| CUDA 11.4 | [![PyPI Version][pypi-ver-114]][pypi-url-114] | [![pypi monthly download][pypi-download-114]][pypi-url-114]|
```spconv``` is a project that provide heavily-optimized sparse convolution implementation with tensor core support.
```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 1.x code](https://github.com/traveller59/spconv/tree/v1.2.1). We won't provide any support for spconv 1.x since it's deprecated. use spconv 2.x if possible. <!--remove this message in spconv 2.2-->
[Spconv 1.x code](https://github.com/traveller59/spconv/tree/v1.2.1). We won't provide any support for spconv 1.x since it's deprecated. use spconv 2.x if possible. <!--remove this message in spconv 2.2-->
...
@@ -99,7 +119,10 @@ The c++ code will be built automatically when you change c++ code in project.
...
@@ -99,7 +119,10 @@ The c++ code will be built automatically when you change c++ code in project.
For NVIDIA Embedded Platforms, you need to specify cuda arch before build: ```export CUMM_CUDA_ARCH_LIST="7.2"``` for xavier.
For NVIDIA Embedded Platforms, you need to specify cuda arch before build: ```export CUMM_CUDA_ARCH_LIST="7.2"``` for xavier.
You need to remove ```cumm``` in ```requires``` section in pyproject.toml after install editable ```cumm``` and before install spconv due to pyproject limit (can't find editable installed ```cumm```).
The network/input/profile code is same as above table.
This table only profile **fp16 gemm kernels** without output tensor create/clear overhead. this table show the performance upper bound of our algorithm.
We can see that the implicit gemm is very fast, gemm only use 4.3ms/11.2ms in network forward. we can achieve better performance in TensorRT + Pure C++.
**NOTE**
When you want to benchmark network in your laptop, don't forget to close all apps except terminals! Other apps will consume GPU resource and make kernels run slower.
## Comparsion with [MinkowskiEngine](https://github.com/NVIDIA/MinkowskiEngine) and [torchsparse](https://github.com/mit-han-lab/torchsparse)
@@ -57,4 +52,4 @@ In my test, ```Implicit Gemm``` is almost 2x faster than ```Native```.
...
@@ -57,4 +52,4 @@ In my test, ```Implicit Gemm``` is almost 2x faster than ```Native```.
TODO
TODO
In my test, ```Implicit Gemm Split Mask``` is slightly faster than ```Implicit Gemm```, but the indice generation is greatly slower, so currently we use ```Implicit Gemm``` by default.
In my test, ```Implicit Gemm Split Mask``` is slightly faster than ```Implicit Gemm```, but the indice generation is slower, so currently we use ```Implicit Gemm``` by default.
