[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [PATCH v13 04/10] accel/tcg: add jit stats and time to TBStatistics
From: |
Wu, Fei |
Subject: |
Re: [PATCH v13 04/10] accel/tcg: add jit stats and time to TBStatistics |
Date: |
Thu, 1 Jun 2023 20:48:40 +0800 |
User-agent: |
Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:102.0) Gecko/20100101 Thunderbird/102.11.0 |
On 6/1/2023 7:51 PM, Alex Bennée wrote:
>
> "Wu, Fei" <fei2.wu@intel.com> writes:
>
>> On 5/30/2023 6:08 PM, Alex Bennée wrote:
>>>
>>> "Wu, Fei" <fei2.wu@intel.com> writes:
>>>
>>>> On 5/30/2023 1:01 PM, Wu, Fei wrote:
>>>>> On 5/30/2023 12:07 PM, Richard Henderson wrote:
>>>>>> On 5/29/23 04:49, Fei Wu wrote:
> <snip>
>> ----------------
>> IN:
>> Priv: 1; Virt: 0
>>
>> 0xffffffff800abe14: 864a mv a2,s2
>> 0xffffffff800abe16: 85ce mv a1,s3
>> 0xffffffff800abe18: 8526 mv a0,s1
>> 0xffffffff800abe1a: 46bd addi a3,zero,15
>> 0xffffffff800abe1c: fffff097 auipc ra,-4096
>> # 0xffffffff800aae1c
>> 0xffffffff800abe20: cc0080e7 jalr ra,ra,-832
>> ------------------------------
>>
>> Look at the tb with phys:0x2abe14, although the first time IR takes
>> 75274ns, but in the second command we can see it takes much less time
>> (Note IR time is accumulated).
>>
>> So if the time for the same TB is not consistent, and the deviation
>> could be dominated by system events such as memory allocation instead of
>> codegen itself (?), I think it's less useful.
>>
>> Alex, regarding dropping time profile, do you mean remove TB_JIT_TIME
>> completely?
>
> I think so - perf would do a better job of separating system events from
> the core code as it has better visibility of the whole system.
>
OK, got it.
Richard, can we reach the agreement here? If yes, I will remove
TB_JIT_TIME and all the time stuffs (dev_time, cpu_exec_time)
completely. We can still add it back if it proves to be useful. The only
concern is that tbstats replaces CONFIG_PROFILER but drops this function
from it.
Thanks,
Fei.