[PATCH] drm/panfrost: Add governor data with pre-defined thresholds

[Steven Price]

On 22/01/2021 10:11, Lukasz Luba wrote:
> 
> 
> On 1/21/21 5:15 PM, Daniel Lezcano wrote:
>> On 21/01/2021 18:04, Lukasz Luba wrote:
>>> The simple_ondemand devfreq governor uses two thresholds to decide about
>>> the frequency change: upthreshold, downdifferential. These two tunable
>>> change the behavior of the governor decision, e.g. how fast to increase
>>> the frequency or how rapidly limit the frequency. This patch adds needed
>>> governor data with thresholds values gathered experimentally in 
>>> different
>>> workloads.
>>>
>>> Signed-off-by: Lukasz Luba <lukasz.luba at arm.com>
>>> ---
>>> Hi all,
>>>
>>> This patch aims to improve the panfrost performance in various 
>>> workloads,
>>> (benchmarks, games). The simple_ondemand devfreq governor supports
>>> tunables to tweak the behaviour of the internal algorithm. The default
>>> values for these two thresholds (90 and 5) do not work well with 
>>> panfrost.
>>> These new settings should provide good performance, short latency for
>>> rising the frequency due to rapid workload change and decent freq slow
>>> down when the load is decaying. Based on frequency change statistics,
>>> gathered during experiments, all frequencies are used, depending on
>>> the load. This provides some power savings (statistically). The highest
>>> frequency is also used when needed.
>>>
>>> Example glmark2 results:
>>> 1. freq fixed to max: 153
>>> 2. these new thresholds values (w/ patch): 151
>>> 3. default governor values (w/o patch): 114
>>>
>>> In future the devfreq framework would expose via sysfs these two
>>> tunables, so they can be adjusted by the middleware based on currently
>>> running workload (game, desktop, web browser, etc). These new values
>>> should be good enough, though.
>>>
>>> Regards,
>>> Lukasz Luba
>>>
>>>   drivers/gpu/drm/panfrost/panfrost_devfreq.c | 10 +++++++++-
>>>   drivers/gpu/drm/panfrost/panfrost_devfreq.h |  2 ++
>>>   2 files changed, 11 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/gpu/drm/panfrost/panfrost_devfreq.c 
>>> b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
>>> index 56b3f5935703..7c5ffc81dce1 100644
>>> --- a/drivers/gpu/drm/panfrost/panfrost_devfreq.c
>>> +++ b/drivers/gpu/drm/panfrost/panfrost_devfreq.c
>>> @@ -130,8 +130,16 @@ int panfrost_devfreq_init(struct panfrost_device 
>>> *pfdev)
>>>       panfrost_devfreq_profile.initial_freq = cur_freq;
>>>       dev_pm_opp_put(opp);
>>> +    /*
>>> +     * Setup default thresholds for the simple_ondemand governor.
>>> +     * The values are chosen based on experiments.
>>> +     */
>>> +    pfdevfreq->gov_data.upthreshold = 45;
>>> +    pfdevfreq->gov_data.downdifferential = 5;
>>> +
>>>       devfreq = devm_devfreq_add_device(dev, &panfrost_devfreq_profile,
>>> -                      DEVFREQ_GOV_SIMPLE_ONDEMAND, NULL);
>>> +                      DEVFREQ_GOV_SIMPLE_ONDEMAND,
>>> +                      &pfdevfreq->gov_data);
>>>       if (IS_ERR(devfreq)) {
>>>           DRM_DEV_ERROR(dev, "Couldn't initialize GPU devfreq\n");
>>>           ret = PTR_ERR(devfreq);
>>> diff --git a/drivers/gpu/drm/panfrost/panfrost_devfreq.h 
>>> b/drivers/gpu/drm/panfrost/panfrost_devfreq.h
>>> index db6ea48e21f9..1e2a4de941aa 100644
>>> --- a/drivers/gpu/drm/panfrost/panfrost_devfreq.h
>>> +++ b/drivers/gpu/drm/panfrost/panfrost_devfreq.h
>>> @@ -4,6 +4,7 @@
>>>   #ifndef __PANFROST_DEVFREQ_H__
>>>   #define __PANFROST_DEVFREQ_H__
>>> +#include <linux/devfreq.h>
>>>   #include <linux/spinlock.h>
>>>   #include <linux/ktime.h>
>>> @@ -17,6 +18,7 @@ struct panfrost_devfreq {
>>>       struct devfreq *devfreq;
>>>       struct opp_table *regulators_opp_table;
>>>       struct thermal_cooling_device *cooling;
>>> +    struct devfreq_simple_ondemand_data gov_data;
>>>       bool opp_of_table_added;
>>>       ktime_t busy_time;
>>
>> I think it is simpler to do:
>>
>> +static struct devfreq_simple_ondemand_data panfrost_ondemand_data = {
>> +       .upthreshold = 45,
>> +       .downdifferential = 5,
>> +};
>>
>> [ ... ]
>>
>>         devfreq = devm_devfreq_add_device(dev, &panfrost_devfreq_profile,
>> -                                         DEVFREQ_GOV_SIMPLE_ONDEMAND,
>> NULL);
>> +                                         DEVFREQ_GOV_SIMPLE_ONDEMAND,
>> +                                         &panfrost_ondemand_data);
>>
>>
> 
> Yes, it's simpler. The driver would probably never have to serve two
> GPUs. I've tried to keep this thing inside the panfrost struct,
> forgetting about it.

The Juno platform with an FPGA attached is the only example I know of 
where a system has multiple Mali GPUs - so it can happen, but it rare.

As it stands a static structure would work because the values are 
constant - but Lukasz mentioned that they would be exported in sysfs in 
the future, in which case they really should be part of the panfrost struct.

Ultimately having a (non-const) static struct like above would mean 
wasting a few bytes on systems with Panfrost loaded but no Mali GPU. 
Having it in struct panfrost means the cost is only for Mali. Admittedly 
it's only a few bytes in this case and often Panfrost will be a module.

Steve

> Steven already reviewed the patch, so it can probably stay.
> I will keep it in mind. Thank you for the comments.
> 
> Regards,
> Lukasz