Voxel count limit warning in spconv 1.2.1 - need configurable maximum voxel threshold

[LexcaliburR]

After upgrading to CUDA CenterPoint library version 1.2.1, we encountered a warning about voxel count limits that was not present in version 1.1.10.

Current Behavior (v1.2.1)

Version 1.2.1 introduced a theoretical maximum voxel count limit of 100,000. When processing point clouds with high voxel counts, the system generates the following， warning:

[TIME] Voxelization:            1.99907 ms
valid_num: 128219
Warning: Found the number of input[128219d] points exceeds the theoretical maximum number[100000], the default processing is to truncate.
[engine.cu:1087] (Verb) Forward[conv0:Float16] SparseConvolution, input0=128219x5[41x1440x1440]Float16() | output0=100000x16[41x1440x1440]Float16(1)
[engine.cu:1087] (Verb) Forward[conv1:Float16] SparseConvolution, input0=100000x16[41x1440x1440]Float16(1) | output0=100000x16[41x1440x1440]Float16(2)
[engine.cu:1087] (Verb) Forward[conv2:Float16] SparseConvolution, input0=100000x16[41x1440x1440]Float16(2) | output0=100000x16[41x1440x1440]Float16(3)

Expected Behavior (v1.1.10)

[TIME] Voxelization:            1.92131 ms
valid_num: 128219
[Verb🚀] Forward[conv0:Float16] SparseConvolution, input0=128219x5[41x1440x1440]Float16() | output0=128219x16[41x1440x1440]Float16(1)
[Verb🚀] Forward[conv1:Float16] SparseConvolution, input0=128219x16[41x1440x1440]Float16(1) | output0=128219x16[41x1440x1440]Float16(2)
[Verb🚀] Forward[conv2:Float16] SparseConvolution, input0=128219x16[41x1440x1440]Float16(2) | output0=128219x16[41x1440x1440]Float16(3)

Impact

This truncation can lead to: Loss of point cloud information Potential degradation in detection accuracy