Gpmark


Exophase said:
Rotozoomer: 169.6 vs 162.5, Wiz + 4.19%

The image Optimus uses is 80x80, presumably 16bpp, which is actually still small enough to fit into dcache. What I don't know is this image is repeated in memory or not. I presume it is, which means that it won't fit in dcache and will therefore exercise memory latency on the reads. This will hold part of the time; where it's zoomed out the Wiz has the advantage, when it's zoomed in the Caanoo does.

It's 128 x 128 btw. It's not repeated more times in memory, it's tiled with an & 127 in the rotozoomer code.

Exophase said:
Rotozoomer far: 73.8 vs 59.3, Wiz + 19.65%

Never fitting in dcache, this will heavily exercise memory latency, especially if the image is not tiled. And we can see a big difference in the results. If whatever is giving the Caanoo the advantage the rest of the time is normalized out (sorry but I think there's something in terms of software configuration that is, either refresh rate or time spent in kernel routines) then Wiz will have an even bigger advantage.

Indeed, it seems that Wiz wins over Caanoo at cache intensive tests.


Exophase said:
Radial blur: 20.9 vs 21.8, Caanoo + 4.13%

This does appear to do memory to memory, but it's very slow, so it probably spends proportionately much more time working the ALU/cache and not latency sensitive operations, unless it's using huge LUTs which I imagine not. So Caanoo wins again.

I am actually using a lot of LUTs here. This should be extremely cache intensive. Why Wiz doesn't win in this case as it did it in the far rotozoomer, I don't know.


Exophase said:
3D Bunny: 5.3 vs 5.1, Wiz + 3.77%

This is non-textured, so it's going to be very ALU intensive. There will be some memory latency involved for fetching vertex data, which probably won't all fit in cache. This will only have a small impact.

Well, it is textured. I use an environment map. Anyway, most of the CPU time is lost to transform all those 3d points and render the massive number triangles so I don't think fetching from a texture in memory would steal much performance here.

In the next post I am going to release the source code. I think it was time to do so, since I don't think I will work on the next version any time soon.
 
Last edited by a moderator:
Thanks for your comments Optimus.

Optimus said:
It's 128 x 128 btw. It's not repeated more times in memory, it's tiled with an & 127 in the rotozoomer code.

Yeah that (32KB, confirmed bit depth) is enough to blow L1 when fully zoomed out.

Optimus said:
I am actually using a lot of LUTs here. This should be extremely cache intensive. Why Wiz doesn't win in this case as it did it in the far rotozoomer, I don't know.

While the LUTs themselves are large, the LUT coverage is limited to this per frame:

fsin1: 0 to 319
fsin2: k1 to (k1 + 239)
fsin3: k3 to (k3 + 239 + 319)

Because k1 is limited to 0 to 503 and k3 is limited to 0 to 880 this makes your ranges limited to:

fsin1: 0 to 319
fsin2: 0 to 742
fsin3: 0 to 1438

Or 2502 entries in total, which takes up 10008 bytes, fitting in L1 cache (arrays are consecutive so there's no concern of causing associativity collisions). Even if it didn't fit, you only grow the coverage by 1 entry in fsin2 and fsin3 respectively per frame, so most would stay resident in cache between frames and the impact to main memory is minimal.


Exophase said:
Well, it is textured. I use an environment map. Anyway, most of the CPU time is lost to transform all those 3d points and render the massive number triangles so I don't think fetching from a texture in memory would steal much performance here.

True, but your environment texture is only 64x64x16bpp (8KB) which fits in cache, and because you do T&L before rendering you have a lot of room for reuse on it. If you had z-buffering on for this test I suspect the Wiz would come out further ahead because that will definitely stress memory latency.

But yeah, it'll probably still by dominated by the math.

So, I was thinking, someone should add to the test a printout of the clock configuration being used. I want to see if Caanoo is using different memory bus dividers between firmware versions and possibly between it and the Wiz. Also want to see the refresh rate used. Caanoo doesn't have any hardware advantages I can think of so its advantages are almost definitely in configuration or systems software quality.
 
Last edited by a moderator:
aho said:
DaveC said:
[...]
One thing I noticed is the Caanoo has a blue gradient background where as the Wiz is black (on black or dark graphics). What it up with that? Is that some weird thing different in the emulator or something? It is quite different, not just a litle.

