[Enhance] Faster implementation of KNN

[Wuziyi616]

This new KNN code is borrowed from PAConv. They use heap data structure to accelerate so it's called knn_heap.

points.shape == [8, 4096, 3]
centers.shape == [8, 64, 3]
k = 32
100 loops:
The current knn needs ~8s, while knn_heap only needs < 0.1s!

Should we still keep the current knn in the codebase? After all, they have the same functionality. And knn is not used in our codebase now so this won't cause any BC-breaking. (Or we can do a import knn_heap as knn trick in mmdet3d/ops/__init__.py?)

[Wuziyi616]

My first time working with custom gpu ops. Please point out any mistakes I made.

[codecov]

Codecov Report

Merging #586 (6abea67) into master (d37c330) will increase coverage by 0.04%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master     #586      +/-   ##
==========================================
+ Coverage   50.77%   50.81%   +0.04%     
==========================================
  Files         197      197              
  Lines       14923    15052     +129     
  Branches     2426     2451      +25     
==========================================
+ Hits         7577     7649      +72     
- Misses       6851     6893      +42     
- Partials      495      510      +15     

Flag	Coverage Δ
unittests	50.81% <0.00%> (+0.04%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmdet3d/ops/knn/knn.py	30.00% <0.00%> (-2.15%)	⬇️
mmdet3d/apis/inference.py	53.62% <0.00%> (-5.58%)	⬇️
mmdet3d/core/bbox/structures/coord_3d_mode.py	63.12% <0.00%> (-1.85%)	⬇️
mmdet3d/core/visualizer/show_result.py	66.38% <0.00%> (-0.57%)	⬇️
mmdet3d/apis/test.py	15.78% <0.00%> (-0.43%)	⬇️
mmdet3d/datasets/nuscenes_dataset.py	40.70% <0.00%> (-0.29%)	⬇️
mmdet3d/apis/__init__.py	100.00% <0.00%> (ø)
mmdet3d/core/bbox/structures/utils.py	97.05% <0.00%> (+0.69%)	⬆️
mmdet3d/models/detectors/single_stage_mono3d.py	16.48% <0.00%> (+1.20%)	⬆️
mmdet3d/datasets/nuscenes_mono_dataset.py	38.27% <0.00%> (+4.94%)	⬆️
... and 1 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update d37c330...6abea67. Read the comment docs.

[ZwwWayne]

If the input and output of these two modules are the same, we can simply replace the original one for acceleration. Otherwise, we might need to prevent BC-breaking.

[Wuziyi616]

If the input and output of these two modules are the same, we can simply replace the original one for acceleration. Otherwise, we might need to prevent BC-breaking.

Yes, I modify the wrapper so that the input and output are the same now. Also, knn is not used in our current codebase (I understand some users may use it in their projects). I will replace the original one.

[yezhen17]

Maybe add a line somewhere telling users the speed difference of two implementations?

[Wuziyi616]

Maybe add a line somewhere telling users the speed difference of two implementations?

I will just replace the old one with the new KNN op since their interface are the same.