Deep Learning Compiler
- Why: to be able to holistically evaluate and optimize the entire network, including across-layer optimizations, which cannot be done by interpreters
- What: compiler that translate a computational graph to target-specific kernels that can be executed in the target platform
- How:
- Graph Level Optimizations
- Layout Optimization
- Kernel Selection and Kernel Generation
- Scheduling
- Tensor Allocation
- How good: out-of-the-box high performance
实例
- Myelin
- TensorRT
- XLA
- PyTorch JIT
- TVM
相关内容
| Traditional Compiler | DLC | |
|---|---|---|
| Input | High-level languages | computational graph |
| Output | Low-level languages like asm | kernels |