Update 5 - lower SDL performance

[Radov]

Hi!

I'm writing an article about AmigaOS 4.1 update 5 (how well it performs on a Sam440ep flex) and did some performance test. All the results will be published in a near time, unfortunately - SDL benchmark gave me 'weird' values that I decided to post here. I know, that this is an old benchmark, but even though: I believe it should act in a different manner…
Why I think that SDL-bench results are weird and should act differently? With Update 5 SDL-Bench results are:
- in some cases almost 50% (s)lower
- differ every time I launch SDL-Bench again (up to 10-20%). With Update 4 values were more stable.

These are my results I've got for on a Sam440ep flex 800 (1GB RAM + Radeon 9200L):
Update 4

Code: Select all

Resolution:  320x240 / 320x240 / 640x480 / 640x480
Slow points     3.41    150.94      0.64     38.65
Fast points   111.84     77.60     30.97     19.42
Rect fill    7474.45   97523.8   2005.89   46545.5
32x32 blits  23953.2    157538   23405.7    157538

Update 5 - first run

Code: Select all

Resolution:  320x240 / 320x240 / 640x480 / 640x480
Slow points     3.37     80.02      0.63     31.01
Fast points   108.89     76.92     30.61     19.35
Rect fill    7314.39   80313.7   1958.87     39767
32x32 blits  23272.2   97523.8   21222.8   99902.4

Update 5 – third run in a row

Code: Select all

Resolution:   320x240 / 320x240 / 640x480 / 640x480
Slow points      2.48     96.39      0.31     33.47
Fast points    101.55     77.27     25.58     19.35
Rect fill        7353   93090.7   1858.87     42226
32x32 blits     23540  117029.8     22021    120471

Could someone confirm this behavior? Is it expected on Update 5? Should it be addressed as a problem?

[Slayer]

I can run this when I get home from work

excuse my ignorance but I haven't had much to do with SDL

Where does one get this SDL-bench executable from? And does it matter if I use a 733MHz Flex? My 800MHz is not currently setup and I see you've used an 800MHz.

So in short I boot up the Flex, open a shell and type on SDL-bench and cut and paste the output?

[Radov]

You can find this benchmark on OS4Depot: link. The usage is very simple, You need only to type "bench >out.txt" to have results saved in file. (A little warning: those that suffer epilepsy should better turn the screen of when the sdlbench is processing;)

Some more details:
Don't know if it matters, SDL seems to open its own screen, but I try this test with WB composition turned "on / 32bit color depth" and "off / 16 bit".

[Slayer]

Pitch = 320
Hardware surfaces avail = 1
Window manager avail = 1
Blitter hardware = 1
Colorkey blit hardware = 0
Alpha blit hardware = 0
Software->Hardware accel = 0
Video memory = 0

320x240 320x240 640x480 640x480
software hardware software hardware
Slow points (frames/sec): 3.18471 156.863 0.653168 39.0244
Fast points (frames/sec): 103.018 77.7879 28.9004 19.4618
Rect fill (rects/sec): 7262.41 178087 1953.27 71859.6
32x32 blits (blits/sec): 23540.2 157538 23141.2 157538

[ChrisH]

If benchmarks are varying more than before, then my first guess would be that you now have something running in the background that you didn't before... or are you doing both tests on "clean" OS4 installations?

[Slayer]

hmm, also, if you warm reset between each activation do they become more stable?

Might be something that is created internally after each activation that causes some residue but not necessarily causes an actual drop in performance, the actual program itself is flawed? Just speculating.

Also I realised I didn't run the test more than once, when I get a chance tonight I'll give it a whirl.

[Radov]

@ ChrisH
Both tests were made in same environment (with the only difference of Update 5 being installed of course;). The only background task I can think of (that may interfere with SDL_bench) is USB driver (with all those notifications). This driver has problems with my 5-1 card reader connected to Sam’s inner USB ports (endless reset loop?). Unfortunately, connecting off the reader didn’t make any difference…

@Slayer
I like your results, they seem to be similar to mine from Update4. There’re still big differences at “hardware Rect fill”, but it can answered very simple: you have Radeon 9250 with fast 128bit RAM, whereas mine card is a slower, 64bit version.

---------------------

OK. It seems that today I solved part of this problem. Accidently. Nearly had a heart attack when a short circuit fried my PSU. Luckily the board is fine and with backup PSU - SDL_bench starts to behave fine….
For some, unknown for me reasons, it is very hard to place graphic cards properly in any of my Sam’s PCI slots. With card at slightly at wrong angle trashes start to appear all over the screen. I wanted to stabilize it a bit using a dry, wooden stick at the best angle to fit PCI slot, but it got me only a fried PSU after a power on….
With new PSU I’ve got stable, similar results a few times in a row. That are my current values:

Code: Select all

                          320x240  320x240  640x480  640x480
                          software hardware software hardware
Slow points (frames/sec):  2.51099  145.455 0.319706  38.4615 
Fast points (frames/sec):  101.668  77.6699  25.8012  19.4175 
   Rect fill (rects/sec):  7301.25   120471  1884.08  46545.5 
 32x32 blits (blits/sec):    23814   146286  23405.7   157538

As you can see – those values are finally close to those from Update4.
So… the case seems to be solved by introducing new Power Supply. It amazes me why this problem has occurred just after Update 5 installation. Probably bad luck. This fried PSU was a new (bought 6 months ago) Be Quiet 300W model and the “backup one” – an old 450W Corsair. Maybe the best solution will be to get a new graphics card…. before something else will fry...