fused_ops", fused_ops); |
| 2760 | nb->Attr("leakyrelu_alpha", leakyrelu_alpha); |
| 2761 | nb->Attr("use_cudnn_on_gpu", use_cudnn_on_gpu); |
| 2762 | |
| 2763 | // Check `explicit_paddings` first because some Conv ops don't have |
| 2764 | // this attribute. |
| 2765 | if (TryGetNodeAttr(orig_node->def(), "explicit_paddings", |
| 2766 | &explicit_paddings) && |
| 2767 | !explicit_paddings.empty()) { |
| 2768 | nb->Attr("explicit_paddings", explicit_paddings); |
| 2769 | } |
| 2770 | |
| 2771 | // Check and set filter attribute. |
| 2772 | Node* filter_node = nullptr; |
| 2773 | TF_CHECK_OK(orig_node->input_node(1, &filter_node)); |
| 2774 | nb->Attr("is_filter_const", filter_node->IsConstant()); |
| 2775 | } |
| 2776 | |
| 2777 | void MklLayoutRewritePass::CopyAttrsFromPadAndFusedConv2D( |
| 2778 | const Node* fused_conv2d, const Node* pad, NodeBuilder* nb, |
| 2779 | bool change_format) { |
| 2780 | DataType T; |
| 2781 | int num_args; |
| 2782 | string data_format; |
| 2783 | string padding; |
| 2784 | std::vector<int32> strides; |
| 2785 | std::vector<int32> dilations; |
| 2786 | float epsilon; |
| 2787 | std::vector<string> fused_ops; |
| 2788 | DataType Tpaddings; |
| 2789 | float leakyrelu_alpha; |
| 2790 | |
| 2791 | // Get all attributes from old node. |
| 2792 | TF_CHECK_OK(GetNodeAttr(fused_conv2d->def(), "T", &T)); |
| 2793 | TF_CHECK_OK(GetNodeAttr(fused_conv2d->def(), "num_args", &num_args)); |
| 2794 | TF_CHECK_OK(GetNodeAttr(fused_conv2d->def(), "strides", &strides)); |
| 2795 | TF_CHECK_OK(GetNodeAttr(fused_conv2d->def(), "padding", &padding)); |
| 2796 | TF_CHECK_OK(GetNodeAttr(fused_conv2d->def(), "data_format", &data_format)); |
| 2797 | TF_CHECK_OK(GetNodeAttr(fused_conv2d->def(), "dilations", &dilations)); |
| 2798 | TF_CHECK_OK(GetNodeAttr(fused_conv2d->def(), "fused_ops", &fused_ops)); |
| 2799 | TF_CHECK_OK(GetNodeAttr(fused_conv2d->def(), "epsilon", &epsilon)); |
| 2800 | TF_CHECK_OK( |
| 2801 | GetNodeAttr(fused_conv2d->def(), "leakyrelu_alpha", &leakyrelu_alpha)); |
| 2802 | TF_CHECK_OK(GetNodeAttr(pad->def(), "Tpaddings", &Tpaddings)); |
| 2803 | |
| 2804 | // Add attributes to new node. |
| 2805 | nb->Attr("T", T); |
| 2806 | nb->Attr("num_args", num_args); |
| 2807 | nb->Attr("strides", strides); |
| 2808 | nb->Attr("padding", padding); |
| 2809 | nb->Attr("data_format", data_format); |
| 2810 | nb->Attr("dilations", dilations); |
| 2811 | nb->Attr("epsilon", epsilon); |
| 2812 | nb->Attr("Tpaddings", Tpaddings); |
| 2813 | nb->Attr("fused_ops", fused_ops); |
| 2814 | nb->Attr("leakyrelu_alpha", leakyrelu_alpha); |
| 2815 | } |
| 2816 | |
| 2817 | void MklLayoutRewritePass::CopyAttrsQuantizedConv2D(const Node* orig_node, |
| 2818 | NodeBuilder* nb, |
| 2819 | bool change_format) { |
| 2820 | DataType Tinput, Tfilter, out_type; |
| 2821 | string padding; |
| 2822 | string data_format("NHWC"); |
| 2823 | std::vector<int32> strides, dilations, padding_list; |
| 2824 | bool has_padding_list = HasNodeAttr(orig_node->def(), "padding_list"); |
| 2825 | |
| 2826 | // Get all attributes from old node. |
| 2827 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "Tinput", &Tinput)); |
| 2828 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "Tfilter", &Tfilter)); |
| 2829 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "out_type", &out_type)); |
| 2830 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "padding", &padding)); |
| 2831 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "strides", &strides)); |
| 2832 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "dilations", &dilations)); |
| 2833 | if (has_padding_list) { |
| 2834 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "padding_list", &padding_list)); |
| 2835 | } |
| 2836 | |
| 2837 | Node* filter_node = nullptr; |
| 2838 | TF_CHECK_OK(orig_node->input_node(1, &filter_node)); |
| 2839 | |
| 2840 | // Add attributes to new node. |
| 2841 | nb->Attr("Tinput", Tinput); |
| 2842 | nb->Attr("Tfilter", Tfilter); |
| 2843 | nb->Attr("out_type", out_type); |
| 2844 | nb->Attr("padding", padding); |
| 2845 | nb->Attr("is_filter_const", filter_node->IsConstant()); |
| 2846 | nb->Attr("strides", strides); |
| 2847 | nb->Attr("dilations", dilations); |
| 2848 | nb->Attr("data_format", data_format); |
| 2849 | if (has_padding_list) { |
| 2850 | nb->Attr("padding_list", padding_list); |
| 2851 | } |
| 2852 | |
| 2853 | // Requantization attr Tbias. |
| 2854 | DataType Tbias; |
| 2855 | Status bias_status = GetNodeAttr(orig_node->def(), "Tbias", &Tbias); |
| 2856 | if (bias_status.ToString() == "OK") nb->Attr( "Tbias", Tbias); |
| 2857 | } |
| 2858 | |
| 2859 | void MklLayoutRewritePass::CopyAttrsQuantizedMatMulWithBiasAndDequantize( |
| 2860 | const Node* orig_node, NodeBuilder* nb, bool change_format) { |
| 2861 | CopyAttrsAll(orig_node, nb, change_format); |
| 2862 | |
| 2863 | // Check and set filter attribute. |
| 2864 | Node* filter_node = nullptr; |
| 2865 | TF_CHECK_OK(orig_node->input_node(1, &filter_node)); |
| 2866 | nb->Attr("is_weight_const", filter_node->IsConstant()); |
| 2867 | } |
| 2868 | |
| 2869 | void MklLayoutRewritePass::CopyAttrsQuantizedMatMulWithBias( |
| 2870 | const Node* orig_node, NodeBuilder* nb, bool change_format) { |
| 2871 | DataType T1, T2, Toutput; |
| 2872 | |
| 2873 | // Get all attributes from old node. |
| 2874 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "T1", &T1)); |
| 2875 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "T2", &T2)); |
| 2876 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "Toutput", &Toutput)); |
| 2877 | |
| 2878 | Node* weight_node = nullptr; |
| 2879 | TF_CHECK_OK(orig_node->input_node(1, &weight_node)); |
| 2880 | |
| 2881 | // Add attributes to new node. |
| 2882 | nb->Attr("T1", T1); |
| 2883 | nb->Attr("T2", T2); |
| 2884 | nb->Attr("Toutput", Toutput); |
| 2885 | nb->Attr("is_weight_const", weight_node->IsConstant()); |
| 2886 | |
| 2887 | // Requantization attr Tbias |
| 2888 | DataType Tbias; |
| 2889 | Status bias_status = GetNodeAttr(orig_node->def(), "Tbias", &Tbias); |
| 2890 | if (bias_status.ToString() == "OK") nb->Attr( "Tbias", Tbias); |
| 2891 | } |
| 2892 | |
| 2893 | void MklLayoutRewritePass::CopyFormatAttrsConv( |
| 2894 | const Node* orig_node, NodeBuilder* nb, const std::vector<int32>& strides, |
| 2895 | const std::vector<int32>& dilations, bool change_format) { |
| 2896 | string data_format; |
| 2897 | |
| 2898 | if (!change_format) { |
| 2899 | nb->Attr("strides", strides); |
| 2900 | nb->Attr("dilations", dilations); |
| 2901 | |
| 2902 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "data_format", &data_format)); |
| 2903 | nb->Attr("data_format", data_format); |
| 2904 | } else { |
| 2905 | std::vector<int32> new_strides; |
| 2906 | std::vector<int32> new_dilations; |
| 2907 | if (strides.size() == 5) { |
| 2908 | // `strides` and `dilations` also need to be changed according to |
| 2909 | // `data_format`. In this case, from `NDHWC` to `NCDHW`. |
| 2910 | new_strides = {strides[NDHWC::dim::N], strides[NDHWC::dim::C], |
| 2911 | strides[NDHWC::dim::D], strides[NDHWC::dim::H], |
| 2912 | strides[NDHWC::dim::W]}; |
| 2913 | |
| 2914 | new_dilations = {dilations[NDHWC::dim::N], dilations[NDHWC::dim::C], |
| 2915 | dilations[NDHWC::dim::D], dilations[NDHWC::dim::H], |
| 2916 | dilations[NDHWC::dim::W]}; |
| 2917 | } else { |
| 2918 | // `strides` and `dilations` also need to be changed according to |
| 2919 | // `data_format`. In this case, from `NHWC` to `NCHW`. |
| 2920 | |
| 2921 | new_strides = {strides[NHWC::dim::N], strides[NHWC::dim::C], |
| 2922 | strides[NHWC::dim::H], strides[NHWC::dim::W]}; |
| 2923 | |
| 2924 | new_dilations = {dilations[NHWC::dim::N], dilations[NHWC::dim::C], |
| 2925 | dilations[NHWC::dim::H], dilations[NHWC::dim::W]}; |
| 2926 | } |
| 2927 | nb->Attr("strides", new_strides); |
| 2928 | nb->Attr("dilations", new_dilations); |
| 2929 | } |
| 2930 | } |
| 2931 | |
| 2932 | void MklLayoutRewritePass::CopyAttrsPooling(const Node* orig_node, |
| 2933 | NodeBuilder* nb, |
| 2934 | bool change_format) { |
| 2935 | DataType T; |
| 2936 | string data_format; |
| 2937 | string padding; |
| 2938 | std::vector<int32> ksize, strides; |
| 2939 | |