That's AMOLED (Wiz) vs TFT (Caanoo). AMOLED got a much higher contrast since there is no back light. Each pixel is its own light source. It also got a much higher viewing angle (about 170° in both directions, I'd say).

Well, AMOLEDs are more expensive, have a shorter lifetime, and a higher failure rate. The smallest imperfection can reduce its lifetime drastically. You can also get burn-in fairly quickly.

It's a pretty new technology. In a few years AMOLEDs will be a lot better. Right now they are somewhat meh... but at least the image is very pretty and amazingly crispy.

---

Looking at those numbers, the Caanoo looks pretty disappointing. Not really much of an upgrade. They should have skipped that one. Without GLES 2.0 and without an FPU it's not very interesting hardware. I don't need another Wiz.
So that is actually in the Caanoo display? Yikes even for an LCD that seems really bad. It truly looks like a blue gradient that is used for backgrounds. I have even seen LCDs with poor contrast but the lighting is at least more even. Oh well I don't have a Caanoo so it don't bother me.
 
Last edited by a moderator:
DaveC said:
[...]
So that is actually in the Caanoo display? Yikes even for an LCD that seems really bad. It truly looks like a blue gradient that is used for backgrounds. I have even seen LCDs with poor contrast but the lighting is at least more even. Oh well I don't have a Caanoo so it don't bother me.

I'd say it only looks this bad in comparison. I'd also say that it looks this blue-ish, because the cam auto-adjusted to a rather warm color temperature. So, you won't perceive it as overly blue-ish in real life (just like you won't perceive a TV as blue-ish, even if the sun-like light it tries to simulate is actually very blue).

The black, however, will be more of a dark gray, since there is still some of the back light shining through.
 
Last edited by a moderator:
GPmark GP2X port

My results

Blitting Test: 158.8
Plasma: 105.5
Rotozoomer: 65.5
Rotozoomer Near: 88.3
Rotozoomer Far: 35.7
Radial Blur: 17.9
3D Bunny: 2.3

Also, you might get better results because sometimes I got a little better and I can't figure what happened now. I am using mmuhack too, I had two versions of mmuhack.o, one from Mame and one from my Rerecycle demo. The rerecycle mmuhack.o was better, made blitting test 188.0 for example orplasma 117.0. The mame made what is there now in the results. I put the faster one in the archive, tested it and was ok. Then I uploaded in GP2X file archive, then I downloaded to test again if it was right and I got the slower results. What the heck? Go, figure, etc, etc.. I checked the mmuhack, it's the right one, I tried with both, it's the same results. Why did I got better, oh forget it.. anyway, I am fed up with these details. Enjoy this version and I might do one for Dingoo too.

p.s. My battery is about to die. Could the results be affected and reduced the very last times I tried it? Maybe. If someone tries it on his GP2X and gets better results (188.0, 117.7, 70.0, 96.6, 37.0, 18.2, 2.3) I will be happy. I might try to fill my bateries and try again later anyway. But anyway, this is just an interesting test to test how much better the Wiz/Caanoo is from GP2X. Strange, the numbers seem low, I remember a plasma went 192 in my very old GP32 but now I don't have a functioning GP32 to verify again. Then, my old releases do these numbers too. I guess I had a better feel of the performance of GP2X than what it's in reality.

p.p.s. Dingoo will be fun to see results. The radial blur makes 30.0fps in Creep Tea port on Dingoo. Winning over every gamepark for some reasons. But not the rest of the effects..
 
Motorola ZN5 port ;)

My results:
4.png


Specifications:
Some Random ARM11 532 MHz, 64 MB RAM + 128 SWAP
Code:
Processor       : Some Random V6 Processor rev 1 (v6l)
BogoMIPS        : 354.70
Features        : swp half thumb fastmult vfp edsp java
CPU implementer : 0x41
CPU architecture: 6TEJ
CPU variant     : 0x1
CPU part        : 0xb36
CPU revision    : 1
Cache type      : write-back
Cache clean     : cp15 c7 ops
Cache lockdown  : format C
Cache format    : Harvard
I size          : 16384
I assoc         : 4
I line length   : 32
I sets          : 128
D size          : 16384
D assoc         : 4
D line length   : 32
D sets          : 128

Hardware        : Motorola Product - SCM-A11 Phone
Revision        : 0021
Serial          : 0000000000000000

