CarSim: Did you implent replays?

Did any one of you already implent replays in your car sim? Currently I'm asking myself which variables to save in a replay file to keep it small and fast but of course near to the original. I thought of saving the game inputs and every 1/10 sec or so the positions & velocities of the cars. What do you think about a replay implentation? Ciao! Roppi

if your sim is completely deterministic (ie the same input will always produce the same simulation) then saving out the user inputs is definitely the best way since the data is quite small. However ensuring that your sim is deterministic can be quite difficult, (random number generators need the same seed, all simulation data needs to be tied to the simulation rate rather than the render speed, you'll need a fixed simulation rate rather than variable). In my sim I save out the inputs and random number seed, mainly as a debug tool but I will also extend it to replays some day. I always get an exact replay, but this is only beacuse I had this approach in mind from the beginning. If you try to fit this into your sim retrospectively, this can be a complete b*****d (I know from experience). If it doesn't play back exactly, it's very difficult to debug when the replayed version first diverges from the original version. Another approach is similar to one we used on a football sim I worked on a couple of years ago. We saved out all "interesting" data (ie player positions position, player animation states, ball position) every frame, and that was used for the replay. The problem is that the data can get quite large (we only had sufficient memory for the previous 10 secs of the game). Of the 2 approaches, I prefer the first, since it can also be used as a debug tool and can also be extended to a networked multiplayer game (since the system is deterministic, the same inputs on the different machines will produce the same game state - you only need to pass around user inputs). But like I said, it can be difficult to implement at first. Mike "MothaBrain" <[email hidden]> wrote in message news:[email hidden]...

Quoted text (10 lines)

> Did any one of you already implent replays in your car sim? > Currently I'm asking myself which variables to save in a replay file > to keep it small and fast but of course near to the original. > > I thought of saving the game inputs and every 1/10 sec or so the > positions & velocities of the cars. What do you think about a replay > implentation? > > Ciao! > Roppi

On 28 Nov 2001 06:09:32 -0800, [email hidden] (MothaBrain) wrote:

Quoted text (7 lines)

>Did any one of you already implent replays in your car sim? >Currently I'm asking myself which variables to save in a replay file >to keep it small and fast but of course near to the original. > >I thought of saving the game inputs and every 1/10 sec or so the >positions & velocities of the cars. What do you think about a replay >implentation?

Quite right. Sort of MPEG. I only store states though (at 10Hz), but it gets big pretty soon. That does enable me to do reverse playing, and superslomo's. (although probably such an MPEG-type replay may be able to as well) GT3 uses only player inputs for example. GPL does states only I think (big replays). The pro of saving game inputs is that it's the same code base as what you would like for network communication. Ruud van Gaal Free car sim : http://www.marketgraph.nl/gallery/racer/ Pencil art : http://www.marketgraph.nl/gallery/

"Mike Stanley" <[email hidden]> wrote in message news:<3G6N7.1809$[email hidden]>... > if your sim is completely deterministic (ie the same input will always > produce the same simulation) then saving out the user inputs is definitely > the best way since the data is quite small. (lots snipped) But this doesn't (in my system anyway) allow you to rewind, does it? I hate replay systems that don't allow you to rewind, what's the point of having a replay if you can't watch that spectacular wipeout (or passing move, if you're into that kind of thing <g>) over and over from different camera angles? I haven't figured out a solution to this yet, though I've been busy on other things. My initial idea was to save a complete state once every n seconds, and that way I could just run forward from the closest state, rather than from the beginning of the recording. Size/speed tradeoff. Jason 379

On 29 Nov 2001 06:24:30 -0800, [email hidden] (J. Perkins) wrote:

Quoted text (7 lines)

>"Mike Stanley" <[email hidden]> wrote in message news:<3G6N7.1809$[email hidden]>... >> if your sim is completely deterministic (ie the same input will always >> produce the same simulation) then saving out the user inputs is definitely >> the best way since the data is quite small. (lots snipped) > >But this doesn't (in my system anyway) allow you to rewind, does >it?

