Function Prepare

tensorflow/lite/kernels/svdf.cc:76–193

}

TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
  const auto* params = reinterpret_cast<TfLiteSVDFParams*>(node->builtin_data);
  OpData* op_data = reinterpret_cast<OpData*>(node->user_data);
  int scratch_tensor_index = op_data->scratch_tensor_index;

  // Check we have all the inputs and outputs we need.
  TF_LITE_ENSURE_EQ(context, node->outputs->size, 1);
  TF_LITE_ENSURE_EQ(context, node->inputs->size, 5);
  op_data->activation_state_tensor_index =
      node->inputs->data[kInputActivationStateTensor];

  const TfLiteTensor* input = GetInput(context, node, kInputTensor);
  const TfLiteTensor* weights_feature =
      GetInput(context, node, kWeightsFeatureTensor);
  const TfLiteTensor* weights_time =
      GetInput(context, node, kWeightsTimeTensor);

  TF_LITE_ENSURE_EQ(context, input->type, kTfLiteFloat32);

  // Check all the parameters of tensor match within themselves and match the
  // input configuration.
  const int rank = params->rank;
  const int batch_size = input->dims->data[0];
  const int num_filters = weights_feature->dims->data[0];
  TF_LITE_ENSURE_EQ(context, num_filters % rank, 0);
  const int num_units = num_filters / rank;
  const int memory_size = weights_time->dims->data[1];
  TF_LITE_ENSURE_EQ(context, input->dims->data[1],
                    weights_feature->dims->data[1]);
  TF_LITE_ENSURE_EQ(context, weights_time->dims->data[0], num_filters);

  const TfLiteTensor* bias = GetOptionalInputTensor(context, node, kBiasTensor);
  if (bias) {
    TF_LITE_ENSURE_EQ(context, bias->dims->data[0], num_units);
  }

  TfLiteTensor* activation_state =
      &context->tensors[op_data->activation_state_tensor_index];
  TfLiteTensor* output = GetOutput(context, node, kOutputTensor);

  // Check the shape of input state tensors.
  TF_LITE_ENSURE_EQ(context, NumDimensions(activation_state), 2);
  TF_LITE_ENSURE_EQ(context, SizeOfDimension(activation_state, 0), batch_size);
  TF_LITE_ENSURE_EQ(context, SizeOfDimension(activation_state, 1),
                    memory_size * num_filters);

  // Resize output.
  TfLiteIntArray* output_size_array = TfLiteIntArrayCreate(2);
  output_size_array->data[0] = batch_size;
  output_size_array->data[1] = num_units;
  TF_LITE_ENSURE_OK(context,
                    context->ResizeTensor(context, output, output_size_array));

  // The weights are of consistent type, so it suffices to check one.
  const bool is_hybrid_op = IsHybridOp(input, weights_feature);

  // Resize scratch.
  TfLiteIntArrayFree(node->temporaries);