Post Reply

Smartreflex - how to improve the algorithm

OP franciscofranco

26th January 2013, 07:57 PM   |  #11  
osm0sis's Avatar
Recognized Developer / Recognized Contributor
Thanks Meter: 14,513
 
8,454 posts
Join Date:Joined: Mar 2012
More
As I've always seen it, the limitations/problems with SmartReflex are as follows:

- Can't calibrate to a lower voltage than 830mV (problematic when 384MHz generally does fine at 775mV and 192MHz generally does fine at 725mV).

- Does not handle frequencies above 1305 well at all (calibrates them the same as 1228, so it must be an upper limit to its bracketing/range, almost as though it sees it as the same as 1228).

- Could definitely be a bit more aggressive safely when everyone can UV the SR values by another 50-100mV (so I'd say another 50mV safely across the board).

No idea how to proceed on any of those, but I thought it might help to make the limitations a bit more clear. I'm looking forward to following the progress of this thread. Great idea, Francisco.
Last edited by osm0sis; 27th January 2013 at 04:09 AM.
The Following 4 Users Say Thank You to osm0sis For This Useful Post: [ View ]
27th January 2013, 10:39 AM   |  #12  
Ashtrix's Avatar
Senior Member
Thanks Meter: 341
 
1,246 posts
Join Date:Joined: Dec 2010
More
Thumbs up Smartreflex improvement
Quote:
Originally Posted by osm0sis

As I've always seen it, the limitations/problems with SmartReflex are as follows:

- Can't calibrate to a lower voltage than 830mV (problematic when 384MHz generally does fine at 775mV and 192MHz generally does fine at 725mV).

- Does not handle frequencies above 1305 well at all (calibrates them the same as 1228, so it must be an upper limit to its bracketing/range, almost as though it sees it as the same as 1228).

- Could definitely be a bit more aggressive safely when everyone can UV the SR values by another 50-100mV (so I'd say another 50mV safely across the board).

No idea how to proceed on any of those, but I thought it might help to make the limitations a bit more clear. I'm looking forward to following the progress of this thread. Great idea, Francisco.

So you mean the Lean's 1.6Ghz SR does also do the same, 1.6GHz calibrates as 1.2GHz range instead of the real application ??
and it would be great if we could overcome the 2nd limitation .
27th January 2013, 03:43 PM   |  #13  
osm0sis's Avatar
Recognized Developer / Recognized Contributor
Thanks Meter: 14,513
 
8,454 posts
Join Date:Joined: Mar 2012
More
Most kernels (probably all but I haven't personally checked) disable SR for frequencies above 1305, but yes I believe that's why it doesn't work. Franco used to have the cutoff for SR be voltage based and on my device 1536 UV's by 100mV stably, so if I left SR on and lowered it too much (75mV) SR would act on it and calibrate it to the same voltage as 1228, which of course was too low and froze and crashed. So yeah, its as though it needs new brackets/ranges added for higher voltages.
Last edited by osm0sis; 27th January 2013 at 03:46 PM.
27th January 2013, 06:59 PM   |  #14  
blackhand1001's Avatar
Senior Member
Thanks Meter: 684
 
1,275 posts
Join Date:Joined: Sep 2011
Donate to Me
Has amazon released the kernel source for the kindle fire hd 8.9? That uses a ti omap clocked at 1.5ghz stock. Maybe a look at the smart reflex code on that would shed some light.
The Following 4 Users Say Thank You to blackhand1001 For This Useful Post: [ View ]
27th January 2013, 10:55 PM   |  #15  
osm0sis's Avatar
Recognized Developer / Recognized Contributor
Thanks Meter: 14,513
 
8,454 posts
Join Date:Joined: Mar 2012
More
Quote:
Originally Posted by supernova_00

I might be pointing out the obvious here but sometimes the obvious isn't so obvious...has anyone contacted Nishanth Menon that is named in the code header to see if TI has a latest and greatest SR? I would assume they would since phones are coming out now with higher CPU speeds then what the current SR could handle around two years ago. I see the guy is pretty active online with posting stuff to open source sites and still deals with smart reflex related things as this site shows: http://marc.info/?a=124043331500007&r=1&w=2
I dunno what that site is or know what any of that stuff means but I just stumbled upon it and just wanted to show he is active with TI and smart reflex work, among other things

Quote:
Originally Posted by blackhand1001

Has amazon released the kernel source for the kindle fire hd 8.9? That uses a ti omap clocked at 1.5ghz stock. Maybe a look at the smart reflex code on that would shed some light.

