Workflow

Workflow#

The diagram below provides a brief overview of how KernelGen assists in Kernel generation.

However, user configurations may vary depending on the selected platform, AI agent, skill, and specific use case. For more information, see KernelGen Web Platform User Guide, KernelGen Operator Development MCP Toolkit User Guide, and KernelGen Skills User Guide.

alt text

The generation process is as follows:

  1. Collect Kernel information: User enters semantic operator definitions into KernelGen, for example, by referring to the ReLU operator definitions. KernelGen collects operator basic parameters from the definitions.

  2. Search code snippets: KernelGen searches code snippets similar to user’s definitions as references and extracts Kernel parameters. During this step, the user can select to use the searched code snippets or not.

  3. Generate Kernel code and CUDA implementation code: KernelGen generates codes of Kernel and CUDA implementation. CUDA implementation is used as a PyTorch benchmark reference.

  4. Test Kernel based on CUDA implementation code：KernelGen tests Kernel based on the PyTorch benchmark, and outputs the test results of Correctness and Speedup Ratio.