[PATCH] perf annotate: Remove hist__account_cycles from callback

[PATCH] perf annotate: Remove hist__account_cycles from callback

[Jin Yao]

The hist__account_cycles is executed when the hist_iter__branch_callback
is called. But it looks it's not necessary. In hist__account_cycles, it
already walks on all branch entries.

This patch moves the hist__account_cycles out of callback, now the data
processing is much faster than before.

For example,
perf record -b ...
perf annotate

The before/after output should be no change.

Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
---
 tools/perf/builtin-annotate.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/tools/perf/builtin-annotate.c b/tools/perf/builtin-annotate.c
index 67f9d9f..77deb3a 100644
--- a/tools/perf/builtin-annotate.c
+++ b/tools/perf/builtin-annotate.c
@@ -159,8 +159,6 @@ static int hist_iter__branch_callback(struct hist_entry_iter *iter,
 	struct perf_evsel *evsel = iter->evsel;
 	int err;
 
-	hist__account_cycles(sample->branch_stack, al, sample, false);
-
 	bi = he->branch_info;
 	err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
 
@@ -199,6 +197,8 @@ static int process_branch_callback(struct perf_evsel *evsel,
 	if (a.map != NULL)
 		a.map->dso->hit = 1;
 
+	hist__account_cycles(sample->branch_stack, al, sample, false);
+
 	ret = hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
 	return ret;
 }
-- 
2.7.4

Re: [PATCH] perf annotate: Remove hist__account_cycles from callback

[Jin, Yao]

Any comments for this patch or any issue is found? In testing, I just 
feel it accelerates the perf annotate data processing.

Thanks
Jin Yao

On 3/13/2019 4:22 AM, Jin Yao wrote:
> The hist__account_cycles is executed when the hist_iter__branch_callback
> is called. But it looks it's not necessary. In hist__account_cycles, it
> already walks on all branch entries.
> 
> This patch moves the hist__account_cycles out of callback, now the data
> processing is much faster than before.
> 
> For example,
> perf record -b ...
> perf annotate
> 
> The before/after output should be no change.
> 
> Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
> ---
>   tools/perf/builtin-annotate.c | 4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/tools/perf/builtin-annotate.c b/tools/perf/builtin-annotate.c
> index 67f9d9f..77deb3a 100644
> --- a/tools/perf/builtin-annotate.c
> +++ b/tools/perf/builtin-annotate.c
> @@ -159,8 +159,6 @@ static int hist_iter__branch_callback(struct hist_entry_iter *iter,
>   	struct perf_evsel *evsel = iter->evsel;
>   	int err;
>   
> -	hist__account_cycles(sample->branch_stack, al, sample, false);
> -
>   	bi = he->branch_info;
>   	err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
>   
> @@ -199,6 +197,8 @@ static int process_branch_callback(struct perf_evsel *evsel,
>   	if (a.map != NULL)
>   		a.map->dso->hit = 1;
>   
> +	hist__account_cycles(sample->branch_stack, al, sample, false);
> +
>   	ret = hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
>   	return ret;
>   }
> 

Re: [PATCH] perf annotate: Remove hist__account_cycles from callback

[Jiri Olsa]

On Wed, Mar 13, 2019 at 04:22:54AM +0800, Jin Yao wrote:
> The hist__account_cycles is executed when the hist_iter__branch_callback
> is called. But it looks it's not necessary. In hist__account_cycles, it
> already walks on all branch entries.
> 
> This patch moves the hist__account_cycles out of callback, now the data
> processing is much faster than before.
> 
> For example,
> perf record -b ...
> perf annotate
> 
> The before/after output should be no change.

so in the old code it was called multiple times per sample,
while after your change it's called just once per sample

it increases cycles stats for symbol, so it's strange
the behaviour is the same.. could you please exaplin this
in more detail?

> 
> Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
> ---
>  tools/perf/builtin-annotate.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/tools/perf/builtin-annotate.c b/tools/perf/builtin-annotate.c
> index 67f9d9f..77deb3a 100644
> --- a/tools/perf/builtin-annotate.c
> +++ b/tools/perf/builtin-annotate.c
> @@ -159,8 +159,6 @@ static int hist_iter__branch_callback(struct hist_entry_iter *iter,
>  	struct perf_evsel *evsel = iter->evsel;
>  	int err;
>  
> -	hist__account_cycles(sample->branch_stack, al, sample, false);
> -

there's similar code in perf report, could you please cover
that as well?

thanks,
jirka

>  	bi = he->branch_info;
>  	err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
>  
> @@ -199,6 +197,8 @@ static int process_branch_callback(struct perf_evsel *evsel,
>  	if (a.map != NULL)
>  		a.map->dso->hit = 1;
>  
> +	hist__account_cycles(sample->branch_stack, al, sample, false);
> +
>  	ret = hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
>  	return ret;
>  }
> -- 
> 2.7.4
> 

Re: [PATCH] perf annotate: Remove hist__account_cycles from callback

[Jin, Yao]

