cumm

CUda Matrix Multiply library.

cumm is developed during learning of CUTLASS, which use too much c++ template and make code unmaintainable. So I develop pccm, use python as meta programming language, to replace c++ template meta programming. Now pccm become a foundational framework of cumm and my other c++ project such as spconv. cumm also contains a python asyncio-based gemm simulator that share same meta program with CUDA code, enable gemm visualization and easy debug experience.

Install

Prebuilt

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 10.2/11.1/11.3/11.4 prebuilt binaries for windows 10/11.

We will offer prebuilts for CUDA versions supported by latest pytorch release. For example, pytorch 1.9 support cuda 10.2 and 11.1, so we support them too.

pip install cumm for CPU-only

pip install cumm-cu102 for CUDA 10.2

pip install cumm-cu111 for CUDA 11.1

pip install cumm-cu113 for CUDA 11.3

pip install cumm-cu114 for CUDA 11.4

Build from source for development (JIT, recommend for develop)

WARNING Use code in tags!!! code in main branch may contain bugs.

The c++ code will be built automatically when you change c++ code in project.

Linux

0. uninstall cumm installed by pip. you must ensure no "cumm" exists in pip list | grep cumm
1. install build-essential, install CUDA
2. git clone https://github.com/FindDefinition/cumm, cd ./cumm, pip install -e .
3. in python, import cumm and wait for build finish.

Windows

0. uninstall spconv and cumm installed by pip. you must ensure no "cumm" exists in pip list | grep cumm
1. install visual studio 2019 or newer. make sure C++ development component is installed. install CUDA
2. set powershell script execution policy
3. start a new powershell, run tools/msvc_setup.ps1
4. git clone https://github.com/FindDefinition/cumm, cd ./cumm, pip install -e .
5. in python, import cumm and wait for build finish.

Build wheel from source

WARNING Use code in tags!!! code in main branch may contain bugs.

WARNING: If CUMM_CUDA_VERSION is set with a CUDA version, following steps will create a wheel named "cumm-cuxxx", not "cumm", this means you must use cumm-cuxxx in dependency of your project which depend on cumm, not cumm. If CUMM_CUDA_VERSION isn't set, cumm will always built with CUDA, so the CUDA must exists in your system. The wheel name will be cumm even if it is built with cuda.

Linux

It's recommend to build Linux packages in official build docker. Build with CUDA support don't need a real GPU.

Build in Official Docker

1. select a cuda version. available: CUDA 10.2, 11.1, 11.3, 11.4, 11.5
2. (Example for CUDA 11.4) git clone https://github.com/FindDefinition/cumm, cd ./cumm, docker run --rm -e PLAT=manylinux2014_x86_64 -e CUMM_CUDA_VERSION=114 -v `pwd`:/io scrin/manylinux2014-cuda:cu114-devel-1.0.0 bash -c "source /etc/bashrc && /io/tools/build-wheels.sh"

Build in your environment

1. install build-essential, install CUDA
2. set env for installed cuda version. for example, export CUMM_CUDA_VERSION="11.4". If you want to build CPU-only, run export CUMM_CUDA_VERSION="". If CUMM_CUDA_VERSION isn't set, you need to ensure cuda libraries are inside OS search path, and the built wheel name will be cumm, otherwise cumm-cuxxx
3. run export CUMM_DISABLE_JIT="1"
4. run python setup.py bdist_wheel+pip install dists/xxx.whl

Windows 10/11

1. install visual studio 2019 or newer. make sure C++ development package is installed. install CUDA
2. set powershell script execution policy
3. start a new powershell, run tools/msvc_setup.ps1
4. set env for installed cuda version. for example, $Env:CUMM_CUDA_VERSION = "11.4". If you want to build CPU-only, run $Env:CUMM_CUDA_VERSION = "". . If CUMM_CUDA_VERSION isn't set, you need to ensure cuda libraries are inside OS search path, and the built wheel name will be cumm, otherwise cumm-cuxxx
5. run $Env:CUMM_DISABLE_JIT = "1"
6. run python setup.py bdist_wheel+pip install dists/xxx.whl

Note

The work is done when the author is an employee at Tusimple.

LICENSE

Apache 2.0