Function Eval

tensorflow/lite/kernels/basic_rnn.cc:205–238

}

TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
  auto* params = reinterpret_cast<TfLiteRNNParams*>(node->builtin_data);

  const TfLiteTensor* input = GetInput(context, node, kInputTensor);
  const TfLiteTensor* input_weights = GetInput(context, node, kWeightsTensor);
  const TfLiteTensor* recurrent_weights =
      GetInput(context, node, kRecurrentWeightsTensor);
  const TfLiteTensor* bias = GetInput(context, node, kBiasTensor);
  TfLiteTensor* hidden_state =
      &context->tensors[node->inputs->data[kHiddenStateTensor]];
  TfLiteTensor* output = GetOutput(context, node, kOutputTensor);

  // We already checked that weight types are consistent, so branch on one.
  switch (input_weights->type) {
    case kTfLiteFloat32:
      return EvalFloat(input, input_weights, recurrent_weights, bias, params,
                       hidden_state, output);
    case kTfLiteUInt8:
    case kTfLiteInt8: {
      // TODO(mirkov): implement eval with quantized inputs as well.
      TfLiteTensor* input_quantized = GetTemporary(context, node, 0);
      TfLiteTensor* hidden_state_quantized = GetTemporary(context, node, 1);
      TfLiteTensor* scaling_factors = GetTemporary(context, node, 2);
      return EvalHybrid(input, input_weights, recurrent_weights, bias, params,
                        input_quantized, hidden_state_quantized,
                        scaling_factors, hidden_state, output);
    }
    default:
      context->ReportError(context, "Type %d not currently supported.",
                           input_weights->type);
      return kTfLiteError;
  }
  return kTfLiteOk;
}

}  // namespace rnn