5fps in duke3d on a 386 is awesome, i got 1.5fps in doom on 486dx-25.

That's awesome! So awesome that I posted it over at Duke4.net forums because there was a discussion there about the lowest spec the game would run on a while back.

Oh, I keep a mirror of 3D Realms' FTP, so if the new owners (Bunch of poseur dickheads) ever take those pages down again, it's in my sig. Edit: It was meant to be anyway. It is now.

I would think the CON system would have a large overhead, makes me wonder how Shadow Warrior could perform or maybe just the build editor 3d view

leileilol wrote:

makes me wonder how Shadow Warrior could perform

I tried to recompile Shadow Warrior, but have same problem https://forums.3drealms.com/vb/showthread.php?t=16657. I don't registered at 3drealms forums and I can't download attached htfile.zip. If anyone registered, please download this file and attach to this thread...

^ here ya go

Very cool. I know that you can turn on framerate with "DNRATE", but is there a way to get an average framerate for a demo loop (like Quake), or do you just have to observe the dynamic framerate on the opening demo?

boxpressed wrote:

Very cool. I know that you can turn on framerate with "DNRATE", but is there a way to get an average framerate for a demo loop (like Quake), or do you just have to observe the dynamic framerate on the opening demo?

I start new game then type DNRATE and observe overage framerate in real game. If i turn off shadows and minimize game window i get ~15-18 fps, so i finished first Duke level on 386 -)

keropi wrote:

^ here ya go

Thank you. Now i compile Shadow Warrior game EXE for run on i386, but Shadow Warrior need more then 9Mb to start. My 386 configuration is 8Mb and i can't increase memory size (don't have 2Mb or 4Mb SIMMs). So, i tried disable memory check, game started but after few seconds crashed with stack or heap overflow messages. If anyone have 386 + 12Mb RAM might try attached EXE...

Tronix wrote:

Thank you. Now i compile Shadow Warrior game EXE for run on i386, but Shadow Warrior need more then 9Mb to start. My 386 configuration is 8Mb and i can't increase memory size (don't have 2Mb or 4Mb SIMMs). So, i tried disable memory check, game started but after few seconds crashed with stack or heap overflow messages. If anyone have 386 + 12Mb RAM might try attached EXE...

Try to use a swap on hard disk (not sure will it work on DOS32A but on DOS4GW it seems to run fine): http://ukrfaq.narod.ru/ru/game/games4mb.htm

Bad quality video recorded from my am386-dx40: http://www.youtube.com/watch?v=oEBtT7vfLNo .

Hey, that's actually quite playable! I'd expect 10fps or lower on a 386 machine.

On my 386DX-40 DNRATE shows 7 to 9 fps without sound (low detail, shadows off, 320x200)
Enabling sound SB 2.0, 8 Voices, 22 kHz and Music on GUS costs 1 fps.

So it's really quite playable.

Tronix wrote:

Hello, […]
Show full quote

Hello,

Today I would like to run Duke Nukem 3D on my Am386-DX40 with 128Kb cache and 8Mb RAM: game console start, display version, display "loading strings" and then I get DOS4GW error: "06h - invalid opcode". CS:EIP contain 0F C8 means BSWAP 486+ instruction. I downloaded Duke3d original source code (very hard to find, now I have lost link again), find all BSWAP instructions in sources and replace them to:

Copy code to clipboard

1  "rol ax, 8",\
2  "rol eax, 16",\
3  "rol ax, 8"\

No idea if these instructions are relevant for performance, but I was thinking if these could be optimized.
A rol ax, 8 could be replaced by xchg ah,al. Both should use 3 cycles on a 386, but the rol instruction uses 3 bytes while xchg reg,reg only uses 2 bytes (or even only 1 byte since it is xchg reg,accum). If the code is in 32 bit mode xchg also won't need a data size prefix byte which rol with a 16 bit register will need. I don't know if any stalls can happen on a 386, but IIRC those are a Pentium/PPro thing.

I wonder how this would perform on a blue lightning machine xD

idspispopd wrote:

No idea if these instructions are relevant for performance, but I was thinking if these could be optimized.

I can't really think of something in the innerloop of a software renderer that would benefit from bswap, so I wonder what they are using it for exactly.
Is it in the innerloop at all?
And if so, what is its purpose? It may well be that they only need a few of the bytes, instead of all of them, so perhaps you don't need to do a full bswap, but only two of the three instructions.

Edit: Browsed the code quickly, and I see it only in two places:
1) A CRC routine for the level data. We need all 4 bytes for this. However, it shouldn't be performance-critical.
2) There's a copybufreverse() function, which uses it for the mirror-effect, to mirror the screen x-wise.

So I think it only affects performance when there are mirror effects in view.
I wonder if there is a faster way to do copybufreverse() on 386.
You could do it with word loads/stores so you'd only need the xchg, not the rol eax, 16. But you'd need two extra mov instructions, so I doubt it's faster.
But you could use stosw/stosd instead of the mov/add which they do.
Perhaps you could also reverse the whole routine, and then use lodsd to read data forward and push eax to write it in reverse order.

All in all I don't think it's going to affect performance that much, because the mirror effect is not used everywhere.

Here is a capture how it runs on one of my 386: https://www.youtube.com/watch?v=Rt6xf43nnFg

^^ Why didnt you use the GUS for sound? Wouldn't you see some benefit from hardware mixing?

The GUS uses software mixing in Duke3D. This has additional overhead since it has to transfer the sound buffer each time. On the 386 it is slower than SB due to this additional overhead.

elianda wrote:

The GUS uses software mixing in Duke3D.

For sounds right? It should use GF1 for music playback, and the performance would be the same as AdLib, since OPL2/3 also does hardware 'mixing'.

Exactly thats why I chose SB for Sound and GUS for Music as seen in the video.