void EvalIntegerSvdfReference()

in src/tensorflow/lite/micro/kernels/svdf_common.cpp [52:192]
113 lines of code
16 McCabe index (conditional complexity)

void EvalIntegerSvdfReference(TfLiteContext* context, TfLiteNode* node,
                              const TfLiteEvalTensor* input_tensor,
                              const TfLiteEvalTensor* weights_feature_tensor,
                              const TfLiteEvalTensor* weights_time_tensor,
                              const TfLiteEvalTensor* bias_tensor,
                              const TfLiteSVDFParams* params,
                              TfLiteEvalTensor* activation_state_tensor,
                              TfLiteEvalTensor* output_tensor,
                              const OpDataSvdf& data) {
  const int n_rank = params->rank;
  const int n_batch = input_tensor->dims->data[0];
  const int n_input = input_tensor->dims->data[1];
  const int n_filter = weights_feature_tensor->dims->data[0];
  const int n_unit = n_filter / n_rank;
  const int n_memory = weights_time_tensor->dims->data[1];

  TFLITE_DCHECK(context != nullptr);
  TFLITE_DCHECK(context->GetScratchBuffer != nullptr);

  int32_t* scratch_tensor = static_cast<int32_t*>(
      context->GetScratchBuffer(context, data.scratch_tensor_index));
  int32_t* scratch_output_tensor = static_cast<int32_t*>(
      context->GetScratchBuffer(context, data.scratch_output_tensor_index));

  // Shift states.
  T* const state_ptr = tflite::micro::GetTensorData<T>(activation_state_tensor);

  // Left shift the activation_state.
  {
    T* new_state_start = state_ptr;
    const T* old_state_start = state_ptr + 1;
    const T* old_state_end = state_ptr + n_batch * n_filter * n_memory;
    while (old_state_start != old_state_end) {
      *new_state_start++ = *old_state_start++;
    }
  }

  // Note: no need to clear the latest activation, matmul is not accumulative.

  // Feature matmul.
  {
    T* state = tflite::micro::GetTensorData<T>(activation_state_tensor);
    const int8_t* input = tflite::micro::GetTensorData<int8_t>(input_tensor);
    const int8_t* weight_feature =
        tflite::micro::GetTensorData<int8_t>(weights_feature_tensor);
    const int32_t output_max = std::numeric_limits<T>::max();
    const int32_t output_min = std::numeric_limits<T>::min();
    T* result_in_batch = state + (n_memory - 1);
    for (int b = 0; b < n_batch; b++) {
      const int8_t* matrix_ptr = weight_feature;
      for (int r = 0; r < n_filter; r++) {
        int32_t dot_prod = 0;
        const int8_t* vector_in_batch = input + b * n_input;
        for (int c = 0; c < n_input; c++) {
          dot_prod +=
              *matrix_ptr++ * (*vector_in_batch++ - data.input_zero_point);
        }
        dot_prod = MultiplyByQuantizedMultiplier(
            dot_prod, data.effective_scale_1_a, data.effective_scale_1_b);
        dot_prod = std::min(std::max(output_min, dot_prod), output_max);
        // The int16 version of the op assumes a zero_point of 0.  This
        // code accounts for the potentially non-zero zero_point for the int8
        // version of the op.
        *result_in_batch = data.activation_state_zero_point + dot_prod;
        result_in_batch += n_memory;
      }
    }
  }

  // Time.
  {
    for (int b = 0; b < n_batch; ++b) {
      int32_t* scratch_ptr_batch = scratch_tensor + b * n_filter;

      // Perform batched vector dot product:
      const T* vector1_ptr =
          tflite::micro::GetTensorData<T>(weights_time_tensor);
      const T* vector2_ptr =
          tflite::micro::GetTensorData<T>(activation_state_tensor) +
          b * n_memory * n_filter;

      for (int i = 0; i < n_filter; i++) {
        *scratch_ptr_batch = 0;
        for (int j = 0; j < n_memory; j++) {
          *scratch_ptr_batch +=
              *vector1_ptr++ *
              (*vector2_ptr++ - data.activation_state_zero_point);
        }
        scratch_ptr_batch++;
      }
    }
  }

  // Reduce, add bias, rescale, activation.
  {
    // Add bias.
    if (bias_tensor) {
      // Vector batch assign:
      const int32_t* bias_data =
          tflite::micro::GetTensorData<int32_t>(bias_tensor);
      for (int i = 0; i < n_batch; ++i) {
        int32_t* output_ptr = scratch_output_tensor + i * n_unit;
        const int32_t* bias_ptr = bias_data;
        for (int j = 0; j < n_unit; ++j) {
          *output_ptr++ = *bias_ptr++;
        }
      }
    } else {
      int32_t* output_ptr = scratch_output_tensor;
      for (int i = 0; i < n_batch * n_unit; ++i) {
        *output_ptr++ = 0;
      }
    }

    // Reduce.
    for (int b = 0; b < n_batch; ++b) {
      int32_t* output_temp_ptr = scratch_output_tensor + b * n_unit;
      int32_t* scratch_ptr_batch = scratch_tensor + b * n_filter;

      // Reduction sum vector
      for (int i = 0; i < n_unit; ++i) {
        for (int j = 0; j < n_rank; ++j) {
          output_temp_ptr[i] += *scratch_ptr_batch++;
        }
      }
    }

    // Rescale.
    const int32_t output_max = std::numeric_limits<int8_t>::max();
    const int32_t output_min = std::numeric_limits<int8_t>::min();
    for (int i = 0; i < n_batch * n_unit; ++i) {
      int32_t x1 = scratch_output_tensor[i];
      int32_t x2 = MultiplyByQuantizedMultiplier(x1, data.effective_scale_2_a,
                                                 data.effective_scale_2_b);
      int32_t x3 = x2 + data.output_zero_point;
      int32_t x4 = std::min(std::max(output_min, x3), output_max);
      tflite::micro::GetTensorData<int8_t>(output_tensor)[i] =
          static_cast<int8_t>(x4);
    }
  }
}