| 2940 | // Get all attributes from old node. |
| 2941 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "T", &T)); |
| 2942 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "ksize", &ksize)); |
| 2943 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "strides", &strides)); |
| 2944 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "padding", &padding)); |
| 2945 | TF_CHECK_OK(GetNodeAttr(orig_node->def(), "data_format", &data_format)); |
| 2946 | |
| 2947 | // Add attributes to new node. |
| 2948 | nb->Attr("T", T); |
| 2949 | nb->Attr("padding", padding); |
| 2950 | |
| 2951 | if (!change_format) { |
| 2952 | nb->Attr("strides", strides); |
| 2953 | nb->Attr("ksize", ksize); |
| 2954 | |
| 2955 | nb->Attr("data_format", data_format); |
| 2956 | } else { |
| 2957 | std::vector<int32> new_strides; |
| 2958 | std::vector<int32> new_ksize; |
| 2959 | if (strides.size() == 5) { |
| 2960 | DCHECK(data_format == "NCDHW"); |
| 2961 | // `strides` and `ksize` also need to be changed according to |
| 2962 | // `data_format`. In this case, from `NDHWC` to `NCDHW`. |
| 2963 | new_strides = {strides[NDHWC::dim::N], strides[NDHWC::dim::C], |
| 2964 | strides[NDHWC::dim::D], strides[NDHWC::dim::H], |
| 2965 | strides[NDHWC::dim::W]}; |
| 2966 | |
| 2967 | new_ksize = {ksize[NDHWC::dim::N], ksize[NDHWC::dim::C], |
| 2968 | ksize[NDHWC::dim::D], ksize[NDHWC::dim::H], |
| 2969 | ksize[NDHWC::dim::W]}; |
| 2970 | |
| 2971 | } else { |
| 2972 | // `strides` and `ksize` also need to be changed according to |
| 2973 | // `data_format`. In this case, from `NHWC` to `NCHW`. |
| 2974 | DCHECK(data_format == "NCHW"); |
| 2975 | new_strides = {strides[NHWC::dim::N], strides[NHWC::dim::C], |
| 2976 | strides[NHWC::dim::H], strides[NHWC::dim::W]}; |
| 2977 | |
| 2978 | new_ksize = {ksize[NHWC::dim::N], ksize[NHWC::dim::C], |
| 2979 | ksize[NHWC::dim::H], ksize[NHWC::dim::W]}; |
| 2980 | } |
| 2981 | nb->Attr("strides", new_strides); |
| 2982 | nb->Attr("ksize", new_ksize); |
| 2983 | } |
| 2984 | } |
| 2985 | |
| 2986 | ////////////////////////////////////////////////////////////////////////// |
| 2987 | // Helper functions related to node merge pass |
| 2988 | ////////////////////////////////////////////////////////////////////////// |
| 2989 | |
| 2990 | Node* MklLayoutRewritePass::CheckForNodeMerge(const Node* a) const { |
| 2991 | // TODO(nhasabni) Add check for type of node similar to CheckForNodeRewrite |
| 2992 | // once we support BiasAddGrad as Mkl layer. |
| 2993 | |
| 2994 | // Search for all matching mergeinfo. |
| 2995 | // We allow more than one match for extensibility. |
| 2996 | std::vector<const MergeInfo*> matching_mi; |
| 2997 | for (auto mi = minfo_.cbegin(); mi != minfo_.cend(); ++mi) { |
| 2998 | if (a->type_string() == mi->op1 || a->type_string() == mi->op2) { |
| 2999 | matching_mi.push_back(&*mi); |
| 3000 | } |
| 3001 | } |
| 3002 | |
| 3003 | for (const MergeInfo* mi : matching_mi) { |
| 3004 | // Get the operand with which 'a' can be merged. |
| 3005 | Node* b = nullptr; |
| 3006 | if ((b = mi->get_node_to_be_merged(a)) == nullptr) { |
| 3007 | continue; |
| 3008 | } |
| 3009 | |
| 3010 | // Get the control edges and input of node |
| 3011 | const int N_in = a->num_inputs(); |
| 3012 | gtl::InlinedVector<Node*, 4> a_control_edges; |
| 3013 | gtl::InlinedVector<std::pair<Node*, int>, 4> a_in(N_in); |
| 3014 | FillInputs(a, &a_control_edges, &a_in); |
| 3015 | |
| 3016 | const int B_in = b->num_inputs(); |
| 3017 | gtl::InlinedVector<Node*, 4> b_control_edges; |
| 3018 | gtl::InlinedVector<std::pair<Node*, int>, 4> b_in(B_in); |
| 3019 | FillInputs(b, &b_control_edges, &b_in); |
| 3020 | |
| 3021 | // Shouldn't merge if a and b have different control edges. |
| 3022 | if (a_control_edges != b_control_edges) { |
| 3023 | continue; |
| 3024 | } else { |
| 3025 | // We found a match. |
| 3026 | return b; |
| 3027 | } |
| 3028 | } |
| 3029 | |
| 3030 | return nullptr; |
| 3031 | } |
| 3032 | |
| 3033 | Status MklLayoutRewritePass::MergeConv2DWithBiasAdd(std::unique_ptr<Graph>* g, |
| 3034 | Node* m, Node* n) { |
| 3035 | CHECK_EQ(((m->type_string() == csinfo_.bias_add && |
| 3036 | n->type_string() == csinfo_.conv2d)) || |
| 3037 | ((n->type_string() == csinfo_.bias_add && |
| 3038 | m->type_string() == csinfo_.conv2d)), |
| 3039 | true); |
| 3040 | |
| 3041 | // If 'm' is BiasAdd, then 'n' is Conv2D. Since Conv2D feeds BiasAdd, |
| 3042 | // BiasAdd is successor node, and Conv2D predecessor node. |
| 3043 | Node* pred = m->type_string() == csinfo_.bias_add ? n : m; |
| 3044 | Node* succ = m->type_string() == csinfo_.bias_add ? m : n; |
| 3045 | |
| 3046 | // 1. Get all attributes from input nodes. |
| 3047 | DataType T_pred, T_succ; |
| 3048 | string padding; |
| 3049 | std::vector<int32> strides; |
| 3050 | std::vector<int32> dilations; |
| 3051 | string data_format_pred, data_format_succ; |
| 3052 | bool use_cudnn_on_gpu; |
| 3053 | TF_CHECK_OK(GetNodeAttr(pred->def(), "T", &T_pred)); |
| 3054 | TF_CHECK_OK(GetNodeAttr(succ->def(), "T", &T_succ)); |
| 3055 | TF_CHECK_OK(GetNodeAttr(pred->def(), "padding", &padding)); |
| 3056 | TF_CHECK_OK(GetNodeAttr(pred->def(), "strides", &strides)); |
| 3057 | TF_CHECK_OK(GetNodeAttr(pred->def(), "dilations", &dilations)); |
| 3058 | TF_CHECK_OK(GetNodeAttr(pred->def(), "data_format", &data_format_pred)); |
| 3059 | TF_CHECK_OK(GetNodeAttr(succ->def(), "data_format", &data_format_succ)); |
| 3060 | TF_CHECK_OK(GetNodeAttr(pred->def(), "use_cudnn_on_gpu", &use_cudnn_on_gpu)); |
| 3061 | // We check to ensure that data formats of both succ and pred are same. |
| 3062 | // We expect them to be same, so we can enforce this as assert. |
| 3063 | // But assert can be too strict, so we enforce this as a check. |
| 3064 | // If the check fails, then we do not merge two nodes. |
| 3065 | // We also do same check for devices. |
| 3066 | if (data_format_pred != data_format_succ || T_pred != T_succ || |
| 3067 | pred->assigned_device_name() != succ->assigned_device_name() || |
| 3068 | pred->def().device() != succ->def().device()) { |
| 3069 | return Status(error::Code::INVALID_ARGUMENT, |
| 3070 | "data_format or T attribute or devices of Conv2D and " |
| 3071 | "BiasAdd do not match. Will skip node merge optimization"); |
| 3072 | } |
| 3073 | |
| 3074 | const int succ_num = succ->num_inputs(); |
| 3075 | gtl::InlinedVector<Node*, 4> succ_control_edges; |
| 3076 | gtl::InlinedVector<std::pair<Node*, int>, 4> succ_in(succ_num); |
| 3077 | FillInputs(succ, &succ_control_edges, &succ_in); |
| 3078 | |
| 3079 | const int pred_num = pred->num_inputs(); |
| 3080 | gtl::InlinedVector<Node*, 4> pred_control_edges; |
| 3081 | gtl::InlinedVector<std::pair<Node*, int>, 4> pred_in(pred_num); |
| 3082 | FillInputs(pred, &pred_control_edges, &pred_in); |
| 3083 | |
| 3084 | // We need to ensure that Conv2D only feeds to BiasAdd (some other operator is |
| 3085 | // not expecting output of Conv2D). If this is not the case, then we cannot |
| 3086 | // merge Conv2D with BiasAdd. |
| 3087 | const int kFirstOutputSlot = 0; |
| 3088 | for (const Edge* e : pred->out_edges()) { |
| 3089 | if (e->src_output() == kFirstOutputSlot && e->dst() != succ) { |
| 3090 | return Status(error::Code::INVALID_ARGUMENT, |
| 3091 | "Conv2D does not feed to BiasAdd, or " |
| 3092 | "it feeds BiasAdd but has multiple outputs. " |
| 3093 | "Will skip node merge optimization"); |
| 3094 | } |
| 3095 | } |
| 3096 | |
| 3097 | // 2. Get inputs from both the nodes. |
| 3098 | // Find the 2 inputs from the conv and the bias from the add Bias. |
| 3099 | // Get operand 0, 1 of conv2D. |
| 3100 | CHECK_EQ(pred->in_edges().size(), 2); // Conv2D must have 2 inputs. |
| 3101 | // Get operand 1 of add_bias |
| 3102 | // BiasAdd must have 2 inputs: Conv, bias |
| 3103 | CHECK_EQ(succ->in_edges().size(), 2); |
| 3104 | |
| 3105 | // We will use the node name of BiasAdd as the name of new node |
| 3106 | // Build new node. We use same name as original node, but change the op |
| 3107 | // name. |
| 3108 | NodeBuilder nb(succ->name(), csinfo_.conv2d_with_bias); |
| 3109 | nb.Input(pred_in[0].first, pred_in[0].second); // In1 of Conv2D |
| 3110 | // pred_in[1] will be 2nd Tensorflow tensor for Conv2D. |
