syntax = "proto3";

package xla.gpu;

import "tensorflow/compiler/xla/xla_data.proto";

// Backend configs for XLA:GPU.
//
// These are metadata that the GPU backend attaches to HloInstrucitons and later
// uses during e.g. codegen.
//
// Remember that proto3 doesn't give clients a way to tell the difference
// between a field not being present and a field having the default value.
// Choose your defaults carefully.
//
// No guarantee is made about the stability of these protos.
//
// See HloInstruction::backend_config() for more info.

// Backend config for a convolution that runs through cudnn.
message CudnnConvBackendConfig {
  // Opaque algorithm number of cudnn algorithm chosen for this conv.
  int64 algorithm = 1;

  // Whether we may use tensor cores when running this conv.  Even if this is
  // true, cudnn may choose not to use tensor cores, e.g. because the GPU or
  // selected algorithm doesn't support it.
  bool tensor_ops_enabled = 2;

  // The scaling factor multiplied with the convolution result.
  double conv_result_scale = 4;

  // Below are the fields related to cuDNN's fused convolution. Refer to
  // CudnnConvParams for their meanings.

  // The requested activation (e.g. relu) after the convolution. It is with type
  // stream_executor::dnn::ActivationMode.
  int64 activation_mode = 3;

  // The scaling factor multiplied with the side input. If no side input buffer
  // is provided, this field must be 0.
  double side_input_scale = 5;
}

// Backend config for the GEMM operation running through cuBLAS.
message GemmBackendConfig {
  // Opaque optional algorithm number. No chosen number indicates that a
  // different cuBLAS API will be used, which does not allow for choosing an
  // algorithm.
  oneof algorithm {
    int64 selected_algorithm = 1;
  }

  double alpha_real = 2;
  double alpha_imag = 9;

  double beta = 3;

  xla.DotDimensionNumbers dot_dimension_numbers = 7;

  int64 batch_size = 8;
}