跳转至

Constraints and Precautions for MindStudio Kernel Performance Prediction

Development Constraints

  • When using the msKPP library to implement operator simulation, pay attention to the following points:
    • Before modeling a simulation operator, import Tensor, Chip, and instructions (in lowercase) required for operator implementation from the msKPP library.
    • Refer to the sample sample_vadd.py or sample_mmad.py in the project. Use the with statement to enable the entry of the operator implementation code. The enable_trace and enable_metrics APIs can enable the trace dotting and instruction statistics functions.

Runtime Constraints

  • Performance modeling results depend on time estimation based on input/output scales. No actual computation is performed, and the results are for peak performance reference only.
  • To generate the instruction proportion pie chart (instruction_cycle_consumption.html), the third-party Python library plotly must be installed in advance:
    pip3 install plotly
    

Security Precautions

  • Ensure that the input data is reliable and secure during secondary development.
  • The tool involves dynamic Python module loading during runtime. Ensure that dependency libraries in the runtime environment are from trusted sources to avoid arbitrary code injection risks.