想查询ParallelMLP.forward中RowParallelLinear.forward的耗时,降序排列。查看耗时最多的事件并分析原因

1.打开https://ui.perfetto.dev 导入Chrome Trace Json文件

2.ParallelMLP.forward下的RowParallelLinear.forward

3.点击Query(SQL),在输入框中输入以下内容,按Ctrl+Enter,显示查询结果

WITH transformers AS (
  SELECT 
    ts AS start_ts, 
    ts + dur AS end_ts
  FROM slice
  WHERE name = 'ParallelMLP.forward'
)

SELECT 
  s.* 
FROM slice s
JOIN transformers t
ON s.ts >= t.start_ts AND (s.ts + s.dur) <= t.end_ts
WHERE s.name = 'RowParallelLinear.forward' ORDER BY s.dur DESC;

4.点击Show timeline,点击Standalone Query Tab中相应的行,Timeline会跳到相应的位置