| 3111 | nb.Input(pred_in[1].first, pred_in[1].second); // In2 of Conv2D |
| 3112 | // In1 of BiasAdd is same as output of Conv2D. |
| 3113 | nb.Input(succ_in[1].first, succ_in[1].second); // In2 of BiasAdd |
| 3114 | |
| 3115 | // Copy attributes from Conv2D to Conv2DWithBias. |
| 3116 | CopyAttrsConvCheckConstFilter(const_cast<const Node*>(pred), &nb); |
| 3117 | |
| 3118 | // Copy the device assigned to old node to new node. |
| 3119 | nb.Device(succ->def().device()); |
| 3120 | |
| 3121 | // Create node. |
| 3122 | Node* new_node; |
| 3123 | TF_CHECK_OK(nb.Finalize(&**g, &new_node)); |
| 3124 | |
| 3125 | // In the following code of this function, an unsorted set is used to make |
| 3126 | // sure no duplicated edges be added into the new node. Therefore, we can |
| 3127 | // pass allow_duplicates = true in AddControlEdge call to skip the O(#edges) |
| 3128 | // check in the routine. |
| 3129 | |
| 3130 | // Incoming data edges from 'pred' node and 'succ' node to new 'new_node' |
| 3131 | // node are already copied in BuildNode. We handle control edges now. |
| 3132 | std::unordered_set<Node*> unique_node; |
| 3133 | for (const Edge* e : pred->in_edges()) { |
| 3134 | if (e->IsControlEdge()) { |
| 3135 | auto result = unique_node.insert(e->src()); |
| 3136 | if (result.second) { |
| 3137 | (*g)->AddControlEdge(e->src(), new_node, true); |
| 3138 | } |
| 3139 | } |
| 3140 | } |
| 3141 | unique_node.clear(); |
| 3142 | |
| 3143 | for (const Edge* e : succ->in_edges()) { |
| 3144 | if (e->IsControlEdge()) { |
| 3145 | auto result = unique_node.insert(e->src()); |
| 3146 | if (result.second) { |
| 3147 | (*g)->AddControlEdge(e->src(), new_node, true); |
| 3148 | } |
| 3149 | } |
| 3150 | } |
| 3151 | unique_node.clear(); |
| 3152 | |
| 3153 | // Incoming edges are fixed, we will fix the outgoing edges now. |
| 3154 | // First, we will fix outgoing control edges from 'pred' node. |
| 3155 | for (const Edge* e : pred->out_edges()) { |
| 3156 | if (e->IsControlEdge()) { |
| 3157 | auto result = unique_node.insert(e->dst()); |
| 3158 | if (result.second) { |
| 3159 | (*g)->AddControlEdge(new_node, e->dst(), true); |
| 3160 | } |
| 3161 | } |
| 3162 | } |
| 3163 | unique_node.clear(); |
| 3164 | |
| 3165 | // Second, we will fix outgoing control and data edges from 'succ' node. |
| 3166 | for (const Edge* e : succ->out_edges()) { |
| 3167 | if (e->IsControlEdge()) { |
| 3168 | auto result = unique_node.insert(e->dst()); |
| 3169 | if (result.second) { |
| 3170 | (*g)->AddControlEdge(new_node, e->dst(), true); |
| 3171 | } |
| 3172 | } else { |
| 3173 | // BiasAdd has only 1 output (at slot 0) and merged node also has only 1 |
| 3174 | // output (at slot 0). |
| 3175 | const int kConv2DWithBiasOutputSlot = 0; |
| 3176 | auto new_edge = (*g)->AddEdge(new_node, kConv2DWithBiasOutputSlot, |
| 3177 | e->dst(), e->dst_input()); |
| 3178 | DCHECK(new_edge); |
| 3179 | } |
| 3180 | } |
| 3181 | |
| 3182 | // Copy device assigned to old node to new node. |
| 3183 | // It's ok to use pred or succ as we have enforced a check that |
| 3184 | // both have same device assigned. |
| 3185 | new_node->set_assigned_device_name(pred->assigned_device_name()); |
| 3186 | |
| 3187 | VLOG(1) << "MklLayoutRewritePass: Merged old node:"<< pred->DebugString() |
| 3188 | << ", and node: "<< succ->DebugString() |
| 3189 | << ", into node:"<< new_node->DebugString(); |
| 3190 | |
| 3191 | (*g)->RemoveNode(succ); |
| 3192 | (*g)->RemoveNode(pred); |
| 3193 | |
| 3194 | return OkStatus(); |
| 3195 | } |
| 3196 | |
| 3197 | Status MklLayoutRewritePass::MergePadWithConv2D(std::unique_ptr<Graph>* g, |
| 3198 | Node* m, Node* n) { |
| 3199 | DCHECK((m->type_string() == csinfo_.pad && |
| 3200 | (n->type_string() == csinfo_.conv2d || |
| 3201 | n->type_string() == csinfo_.fused_conv2d)) || |
| 3202 | (n->type_string() == csinfo_.pad && |
| 3203 | (m->type_string() == csinfo_.conv2d || |
| 3204 | m->type_string() == csinfo_.fused_conv2d))); |
| 3205 | |
| 3206 | bool is_fused_conv2d = n->type_string() == csinfo_.fused_conv2d || |
| 3207 | m->type_string() == csinfo_.fused_conv2d; |
| 3208 | // Conv2D is successor node, and Pad predecessor node. |
| 3209 | Node* pred = m->type_string() == csinfo_.pad ? m : n; |
| 3210 | Node* succ = m->type_string() == csinfo_.pad ? n : m; |
| 3211 | |
| 3212 | // 1. Get all attributes from input nodes. |
| 3213 | DataType T_pred, T_succ; |
| 3214 | string padding; |
| 3215 | std::vector<int32> strides; |
| 3216 | std::vector<int32> dilations; |
| 3217 | string data_format_pred, data_format_succ; |
| 3218 | |
| 3219 | TF_CHECK_OK(GetNodeAttr(pred->def(), "T", &T_pred)); |
| 3220 | TF_CHECK_OK(GetNodeAttr(succ->def(), "T", &T_succ)); |
| 3221 | TF_CHECK_OK(GetNodeAttr(succ->def(), "padding", &padding)); |
| 3222 | TF_CHECK_OK(GetNodeAttr(succ->def(), "strides", &strides)); |
| 3223 | TF_CHECK_OK(GetNodeAttr(succ->def(), "dilations", &dilations)); |
| 3224 | // Check if the devices of both succ and pred are the same. |
| 3225 | // Assert is not used because it can be too strict. |
| 3226 | // Don't need to check for data formats because it is not available in Pad. |
| 3227 | if (T_pred != T_succ || |
| 3228 | pred->assigned_device_name() != succ->assigned_device_name() || |
| 3229 | pred->def().device() != succ->def().device()) { |
| 3230 | return Status(error::Code::INVALID_ARGUMENT, |
| 3231 | "T attribute or devices of Conv2D and " |
| 3232 | "Pad do not match. Will skip node merge optimization"); |
| 3233 | } |
| 3234 | |
| 3235 | const int succ_num = succ->num_inputs(); |
| 3236 | gtl::InlinedVector<Node*, 4> succ_control_edges; |
| 3237 | gtl::InlinedVector<std::pair<Node*, int>, 4> succ_in(succ_num); |
| 3238 | FillInputs(succ, &succ_control_edges, &succ_in); |
| 3239 | |
| 3240 | const int pred_num = pred->num_inputs(); |
| 3241 | gtl::InlinedVector<Node*, 4> pred_control_edges; |
| 3242 | gtl::InlinedVector<std::pair<Node*, int>, 4> pred_in(pred_num); |
| 3243 | FillInputs(pred, &pred_control_edges, &pred_in); |
| 3244 | |
| 3245 | // We need to ensure that Pad only feeds to Conv2D (some other operator is |
| 3246 | // not expecting output of Pad). If this is not the case, then we cannot |
| 3247 | // merge Conv2D with Pad. |
| 3248 | const int kFirstOutputSlot = 0; |
| 3249 | for (const Edge* e : pred->out_edges()) { |
| 3250 | if (e->src_output() == kFirstOutputSlot && e->dst() != succ) { |
| 3251 | return Status(error::Code::INVALID_ARGUMENT, |
| 3252 | "Pad does not feed to Conv2D, or " |
| 3253 | "it feeds Conv2D but has multiple outputs. " |
| 3254 | "Will skip node merge optimization"); |
| 3255 | } |
| 3256 | } |
| 3257 | |
| 3258 | // 2. Get inputs from both the nodes. |
| 3259 | |
| 3260 | // Pad must have 2 data inputs: "input" and paddings. |
| 3261 | int PadDataInputEdges = 0; |
| 3262 | for (const Edge* e : pred->in_edges()) { |
| 3263 | if (!e->IsControlEdge()) { |
| 3264 | PadDataInputEdges++; |
| 3265 | } |
| 3266 | } |
| 3267 | DCHECK_EQ(PadDataInputEdges, 2); |
| 3268 | |
| 3269 | // Conv2D must have 2 data inputs: Pad output and Filter |
| 3270 | // FusedConv2D have 3 data inputs: Pad output, Filter and Args; |
| 3271 | int ConvDataInputEdges = 0; |
| 3272 | for (const Edge* e : succ->in_edges()) { |
| 3273 | if (!e->IsControlEdge()) { |
| 3274 | ConvDataInputEdges++; |
| 3275 | } |
| 3276 | } |
| 3277 | |
| 3278 | DCHECK_EQ(ConvDataInputEdges, is_fused_conv2d ? 3 : 2); |
| 3279 | |
| 3280 | // We will use the node name of Conv2D as the name of new node |
| 3281 | // Build new node. We use same name as original node, but change the op |
| 3282 | // name. |
| 3283 | |
| 3284 | NodeBuilder nb(succ->name(), is_fused_conv2d ? csinfo_.pad_with_fused_conv2d |
| 3285 | : csinfo_.pad_with_conv2d); |
| 3286 | nb.Input(pred_in[0].first, pred_in[0].second); // In1 (input data) of Pad |
| 3287 | // pred_in[1] will be 2nd Tensorflow tensor for Conv2D. |
| 3288 | nb.Input(succ_in[1].first, succ_in[1].second); // In2 (filter) of conv2d |
| 3289 | // In1 of Conv2D is same as output of Pad. |
| 3290 | // Thus, only need to add In2 of Conv2D |
| 3291 | |
| 3292 | if (is_fused_conv2d) { |
| 3293 | // FusedConv2D has one additional input, args |
