```spconv``` is a project that provide heavily-optimized sparse convolution implementation with tensor core support. check [benchmark](docs/BENCHMARK.md) to see how fast spconv 2.x runs.
```spconv``` is a project that provide heavily-optimized sparse convolution implementation with tensor core support. check [benchmark](docs/BENCHMARK.md) to see how fast spconv 2.x runs.
See [dev plan](docs/SPCONV_DEVELOP_PLAN.md). A complete guide of spconv development will be released soon.
See [dev plan](docs/SPCONV_DEVELOP_PLAN.md). A complete guide of spconv development will be released soon.
## Usage
## Usage
Firstly you need to use ```import spconv.pytorch as spconv``` in spconv 2.x.
Firstly you need to use ```import spconv.pytorch as spconv``` in spconv 2.x.
...
@@ -78,7 +76,7 @@ Don't forget to check [performance guide](docs/PERFORMANCE_GUIDE.md).
...
@@ -78,7 +76,7 @@ Don't forget to check [performance guide](docs/PERFORMANCE_GUIDE.md).
## Install
## Install
You need to install python >= 3.7 first to use spconv 2.x.
You need to install python >= 3.6 (>=3.7 for windows) first to use spconv 2.x.
You need to install CUDA toolkit first before using prebuilt binaries or build from source.
You need to install CUDA toolkit first before using prebuilt binaries or build from source.
...
@@ -86,7 +84,9 @@ You need at least CUDA 10.2 to build and run spconv 2.x. We won't offer any supp
...
@@ -86,7 +84,9 @@ You need at least CUDA 10.2 to build and run spconv 2.x. We won't offer any supp
### Prebuilt
### Prebuilt
We offer python 3.7-3.10 and cuda 10.2/11.1/11.3/11.4 prebuilt binaries for linux (manylinux) and windows 10/11.
We offer python 3.6-3.10 and cuda 10.2/11.1/11.3/11.4 prebuilt binaries for linux (manylinux).
We offer python 3.7-3.10 and cuda 10.2/11.1/11.4 prebuilt binaries for windows 10/11.
We will provide prebuilts for CUDA versions supported by latest pytorch release. For example, pytorch 1.10 provide cuda 10.2 and 11.3 prebuilts, so we provide them too.
We will provide prebuilts for CUDA versions supported by latest pytorch release. For example, pytorch 1.10 provide cuda 10.2 and 11.3 prebuilts, so we provide them too.
*```x.features = F.relu(x.features)``` now raise error. use ```x = x.replace_feature(F.relu(x.features))``` instead.
*```x.features = F.relu(x.features)``` now raise error. use ```x = x.replace_feature(F.relu(x.features))``` instead.
* weight layout has been changed to RSKC (native algorithm) or KRSC (implicit gemm), no longer RSCK (spconv 1.x). RS is kernel size, C is input channel, K is output channel.
* weight layout has been changed to RSKC (native algorithm) or KRSC (implicit gemm), no longer RSCK (spconv 1.x). RS is kernel size, C is input channel, K is output channel.
* all util ops are removed (pillar scatter/nms/rbbox_iou...)
* all util ops are removed (pillar scatter/nms/rbbox_iou...)
* VoxelGenerator has been replaced by Point2VoxelGPU[1-4]d/Point2VoxelCPU[1-4]d.
* VoxelGenerator has been replaced by ```spconv.pytorch.utils.PointToVoxel``` (torch API) or Point2VoxelGPU[1-4]d/Point2VoxelCPU[1-4]d (tv.Tensor API).
* spconv < 2.1 don't support CPU. spconv 2.1+ support cpu for debug usage.
* spconv < 2.1 don't support CPU. spconv 2.1+ support cpu for debug usage.
* test spconv 1.x model in spconv 2.x: set environment variable before run program. Linux: ```export SPCONV_FILTER_HWIO="1"```, Windows powershell: ```$Env:SPCONV_FILTER_HWIO = "1"```. **WARNING** test spconv 1.x model don't support implicit gemm algorithm, you need to train from scratch with spconv 2.x and select ConvAlgo.MaskSplitImplicitGemm.
* test spconv 1.x model in spconv 2.x: Firstly set environment variable before run program, Then set all ```algo``` in conv/pool to ```ConvAlgo.Native```. Linux: ```export SPCONV_FILTER_HWIO="1"```, Windows powershell: ```$Env:SPCONV_FILTER_HWIO = "1"```. **WARNING** test spconv 1.x model don't support implicit gemm algorithm.
