Awesome

grain2

Autograd and GPGPU library for dynamic neural networks in D. This project is a successor of grain.

features

• @nogc support
• heterogeneous device (CPU/CUDA/OpenCL) support
• multiple CUDA device support
• see also document

requirements

• dpp master
• CPU backend
  • BLAS library (e.g., OpenBLAS, MKL)
• CUDA backend
  • CUDA header/library v10-
  • CUDNN header/library v7-
• OpenCL backend
  • OpenCL header/library v1.2

development

• test dub test grain2:core and dub test grain2:cuda
• generate ddoc make doc

roadmap

• heterogeneous multi-dimentional structure (grain.tensor.Tensor)
• reference count-based memory management
• multiple CUDA devices support
• core autograd (forward/backward chain)
• dataset API (MNIST, PTB, etc)
• more memory manegement policy (garbage collection, memory pool)
• JIT compilation of computational graph
• OpenCL/HIP support

trouble shooting

• Q. My local CUDA/OpenCL libraries are different from files under grain2/dpp/build
  • A. Install dpp (commit 821a5d4) and regenerate them by $ cd dpp; make -j cuda (or opencl)