| 3294 | std::vector<NodeBuilder::NodeOut> args; |
| 3295 | int num_args = 1; |
| 3296 | TF_CHECK_OK(GetNodeAttr(succ->def(), "num_args", &num_args)); |
| 3297 | for (int i = 0; i < num_args; i++) { |
| 3298 | args.emplace_back(succ_in[2 + i].first, succ_in[2 + i].second); |
| 3299 | } |
| 3300 | nb.Input(gtl::ArraySlice<NodeBuilder::NodeOut>{ |
| 3301 | args}); // In3 (args) of FusedConv2D |
| 3302 | nb.Input(pred_in[1].first, pred_in[1].second); // In2 (paddings) of Pad |
| 3303 | // Copy attributes from Pad and FusedConv2D to PadWithFusedConv2D. |
| 3304 | CopyAttrsFromPadAndFusedConv2D(const_cast<const Node*>(succ), |
| 3305 | const_cast<const Node*>(pred), &nb); |
| 3306 | } else { |
| 3307 | nb.Input(pred_in[1].first, pred_in[1].second); // In2 (paddings) of Pad |
| 3308 | // Copy attributes from Pad and conv2D to PadWithConv2D. |
| 3309 | CopyAttrsFromPadAndConv2D(const_cast<const Node*>(succ), |
| 3310 | const_cast<const Node*>(pred), &nb); |
| 3311 | } |
| 3312 | |
| 3313 | // Copy the device assigned to old node to new node. |
| 3314 | nb.Device(succ->def().device()); |
| 3315 | |
| 3316 | // Create node. |
| 3317 | Node* new_node; |
| 3318 | TF_CHECK_OK(nb.Finalize(&**g, &new_node)); |
| 3319 | // No need to check if new_node is null because it will be null only when |
| 3320 | // Finalize fails. |
| 3321 | |
| 3322 | // Incoming data edges from 'pred' node and 'succ' node to new 'new_node' |
| 3323 | // node are already copied in BuildNode. |
| 3324 | // We handle control edges now. |
| 3325 | for (const Edge* e : pred->in_edges()) { |
| 3326 | if (e->IsControlEdge()) { |
| 3327 | // Don't allow duplicate edge |
| 3328 | (*g)->AddControlEdge(e->src(), new_node, false); |
| 3329 | } |
| 3330 | } |
| 3331 | for (const Edge* e : succ->in_edges()) { |
| 3332 | if (e->IsControlEdge()) { |
| 3333 | // Don't allow duplicate edge |
| 3334 | (*g)->AddControlEdge(e->src(), new_node, false); |
| 3335 | } |
| 3336 | } |
| 3337 | |
| 3338 | // Incoming edges are fixed, we will fix the outgoing edges now. |
| 3339 | // First, we will fix outgoing control edges from 'pred' node. |
| 3340 | for (const Edge* e : pred->out_edges()) { |
| 3341 | if (e->IsControlEdge()) { |
| 3342 | // Don't allow duplicate edge |
| 3343 | (*g)->AddControlEdge(new_node, e->dst(), false); |
| 3344 | } |
| 3345 | } |
| 3346 | |
| 3347 | // Second, we will fix outgoing control and data edges from 'succ' node. |
| 3348 | for (const Edge* e : succ->out_edges()) { |
| 3349 | if (e->IsControlEdge()) { |
| 3350 | // Allow duplicate while adding control edge as it would fail (return |
| 3351 | // NULL) if we try to add duplicate edge. |
| 3352 | (*g)->AddControlEdge(new_node, e->dst(), false); |
| 3353 | } else { |
| 3354 | // Conv2D has only 1 output (at slot 0) and merged node also has only 1 |
| 3355 | // output (at slot 0). |
| 3356 | const int kPadWithConv2DOutputSlot = 0; |
| 3357 | (*g)->AddEdge(new_node, kPadWithConv2DOutputSlot, e->dst(), |
| 3358 | e->dst_input()); |
| 3359 | } |
| 3360 | } |
| 3361 | |
| 3362 | // Copy device assigned to old node to new node. |
| 3363 | // It's ok to use pred or succ as we have enforced a check that |
| 3364 | // both have same device assigned. |
| 3365 | new_node->set_assigned_device_name(pred->assigned_device_name()); |
| 3366 | |
| 3367 | VLOG(1) << "MklLayoutRewritePass: Merged old node:"<< pred->DebugString() |
| 3368 | << ", and node: "<< succ->DebugString() |
| 3369 | << ", into node:"<< new_node->DebugString(); |
| 3370 | |
| 3371 | (*g)->RemoveNode(succ); |
| 3372 | (*g)->RemoveNode(pred); |
| 3373 | |
| 3374 | return OkStatus(); |
| 3375 | } |
| 3376 | |
| 3377 | Status MklLayoutRewritePass::MergeConv2DBackpropFilterWithBiasAddGrad( |
| 3378 | std::unique_ptr<Graph>* g, Node* m, Node* n) { |
| 3379 | CHECK_EQ(((m->type_string() == csinfo_.bias_add_grad && |
| 3380 | n->type_string() == csinfo_.conv2d_grad_filter)) || |
| 3381 | ((n->type_string() == csinfo_.bias_add_grad && |
| 3382 | m->type_string() == csinfo_.conv2d_grad_filter)), |
| 3383 | true); |
| 3384 | |
| 3385 | // If 'm' is BiasAddGrad, then 'n' is BackpropFilter. |
| 3386 | Node* badd = m->type_string() == csinfo_.bias_add_grad ? m : n; |
| 3387 | Node* fltr = m->type_string() == csinfo_.bias_add_grad ? n : m; |
| 3388 | |
| 3389 | // Sanity check for attributes from input nodes. |
| 3390 | DataType T_b, T_f; |
| 3391 | string data_format_b, data_format_f; |
| 3392 | TF_CHECK_OK(GetNodeAttr(badd->def(), "T", &T_b)); |
| 3393 | TF_CHECK_OK(GetNodeAttr(fltr->def(), "T", &T_f)); |
| 3394 | TF_CHECK_OK(GetNodeAttr(badd->def(), "data_format", &data_format_b)); |
| 3395 | TF_CHECK_OK(GetNodeAttr(fltr->def(), "data_format", &data_format_f)); |
| 3396 | if (data_format_b != data_format_f || T_b != T_f || |
| 3397 | badd->assigned_device_name() != fltr->assigned_device_name() || |
| 3398 | badd->def().device() != fltr->def().device()) { |
| 3399 | return Status(error::Code::INVALID_ARGUMENT, |
| 3400 | "data_format or T attribute or devices of " |
| 3401 | "Conv2DBackpropFilter and BiasAddGrad do not match. " |
| 3402 | "Will skip node merge optimization"); |
| 3403 | } |
| 3404 | |
| 3405 | // We will use the node name of Conv2DBackpropFilter as the name of new node. |
| 3406 | // This is because BackpropFilterWithBias is going to emit bias output also. |
| 3407 | NodeBuilder nb(fltr->name(), csinfo_.conv2d_grad_filter_with_bias); |
| 3408 | // Since Conv2DBackpropFilterWithBias has same number of inputs as |
| 3409 | // Conv2DBackpropFilter, we can just copy input edges directly. We don't need |
| 3410 | // to copy any data input of BiasAddGrad because that input also goes to |
| 3411 | // Conv2DBackpropFilter. |
| 3412 | const int fltr_ins = fltr->num_inputs(); |
| 3413 | gtl::InlinedVector<Node*, 4> fltr_control_edges; |
| 3414 | gtl::InlinedVector<std::pair<Node*, int>, 4> fltr_in_edges(fltr_ins); |
| 3415 | FillInputs(fltr, &fltr_control_edges, &fltr_in_edges); |
| 3416 | for (int idx = 0; idx < fltr_ins; idx++) { |
| 3417 | nb.Input(fltr_in_edges[idx].first, fltr_in_edges[idx].second); |
| 3418 | } |
| 3419 | |
| 3420 | // Copy attributes from Conv2DBackpropFilter. |
| 3421 | CopyAttrsConv(const_cast<const Node*>(fltr), &nb); |
| 3422 | |
| 3423 | // Copy the device assigned to old node to new node. |
| 3424 | nb.Device(fltr->def().device()); |
| 3425 | |
| 3426 | // Create node. |
| 3427 | Node* new_node; |
| 3428 | TF_CHECK_OK(nb.Finalize(&**g, &new_node)); |
| 3429 | |
| 3430 | // In the following code of this function, an unsorted set is used to make |
| 3431 | // sure no duplicated edges be added into the new node. Therefore, we can |
| 3432 | // pass allow_duplicates = true in AddControlEdge call to skip the O(#edges) |
| 3433 | // check in the routine. |
| 3434 | |
| 3435 | // Incoming data edges from BiasAddGrad node and Conv2DBackpropFilter node to |
| 3436 | // new 'new_node' node are already copied in BuildNode. We handle control |
| 3437 | // edges now. |
| 3438 | std::unordered_set<Node*> unique_node; |
| 3439 | for (const Edge* e : badd->in_edges()) { |
| 3440 | if (e->IsControlEdge()) { |
| 3441 | auto result = unique_node.insert(e->src()); |
| 3442 | if (result.second) { |
| 3443 | (*g)->AddControlEdge(e->src(), new_node, true); |
| 3444 | } |
| 3445 | } |
| 3446 | } |
| 3447 | unique_node.clear(); |
| 3448 | for (const Edge* e : fltr->in_edges()) { |
| 3449 | if (e->IsControlEdge()) { |
| 3450 | auto result = unique_node.insert(e->src()); |
| 3451 | if (result.second) { |
| 3452 | (*g)->AddControlEdge(e->src(), new_node, true); |
| 3453 | } |
| 3454 | } |
| 3455 | } |
| 3456 | unique_node.clear(); |
| 3457 | |
| 3458 | // Incoming edges are fixed, we will fix the outgoing edges now. |
| 3459 | // First, we will fix outgoing control edges from 'badd' node. |
| 3460 | // Conv2DBackpropFilter has 1 output -- filter_grad. |
| 3461 | // Conv2DBackpropFilterWithBias has 2 outputs -- filter_grad and |
| 3462 | // bias_grad. But filter_grad is at same slot number (0) in both the |
| 3463 | // nodes. bias_grad is at slot number 1 in Conv2DBackpropFilterWithBias, while |
| 3464 | // it is at slot number 0 in BiasAddGrad. |
| 3465 | const int kMergedNodeFilterGradOutputIdx = 0; |
