S615 lab tests?

[zxcvbad]

Hey,

I wonder if Lordus had a chance to test performance on snapdragon 615 running Drastic.
I've got my hands on this SoC and results are pretty weird. In most cases "bigger" A53s cores (those clocked @1.5GHz) using a range of 499-800Mhz, while lower clocked cores having a range of 900-1000Mhz.
I wonder if it's supposed to be like that, where DraStic mainly utilize GPU so this could potentially be a main bottleneck on S615 (A405)?
Some time ago I did tests on One M9, completely switching off A57s, and A53s were giving me pretty solid results. In case of S615 situation is terrible, sound/video lags everywhere and pretty identical even if not to use higher res.

P.S. I found that v100 Adreno 405 driver, that I have on this device having twice as higher overhead than a recent v136 that is used on One A9, so it looks like driver potentially might be a bottleneck.. Unless DraStic isn't rely too much on GPU..

[Exophase]

DraStic shouldn't be GPU limited unless you're using a very expensive shader.

The one place where there's a lot of GPU communication is in uploading the textures that contain the emulated DS screens. Especially in hires mode where these are 4x larger. We've seen before issues where the Qualcomm drivers have weird problems with the texture uploads, where it would for example insert long stalls (that may have been frame syncs) and this was only fixed by adding a thread sleep after texture uploads.

Does using 16bpp color change anything? That makes the texture uploads half as large.

You could also try disabling the thread sleep which is in Advanced -> Sleep in rendering.

[zxcvbad]

16bpp doesn't have any effect, hires either.
There's interesting thing appeared in Antutu 6.0 bench,
GPU is having a ridiculous scores, they're around Adreno 205. I guess it's something with v100 driver. I'll play with settings and and will post an update later

[Lordus]

You can also try if different audio latency settings make any difference, as well as the number of threads.
To change the thread count just create a text file named 'threads.cfg' and write the number of threads you want (1-4) into it, then copy it to /sdcard/DraStic/config.
You can check if the config file is working by looking at the debug info at the bottom of the advanced settings.

[zxcvbad]

Did some additional tests:

1. Can confirm A53s on S810 appeared to be doing great even without frameskip.
2. Disabling sleep in rendering thread slightly improve situation (just a bit)
3. Audio latency doesn't have effect on performance

I've accidentally noticed a huge performance difference on Xperia M4 Aqua comparing to any other Snapdragon 615 device.

Antutu 6.0 (both are S615/v100 driver)
Garden bench:
M4: 751
Vivo X5 Pro: 4346
Maroon bench:
M4: 1020
X5 Pro: 3944

GFXbench
Overhead2 bench: (drivers v100 vs v136 on A9)
M4: onscreen 5.6 offscreen 5.9
X5 Pro: onscreen 5.0 offscreen 5.4
One A9: onscreen 9.7 offscreen 9.9*

* One A9 has Snapdragon 617 but it's almost identical SoC with a revised modem,
CPU/GPU are the same and clocked the same with S615

Geekbench 3 (average scores)
M4 Aqua: single 595 multi 2235
Vivo X5 Pro: single 686 multi 3587
One A9: single 745 multi 3103

At this point it's really hard to make a conclusion, neither to name what's the main bottleneck we have here, performance results are so different it's actually weird.
It looks like it's not just a driver, since in overhead2 test both X5 Pro and M4 Aqua are having pretty same results (assuming v100 driver much slower than a newer v136 that A9 have) but in Antutu and Geekbench 3 X5 Pro doing drastically better than M4.

What do you think?

[Exophase]

I would guess either something is eating a bunch of CPU time or the CPU clocks aren't as high as they're supposed to be. You're seeing pretty consistently bad results from CPU benchmarks, including ones like GB which are pretty purely CPU based.

Good thing to start with is running something to measure CPU utilization, for example an app like this: https://play.google.com/store/apps/deta ... itor&hl=en You can check if something is taking a lot of CPU time just while you're otherwise idling.