Both good ideas!
28th January 2013, 10:03 AM   |  #16  
Ashtrix's Avatar
Senior Member
Thanks Meter: 341
 
1,246 posts
Join Date:Joined: Dec 2010
More
Quote:
Originally Posted by osm0sis

Most kernels (probably all but I haven't personally checked) disable SR for frequencies above 1305, but yes I believe that's why it doesn't work. Franco used to have the cutoff for SR be voltage based and on my device 1536 UV's by 100mV stably, so if I left SR on and lowered it too much (75mV) SR would act on it and calibrate it to the same voltage as 1228, which of course was too low and froze and crashed. So yeah, its as though it needs new brackets/ranges added for higher voltages.

According to SR v1 by Imoseyon & his explanation says he had modified the SR code to override the below limits
1) VDDMIN - 830mV for all three domains,
2) VDDMAX - 1410mV for MPU, 1260mV for IVA, and 1200mV for CORE.

So, i think do already have a better version of SR over the TI's stock SR from Lean but only a few phones can handle the SR calibration for the 1.5+ GHz freq .
Last edited by Ashtrix; 28th January 2013 at 10:07 AM.
The Following User Says Thank You to Ashtrix For This Useful Post: [ View ]
28th January 2013, 07:28 PM   |  #17  
blackhand1001's Avatar
Senior Member
Thanks Meter: 684
 
1,275 posts
Join Date:Joined: Sep 2011
Donate to Me
Re: Smartreflex - how to improve the algorithm
Quote:
Originally Posted by Ashtrix

According to SR v1 by Imoseyon & his explanation says he had modified the SR code to override the below limits
1) VDDMIN - 830mV for all three domains,
2) VDDMAX - 1410mV for MPU, 1260mV for IVA, and 1200mV for CORE.

So, i think do already have a better version of SR over the TI's stock SR from Lean but only a few phones can handle the SR calibration for the 1.5+ GHz freq .

That's because it doesn't seem to actually calibrate properly for 1.5ghz. That's why I suggested looking at device like the kndl fire hd 8.9 which is 1.5 GHz stock. TI most likely modified the smartreflex code to work with that frequency.

Sent from my Galaxy Nexus using xda premium
The Following 3 Users Say Thank You to blackhand1001 For This Useful Post: [ View ]
22nd April 2013, 11:49 AM   |  #18  
Charmacas's Avatar
Senior Member
Flag Hamburg
Thanks Meter: 766
 
1,034 posts
Join Date:Joined: Jan 2010
Did you guys already give up the discussion or is there nothing more to discuss?

Sent from my Galaxy Nexus with Tapatalk 2
The Following 2 Users Say Thank You to Charmacas For This Useful Post: [ View ]
18th August 2013, 08:28 AM   |  #19  
Senior Member
Flag Green Bay, WI
Thanks Meter: 2,242
 
669 posts
Join Date:Joined: Dec 2010
Donate to Me
More
The Kindle Fire has source code released: http://www.amazon.com/gp/help/custom...deId=200203720

Inside the downloaded tar are multiple tars, the kernel is in omap.tar

There appears to be a few difference in the SR code, but it doesn't look like much. EDIT: This was through a simple compare of the 1p5 file, I'll check the other possibly relevant files.

EDIT2: Yeah I didn't notice much of a change in the SR stuff, however I'm not as familiar with this code as some other people here might be, so don't trust my judgement!
Last edited by MWisBest; 18th August 2013 at 08:59 AM.
The Following 2 Users Say Thank You to MWisBest For This Useful Post: [ View ]
4th November 2013, 10:06 PM   |  #20  
Maromi's Avatar
Senior Member
Thanks Meter: 197
 
293 posts
Join Date:Joined: Jul 2012
More
some users were able to run 4460 @1.5GHz with smartreflex enabled. luckily Hashcode had one too and did this http://forum.xda-developers.com/show...5&postcount=93

The Following User Says Thank You to Maromi For This Useful Post: [ View ]
Post Reply Subscribe to Thread
Previous Thread Next Thread
Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes


Top Threads in Galaxy Nexus Developer Discussion [Developers Only] by ThreadRank