| 3466 | const int kMergedNodeBiasGradOutputIdx = 1; |
| 3467 | |
| 3468 | for (const Edge* e : badd->out_edges()) { |
| 3469 | if (e->IsControlEdge()) { |
| 3470 | auto result = unique_node.insert(e->dst()); |
| 3471 | if (result.second) { |
| 3472 | (*g)->AddControlEdge(new_node, e->dst(), true); |
| 3473 | } |
| 3474 | } else { |
| 3475 | auto new_edge = (*g)->AddEdge(new_node, kMergedNodeBiasGradOutputIdx, |
| 3476 | e->dst(), e->dst_input()); |
| 3477 | DCHECK(new_edge); |
| 3478 | } |
| 3479 | } |
| 3480 | unique_node.clear(); |
| 3481 | |
| 3482 | // Second, we will fix outgoing control and data edges from 'fltr' node. |
| 3483 | for (const Edge* e : fltr->out_edges()) { |
| 3484 | if (e->IsControlEdge()) { |
| 3485 | auto result = unique_node.insert(e->dst()); |
| 3486 | if (result.second) { |
| 3487 | (*g)->AddControlEdge(new_node, e->dst(), true); |
| 3488 | } |
| 3489 | } else { |
| 3490 | auto new_edge = (*g)->AddEdge(new_node, kMergedNodeFilterGradOutputIdx, |
| 3491 | e->dst(), e->dst_input()); |
| 3492 | DCHECK(new_edge); |
| 3493 | } |
| 3494 | } |
| 3495 | |
| 3496 | // Copy device assigned to old node to new node. |
| 3497 | // It's ok to use badd or fltr as we have enforced a check that |
| 3498 | // both have same device assigned. |
| 3499 | new_node->set_assigned_device_name(badd->assigned_device_name()); |
| 3500 | |
| 3501 | VLOG(1) << "MklLayoutRewritePass: Merged old node:"<< badd->DebugString() |
| 3502 | << ", and node: "<< fltr->DebugString() |
| 3503 | << ", into node:"<< new_node->DebugString(); |
| 3504 | |
| 3505 | (*g)->RemoveNode(badd); |
| 3506 | (*g)->RemoveNode(fltr); |
| 3507 | |
| 3508 | return OkStatus(); |
| 3509 | } |
| 3510 | |
| 3511 | Status MklLayoutRewritePass::MergeNode(std::unique_ptr<Graph>* g, Node* m, |
| 3512 | Node* n) { |
| 3513 | DCHECK(m); |
| 3514 | DCHECK(n); |
| 3515 | |
| 3516 | if (((m->type_string() == csinfo_.bias_add && |
| 3517 | n->type_string() == csinfo_.conv2d)) || |
| 3518 | ((n->type_string() == csinfo_.bias_add && |
| 3519 | m->type_string() == csinfo_.conv2d))) { |
| 3520 | return this->MergeConv2DWithBiasAdd(g, m, n); |
| 3521 | } |
| 3522 | if ((m->type_string() == csinfo_.pad && |
| 3523 | (n->type_string() == csinfo_.conv2d || |
| 3524 | (n->type_string() == csinfo_.fused_conv2d && FusedConv2DRewrite(n)))) || |
| 3525 | (n->type_string() == csinfo_.pad && |
| 3526 | (m->type_string() == csinfo_.conv2d || |
| 3527 | (m->type_string() == csinfo_.fused_conv2d && FusedConv2DRewrite(m))))) { |
| 3528 | return this->MergePadWithConv2D(g, m, n); |
| 3529 | } |
| 3530 | |
| 3531 | if (((m->type_string() == csinfo_.bias_add_grad && |
| 3532 | n->type_string() == csinfo_.conv2d_grad_filter)) || |
| 3533 | ((n->type_string() == csinfo_.bias_add_grad && |
| 3534 | m->type_string() == csinfo_.conv2d_grad_filter))) { |
| 3535 | return this->MergeConv2DBackpropFilterWithBiasAddGrad(g, m, n); |
| 3536 | } |
| 3537 | |
| 3538 | return Status(error::Code::UNIMPLEMENTED, |
| 3539 | "Unimplemented case for node merge optimization."); |
| 3540 | } |
| 3541 | |
| 3542 | ////////////////////////////////////////////////////////////////////////// |
| 3543 | // Helper functions for node rewrite |
| 3544 | ////////////////////////////////////////////////////////////////////////// |
| 3545 | |
| 3546 | Status MklLayoutRewritePass::RewriteNodeForLayoutPropagation( |
| 3547 | std::unique_ptr<Graph>* g, const Node* orig_node, Node** new_node, |
| 3548 | const RewriteInfo* ri) { |
| 3549 | // Get all data inputs. |
| 3550 | int num_data_inputs = orig_node->in_edges().size(); |
| 3551 | // Drop count for control edges from inputs |
| 3552 | for (const Edge* e : orig_node->in_edges()) { |
| 3553 | if (e->IsControlEdge()) { |
| 3554 | num_data_inputs--; |
| 3555 | } |
| 3556 | } |
| 3557 | |
| 3558 | gtl::InlinedVector<Node*, 4> control_edges; |
| 3559 | gtl::InlinedVector<std::pair<Node*, int>, 4> inputs(num_data_inputs); |
| 3560 | FillInputs(orig_node, &control_edges, &inputs); |
| 3561 | |
| 3562 | // Build new node. We use same name as original node, but change the op name. |
| 3563 | NodeBuilder nb(orig_node->name().c_str(), ri->new_name.c_str()); |
| 3564 | // Copy user-specified device assigned to original node to new node. |
| 3565 | nb.Device(orig_node->def().device()); |
| 3566 | // Set up new inputs to the rewritten node. |
| 3567 | Status s = SetUpInputs(g, inputs, &nb, orig_node); |
| 3568 | if (s != OkStatus()) { |
| 3569 | return s; |
| 3570 | } |
| 3571 | |
| 3572 | const bool kPartialCopyAttrs = false; |
| 3573 | ri->copy_attrs(const_cast<const Node*>(orig_node), &nb, kPartialCopyAttrs); |
| 3574 | |
| 3575 | // Set the Mkl layer label for this op. |
| 3576 | if (DataTypeIsQuantized(orig_node->input_type(0)) || |
| 3577 | DataTypeIsQuantized(orig_node->output_type(0))) { |
| 3578 | nb.Attr("_kernel", mkl_op_registry::kMklQuantizedOpLabel); |
| 3579 | } else { |
| 3580 | nb.Attr("_kernel", mkl_op_registry::kMklLayoutDependentOpLabel); |
| 3581 | } |
| 3582 | // Finalize graph and get new node. |
| 3583 | s = nb.Finalize(&**g, new_node); |
| 3584 | if (s != OkStatus()) { |
| 3585 | return s; |
| 3586 | } |
| 3587 | |
| 3588 | // In the following code of this function, an unsorted set is used to make |
| 3589 | // sure no duplicated edges be added into the new node. Therefore, we can |
| 3590 | // pass allow_duplicates = true in AddControlEdge call to skip the O(#edges) |
| 3591 | // check in the routine. |
| 3592 | |
| 3593 | // Incoming data edges from 'orig_node' node to new 'new_node' node are |
| 3594 | // already copied in BuildNode. We need to handle control edges now. |
| 3595 | std::unordered_set<Node*> unique_node; |
| 3596 | for (const Edge* e : orig_node->in_edges()) { |
| 3597 | if (e->IsControlEdge()) { |
| 3598 | auto result = unique_node.insert(e->src()); |
| 3599 | if (result.second) { |
| 3600 | (*g)->AddControlEdge(e->src(), *new_node, true); |
| 3601 | } |
| 3602 | } |
| 3603 | } |
| 3604 | unique_node.clear(); |
| 3605 | |
| 3606 | // Copy outgoing edges from 'orig_node' node to new |
| 3607 | // 'new_node' node, since the output also follows same ordering among |
| 3608 | // Tensorflow tensors and Mkl tensors. We need to connect Tensorflow |
| 3609 | // tensors appropriately. Specifically, nth output of the original node |
| 3610 | // will become 2*nth output of the Mkl node for the interleaved ordering |
| 3611 | // of the tensors. For the contiguous ordering of the tensors, it will be n. |
| 3612 | // GetTensorDataIndex provides this mapping function. |
| 3613 | for (const Edge* e : orig_node->out_edges()) { |
| 3614 | if (e->IsControlEdge()) { |
| 3615 | auto result = unique_node.insert(e->dst()); |
| 3616 | if (result.second) { |
| 3617 | (*g)->AddControlEdge(*new_node, e->dst(), true); |
| 3618 | } |
| 3619 | } else { |
| 3620 | auto new_edge = (*g)->AddEdge( |
| 3621 | *new_node, |
| 3622 | GetTensorDataIndex(e->src_output(), e->src()->num_outputs()), |
| 3623 | e->dst(), e->dst_input()); |
| 3624 | DCHECK(new_edge); |
| 3625 | } |
| 3626 | } |
| 3627 | return OkStatus(); |
| 3628 | } |
| 3629 | |
| 3630 | Status MklLayoutRewritePass::RewriteNodeForJustOpNameChange( |
| 3631 | std::unique_ptr<Graph>* g, const Node* orig_node, Node** new_node, |
| 3632 | const RewriteInfo* ri) { |
| 3633 | // Get all data inputs. |
| 3634 | int num_data_inputs = orig_node->in_edges().size(); |
| 3635 | // Drop count for control edges from inputs |
| 3636 | for (const Edge* e : orig_node->in_edges()) { |
| 3637 | if (e->IsControlEdge()) { |
| 3638 | num_data_inputs--; |
| 3639 | } |
| 3640 | } |
| 3641 | gtl::InlinedVector<Node*, 4> control_edges; |
| 3642 | gtl::InlinedVector<std::pair<Node*, int>, 4> inputs(num_data_inputs); |
| 3643 | FillInputs(orig_node, &control_edges, &inputs); |
| 3644 | |
| 3645 | // Build new node. We use same name as original node, but change the op name. |
| 3646 | NodeBuilder nb(orig_node->name().c_str(), ri->new_name.c_str()); |
| 3647 | // Copy user-specified device assigned to original node to new node. |
| 3648 | nb.Device(orig_node->def().device()); |
| 3649 | |
| 3650 | Status s = CopyInputs(orig_node, inputs, &nb); |
| 3651 | if (s != OkStatus()) { |
| 3652 | return s; |
| 3653 | } |