On 3/14/2019 8:04 PM, Jiri Olsa wrote:
> On Wed, Mar 13, 2019 at 04:22:54AM +0800, Jin Yao wrote:
>> The hist__account_cycles is executed when the hist_iter__branch_callback
>> is called. But it looks it's not necessary. In hist__account_cycles, it
>> already walks on all branch entries.
>>
>> This patch moves the hist__account_cycles out of callback, now the data
>> processing is much faster than before.
>>
>> For example,
>> perf record -b ...
>> perf annotate
>>
>> The before/after output should be no change.
> 
> so in the old code it was called multiple times per sample,
> while after your change it's called just once per sample
> 
> it increases cycles stats for symbol, so it's strange
> the behaviour is the same.. could you please exaplin this
> in more detail?
> 

In __symbol__account_cycles,
ch[offset].num_aggr++;
ch[offset].cycles_aggr += cycles;

In annotation__compute_ipc,
al = notes->offsets[offset];
al->cycles = ch->cycles_aggr / ch->num_aggr;

So the num_aggr and cycles_aggr are both increased n times, then the 
final result is the same.

>>
>> Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
>> ---
>>   tools/perf/builtin-annotate.c | 4 ++--
>>   1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/tools/perf/builtin-annotate.c b/tools/perf/builtin-annotate.c
>> index 67f9d9f..77deb3a 100644
>> --- a/tools/perf/builtin-annotate.c
>> +++ b/tools/perf/builtin-annotate.c
>> @@ -159,8 +159,6 @@ static int hist_iter__branch_callback(struct hist_entry_iter *iter,
>>   	struct perf_evsel *evsel = iter->evsel;
>>   	int err;
>>   
>> -	hist__account_cycles(sample->branch_stack, al, sample, false);
>> -
> 
> there's similar code in perf report, could you please cover
> that as well?
> 

Sure, let me have a try.

Thanks
Jin Yao

> thanks,
> jirka
> 
>>   	bi = he->branch_info;
>>   	err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
>>   
>> @@ -199,6 +197,8 @@ static int process_branch_callback(struct perf_evsel *evsel,
>>   	if (a.map != NULL)
>>   		a.map->dso->hit = 1;
>>   
>> +	hist__account_cycles(sample->branch_stack, al, sample, false);
>> +
>>   	ret = hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
>>   	return ret;
>>   }
>> -- 
>> 2.7.4
>>

Re: [PATCH] perf annotate: Remove hist__account_cycles from callback

[Jin, Yao]

On 3/14/2019 10:17 PM, Jin, Yao wrote:
> 
> 
> On 3/14/2019 8:04 PM, Jiri Olsa wrote:
>> On Wed, Mar 13, 2019 at 04:22:54AM +0800, Jin Yao wrote:
>>> The hist__account_cycles is executed when the hist_iter__branch_callback
>>> is called. But it looks it's not necessary. In hist__account_cycles, it
>>> already walks on all branch entries.
>>>
>>> This patch moves the hist__account_cycles out of callback, now the data
>>> processing is much faster than before.
>>>
>>> For example,
>>> perf record -b ...
>>> perf annotate
>>>
>>> The before/after output should be no change.
>>
>> so in the old code it was called multiple times per sample,
>> while after your change it's called just once per sample
>>
>> it increases cycles stats for symbol, so it's strange
>> the behaviour is the same.. could you please exaplin this
>> in more detail?
>>
> 
> In __symbol__account_cycles,
> ch[offset].num_aggr++;
> ch[offset].cycles_aggr += cycles;
> 
> In annotation__compute_ipc,
> al = notes->offsets[offset];
> al->cycles = ch->cycles_aggr / ch->num_aggr;
> 
> So the num_aggr and cycles_aggr are both increased n times, then the 
> final result is the same.
> 

There might be another potential issue in annotation__count_and_fill.

/* Hide data when there are too many overlaps. */
if (ch->reset >= 0x7fff || ch->reset >= ch->num / 2)
	return;

When ch->reset is 0 and ch->num is 1, it returns directly. But for this 
case, the data doesn't need to be hidden since there is no overlap 
happening.

I will post v2 and add this fix.

Thanks
Jin Yao

>>>
>>> Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
>>> ---
>>>   tools/perf/builtin-annotate.c | 4 ++--
>>>   1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/tools/perf/builtin-annotate.c 
>>> b/tools/perf/builtin-annotate.c
>>> index 67f9d9f..77deb3a 100644
>>> --- a/tools/perf/builtin-annotate.c
>>> +++ b/tools/perf/builtin-annotate.c
>>> @@ -159,8 +159,6 @@ static int hist_iter__branch_callback(struct 
>>> hist_entry_iter *iter,
>>>       struct perf_evsel *evsel = iter->evsel;
>>>       int err;
>>> -    hist__account_cycles(sample->branch_stack, al, sample, false);
>>> -
>>
>> there's similar code in perf report, could you please cover
>> that as well?
>>
> 
> Sure, let me have a try.
> 
> Thanks
> Jin Yao
> 
>> thanks,
>> jirka
>>
>>>       bi = he->branch_info;
>>>       err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
>>> @@ -199,6 +197,8 @@ static int process_branch_callback(struct 
>>> perf_evsel *evsel,
>>>       if (a.map != NULL)
>>>           a.map->dso->hit = 1;
>>> +    hist__account_cycles(sample->branch_stack, al, sample, false);
>>> +
>>>       ret = hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
>>>       return ret;
>>>   }
>>> -- 
>>> 2.7.4
>>>