ascendc-evaluation

To validate AscendC custom operators, run the evaluation workflow to compare Model and ModelNew outputs for parity. It checks correctness by running scripts against reference implementations in your output directory.

Benchmark AscendC operator performance by measuring latency and memory to compare reference vs custom implementations. The workflow supports cross-backend performance comparison for both Triton and AscendC backends.

Yes, you need torch_npu and torch dependencies to evaluate AscendC custom operators. Environment setup scripts are provided to prepare the necessary dependencies for running correctness checks and benchmarks.

The required directory structure for AscendC operator evaluation is output/<op_name>/. This directory must contain both reference and custom implementations to execute correctness verification and performance benchmarking.

Yes, you can compare Triton and AscendC backends during operator evaluation. The workflow provides end-to-end script execution and result reporting to facilitate cross-backend performance comparison and integration testing.