Overview

The operator compilation method described in this section allows you to manually configure compilation options using the BiSheng command lines and CMake, or implement compilation by writing CMake scripts. You can place the main.cpp file on the host and the kernel function on the device in the same implementation file to implement heterogeneous compilation.

• Currently, this compilation method supports only the following models:
  • Atlas A3 training products/Atlas A3 inference products
  • Atlas A2 training products/Atlas A2 inference products
  • Atlas inference products
• The current version does not support CPU twin debugging and NPU simulation debugging.
• If there are multiple kernel functions in the same compilation unit, the kernel type cannot be automatically inferred. You need to manually set the kernel type.
  • Particularly, for the following models, automatic inference is not supported when the kernel type is not set, regardless of whether there are multiple kernel functions in the same compilation unit. You are advised to manually set the kernel type.
    • Atlas inference products
  • For the Atlas inference products, the kernel type cannot be set to KERNEL_TYPE_MIX_VECTOR_CORE.