"Yes, but how many polygons?" An artist blog entry with interesting numbers

[TapamN]

I guess it really depends on the model topology and how much of it can be shared between verts. Heres a snippet of the human readable model format kamui2 for dreamcast could export to( its less a model format and seems more like c code imo that sets up everything from lighting, materials and so on). This is a teapot which has 25,600 triangles that I exported to the dreamcast format, as you can see it has half the amount of vertices compared to triangles. I guess this is what the emu author meant.

An emulator won't report fewer vertices than ~1 per per polygon. The hardware only supports triangle strips, which asymptote to 1 vertex per tri. Unless the emulator scans every vertex the CPU feeds the PVR, finds the matching vertices, then recounts only the unique vertices, I cannot think of a situation where the emulator will report fewer vertices than triangles.

If you try to capture what an emulator draws, you might get more polygons than emulator-reported vertices because the emulator has to redraw stuff to support certain effects correctly, like transparency sorting and modifier volumes, but that's not relevant to real hardware performance.

 

[Cloofoofoo]

An emulator won't report fewer vertices than ~1 per per polygon. The hardware only supports triangle strips, which asymptote to 1 vertex per tri. Unless the emulator scans every vertex the CPU feeds the PVR, finds the matching vertices, then recounts only the unique vertices, I cannot think of a situation where the emulator will report fewer vertices than triangles.

If you try to capture what an emulator draws, you might get more polygons than emulator-reported vertices because the emulator has to redraw stuff to support certain effects correctly, like transparency sorting and modifier volumes, but that's not relevant to real hardware performance.

Always good to hear from the expert. Ill put his quote on the bottom, its ps note where he says may render more polygons than verts. The teapot code snippet was just a teapot model converted to the pvr .h for use with kamui2. I noticed the text based formats like .h and nja list the vert / triangles count and the vertices count tends to much much lower down to half at best. Reason for that?

DoA2 25000 - 35000 (@60fps), most (if not only) really optimized DC game I'd say.
F355 7000 - 16000, usually about 12000 (@60fps)
Shenmue II during game about 10000 - 15000, in some scenes up to 25000 (@30fps)
Le Mans during track 8000 - 9000, ~25000 at start then all opponent cars and tribunes visible (@30fps)

speaking in general - DC's GPU may render way more polygons, and it may look way better if vertex colors had applied nice lighting, but DC CPU was too weak to transform significant amounts of geometry and calculate good lighting effects.

PS: note - numbers is vertex counts, it is some less than actual triangle polygons count.

 

[TapamN]

Always good to hear from the expert. Ill put his quote on the bottom, its ps note where he says may render more polygons than verts. The teapot code snippet was just a teapot model converted to the pvr .h for use with kamui2. I noticed the text based formats like .h and nja list the vert / triangles count and the vertices count tends to much much lower down to half at best. Reason for that?

It sounds like he means that the PVR is capable of drawing more polygons than the games are pushing (i.e. they're CPU bound, not GPU bound), not that they're drawing more tris than vertices.

The verts per tri stuff is just how the math works out.

With triangle strips, you start with one triangle requiring three vertices (3v/1t = 3 v/t), then adding another triangle to the strip only requires one vertex (4v/2t = 2 v/t), another vertex adds another tri (5v/3t = 1.6... v/t). It gets closer and closer to one vertex per triangle as you increase the strip length.

When you have more than one strip, you can share more vertices. Imagine you have several strips and you put them next to each other. Not connecting the ends of the strips into one long strip, but like this:

|\|\|\|\|
|\|\|\|\|
|\|\|\|\|

The vertices in that make up the middle horizontal strip are all shared with the strips above and below, so you can get that strip for "free" and don't need to store or calculate those vertices.

A 1x1 grid of quads has a 2x2 grid of vertices (4v/2t = 2 v/t). Going up to a 2x2 grid of quads needs a 3x3 grid of vertices (9v/8t = 1.125 v/t). A 3x3 grid of quads is 16v / 18t = 0.8... v/t. A 4x4 grid is 25v / 32t = 0.78125 v/t.