Code:
MemTotal:        55288 kB
MemFree:          2136 kB
Buffers:          2364 kB
Cached:          18884 kB
SwapCached:       2916 kB
Active:          34596 kB
Inactive:         5496 kB
HighTotal:           0 kB
HighFree:            0 kB
LowTotal:        55288 kB
LowFree:          2136 kB
SwapTotal:      110032 kB
SwapFree:        93728 kB
Dirty:               0 kB
Writeback:           0 kB
Mapped:          29444 kB
Slab:             5820 kB
CommitLimit:    137676 kB
Committed_AS:   163308 kB
PageTables:       2160 kB
VmallocTotal:   196608 kB
VmallocUsed:      3416 kB
VmallocChunk:  4112384 kB
 
Caanoo Firmware 1.6.0


blitting 298.2
plasma 216.6
rotozoomer 146.0
rotozomer near 215.9
rotozoomer far 56.7
radial blur 19.4
3D bunny 4.7


So it has the same performance drop as the 1.5.0 had.
 
Hmm,. the ZN5 zipfile also contains a Dingux port.
Has someone tested it? I am really curious about the results of Dingoo compared to the Gameparks. (My Dingoo has a broken LCD, so I can only watch stuff with TV-Out on my tv monitor and thus only native stuff)
 
Optimus said:
Hmm,. the ZN5 zipfile also contains a Dingux port.
Has someone tested it? I am really curious about the results of Dingoo compared to the Gameparks. (My Dingoo has a broken LCD, so I can only watch stuff with TV-Out on my tv monitor and thus only native stuff)

I tried to run it in Dingux, but it only throw me back to gmenu2x
 
Last edited by a moderator:
Sorry, the version for dingoo does not work.
On our new SDL library with support for IPU here are the results:
My Blog

IMG_0004.jpg


Code:
blitting test   235,5
plasma	        150,8
rotozoomer	177,3
rotozoomer near	193,0
rotozoomer far	137,9
radial blur	23.9
3d bunny	5.9
 
This is my result on GP2X F200 (firm Open2x DR7), battery full

My results

Blitting Test: 200.0
Plasma: 127.7
Rotozoomer: 82.6
Rotozoomer Near: 120.1
Rotozoomer Far: 39.0
Radial Blur: 17.6
3D Bunny: 2.5

I have to rebuild the test from the sources because the version on the archive don't work for me.
Compiled statically with codeblocks no change to version 0.4 of sources.
If someone need this let me know.

Thanks a lot for this.

@EXL very nice works/ports in your blog. :) I don't understand a word of russian but it's an impressive amount of ports for motorola devices.
 
An update with some Pandora test.
I have compiled using the same sources under Codeblocks (using VBox image by Sebt3...thank's a lot mate ;) ) with minor change to support exit with window controls.
Compiled with only this flags : -O2 -s

I made some tests :

Pandora 320*240 windowed
My results

Blitting Test: 783.3
Plasma: 427.5
Rotozoomer: 365.0
Rotozoomer Near: 446.9
Rotozoomer Far: 245.8
Radial Blur: 69.7
3D Bunny: 11.1

**********************
Pandora 640*480 windowed
My results

Blitting Test: 226.2
Plasma: 116.0
Rotozoomer: 93.2
Rotozoomer Near: 123.9
Rotozoomer Far: 64.8
Radial Blur: 15.6
3D Bunny: 10.5

***********************

and a surprise !!!
GP2X version under GINGE :)
My results

Blitting Test: 960.3
Plasma: 517.5
Rotozoomer: 401.1
Rotozoomer Near: 513.3
Rotozoomer Far: 248.5
Radial Blur: 66.4
3D Bunny: 9.9


Interesting is the result of GINGE GP2X version, in some test, is more faster than the Pandora itself :blink:

Forgot to say that i'm under HotFix 5 .
 
Hey, that's interesting. I was always curious to see how the Pandora performs in comparison to gamepark consoles but I don't have one so that I could port.

Maybe the native Pandora versions are worse, because as you write, they are running windowed (not fullscreen?).
Whatever, the results are impressive already.
 
Hi everyone! I've compiled GPMark test on multiple platforms and decided to check them by their performance!
The results of the test you can look at this link:
http://motodev.do.am/news/gpmark_multiplatform/2011-07-13-240
 
Sharp Zaurus C-1000, PXA270@520MHZ:

Blitting test: 441.2
Plasma: 243.7
Rotozoomer: 216.3
Rotozoomer near: 247.2
Rotozoomer far: 218.2
Radial blur: 35.8
3D bunny: 3.7
 
Back
Top