How can min_area_polygons be backpropagated?

[nijkah]

RepPointHead series use mmcv.ops.min_area_polygons.
e.g.

mmrotate/mmrotate/models/dense_heads/oriented_reppoints_head.py

Lines 478 to 512 in 4bbef14

	polygons_pred_init = min_area_polygons(pos_pts_pred_init)
	polygons_pred_refine = min_area_polygons(pos_pts_pred_refine)
	sampling_pts_pred_init = self.sampling_points(
	polygons_pred_init, 10, device=device)
	sampling_pts_pred_refine = self.sampling_points(
	polygons_pred_refine, 10, device=device)
	sampling_pts_gt = self.sampling_points(pos_bbox_gt, 10, device=device)
	# quality of orientation
	qua_ori_init = self.ori_qua_weight * ChamferDistance2D(
	sampling_pts_gt, sampling_pts_pred_init)
	qua_ori_refine = self.ori_qua_weight * ChamferDistance2D(
	sampling_pts_gt, sampling_pts_pred_refine)
	# quality of localization
	qua_loc_init = self.loss_bbox_refine(
	pos_pts_pred_init,
	pos_bbox_gt,
	pos_bbox_weight,
	avg_factor=self.loss_cls.loss_weight,
	reduction_override='none')
	qua_loc_refine = self.loss_bbox_refine(
	pos_pts_pred_refine,
	pos_bbox_gt,
	pos_bbox_weight,
	avg_factor=self.loss_cls.loss_weight,
	reduction_override='none')
	# quality of classification
	qua_cls = qua_cls.sum(-1)
	# weighted inti-stage and refine-stage
	qua = qua_cls + self.init_qua_weight * (
	qua_loc_init + qua_ori_init) + (1.0 - self.init_qua_weight) * (
	qua_loc_refine + qua_ori_refine) + qua_poc

When I check the implementation of min_area_polygons, there is no backward kernel for this operation.

• https://github.com/open-mmlab/mmcv/blob/master/mmcv/ops/csrc/common/cuda/min_area_polygons_cuda.cuh
• https://github.com/open-mmlab/mmcv/blob/master/mmcv/ops/csrc/pytorch/min_area_polygons.cpp

I wonder how this can be backpropagated without writing backward kernel?
Is it just used as 'weight' for differentiable output?