Clojure bindings and exploration of the tvm library.
TVM is a high performance compiler for ND numeric code. In it's simplest form, it works via 4 steps:
- Define an AST.
- Schedule the AST, doing things such as tiling and operation or caching a partial result in GPU shared memory. This allows us to make transformations to the algorithm which allow us to map the algorithm to specific hardware such as GPU's, FPGA's, web-based backends such as wasm, graphics backends such as OpenGL and Vulkan and low powered IoT platforms such as microcontrollers. These transformations are are guaranteed not to break the algorithm so they are very safe from a correctness viewpoint but also allow powerful vectorizing, SIMD, and SIMT optimizations.
- Compile the AST to a specific hardware profile. TVM has backends to a wide variety of hardware including, as mentioned, extremely optimized versions for x86 and ARM CPUs, Cuda, and OpenCL.
- Load your function and call it.
Getting all the source
At top level:
git submodule update --init --recursive
sudo apt install make g++ cmake llvm-dev libopenblas-dev ## opencl support (nvidia-cuda includes this) sudo apt install ocl-icd-* opencl-headers ## Cuda support sudo apt install nvidia-cuda-toolkit ## intel graphics adapter support sudo apt install beignet beignet-opencl-icd ## MKL support if you choose. I don't use it generally so this is very optional. curl https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS-2019.PUB | sudo apt-key add - sudo sh -c 'echo deb https://apt.repos.intel.com/mkl all main > /etc/apt/sources.list.d/intel-mkl.list' sudo apt-get update ## Find the version of mkl...I would take the latest. apt-cache search mkl-64bit ## ... sudo apt-get install intel-mkl-64bit-2019.5-075 mkdir -p tvm/build # Config setup for intel and such. # Base config.cmake file only has support for opencl. If you want # CUDA, CUDNN, or MKL I suggest you edit the config file after you copy # it. cp config.cmake tvm/build/ cd tvm/build cmake .. make -j8
This will copy the libs into a platform-specific directory that jna should find.
Another options is to install the tvm libs themselves. We recommend this pathway as then the tvm libraries will work with the python bindings. In fact, it can be worth it to install the python bindings as there are a lot of examples in python that are instructive to work with.
Distributed under the Eclipse Public License either version 1.0 or (at your option) any later version.