Fast Ops for the DiffKt Tensor library

Dev setup

Local prereqs

Install onednn (formerly mkl-dnn): brew install onednn

Install openMP: brew install libomp

Install eigen (Optional): brew install eigen

Install MKL (Optional) on mac:

1. download Intel oneAPI Base Toolkit : https://software.intel.com/content/www/us/en/develop/tools/oneapi/all-toolkits.html#base-kit
2. install Intel oneAPI Base Toolkit : select Custom Installation -> un-select all but Intel oneAPI Math Kernel Library. Then consent, install, finish.
3. download Intel oneAPI HPC Toolkit : https://software.intel.com/content/www/us/en/develop/tools/oneapi/all-toolkits.html#base-kit
4. install Intel oneAPI HPC Toolkit : select Custom Installation -> un-select Intel Fortan Compiler Classic. Then continue, skip.
5. run source /opt/intel/oneapi/setvars.sh (note that setvars.sh only sets up the environment variables for the terminal it runs in. So we need to run this every time before using MKL in a new terminal.)

Devserver prereqs

Install DNNL:

# from <repo root>/cpp/ops
wget $(fwdproxy-config wget) https://github.com/oneapi-src/oneDNN/releases/download/v2.1/dnnl_lnx_2.1.0_cpu_gomp.tgz
tar -zxvf dnnl_lnx_2.1.0_cpu_gomp.tgz
mv dnnl_lnx_2.1.0_cpu_gomp dnnl

Or you can grab a more recent release from here.

Setup

To build:

# One-time or if you update the cmake version or local prereq version
rm -rf build
mkdir build && cd build

cmake ..

# Every time
make

This should produce libsparseops_jni and libops_jni in {repo_root}/kotlin/api/src/main/resources.

To auto-format code:

# From <repo root>/cpp/ops:
./scripts/format.sh

Upgrading DNNL

Update onednn: brew upgrade onednn

Download a specific version of onednn (formerly mkl-dnn):

1. brew search mkl-dnn will show PR links for older brew formulas to download mkl-dnn. If no PRs show up, you'll have to manually navigate to the correct github commit.
2. Navigate to desired PR in browser and navigate to the raw mkl-dnn.rb file.
3. Download the raw mkl-dnn.rb file. echo $(curl -fsSL <url to mkl-dnn.rb>) > mkl-dnn.rb. For example to download version 1.4: echo -e "$(curl -fsSL https://raw.githubusercontent.com/chenrui333/homebrew-core/93fd9e0e95cbfd065ef088bf5500129d606a1b38/Formula/mkl-dnn.rb)" > mkl-dnn.rb.
4. brew install --build-from-source mkl-dnn.rb

For newer versions (1.6.2+), replace mkl-dnn with onednn in the instructions above.

Select the sparse computation implementation

Three options are:

• MKL : a high performance parallel library.
• Eigen : a easy-to-use open-source, but not fully parallelized, BLAS library.
• OMP : parallel implementations with OpenMP. It doesn't rely on any third party library.

See section Local prereqs for installation instructions for Eigen and MKL.

Select with cmake:

cmake variable SPARSE_LIB defines which implementation to use.

• cmake -DSPARSE_LIB=MKL .., choose MKL
• cmake -DSPARSE_LIB=EIGEN .., choose Eigen
• cmake -DSPARSE_LIB=OMP .., choose OMP
• cmake .., then SPARSE_LIB=NONE. That is the user doesn't give preference on which implementation to use. Then predefined preference will be applied. Currently it's: MKL > Eigen > OMP. If a preferred third party library is not installed, then check on the next preferred choice. For example, if MKL is not installed, then choose Eigen. If Eigen is also not installed, then choose OMP.