Speed differences between 5.17 and 4.02

I’ve always had the impression that RO 5 ran faster under RPCEmu than RO 4 (both 5.17 and 4.02 are installed on my PC), and this was confirmed yesterday when I ran Kuemmel’s !Firebench app: three runs of RPCEmu 0.8.8/5.17 gave an average of 14.3 secs, while three following runs of 0.8.9/4.02 gave 17.57 secs. The peak MIPS rate was higher under 5.17 too (over 1000), which I’ve also noticed before. Base PC system is a 3.4GHz Core i7 Dell desktop running W7 (64-bit). The observant will notice that 5.17 is running under an earlier version of RPCEmu; I don’t know whether this makes a difference and I don’t plan to upgrade as I use 5.17 but rarely. I presume 5.17’s advantage is down to being fully 32-bit, maybe architectural changes also play a part.

PS: the screen mode was identical (1680 × 1050 × 16M x 60Hz) in both cases. Both tests used the Recompiler version of RPCEmu.

I presume 5.17’s advantage is down to being fully 32-bit

You’re probably right. From an emulators point of view, it’s a lot easier to emulate a 32bit system than a 26bit system, due to the PC and PSR being seperate registers. Since the PC changes for every instruction executed, being able to reduce the amount of work needed to keep it up to date can have a significant impact on performance.

Just remember that in this case 5.17 is only faster because you’re running it under an emulator. If you tried the same thing on a real machine you’d probably find that the Firebench results are practically the same.

I recently ran Richard Spencer’s benchmarking app romark 1.01 on both my RPCEmu installations (0.8.8/5.17 and 0.8.9/4.02 respectively). Some of the differences – notably HD read/writes and FS read/writes – were quite startling. As I am completely unable to understand the Textile instructions for importing a chart, here are the results (0.8.8/5.17 first, % advantage of 0.8.8/5.17 last):

1. 572562; 592026; 97%
2. 9139; 9380; 97%
3. 3665; 3053; 120%
4. 11389; 6429; 177%
5. 4198; 3060; 137%
6. 4831; 3591; 135%
7. 99020; 21340; 464%
8. 74777; 25683; 291%
9. 3525; 2322; 152%
10. 3006; 1861; 162%
[Key:
1. = Processor-looped instructions (cache);
2. = Memory-Multiple register transfer;
3. = Rectangle Copy – Graphics acceleration test;
4. = Icon plotting – 16 colour sprite with mask;
5. = Draw path – Stroke narrow line;
6. = Draw fill – Plot filled shape;
7. = HD Read – Block load 1MB file;
8. = HD Write – Block save 1MB file;
9. = FS Read – Byte stream file in;
10. = FS Write – Byte stream file out.

The base system remains the same in both cases (3.4GHz Core i7 quad-core running W7 (64-bit)). The only other difference is RPCEmu’s memory allocation; 4.02 has 256MB, 5.17 has 128MB. Both versions display at 1680 × 1050 × 16M. The above would imply that RO 5.XX versions of RPCEmu have a substantial speed edge (on any given base system) over other emulated ROS versions, presumably including VRPC….

It would be nice if you could benchmark the two OS versions against the same underlying RPCEmu setup.

I probably will, as and when I upgrade my 0.8.8/5.17 setup to 0.8.9/5.19. In the meantime, you might be interested in these results (run on the same underlying RPCEmu version):

http://www.riscos.info/pipermail/rpcemu/2012-October/001740.html