Indeed. And be careful when trying to NOT save intermediate states. Different FPU's can (and thus WILL) give different results; so saving a replay on one machine and playing it back on another may screw up (on a Nurburgring-sized lap or race). > I hate replay systems that don't allow you to rewind, what's >the point of having a replay if you can't watch that spectacular >wipeout (or passing move, if you're into that kind of thing <g>) >over and over from different camera angles? Right. A big flaw in GT3 (just as the lack of damage is). > I haven't figured out >a solution to this yet, though I've been busy on other things. >My initial idea was to save a complete state once every n seconds, >and that way I could just run forward from the closest state, >rather than from the beginning of the recording. Size/speed >tradeoff. I think this idea should work. I'm still trying to work out how to reverse things like smoke and spark particles though, without saving too much (state) data. Perhaps in that case, doing it like you described is the *only* sensible (and relatively simple, read codeable) thing. But even then, the particle states (and things like laptimes etc) must be saved as well. But once every couple of replay frames shouldn't be that bad. Anyway, memory becomes bigger and bigger, so it will be less and less of an issue (except on consoles for the time being; that's why GT3 selected storing only the inputs probably). Ruud van Gaal Free car sim : http://www.marketgraph.nl/gallery/racer/ Pencil art : http://www.marketgraph.nl/gallery/

Hee hee, yes, trying to retrofit a "control-stream" replay onto a game that used a compressed-state-stream replay is a real pain We had to do that when we wrote Nascar Heat for the PS2, because a state-based replay is completely impractical for a console system, unless you only need 5 seconds of replay... it's a good demonstration of chaotic principles, as an error in the tenth decimal place will quickly propogate and the player car ends up against a wall with the tires spinning and the wheel waggling purposefully back and forth. When we wrote Heat for the Xbox, we were able to use a state-based replay using the hard-disk, which is nice because than you can shuttle back and forth and all that nice stuff. "Mike Stanley" <[email hidden]> wrote in message news:3G6N7.1809$[email hidden]... > if your sim is completely deterministic (ie the same input will always > produce the same simulation) then saving out the user inputs is definitely > the best way since the data is quite small. However ensuring that your sim > is deterministic can be quite difficult, (random number generators needthe

Quoted text (8 lines)

> same seed, all simulation data needs to be tied to the simulation rate > rather than the render speed, you'll need a fixed simulation rate rather > than variable). > > In my sim I save out the inputs and random number seed, mainly as a debug > tool but I will also extend it to replays some day. I always get an exact > replay, but this is only beacuse I had this approach in mind from the > beginning. If you try to fit this into your sim retrospectively, this can

be

Quoted text (7 lines)

> a complete b*****d (I know from experience). If it doesn't play back > exactly, it's very difficult to debug when the replayed version first > diverges from the original version. > > Another approach is similar to one we used on a football sim I worked on a > couple of years ago. We saved out all "interesting" data (ie player > positions position, player animation states, ball position) every frame,

and > that was used for the replay. The problem is that the data can get quite > large (we only had sufficient memory for the previous 10 secs of thegame). > > Of the 2 approaches, I prefer the first, since it can also be used as a > debug tool and can also be extended to a networked multiplayer game (since > the system is deterministic, the same inputs on the different machineswill > produce the same game state - you only need to pass around user inputs).But

Quoted text (18 lines)

> like I said, it can be difficult to implement at first. > > Mike > > "MothaBrain" <[email hidden]> wrote in message > news:[email hidden]... > > Did any one of you already implent replays in your car sim? > > Currently I'm asking myself which variables to save in a replay file > > to keep it small and fast but of course near to the original. > > > > I thought of saving the game inputs and every 1/10 sec or so the > > positions & velocities of the cars. What do you think about a replay > > implentation? > > > > Ciao! > > Roppi > >

Well, I think the replay will only be satisfiying with a compromise of state/input-information. I think I'll run my physics code in replays as usual as in the game, with collisions etc. and will correct the states and send inputs with 4Hz or so. I I suppose that with a running physics system the result shall be quite realistic. But another prob arises with temporary objects like sparks while collisions, I don't know how to handle things like that. (spawning/deleting) Maybe it works anyway with physics like I said because collisions are tested and that's why sparks should be spawned as usual. Well, think I've got to test this first to have an opinion :) Ciao! Roppi