Curious about the bechmark in llama.cpp

[YaoCheng-Chang]

Since I try to estimate the HW device efficiency with running LLM in llama.cpp, I would like to understand the meanings of these time clearly.

Can I define that the prefill time is equal to the sum of sampling time, load time, and prompt eval time ?
My opinion is these benchmark are calculated before the first token is generated. However, if i lost the time of generating the first token in the eval time.

If someone have any opinion, please let me know. Thanks !