AMOS under pistorm

[huggy70]

Hi,


i wanted to do some speed benchmarks with good old Amos under Pistorm/Pistorm32lite, so i wrote some simple code for prime number generation with Eratosthenes sieve and was a little surprised of the outcome, here are the results:


6543 Prime numbers calculation

A1200 EMU68 850Mips (Sysinfo)
340ms
compiled
80ms

A500 Musashi 20,5Mips
2300ms
compiled
560ms

A500 stock 0,73 Mips
52960ms


Any idea why Amos under EMU68 is only ~7 times faster than Musashi?


I did also a Test under winuae showing 350Mips but the compiled code needs only 40ms

[derSammler]

Because Mips in SysInfo is just crap. It does not represent any real value that shows actual speed. Best example is 040 vs. 060: a 50 MHz 060 has 80 Mips, but SysInfo will tell you it has only half as much - making it show as if a 040 at the same clock speed would not be any slower.

Also, Emu68 is JiT. Speed completely depends on the executed code. The more complex the code is, the slower it is, as there is not much translated code that can be re-used. Or in other words: Emu68 does not perform at 850 Mips only because SysInfo shows that number. It can in a very tight loop using few instructions. But that's not what most code is.

[huggy70]

yes it seems these sysinfo numbers are strange.

In addition with Sysinfo i only get 460 DMIPS and under winuae jit enabled (showing 350MIPS) i get nearly 2000DMIPS.

So as for the prime number code it behaves like 140MIPS then although this is a really simple code in a tight loop but who knows what AMOS/Compiler makes of it.

[VladR]

Originally Posted by derSammler

Because Mips in SysInfo is just crap. It does not represent any real value that shows actual speed.

Is there any comprehensive 3D benchmark somewhere which shows all Amiga CPUs with their respective SysInfo MIPS and the game's performance ?

These days I only have time for some quick tweaks, not a full game dev (for now), but I'm tweaking my RTG flatshader that has a benchmarking component (which I only ran on WinUAE, Vampire V2 and V4 so far).


I'm hoping we can crossreference the SysInfo MIPS with a real-world 3D engine performance. Hopefully across as many Amiga configs as possible.


Then we'll see how meaningful the MIPS figure from SysInfo really is.

The actual non-sysinfo MIPS is a meaningful figure for 3D engine, as it averages all ops/instructions over a second, hence there's no potential for figure being skewed due to short loop or something, as the whole 3D pipeline has a lot of different stages.

I'm planning on having checkboxes for all separate stages of the pipeline, this way you can just benchmark Transform stage or scanline traversal, or scanline fill (or framebuffer update on/off - e.g. just a render into RAM).