Architecture

• common
  • uni hosts the common headers that are used in bolt.
  • gcl hosts the setup of GPU environment.
  • memory hosts the memory data management.
  • model_spec hosts the bolt model definition.
• model_tools
  • X2bolt : a general converter of converting different deep learning models to bolt models.
  • model_optimizer : a static computing graph optimizer to fuse the operators and simplify the calculation graph.
  • post_training_quantization : a post training quantization tool.
• compute
  • blas_enhance hosts the fast implementation of matrix-matrix multiplication and matrix-vector multiplication of FP32, FP16 and INT8. It is referenced by some of the operators in tensor.
  • tensor hosts the implementation for all kinds of operators defined by bolt.
  • image hosts common preprocessing routines for image inputs (e.g. bilinear interpolation).
• inference
  • engine hosts the inference engine of neural networks.
  • flow hosts the multi-backends(CPU+GPU) heterogeneous device schedule for time series data.
  • examples gives some application examples (Network Benchmark, ImageNet classification).
• kit
  • kit provides some application demos.

For API, Flow and operator development, please refer to DEVELOPER.md.