| 3654 | |
| 3655 | std::vector<NodeBuilder::NodeOut> workspace_tensors; |
| 3656 | bool are_workspace_tensors_available = false; |
| 3657 | if (IsWorkspaceCheckNeeded(orig_node)) { |
| 3658 | AddWorkSpaceEdgeIfNeeded(g, orig_node, &nb, &workspace_tensors, |
| 3659 | &are_workspace_tensors_available); |
| 3660 | if (are_workspace_tensors_available) { |
| 3661 | DCHECK_EQ(workspace_tensors.size(), 1); |
| 3662 | nb.Input(workspace_tensors[0].node, workspace_tensors[0].index); |
| 3663 | } |
| 3664 | } |
| 3665 | |
| 3666 | if (!NativeFormatEnabled()) { |
| 3667 | ri->copy_attrs(const_cast<const Node*>(orig_node), &nb, true); |
| 3668 | } else { |
| 3669 | ri->copy_attrs(const_cast<const Node*>(orig_node), &nb, false); |
| 3670 | } |
| 3671 | |
| 3672 | if (DataTypeIsQuantized(orig_node->input_type(0)) || |
| 3673 | DataTypeIsQuantized(orig_node->output_type(0))) { |
| 3674 | nb.Attr("_kernel", mkl_op_registry::kMklQuantizedOpLabel); |
| 3675 | } else { |
| 3676 | nb.Attr("_kernel", mkl_op_registry::kMklNameChangeOpLabel); |
| 3677 | } |
| 3678 | |
| 3679 | // Finalize graph and get new node. |
| 3680 | s = nb.Finalize(&**g, new_node); |
| 3681 | if (s != OkStatus()) { |
| 3682 | return s; |
| 3683 | } |
| 3684 | |
| 3685 | // In the following code of this function, an unsorted set is used to make |
| 3686 | // sure no duplicated edges be added into the new node. Therefore, we can |
| 3687 | // pass allow_duplicates = true in AddControlEdge call to skip the O(#edges) |
| 3688 | // check in the routine. |
| 3689 | |
| 3690 | // Incoming data edges from 'orig_node' node to new 'new_node' node are |
| 3691 | // already copied in BuildNode. We need to handle control edges now. |
| 3692 | std::unordered_set<Node*> unique_node; |
| 3693 | for (const Edge* e : orig_node->in_edges()) { |
| 3694 | if (e->IsControlEdge()) { |
| 3695 | auto result = unique_node.insert(e->src()); |
| 3696 | if (result.second) { |
| 3697 | (*g)->AddControlEdge(e->src(), *new_node, true); |
| 3698 | } |
| 3699 | } |
| 3700 | } |
| 3701 | unique_node.clear(); |
| 3702 | |
| 3703 | // Transfer outgoing edges from 'orig_node' node to new 'new_node' node. |
| 3704 | for (const Edge* e : orig_node->out_edges()) { |
| 3705 | if (e->IsControlEdge()) { |
| 3706 | auto result = unique_node.insert(e->dst()); |
| 3707 | if (result.second) { |
| 3708 | (*g)->AddControlEdge(*new_node, e->dst(), true); |
| 3709 | } |
| 3710 | } else { |
| 3711 | auto result = |
| 3712 | (*g)->AddEdge(*new_node, e->src_output(), e->dst(), e->dst_input()); |
| 3713 | DCHECK(result != nullptr); |
| 3714 | } |
| 3715 | } |
| 3716 | |
| 3717 | return OkStatus(); |
| 3718 | } |
| 3719 | |
| 3720 | Status MklLayoutRewritePass::RewriteNode(std::unique_ptr<Graph>* g, |
| 3721 | Node* orig_node, |
| 3722 | const RewriteInfo* ri) { |
| 3723 | DCHECK(ri != nullptr); |
| 3724 | DCHECK(orig_node != nullptr); |
| 3725 | |
| 3726 | VLOG(1) << "MklLayoutRewritePass: Original node:"<< orig_node->DebugString(); |
| 3727 | |
| 3728 | Status ret_status = OkStatus(); |
| 3729 | Node* new_node = nullptr; |
| 3730 | if (ri->rewrite_cause == kRewriteForLayoutPropagation) { |
| 3731 | ret_status = RewriteNodeForLayoutPropagation(g, orig_node, &new_node, ri); |
| 3732 | } else if (ri->rewrite_cause == kRewriteForOpNameChange) { |
| 3733 | ret_status = RewriteNodeForJustOpNameChange(g, orig_node, &new_node, ri); |
| 3734 | } else { |
| 3735 | ret_status = Status(error::Code::INVALID_ARGUMENT, |
| 3736 | "Unsupported rewrite cause found." |
| 3737 | "RewriteNode will fail."); |
| 3738 | } |
| 3739 | TF_CHECK_OK(ret_status); |
| 3740 | |
| 3741 | // Copy the runtime device assigned from original code to new node. |
| 3742 | new_node->set_assigned_device_name(orig_node->assigned_device_name()); |
| 3743 | |
| 3744 | // Delete original node and mark new node as rewritten. |
| 3745 | (*g)->RemoveNode(orig_node); |
| 3746 | |
| 3747 | VLOG(1) << "MklLayoutRewritePass: New node:"<< new_node->DebugString(); |
| 3748 | return ret_status; |
| 3749 | } |
| 3750 | |
| 3751 | // TODO(mdfaijul): Is there any other elegant way to check for quantized ops |
| 3752 | // having attributes other than "T"? |
| 3753 | // Current implementation reflects only QuantizedConv2D and its fused Ops. |
| 3754 | const MklLayoutRewritePass::RewriteInfo* |
| 3755 | MklLayoutRewritePass::CheckForQuantizedNodeRewrite(const Node* n) const { |
| 3756 | DataType T1, T2; |
| 3757 | DataType Tinput, Tfilter; |
| 3758 | bool type_attrs_present = false; |
| 3759 | |
| 3760 | if (TryGetNodeAttr(n->def(), "Tinput", &Tinput) && |
| 3761 | TryGetNodeAttr(n->def(), "Tfilter", &Tfilter) && |
| 3762 | mkl_op_registry::IsMklQuantizedOp( |
| 3763 | mkl_op_registry::GetMklOpName(n->type_string()), Tinput, Tfilter)) { |
| 3764 | type_attrs_present = true; |
| 3765 | } else if (TryGetNodeAttr(n->def(), "T1", &T1) && |
| 3766 | TryGetNodeAttr(n->def(), "T2", &T2) && |
| 3767 | mkl_op_registry::IsMklQuantizedOp( |
| 3768 | mkl_op_registry::GetMklOpName(n->type_string()), T1, T2)) { |
| 3769 | type_attrs_present = true; |
| 3770 | } |
| 3771 | |
| 3772 | if (type_attrs_present) { |
| 3773 | for (auto ri = rinfo_.cbegin(); ri != rinfo_.cend(); ++ri) { |
| 3774 | if (n->type_string().compare(ri->name) == 0 && ri->rewrite_rule(n)) { |
| 3775 | return &*ri; |
| 3776 | } |
| 3777 | } |
| 3778 | } |
| 3779 | |
| 3780 | return nullptr; |
| 3781 | } |
| 3782 | |
| 3783 | const MklLayoutRewritePass::RewriteInfo* |
| 3784 | MklLayoutRewritePass::CheckForNodeRewrite(const Node* n) const { |
| 3785 | DCHECK(n); |
| 3786 | |
| 3787 | // QuantizedOps may have attributes other than "T", so decoupled the check |
| 3788 | // with a function, CheckForQuantizedNodeRewrite(const Node*). |
| 3789 | const RewriteInfo* ri = CheckForQuantizedNodeRewrite(n); |
| 3790 | if (ri != nullptr) return ri; |
| 3791 | |
| 3792 | // First check if node along with its type is supported by MKL layer. |
| 3793 | // We do not want to rewrite an op into Mkl op if types are not supported. |
| 3794 | // E.g., MklRelu does not support INT32. So we cannot rewrite Relu to |
| 3795 | // MklRelu if type is INT32. |
| 3796 | DataType T; |
| 3797 | if (!TryGetNodeAttr(n->def(), "T", &T)) { |
| 3798 | return nullptr; |
| 3799 | } |
| 3800 | |
| 3801 | // We make an exception for Conv2DGrad and MaxPool related ops as |
| 3802 | // the corresponding MKL ops currently do not support the case |
| 3803 | // of padding == EXPLICIT yet. |
| 3804 | // TODO(intel): support `EXPLICIT` padding for ConvGrad |
| 3805 | if (n->type_string() == csinfo_.conv2d_grad_input || |
| 3806 | n->type_string() == csinfo_.conv2d_grad_filter || |
| 3807 | n->type_string() == csinfo_.depthwise_conv2d_grad_filter || |
| 3808 | n->type_string() == csinfo_.depthwise_conv2d_grad_input || |
| 3809 | n->type_string() == csinfo_.conv3d_grad_filter || |
| 3810 | n->type_string() == csinfo_.conv3d_grad_filter || |
| 3811 | n->type_string() == csinfo_.max_pool || |
| 3812 | n->type_string() == csinfo_.max_pool_grad || |
| 3813 | n->type_string() == csinfo_.max_pool3d || |
| 3814 | n->type_string() == csinfo_.max_pool3d_grad) { |
| 3815 | string padding; |
| 3816 | TF_CHECK_OK(GetNodeAttr(n->def(), "padding", &padding)); |
| 3817 | if (padding == "EXPLICIT") return nullptr; |
| 3818 | } |
| 3819 | |
| 3820 | // We make an exception for __MklDummyConv2DWithBias, |
| 3821 | // __MklConv2DBackpropFilterWithBias, and __MklDummyPadWithConv2D since their |
| 3822 | // names do not match Mkl node names. |
| 3823 | if (n->type_string() != csinfo_.conv2d_with_bias && |
| 3824 | n->type_string() != csinfo_.pad_with_conv2d && |
| 3825 | n->type_string() != csinfo_.pad_with_fused_conv2d && |
| 3826 | n->type_string() != csinfo_.conv2d_grad_filter_with_bias && |
| 3827 | n->type_string() != csinfo_.fused_batch_norm_ex && |
| 3828 | n->type_string() != csinfo_.fused_conv2d && |
| 3829 | n->type_string() != csinfo_.fused_depthwise_conv2d && |
| 3830 | n->type_string() != csinfo_.fused_matmul && |
| 3831 | n->type_string() != csinfo_.fused_conv3d && |
| 3832 | !mkl_op_registry::IsMklOp(mkl_op_registry::GetMklOpName(n->type_string()), |
| 3833 | T)) { |
| 3834 | return nullptr; |
| 3835 | } |
| 3836 | |
