Function BroadcastAdd4DSlow

tensorflow/lite/kernels/internal/reference/add.h:178–216

TODO(jiawen): We can implement BroadcastAdd on buffers of arbitrary dimensionality if the runtime code does a single loop over one dimension that handles broadcasting as the base case. The code generator would then generate max(D1, D2) nested for loops. TODO(benoitjacob): BroadcastAdd is intentionally duplicated from reference_ops.h. Once an optimized version is implemented and NdArrayDesc is n

// is no longer referenced in this file, move NdArrayDesc<T> from types.h to
// reference_ops.h.
inline void BroadcastAdd4DSlow(const ArithmeticParams& params,
                               const RuntimeShape& input1_shape,
                               const float* input1_data,
                               const RuntimeShape& input2_shape,
                               const float* input2_data,
                               const RuntimeShape& output_shape,
                               float* output_data) {
  NdArrayDesc<4> desc1;
  NdArrayDesc<4> desc2;
  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
                                      &desc2);
  const RuntimeShape extended_output_shape =
      RuntimeShape::ExtendedShape(4, output_shape);

  // In Tensorflow, the dimensions are canonically named (batch_number, row,
  // col, channel), with extents (batches, height, width, depth), with the
  // trailing dimension changing most rapidly (channels has the smallest stride,
  // typically 1 element).
  //
  // In generated C code, we store arrays with the dimensions reversed. The
  // first dimension has smallest stride.
  //
  // We name our variables by their Tensorflow convention, but generate C code
  // nesting loops such that the innermost loop has the smallest stride for the
  // best cache behavior.
  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
          output_data[Offset(extended_output_shape, b, y, x, c)] =
              ActivationFunctionWithMinMax(
                  input1_data[SubscriptToIndex(desc1, b, y, x, c)] +
                      input2_data[SubscriptToIndex(desc2, b, y, x, c)],
                  params.float_activation_min, params.float_activation_max);
        }
      }
    }
  }
}

inline void BroadcastAdd4DSlow(const ArithmeticParams& params,
                               const RuntimeShape& input1_shape,