tensorpipe_utils.h

#pragma once

#ifdef USE_TENSORPIPE

#include <torch/csrc/distributed/rpc/utils.h>

namespace tensorpipe {
class Message;
class Allocation;
class Descriptor;
} // namespace tensorpipe

namespace torch {
namespace distributed {
namespace rpc {

TORCH_API const c10::Stream& getStreamForDevice(
    const std::vector<c10::Stream>& streams,
    const c10::Device& device);

// Inspired by c10/core/impl/DeviceGuardImplInterface.h.

class TensorpipeDeviceTypeConverter {
 public:
  // Ideally we'd want this to also return a tensorpipe::Message::Tensor object
  // but we cannot forward-declare that class (because it's nested), and we
  // cannot include the TensorPipe headers because it's a private dependency.
  // Thus we bend over backwards and entrust this method with appending that
  // object to the `tensors` field of the tensorpipe::Message object we pass.
  virtual std::optional<std::vector<char>> prepareTensorForSending(
      const c10::Storage& storage,
      const std::vector<c10::Stream>& streams,
      tensorpipe::Message& message) const = 0;

  // Same as above: this method cannot return a tensorpipe::Allocation::Tensor,
  // thus it appends it to the `tensors` field of the tensorpipe::Allocation.
  virtual at::DataPtr allocateTensorForReceiving(
      c10::DeviceIndex deviceIndex,
      size_t length,
      const std::vector<c10::Stream>& streams,
      tensorpipe::Allocation& allocation) const = 0;

  virtual ~TensorpipeDeviceTypeConverter() = default;
};

extern TORCH_API std::array<
    std::atomic<const TensorpipeDeviceTypeConverter*>,
    static_cast<size_t>(DeviceType::COMPILE_TIME_MAX_DEVICE_TYPES)>
    device_type_converter_registry;

class TORCH_API TensorpipeDeviceTypeConverterRegistrar {
 public:
  TensorpipeDeviceTypeConverterRegistrar(
      DeviceType,
      const TensorpipeDeviceTypeConverter*);
};

#define C10_REGISTER_TENSORPIPE_DEVICE_TYPE_CONVERTER(                     \
    DevType, TensorpipeDeviceTypeConverter)                                \
  static ::torch::distributed::rpc::TensorpipeDeviceTypeConverterRegistrar \
      C10_ANONYMOUS_VARIABLE(g_##DeviceType)(                              \
          ::c10::DeviceType::DevType, new TensorpipeDeviceTypeConverter());

inline const TensorpipeDeviceTypeConverter* getDeviceTypeConverter(
    DeviceType type) {
  return device_type_converter_registry[static_cast<size_t>(type)].load();
}

// A struct that holds pointers that keep alive all the memory that will be
// accessed by TensorPipe during a write operation.
struct TensorpipeWriteBuffers {
  // Allocate on heap so pointers stay valid as we move the holder.
  std::unique_ptr<MessageType> type;
  std::unique_ptr<int64_t> id;
  std::vector<char> payload;
  std::vector<char> pickle;
  // This contains the original tensors and the clones of the sparse tensors.
  std::vector<torch::Tensor> tensors;
  // This contains the copies of the data of the tensors that didn't own their
  // memory, e.g., the ones created from torch::from_blob() with no deleter.
  std::vector<std::vector<char>> copiedTensors;
};

// A struct that holds pointers that keep alive all the memory that will be
// accessed by TensorPipe during a read operation.
struct TensorpipeReadBuffers {
  // Allocate on heap so pointers stay valid as we move the holder.
  std::unique_ptr<MessageType> type;
  std::unique_ptr<int64_t> id;
  std::vector<char> payload;
  std::vector<char> pickle;
  std::vector<c10::DataPtr> tensors;
};

// Convert an RPC message into a TensorPipe message, plus a holder to all the
// data that must be kept alive while the write is performed asynchronously.
TORCH_API std::tuple<tensorpipe::Message, TensorpipeWriteBuffers>
tensorpipeSerialize(
    c10::intrusive_ptr<Message> rpcMessage,
    std::vector<c10::Device> devices,
    const std::vector<c10::Stream>& streams);

// Allocate the buffers that will hold the incoming data. They will be managed
// by the returned holder, which must be kept alive until the asynchronous read
// has finished. Pointers to these buffers will be stored in the returned
// tensorpipe::Allocation struct.
TORCH_API std::pair<tensorpipe::Allocation, TensorpipeReadBuffers>
tensorpipeAllocate(
    const tensorpipe::Descriptor& tpDescriptor,
    const std::vector<c10::Stream>& streams);

// Convert a TensorPipe message back into an RPC message. This requires the data
// to be available and can thus only be performed once the asynchronous read has
// completed. The holder can be destroyed once this function returns.
TORCH_API c10::intrusive_ptr<Message> tensorpipeDeserialize(
    tensorpipe::Descriptor&& tpDescriptor,
    TensorpipeReadBuffers&& holder);

} // namespace rpc
} // namespace distributed
} // namespace torch

#endif // USE_TENSORPIPE