Add eval_backend_perf to further refactor previous test_compiler

[JewelRoam]

PR Category

Feature

Description

Usage:

python3 -m graph_net_bench.torch.eval_backend_diff \
    --model-path-list $model_list \
    --eval-config=$(base64 -w 0 <<EOF
{
    "ref_env":  {
        "seed": int,
        "compiler": str,
        "device": str,
        "op_lib": str,
        "warmup": int,
        "trials": int,
        "log_prompt": str,
        "model_path_prefix": str,
        "backend_config": dict
    },
    "target_env": {
        "seed": int,
        "compiler": str,
        "device": str,
        "op_lib": str,
        "warmup": int,
        "trials": int,
        "log_prompt": str,
        "model_path_prefix": str,
        "backend_config": dict
    }
}
EOF
) \

TODO

目前为能跑通的初始版本，有很大优化空间：

• 添加output文件传递和rpc支持，取代原有test_device
• 改进函数及config中参数的语义
• 把编译器和算子库的注册改为动态链接
• 添加可选use-subprocess支持，优化args传递，改进效率
• ……待补充

[paddle-bot]

Thanks for your contribution!