| 3837 | // We now check if rewrite rule applies for this op. If rewrite rule passes |
| 3838 | // for this op, then we rewrite it to Mkl op. |
| 3839 | // Find matching RewriteInfo and then check that rewrite rule applies. |
| 3840 | for (auto ri = rinfo_.cbegin(); ri != rinfo_.cend(); ++ri) { |
| 3841 | if (n->type_string().compare(ri->name) == 0 && ri->rewrite_rule(n)) { |
| 3842 | return &*ri; |
| 3843 | } |
| 3844 | } |
| 3845 | |
| 3846 | // Else return not found. |
| 3847 | return nullptr; |
| 3848 | } |
| 3849 | |
| 3850 | ////////////////////////////////////////////////////////////////////////// |
| 3851 | // Helper functions for node fusion |
| 3852 | ////////////////////////////////////////////////////////////////////////// |
| 3853 | Status MklLayoutRewritePass::FuseTransposeMklOpTranspose( |
| 3854 | std::unique_ptr<Graph>* g, std::vector<Node*>& nodes, |
| 3855 | std::function<void(const Node*, NodeBuilder* nb, bool)> copy_attrs, |
| 3856 | string data_format) { |
| 3857 | Node* transpose_to_nhwc = nodes[0]; |
| 3858 | Node* mklop = nodes[1]; |
| 3859 | Node* transpose_to_nchw = nodes[2]; |
| 3860 | |
| 3861 | const int transpose_nhwc_num_inputs = transpose_to_nhwc->num_inputs(); |
| 3862 | gtl::InlinedVector<Node*, 4> transpose_nhwc_control_edges; |
| 3863 | gtl::InlinedVector<std::pair<Node*, int>, 4> transpose_nhwc_in( |
| 3864 | transpose_nhwc_num_inputs); |
| 3865 | FillInputs(transpose_to_nhwc, &transpose_nhwc_control_edges, |
| 3866 | &transpose_nhwc_in); |
| 3867 | |
| 3868 | const int mklop_num_inputs = mklop->num_inputs(); |
| 3869 | gtl::InlinedVector<Node*, 4> mklop_control_edges; |
| 3870 | gtl::InlinedVector<std::pair<Node*, int>, 4> mklop_in(mklop_num_inputs); |
| 3871 | FillInputs(mklop, &mklop_control_edges, &mklop_in); |
| 3872 | |
| 3873 | const int transpose_nchw_num_inputs = transpose_to_nchw->num_inputs(); |
| 3874 | gtl::InlinedVector<Node*, 4> transpose_nchw_control_edges; |
| 3875 | gtl::InlinedVector<std::pair<Node*, int>, 4> transpose_nchw_in( |
| 3876 | transpose_nchw_num_inputs); |
| 3877 | FillInputs(transpose_to_nchw, &transpose_nchw_control_edges, |
| 3878 | &transpose_nchw_in); |
| 3879 | |
| 3880 | // We use same name as original node, but change the op |
| 3881 | // type. |
| 3882 | NodeBuilder nb(mklop->name(), mklop->type_string()); |
| 3883 | |
| 3884 | // Storing the output slots of the input nodes. |
| 3885 | for (int i = 0; i < mklop_num_inputs; i++) { |
| 3886 | if (mklop_in[i].first == transpose_to_nhwc) { |
| 3887 | // Fill "x": |
| 3888 | nb.Input(transpose_nhwc_in[0].first, transpose_nhwc_in[0].second); |
| 3889 | } else { |
| 3890 | // Fill inputs other than "x": |
| 3891 | nb.Input(mklop_in[i].first, mklop_in[i].second); |
| 3892 | } |
| 3893 | } |
| 3894 | |
| 3895 | copy_attrs(const_cast<const Node*>(mklop), &nb, true); |
| 3896 | nb.Attr("data_format", data_format); |
| 3897 | |
| 3898 | // Copy the device assigned to old node to new node. |
| 3899 | nb.Device(mklop->def().device()); |
| 3900 | |
| 3901 | // Create node. |
| 3902 | Node* new_node; |
| 3903 | TF_CHECK_OK(nb.Finalize(&**g, &new_node)); |
| 3904 | // No need to check if new_node is null because it will be null only when |
| 3905 | // Finalize fails. |
| 3906 | |
| 3907 | // Fill outputs. |
| 3908 | for (const Edge* e : transpose_to_nchw->out_edges()) { |
| 3909 | if (!e->IsControlEdge()) { |
| 3910 | const int kTransposeWithMklOpOutputSlot = 0; |
| 3911 | auto new_edge = (*g)->AddEdge(new_node, kTransposeWithMklOpOutputSlot, |
| 3912 | e->dst(), e->dst_input()); |
| 3913 | DCHECK(new_edge); |
| 3914 | } |
| 3915 | } |
| 3916 | |
| 3917 | // Copy device assigned to old node to new node. |
| 3918 | new_node->set_assigned_device_name(mklop->assigned_device_name()); |
| 3919 | |
| 3920 | // Copy requested_device and assigned_device_name_index |
| 3921 | new_node->set_requested_device(mklop->requested_device()); |
| 3922 | new_node->set_assigned_device_name_index(mklop->assigned_device_name_index()); |
| 3923 | |
| 3924 | (*g)->RemoveNode(transpose_to_nhwc); |
| 3925 | (*g)->RemoveNode(mklop); |
| 3926 | (*g)->RemoveNode(transpose_to_nchw); |
| 3927 | |
| 3928 | return OkStatus(); |
| 3929 | } |
| 3930 | |
| 3931 | Status MklLayoutRewritePass::FuseNode( |
| 3932 | std::unique_ptr<Graph>* g, std::vector<Node*>& nodes, |
| 3933 | const MklLayoutRewritePass::FusionInfo fi) { |
| 3934 | return fi.fuse_func(g, nodes, fi.copy_attrs); |
| 3935 | } |
| 3936 | |
| 3937 | std::tuple<bool, std::vector<Node*>, const MklLayoutRewritePass::FusionInfo> |
| 3938 | MklLayoutRewritePass::CheckForNodeFusion(Node* a) const { |
| 3939 | // Stores matched nodes, in the same order as node_checkers. |
| 3940 | std::vector<Node*> nodes; |
| 3941 | |
| 3942 | for (auto fi = finfo_.begin(); fi != finfo_.end(); ++fi) { |
| 3943 | // |
| 3944 | // Make sure node "a" and its succeeding nodes (b, c ...), match the pattern |
| 3945 | // defined in fusion info (ops[0], ops[1], ...), |
| 3946 | // a.k.a. "a->b->c" matches "op1->op2->op3" |
| 3947 | // |
| 3948 | |
| 3949 | // Stores the first unvisited outgoing edge of each matched node in "nodes". |
| 3950 | std::stack<EdgeSet::const_iterator> current_neighbor_stack; |
| 3951 | nodes.clear(); |
| 3952 | |
| 3953 | auto node_checker = fi->node_checkers.begin(); |
| 3954 | if (a != nullptr && (*node_checker)(a)) { |
| 3955 | nodes.push_back(a); |
| 3956 | current_neighbor_stack.push(a->out_edges().begin()); |
| 3957 | ++node_checker; |
| 3958 | } |
| 3959 | |
| 3960 | while (!nodes.empty()) { |
| 3961 | auto& current_neighbor_iter = current_neighbor_stack.top(); |
| 3962 | |
| 3963 | if (current_neighbor_iter != nodes.back()->out_edges().end()) { |
| 3964 | // Found an unvisited edge. Goes through the edge to get the neighbor. |
| 3965 | Node* neighbor_node = (*current_neighbor_iter)->dst(); |
| 3966 | ++current_neighbor_stack.top(); // Retrieves the next unvisited edge. |
| 3967 | |
| 3968 | if ((*node_checker)(neighbor_node)) { |
| 3969 | // Found a match. Stores the node and moves to the next checker. |
| 3970 | nodes.push_back(neighbor_node); |
| 3971 | current_neighbor_stack.push(neighbor_node->out_edges().begin()); |
| 3972 | if (++node_checker == fi->node_checkers.end()) { |
| 3973 | return make_tuple(true, nodes, *fi); |
| 3974 | } |
| 3975 | } |
| 3976 | } else { |
| 3977 | // Removes the current node since none of its neighbor leads to a |
| 3978 | // further match. |
| 3979 | nodes.pop_back(); |
| 3980 | current_neighbor_stack.pop(); |
| 3981 | --node_checker; |
| 3982 | } |
| 3983 | } |
| 3984 | } |
| 3985 | |
| 3986 | return make_tuple(false, std::vector<Node*>(), FusionInfo()); |
| 3987 | } |
| 3988 | |
| 3989 | /////////////////////////////////////////////////////////////////////////////// |
| 3990 | // Post-rewrite Mkl metadata fixup pass |
| 3991 | /////////////////////////////////////////////////////////////////////////////// |
| 3992 | bool MklLayoutRewritePass::FixMklMetaDataEdgeIfNeeded(std::unique_ptr<Graph>* g, |
| 3993 | const Edge* e_data, |
| 3994 | const Edge* e_metadata) { |
| 3995 | if (g == nullptr || e_data == nullptr || e_metadata == nullptr) { |
| 3996 | return false; |
| 3997 | } |
| 3998 | |
| 3999 | Node* n_data = e_data->src(); |
| 4000 | int n_data_op_slot = e_data->src_output(); |
| 4001 | int n_metadata_op_slot = |
| 4002 | GetTensorMetaDataIndex(n_data_op_slot, n_data->num_outputs()); |
| 4003 | |
| 4004 | // If the source of meta edge is a constant node (producing dummy Mkl metadata |
| 4005 | // tensor), then we will need to fix. |
| 4006 | if (IsConstant(e_metadata->src())) { |
| 4007 | Node* e_metadata_dst = e_metadata->dst(); |
| 4008 | int e_metadata_in_slot = e_metadata->dst_input(); |
| 4009 | auto new_edge = (*g)->AddEdge(n_data, n_metadata_op_slot, e_metadata_dst, |
| 4010 | e_metadata_in_slot); |
| 4011 | DCHECK(new_edge); |
| 4012 | |
| 4013 | (*g)->RemoveEdge(e_metadata); |
| 4014 | return true; |
| 4015 | } |
| 4016 | |
| 4017 | return false; |
| 4018 | } |
| 4019 | |
| 4020 | bool MklLayoutRewritePass::FixMklMetaDataEdges(std::unique_ptr<Graph>* g, |
| 4021 | Node* n) { |
| 4022 | bool result = false; |
| 4023 | |
| 4024 | // If graph node is not Mkl node, then return. |
| 4025 | DataType T = DT_INVALID; |
