we need to build pytorch from source first. To do that, follow the instruction here https://github.com/pytorch/pytorch#from-source.
Instruction for building on devfair:
module purge
module load cudnn/v7.4-cuda.10.0
module load cuda/10.0
# create a fresh conda environment with python3
conda create --name [your env name] python=3.7
conda activate [your env name] # Or source activate [your env name], depending on conda version.
conda install numpy pyyaml mkl mkl-include setuptools cmake cffi typing
conda install -c pytorch magma-cuda100
# clone the repo
# Note: put the repo onto /scratch partition for MUCH FASTER building speed.
git clone --recursive https://github.com/pytorch/pytorch
cd pytorch
export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
# set cuda arch list so that the built binary can be run on both pascal and volta
TORCH_CUDA_ARCH_LIST="6.0;7.0" python setup.py install
Then to build this repo, first change this line to the path of your newly installed pytorch
mkdir build
cd build
cmake ..
make
Note that we need to set the following before running any multi-threading program that uses torch::Tensor. Otherwise a simple tensor operation will use all cores by default.
export OMP_NUM_THREADS=1
Todo:
- coding convention & clang-format file
- split into .h and .cc properly
Todo++:
- improve performance by changing to lock-free structure
- multi-buffer
- we should be ablt to test environment in C++ alone
To explain:
- pybind shared_ptr
- pybind keep_alive
- torch accessor can still go out of bound, i.e. extremely wired segfault, we need to think about it