"""
Copyright (c) OpenMMLab. All rights reserved.
Copyright (c) Huawei Technologies Co., Ltd. 2024. All rights reserved.
Modification by: Huawei Developers
Modification date: 2024-10-06
Modification Description:
Modification 1. Add support for Ascend NPU
"""
import torch
import torch_npu
from torch.autograd import Function
import mx_driving._C
class GaussianFunction(Function):
@staticmethod
def forward(
ctx,
boxes: torch.Tensor,
out_size_factor,
gaussian_overlap,
min_radius,
voxel_size_x,
voxel_size_y,
pc_range_x,
pc_range_y,
feature_map_size_x,
feature_map_size_y,
norm_bbox=True,
with_velocity=True,
flip_angle=False,
max_objs=500,
):
if (torch.numel(boxes) == 0):
raise Exception("Error! Input Tensor can not be an empty Tensor.\n")
result = mx_driving._C.npu_gaussian(
boxes,
out_size_factor,
gaussian_overlap,
min_radius,
voxel_size_x,
voxel_size_y,
pc_range_x,
pc_range_y,
feature_map_size_x,
feature_map_size_y,
norm_bbox,
with_velocity,
flip_angle,
max_objs
)
return result
npu_gaussian = GaussianFunction.apply