2D benchmark

[StephC]

[Presto]

[StephC]

I've got better results with a 300Mhz Axim...

Hmm...

[StephC]

Presto : have you experienced some kind of short freezing of the display during the test ?

[Presto]

[Johan]

Unfortunately IMHO this test is of no use to anyone except for StephC. For example:

1. FlipScreen
What exactly is it you are measuring here? Do you check if the device is double buffered? If you simply do a fast memcopy to the GX backbuffer and then call GXEndDraw/GXBeginDraw your application will run slower on double buffered devices (where you only need to call GXEndDraw/GXBeginDraw for the flip and would need to measure that). I argue that separate time measures for (a) memory block copy to the GX backbuffer, and (b) calling a GXEndDraw/GXBeginDraw, would be needed. We also need to know if you use memcpy or a fast ldmia/stdmia block copy.

2. 512 ColorKeyed Sprites
How large are these sprites? How are they stored? RLE-encoded? How large is the bitmap image where they are stored? We would need the exact width and height of the source image if the images are not RLE-stored. If they are RLE-stored then this test measures lots of other stuff such as processor performance. Also, are you reading and writing multiple pixels simultaneously using 32-bit reads and writes? If so, how many of the 512 sprites are stored on aligned memory addresses?

Above I have only surfaced some of my questions I would need answers to relating two of these tests for the results to be of any use to me. Perhaps other people find the results more useful than I do.

[sponge]

I'm guessing it's more useful to consumers (in the sense of 3DMark) than a developer looking to optimize code.

[refractor]

"Lies, damn lies, and statistics" - Benjamin Disraeli

Ok. So far as a "consumer" benchmark goes, this one is as good, or as bad, as the standard consumer "gaming" benchmarks. It gives users numbers to compare... for certain, perhaps esoteric, tasks.

Essentially, I agree with Digby/Johan on this. I'm far more interested in "raw" crunching power/memory bandwidth than I am in "synthetic" numbers like how many RLE encoded sprites a certain routine can blit to a screen per second. It really doesn't matter if the routine is "optimised" or not - if that's an unknown quantity then so is everything else.

I'd prefer a more comprehensive "raw" set of benchmarks - cache miss penalties, the actual r/w bandwidth of cached and uncached areas, right through to the more "generic" dhrystone, etc, tests (which although synthetic aren't entirely esoteric).

But then I'm a developer.

[Kzinti]

Johan/Digby/Refractor: I understand that this benchmark is not very useful to you, but I don't think that telling Stephane his benchmark is useless is very constructive.

It does have its uses and I question why you gang on him... This benchmark may indeed be very useful to him and for a lot of other people who wants to know what device is the fastest. I sure do appreciate his work on this as no one else has proposed anything.

Any benchmark will have it's flaws (and they are ALL synthetics), there is simply no way around it.

[Kzinti]

Where is "benchmark.tga" being saved? I've been looking for 10 minutes now!

[refractor]

[Johan]

[Presto]

[StephC]

[Presto]