This release will focus on the core framework, it plans to support the following features:
- CPU X86: high performance CodeGenC_X86 and CodeGenLLVM_X86 backend,
- CUDA: the GPU related schedule, the base performance,
- The core framework:
- a well-encapsulated JIT framework that can call CINN expression or external functions
- full support for extern-Call, the ability to trigger the MKL or CUBLAS functions,
- a computation definition and schedule framework to provide the ability to define new algorithm and performance automatically tune.
- API
- python api for HLIR and CINN(optional)
- C++ APIs
- Documents to introduce
- the basic concepts
- the architecture
- the usage of the APIs
- example to develop new HLIR/primitive instructions
- Some basic primitives:
- the actives,
- Dot,
- some binary ones, such as Add, Sub and so on,
- Conv
- provide a real model and its benchmark