[TkSilver]

Hey,

I wonder if Lordus had a chance to test performance on snapdragon 615 running Drastic.
I've got my hands on this SoC and results are pretty weird. In most cases "bigger" A53s cores (those clocked @1.5GHz) using a range of 499-800Mhz, while lower clocked cores having a range of 900-1000Mhz.
I wonder if it's supposed to be like that, where DraStic mainly utilize GPU so this could potentially be a main bottleneck on S615 (A405)?
Some time ago I did tests on One M9, completely switching off A57s, and A53s were giving me pretty solid results. In case of S615 situation is terrible, sound/video lags everywhere and pretty identical even if not to use higher res.

P.S. I found that v100 Adreno 405 driver, that I have on this device having twice as higher overhead than a recent v136 that is used on One A9, so it looks like driver potentially might be a bottleneck.. Unless DraStic isn't rely too much on GPU..

For the most part yes, your phone is suposed to be utilizing the "weaker" cores more and the more power hungry higher clocked cores less. Unlike a desktop scheduler android cpu scheduling almost universally prioritizes energy efficiency over performance (you know so they can keep giving us smaller and smaller batteries all in the name of design aesthetics)

Now how fast a processor core ramps up in speed, how long it stays at the new clock, how it clocks down, which processor takes over and when, and many many other things all will effect real wprld performance and may be different from manufacturer to manufacturer. If you have a rooted device and either know what you are doing or don't mind the chance of wrecking your device you can easily adjust your phone's governor to suite your personal tastes beter then usually comes stock.

Just keep in mind that the a53 architecture is designed to be a power efficient design and not a high performance design. A good example would be the U line of intel core processors designed for laptops battery life over pure performance of the desktop versions.

[Exophase]

Digging around more, I get the feeling that Xperia M4 Aqua just has problems (also there was another thread here with the performance issue in DraStic thttp://drastic-ds.com/viewtopic.php?f=7&t=2810) Here's some more on the device and it doesn't sound good:

http://forum.xda-developers.com/m4-aqua ... a-t3197676
https://talk.sonymobile.com/t5/Xperia-M ... -p/1024589 (post by cp9)
https://www.reddit.com/r/SonyXperia/com ... r_battery/ (post by maky168)

It's hard to find examples of DraStic running on other Qualcomm Cortex-A53 devices, but here's one on Snapdragon 410 (if the text is honest about the phone anyway):

https://www.youtube.com/watch?v=xhr-yjQCNoY

And even in the map area where the CPU and GPU load is decently high it still seems to run full speed with hires. And this is only a 4x1.2GHz Cortex-A53 device.

You say you get the same performance with hires rendering on or off, that seems like a big flag because the difference in CPU utilization between the two is usually huge. It could be that maybe it's being limited not by the GPU per se but by some other mechanism (such as vsync) that hard locks the frame rate at well below 60Hz and enforces this with heavy stalls. Are your tests with auto frame skip enabled? If not, turn it on and see how many frames it's skipping to keep up by looking at the frame percentage display (it's the value on the right). Maybe check what the LCD refresh rate is with an app like System Info Droid.

[zxcvbad]

OS monitor seems to show pretty standard numbers.
Idling utilize about 3-5% of CPU, what's odd it's that it mainly comes from "settings" process where you get thread: 400+, this is likely Xperia firmware specific since on Nexus 9 (6.0) this process is set to 0/thread 0.
At this point I'm not even sure this process could be a bottleneck, too many things are badly affected overall.
frequency?, S615 seems to be pretty standardized SoC, Just checked AIDA64, GB, CPUz they all reporting standard clocking range CPU@1.5GHz - GPU550MHz.

Another fancy thing I've noticed on E2303 cores 0-3 are clocked @1.5Ghz and cores 4-7@1.1Ghz, so "little" cores are clocked higher than "bigger", I know you can't really call 0-3 bigger cores since S615 is all A53s but on S808/810 cores 4-7 are always bigger ones and I suppose on Helio X10 too. I don't have other S615 devices so I can't really see if it's something to do with SoC specific or/possibly OEM specific.

OS monitor idle:
http://m.imgur.com/g4wUhI4

Cores 0-3 @1.5Ghz and 4-7 @1.1GHz: http://m.imgur.com/Icm3Nde

[Exophase]

Yeah, I think the CPUs are running at a much lower clock speed with DraStic because something else is limiting performance for that particular application. There's a theme of games running poorly on this device throughout comments online. Possibly with a lot of degradation after some software update.