| 4026 | if (!TryGetNodeAttr(n->def(), "T", &T) || |
| 4027 | !mkl_op_registry::IsMklOp(n->type_string(), T, false)) { |
| 4028 | return result; |
| 4029 | } |
| 4030 | |
| 4031 | // If it is Mkl node, then check if the input edges to this node that carry |
| 4032 | // Mkl metadata are linked up correctly with the source node. |
| 4033 | |
| 4034 | // For Mkl nodes, we generate twice the number of input tensors (n for Mkl |
| 4035 | // data tensors + n for Mkl metadata tensors). We need to check for correct |
| 4036 | // connection of n metadata tensors only. |
| 4037 | int num_data_inputs = n->num_inputs() / 2; |
| 4038 | for (int idx = 0; idx < num_data_inputs; idx++) { |
| 4039 | // Get the edge connecting input slot with index (idx). |
| 4040 | const Edge* e = nullptr; |
| 4041 | TF_CHECK_OK(n->input_edge(idx, &e)); |
| 4042 | |
| 4043 | // If e is control edge, then skip. |
| 4044 | if (e->IsControlEdge()) { |
| 4045 | continue; |
| 4046 | } |
| 4047 | |
| 4048 | // Check that the source node for edge 'e' is Mkl node. If it is not an Mkl |
| 4049 | // node, then we don't need to do anything. |
| 4050 | Node* e_src = e->src(); |
| 4051 | if (TryGetNodeAttr(e_src->def(), "T", &T) && |
| 4052 | mkl_op_registry::IsMklOp(e_src->type_string(), T, false)) { |
| 4053 | // Source node for edge 'e' is Mkl node. |
| 4054 | // Destination node and destination input slot of e is node 'n' and 'idx' |
| 4055 | // resp. |
| 4056 | CHECK_EQ(e->dst(), n); |
| 4057 | CHECK_EQ(e->dst_input(), idx); |
| 4058 | |
| 4059 | // Let's get edge that carries Mkl metadata corresponding to Mkl data edge |
| 4060 | // 'e'. For that, let's first get the input slot of 'n' where the meta |
| 4061 | // edge will feed the value. |
| 4062 | int e_meta_in_slot = |
| 4063 | GetTensorMetaDataIndex(e->dst_input(), n->num_inputs()); |
| 4064 | const Edge* e_meta = nullptr; |
| 4065 | TF_CHECK_OK(n->input_edge(e_meta_in_slot, &e_meta)); |
| 4066 | |
| 4067 | // Let's check if we need to fix this meta edge. |
| 4068 | if (FixMklMetaDataEdgeIfNeeded(g, e, e_meta)) { |
| 4069 | result = true; |
| 4070 | } |
| 4071 | } |
| 4072 | } |
| 4073 | |
| 4074 | return result; |
| 4075 | } |
| 4076 | |
| 4077 | /////////////////////////////////////////////////////////////////////////////// |
| 4078 | // Run function for the pass |
| 4079 | /////////////////////////////////////////////////////////////////////////////// |
| 4080 | |
| 4081 | bool MklLayoutRewritePass::RunPass(std::unique_ptr<Graph>* g) { |
| 4082 | bool result = false; |
| 4083 | DCHECK(g); |
| 4084 | |
| 4085 | DumpGraph("Before running MklLayoutRewritePass", &**g); |
| 4086 | |
| 4087 | std::vector<Node*> order; |
| 4088 | GetReversePostOrder(**g, &order); // This will give us topological sort. |
| 4089 | for (Node* n : order) { |
| 4090 | // If node is not an op or it cannot run on CPU device, then skip. |
| 4091 | if (!n->IsOp() || !CanOpRunOnCPUDevice(n)) { |
| 4092 | continue; |
| 4093 | } |
| 4094 | |
| 4095 | Node* m = nullptr; |
| 4096 | if ((m = CheckForNodeMerge(n)) != nullptr && CanOpRunOnCPUDevice(m)) { |
| 4097 | // Check if the node 'n' can be merged with any other node. If it can |
| 4098 | // be 'm' contains the node with which it can be merged. |
| 4099 | string n1_name = n->name(); |
| 4100 | string n2_name = m->name(); |
| 4101 | |
| 4102 | VLOG(1) << "MklLayoutRewritePass: Scheduled nodes "<< n1_name << " and " |
| 4103 | << n2_name << " for merging"; |
| 4104 | |
| 4105 | if (MergeNode(g, n, m) == OkStatus()) { |
| 4106 | VLOG(1) << "MklLayoutRewritePass: Merged nodes "<< n1_name << " and " |
| 4107 | << n2_name; |
| 4108 | result = true; |
| 4109 | } |
| 4110 | } |
| 4111 | } |
| 4112 | |
| 4113 | DumpGraph("After running MklLayoutRewritePass(NodeMerge)", &**g); |
| 4114 | |
| 4115 | order.clear(); |
| 4116 | GetReversePostOrder(**g, &order); // This will give us topological sort. |
| 4117 | for (Node* n : order) { |
| 4118 | // If node is not an op or it cannot run on CPU device, then skip. |
| 4119 | if (!n->IsOp() || !CanOpRunOnCPUDevice(n)) { |
| 4120 | continue; |
| 4121 | } |
| 4122 | |
| 4123 | auto check_result = CheckForNodeFusion(n); |
| 4124 | bool found_pattern = std::get<0>(check_result); |
| 4125 | std::vector<Node*> nodes = std::get<1>(check_result); |
| 4126 | const FusionInfo fi = std::get<2>(check_result); |
| 4127 | |
| 4128 | // if "found_pattern" is true, we can do the fusion. |
| 4129 | if (found_pattern) { |
| 4130 | if (FuseNode(g, nodes, fi) == OkStatus()) { |
| 4131 | result = true; |
| 4132 | } |
| 4133 | } |
| 4134 | } |
| 4135 | DumpGraph("After running MklLayoutRewritePass(NodeFusion)", &**g); |
| 4136 | |
| 4137 | order.clear(); |
| 4138 | GetReversePostOrder(**g, &order); // This will give us topological sort. |
| 4139 | for (Node* n : order) { |
| 4140 | // If node is not an op or it cannot run on CPU device, then skip. |
| 4141 | if (!n->IsOp() || !CanOpRunOnCPUDevice(n)) { |
| 4142 | continue; |
| 4143 | } |
| 4144 | |
| 4145 | const RewriteInfo* ri = nullptr; |
| 4146 | // We will first search if node is to be rewritten. |
| 4147 | if ((ri = CheckForNodeRewrite(n)) != nullptr) { |
| 4148 | string node_name = n->name(); |
| 4149 | string op_name = n->type_string(); |
| 4150 | |
| 4151 | VLOG(1) << "MklLayoutRewritePass: Scheduled node "<< node_name |
| 4152 | << " with op "<< op_name << " for rewrite using" |
| 4153 | << " layout optimization."; |
| 4154 | |
| 4155 | if (RewriteNode(g, n, ri) == OkStatus()) { |
| 4156 | VLOG(1) << "MklLayoutRewritePass: rewrote node "<< node_name |
| 4157 | << " with op "<< op_name << " for Mkl layout optimization."; |
| 4158 | result = true; |
| 4159 | } |
| 4160 | } |
| 4161 | } |
| 4162 | |
| 4163 | DumpGraph("After running MklLayoutRewritePass(NodeMerge+Rewrite)", &**g); |
| 4164 | |
| 4165 | if (!NativeFormatEnabled()) { |
| 4166 | order.clear(); |
| 4167 | GetReversePostOrder(**g, &order); // This will give us topological sort. |
| 4168 | for (Node* n : order) { |
| 4169 | // If node is not an op or it cannot run on CPU device, then skip. |
| 4170 | if (!n->IsOp() || !CanOpRunOnCPUDevice(n)) { |
| 4171 | continue; |
| 4172 | } |
| 4173 | if (FixMklMetaDataEdges(g, n)) { |
| 4174 | string node_name = n->name(); |
| 4175 | string op_name = n->type_string(); |
| 4176 | |
| 4177 | VLOG(1) << "MklLayoutRewritePass: fixed metadata edges for node " |
| 4178 | << node_name << " with op "<< op_name; |
| 4179 | result = true; |
| 4180 | } |
| 4181 | } |
| 4182 | DumpGraph("After running MklLayoutRewritePass(NodeMerge+Rewrite+Fixup)", |
| 4183 | &**g); |
| 4184 | } |
| 4185 | |
| 4186 | return result; |
| 4187 | } |
| 4188 | |
| 4189 | bool RunMklLayoutRewritePass(std::unique_ptr<Graph>* g) { |
| 4190 | return MklLayoutRewritePass().RunPass(g); |
| 4191 | } |
| 4192 | |
| 4193 | Status MklLayoutRewritePass::Run(const GraphOptimizationPassOptions& options) { |
| 4194 | if (options.graph == nullptr && options.partition_graphs == nullptr) { |
| 4195 | return OkStatus(); |
| 4196 | } |
| 4197 | if (!IsMKLEnabled()) { |
| 4198 | VLOG(2) << "TF-MKL: MKL is not enabled"; |
| 4199 | return OkStatus(); |
| 4200 | } |
| 4201 | |
| 4202 | auto process_graph = [&](std::unique_ptr<Graph>* g) { |
| 4203 | // Get the ownership of a graph |
| 4204 | std::unique_ptr<Graph>* ng = std::move(g); |
| 4205 | RunPass(ng); |
| 4206 | // Return the ownership of a graph back |
| 4207 | g->reset(ng->release()); |
| 4208 | }; |
| 4209 | |
| 4210 | if (kMklLayoutRewritePassGroup != |
| 4211 | OptimizationPassRegistry::POST_PARTITIONING) { |
| 4212 | // For any pre-partitioning phase, a graph is stored in options.graph. |
| 4213 | process_graph(options.graph); |
| 4214 | } else { |
| 4215 | // For post partitioning phase, graphs are stored in |
| 4216 | // options.partition_graphs. |
| 4217 | for (auto& pg : *options.partition_graphs) { |
| 4218 | process_graph(&pg.second); |
| 4219 | } |
| 4220 | } |
| 4221 | |
| 4222 | return OkStatus(); |
| 4223 | } |
| 4224 | |
| 4225 | } // namespace tensorflow |
| 4226 | |
| 4227 | #endif // INTEL_MKL |
| 4228 |
Generated on 2022-Sep-26 from project tensorflow revision f2e850b6cce
Powered by 
Code Browser 2.1
Generator usage only permitted with license.