fixed wrong cur_geom_feats address in file of bev_pool_cuda.cu and add QuickCumsumMean function in torch

[ZouJiu1]

Contribution

I fix a problem with wrong cur_geom_feats address in file of bev_pool_cuda.cu.

I also add a new function QuickCumsumMean with pytorch implement.

Problem

The result is different between the output of function QuickCumsumCuda and the output of function QuickCumsum in file bev_pool.py.

The result is also different between the output of function QuickCumsumMeanCuda and the output of function QuickCumsumMean in file bev_pool.py.

the reason is the wrong implement of file bev_pool_cuda.cu

Reproduce

To reproduce the wrong different result, you can follow the instruction of file README.md https://github.com/ZouJiu1/bevPool/blob/master/README.md

step 1

reproduce the problem

download the file bev_pool_cuda.cu

mv bev_pool/src/bev_pool_cuda.cu bev_pool/src/bev_pool_cuda_tmp.cu

git clone https://github.com/ZouJiu1/bevPool.git

cd bevPool

pip install -e .

cd bev_pool

python3 bev_pool.py

by checking or debuging or printing the variables like error, errorMean, www, tmp, tmp2 in file bev_pool.py, you can reproduce the problem

step 2

fix the problem

mv bev_pool/src/bev_pool_cuda_tmp.cu bev_pool/src/bev_pool_cuda.cu

cd bevPool

pip install -e .

cd bev_pool

python3 bev_pool.py

by checking variables in file bev_pool.py, you will find the problem is fixed.

Environment:

Ubuntu22.04
Python 3.10.12
torch 2.2.2+cu118
torchvision 0.17.2+cu118
mmcv 2.1.0
mmengine 0.10.7
mmdet 3.3.0
mmdet3d 1.4.0
cuda 11.8
cudnn 8.9.7
gcc g++ 11.4.0

others:

hustvl/MapTR#212
autowarefoundation/autoware_universe#12491
open-mmlab/mmdetection3d#3143
#4
hustvl/VMA#21

[kdh044]

Subject: Question about Fisheye3DOD checkpoints

Hi, I saw your recent PR on Fisheye3DOD (the bev_pool_cuda.cu fix) — nice work.
I need the pretrained checkpoints (fisheye_bevdet.pth, fisheye_petr.pth) for my research, but Baidu Netdisk isn't accessible outside mainland China without a local phone number.
If you happened to download them while testing, would you mind sharing via Google Drive or WeTransfer? No worries at all if not.
Thanks.