API 清单

Driving SDK提供下列高性能算子，列表中Released标注为N代表使用场景受限，仅针对本仓库的模型，标注为beta的算子为初版，未经过版本测试。

api类型	api名称	Released	A2	A3	A5
通用	hypot		✓	✓	✓
	scatter_max		✓	✓	✓
	three_interpolate		✓	✓
	knn		✓	✓	✓
	three_nn		✓	✓	✓
	scatter_mean	✓	✓	✓	✓
	assign_score_withk	✓	✓	✓	✓
	furthest_point_sampling		✓	✓	✓
	furthest_point_sample_with_dist		✓	✓	✓
	group_points		✓	✓	✓
	unique_voxel	✓	✓	✓	✓
	scatter_add	✓	✓	✓	✓
	radius		✓	✓	✓
	npu_unique[beta]		✓	✓
	graph_softmax		✓	✓	✓
	sigmoid_focal_loss				✓
	npu_fake_tensor_quant[beta]		✓	✓
	npu_fake_tensor_quant_inplace[beta]		✓	✓
	npu_fake_tensor_quant_with_axis[beta]		✓	✓
采样	roipoint_pool3d	✓	✓	✓	✓
	roiaware_pool3d		✓	✓
	npu_voxel_pooling_train		✓	✓	✓
	bev_pool_v1		✓	✓	✓
	bev_pool_v2		✓	✓	✓
	bev_pool_v3		✓	✓	✓
	roi_align_rotated		✓	✓
	border_align		✓	✓	✓
	pixel_group		✓	✓	✓
	grid_sampler2d_v2		✓	✓
	grid_sampler3d_v1		✓	✓
	npu_index_select[beta]		✓	✓
体素化	voxelization		✓	✓	✓
	dynamic_scatter		✓	✓	✓
检测	boxes_overlap_bev	✓	✓	✓	✓
	box_iou_quadri		✓	✓	✓
	box_iou_rotated[beta]		✓	✓	✓
	boxes_iou_bev	✓	✓	✓	✓
	nms3d		✓	✓	✓
	nms3d_normal		✓	✓	✓
	nms3d_on_sight	✓	✓	✓	✓
	npu_rotated_iou[beta]		✓	✓
	points_in_boxes_all		✓	✓	✓
	points_in_box		✓	✓	✓
	diff_iou_rotated_2d	✓	✓	✓	✓
稀疏	SparseConv3d		✓	✓	✓
	SubMConv3d		✓	✓	✓
	SparseInverseConv3d		✓	✓
融合	multi_scale_deformable_attn		✓	✓	✓
	deformable_aggregation		✓	✓	✓
	deform_conv2d		✓	✓
	modulated_deform_conv2d		✓	✓
	npu_add_relu	✓	✓	✓	✓
	npu_max_pool2d	✓	✓	✓
	geometric_kernel_attention		✓	✓	✓
	npu_gaussian		✓	✓	✓
	npu_draw_gaussian_to_heatmap[beta]		✓	✓	✓
	npu_assign_target_of_single_head		✓	✓	✓
	npu_fused_bias_leaky_relu	✓	✓	✓	✓
	npu_batch_matmul		✓	✓	✓
	cartesian_to_frenet		✓	✓	✓
	cal_anchors_heading		✓	✓	✓