pytorch

cffa70d3 - Merge channel hierarchies. (#54333)

Merge channel hierarchies. (#54333) Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/54333 Pull Request resolved: https://github.com/pytorch/tensorpipe/pull/326 Pull Request resolved: https://github.com/pytorch/tensorpipe/pull/312 This is a first step towards cross-device type transfers: eventually, channels will not connect devices of a given type between two hosts, but possibly heterogeneous pairs of devices. Hence, the distinction between CPU-to-CPU and GPU-to-GPU channels will not make much sense anymore, and we can afford to simplify the Pipe's code quite bit. The main change here is that the `channel::Channel` and `channel::Context` classes are not templated (on the buffer type) anymore. Instead, a channel's `send`/`recv` methods act on generic `Buffer`s and the actual unpacking is done in the `ChannelBoilerplate`. The `channel::CpuContext`/`channel::CudaContext` (respectively `channel::CudaContext`/`channel::CudaChannel`) aliases now simply resolve to `channel::Context` (respectively `channel::Channel`). A subsequent diff will get rid of the aliases altogether. The Pipe is being simplified: all the duplication due to having separate hierarchies is gone, which gets rid of a lot of boiler plate template code. Note that previously, two channels with the same name could potentially coexist, provided one was a CPU channel and the other a GPU channel. This is not the case anymore, though it should not matter. In its current state, the Pipe still needs to pick a channel based on whether that channel acts on CPU or GPU buffers. This is solved by introducing the temporary method `bool channel::Context::supportsDeviceType(DeviceType t)`. When iterating through available channels to select one for a given tensor, the Pipe now discards channels that do not support the tensor's `DeviceType`. This leads to having a single ordered list of channels, which in practice is two separate lists (one for CPU, one for GPU) merged together. This will change soon as we initialize only one channel per `DeviceType`. Test Plan: Imported from OSS Reviewed By: lw Differential Revision: D26958187 Pulled By: beauby fbshipit-source-id: 3e3f7921166892d468fa78cfad3199277588021c