|
Mila
Deep Neural Network Library
|
Device-to-backend map behind the TensorOps alias template. More...
Device-to-backend map behind the TensorOps alias template.
Public because TensorOps is: a consumer instantiating a device-neutral entry point resolves the alias through this map. Adding a backend means specializing it.
Each backend specializes this in its own module with a single type alias, for the same reason as Compute::ExecutionContextTraits: MSVC 14.51 will not complete a merely reachable explicit specialization when the dereference is dependent, and the device-neutral entry points in Dnn.TensorOps all dispatch through a dependent TensorOps<TMemoryResource::device_type>. Holding the specializations here lets Dnn.TensorOps publish the map without publishing the backends.
| TDevice | Compute device type to map. |