Share performance profiling

[njzjz]

I just used TensorFlow's new profiling tool (tf.profiler.experimental.Profile) on the DeePMD-kit. I tested on examples/water/se_e2_a. Here's results:

overall

Average Step Time (lower is better) (σ = 19.4 ms) 19.0 ms
Kernel Launch Time (σ = 4.3 ms) 2.6 ms
Device Compute Time (σ = 16.5 ms) 16.4 ms

GPU Kernel Stats

data.txt

Memory

TF stats

Trace viewer

[amcadmus]

@denghuilu @iProzd @pkulzy Please take a look