Changelog

[2.1.13] - 2021-?-?

Added

Add some ops from spconv 1.x, see spconv.utils for more details.
Add some debug tool for users to attach more info in issue.

[2.1.12] - 2021-11-23

Added

Add a method for voxel generator to get pc_voxel_id, which is usually used in semantic segmentation

Fixed

Fix a bug in cuda voxel generater when max_voxels is smaller than real number of voxels

[2.1.11] - 2021-11-22

Fixed

Fixed a bug Volta kernels (TITAN V, Tesla V100), backward weight kernels use f16 as accumulator. we should use f32.
Fixed a corner case when user use kernel size = 1x1 but stride != 1.
Fixed a corner case when input feature is non-contiguous when maxpool.

[2.1.10] - 2021-11-19

Fixed

Fixed a bug in utils.PointToVoxel, shouldn't get cuda stream in cpu code

[2.1.9] - 2021-11-18

Removed

Remove a wrong assert

[2.1.8] - 2021-11-15

Added

Add support for pytorch 1.5

[2.1.7] - 2021-11-11

Fixed

Fix a bug when net have inverse and run inference in eval mode.

[2.1.6] - 2021-11-10

Fixed

Fix missing -fopenmp in linker for CPU only

Removed

remove stale comment sending in CI

[2.1.5] - 2021-11-10

Added

Add cuda profile tool
Add python 36 support

Changed

Format all code

Removed

remove a unnecessary device sync and slightly improve performance.

[2.1.4] - 2021-11-10

Fixed

Fix a bug of SparseInverseConv3d

[2.1.3] - 2021-11-08

Fixed

Fix a bug of CPU only package

[2.1.2] - 2021-11-06

Fixed

Fix a bug of python 3.7

[2.1.0] - 2021-10-31

Added

add implicit gemm algorithm for all kind of convolution with kernel volume <= 32. this algorithm is very fast with float16.
add pytorch wrapper for voxel generator
add CPU support and CPU-only build.

[2.0.2] - 2021-10-26

Fixed

Fix a serious bug that do nothing with non-spconv layers in SparseSequential
Fix a bug of ProxyableClassMeta

[2.0.0] - 2021-10-16

Changed

Change build system from cmake to pccm.
Change pytorch python code to spconv.pytorch
Rewrite All c++ code.

[1.2.1] - 2020-06-04

Changed

The subm indice pair generation speed is greatly increased by two tricks: 1. most subm conv use only kernelsize=3, so we can unroll loops to get 100% performance increase. 2. subm indice pairs have a property: indicePairs[0, i] = indicePairs[1, kernelVolume - i - 1], so we can get another 100% performance increase.

[1.2.0] - 2020-05-28

Added

add batch gemm support. small performance increasement but more gpu memory usage. you can use algo=spconv.ConvAlgo.Batch to use it.

Changed

replace most of 'functor' with c++14 dispatch in c++ code.

Fixed

change gather/scatterAdd kernel parameter to support large points.