| 319 | } |
| 320 | |
| 321 | bool IsMultiDevice(const FunctionDef* fdef) { |
| 322 | if (fdef == nullptr) { |
| 323 | // Primitive op. |
| 324 | return false; |
| 325 | } |
| 326 | |
| 327 | // Run all functions as multi-device. |
| 328 | return true; |
| 329 | |
| 330 | // We can eliminate some overhead by running simple functions using regular |
| 331 | // CallOp kernel. However, it is tricky to figure out which functions should |
| 332 | // be run using CallOp. Also, currently CallOp runs neither optimization |
| 333 | // passes (needed for TPU/XLA) nor grappler. |
| 334 | // Here are some cases where a function should be run in multi-device mode: |
| 335 | // - Function takes at least two resources on different devices. |
| 336 | // - Function takes a resource on deviceA and a body op explicitly placed |
| 337 | // on deviceB. |
| 338 | // - Function has a colocation constraint. |
| 339 | // - Function has an explicit device annotation (which might not be using |
| 340 | // full canonical device name) different from op_device. Note that false |
| 341 | // positives are ok. |
| 342 | // - Function has a node or a (node) attribute that can potentially make |
| 343 | // the function multi-device after a rewrite pass (e.g. various XLA/TPU |
| 344 | // special nodes and attributes) |
| 345 | } |
| 346 | |
| 347 | Status GetDeviceForInput(const EagerContext* ctx, TensorHandle* tensor_handle, |
| 348 | Device** result) { |
no outgoing calls
no test coverage detected