Rockchip RK3399 (ROCKPro64) boots to multiuser

[Greg V]

On Wed, Aug 15, 2018 at 11:57 PM, Greg V <greg at unrelenting.technology> 
wrote:
> 
> 
> On Wed, Aug 15, 2018 at 11:44 PM, Emmanuel Vadot 
> <manu at bidouilliste.com> wrote:
>> On Wed, 15 Aug 2018 22:41:34 +0300
>> Greg V <greg at unrelenting.technology> wrote:
>>>  Alright everyone, good news ? I managed to reclock the CPU!!!
>>> 
>>>  The patch is now at https://reviews.freebsd.org/D16732
>> 
>>  Thanks a lot !!
>>  I'll have a deeper look when I'm back from BSDCam.
>> 
>>>  (and I think the style is more correct now. Though it's really 
>>> fscking
>>>  silly that the style doesn't like making "table-like" structures 
>>> look
>>>  like tables, i.e. with one-line "rows".)
>>> 
>>>  Plus the hack you need to reclock the CPU right now at
>>>  https://gist.github.com/myfreeweb/88cb9340652f56498f4be770c77b9d61
>>> 
>>>  (the hack allows cpufreq_dt to deal with clock only, no voltage ?
>>>  since we don't have all the drivers for voltage.)
>> 
>>  Are you able to switch to any frequency with that ?
>>  I would expect the cpu to hang if the voltage is too low or too 
>> high.
>> (I encounter that on RK3328)
> 
> Yeah — I maxed the clocks for both big and LITTLE cores and got 
> pretty great performance.
> 
> e.g. unixbench dhrystone index with cpuset to a big core: 804 — 
> which is more than the 737 I got on Scaleway's ThunderX VPS!
> ThunderX is still way better on unixbench's other tests though.
> Not that unixbench is a great test…
> 
> Compiling neovim also took *way* less time than on RPi/ROCK64.
> 
> So, I think the big cores' voltage regulator (silergy,syr827) might 
> just default to the highest voltage.
> The chip gets rather warm when just idling in FreeBSD…

Update: tried porting the fanpwr driver from OpenBSD:

https://gist.github.com/myfreeweb/584de9b746a328e10c904395afe8a48f

Reports 1.0V on boot. For some reason, cpufreq doesn't see the 
regulator though — any idea why could that be??
(cpufreq_dt shouldn't require the controller and regulator to be 
separate nodes, right? There are other drivers like sy8106a where it's 
all one node…)

Also, overclocked to 2.184GHz, still works great (benchmark score went 
up again.)

I guess either the syr827 is not actually running 1.0 V, or the 
provided table is waaaay overvolted, or I won the silicon lottery and 
my chip is just that good.
Maybe I should write an efuse driver to look at the leakage 
measurements…