For a triangle strip with X triangles, you can calculate the vertices per tri with the formula (X + 2) / X). For a grid of quads with dimensions X quads by Y quads, the formula for verts per tri is ((X+1)*(Y+1) / (X*Y*2)).

 

[Cloofoofoo]

It sounds like he means that the PVR is capable of drawing more polygons than the games are pushing (i.e. they're CPU bound, not GPU bound), not that they're drawing more tris than vertices.

The verts per tri stuff is just how the math works out.

With triangle strips, you start with one triangle requiring three vertices (3v/1t = 3 v/t), then adding another triangle to the strip only requires one vertex (4v/2t = 2 v/t), another vertex adds another tri (5v/3t = 1.6... v/t). It gets closer and closer to one vertex per triangle as you increase the strip length.

When you have more than one strip, you can share more vertices. Imagine you have several strips and you put them next to each other. Not connecting the ends of the strips into one long strip, but like this:

|\|\|\|\|
|\|\|\|\|
|\|\|\|\|

The vertices in that make up the middle horizontal strip are all shared with the strips above and below, so you can get that strip for "free" and don't need to store or calculate those vertices.

A 1x1 grid of quads has a 2x2 grid of vertices (4v/2t = 2 v/t). Going up to a 2x2 grid of quads needs a 3x3 grid of vertices (9v/8t = 1.125 v/t). A 3x3 grid of quads is 16v / 18t = 0.8... v/t. A 4x4 grid is 25v / 32t = 0.78125 v/t.

For a triangle strip with X triangles, you can calculate the vertices per tri with the formula (X + 2) / X). For a grid of quads with dimensions X quads by Y quads, the formula for verts per tri is ((X+1)*(Y+1) / (X*Y*2)).

ok I get the concept of the strips, so in case of my stripfied(?)teapot model with 25k triangles and 12k vertices, running single or multiple of these on actual hardware I would still end up with less vertices than triangles then for the entire scene?

 

[TapamN]

ok I get the concept of the strips, so in case of my stripfied(?)teapot model with 25k triangles and 12k vertices, running single or multiple of these on actual hardware I would still end up with less vertices than triangles then for the entire scene?

The model format does not store duplicate vertices shared by multiple triangles, so it only needs to store about one vertex per two triangles. The format the GPU accepts requires some duplicate vertices to be sent, so its best case around one vertex per triangle.

 

[Cloofoofoo]

Yume Baken 99 - Dreamcast
Seems to support up to 18 horses but I am not if you have to download some kind of race file or something, otherwise only seems to be only 1 default race.

Horse - 665 triangles

Jockey - 270 triangles

Track - 6,139 triangles

 

[xaeroxcore]

DoA2 25000 - 35000 (@60fps), most (if not only) really optimized DC game I'd say.
F355 7000 - 16000, usually about 12000 (@60fps)
Shenmue II during game about 10000 - 15000, in some scenes up to 25000 (@30fps)
Le Mans during track 8000 - 9000, ~25000 at start then all opponent cars and tribunes visible (@30fps)

TapamN, do you know numbers of PSP Games...If they are similar, may be we can conclude that at least geometry wise, PSP games are on dreamcast reach.​

 

[Cloofoofoo]

DoA2 25000 - 35000 (@60fps), most (if not only) really optimized DC game I'd say.
F355 7000 - 16000, usually about 12000 (@60fps)
Shenmue II during game about 10000 - 15000, in some scenes up to 25000 (@30fps)
Le Mans during track 8000 - 9000, ~25000 at start then all opponent cars and tribunes visible (@30fps)

TapamN, do you know numbers of PSP Games...If they are similar, may be we can conclude that at least geometry wise, PSP games are on dreamcast reach.​

I dont even seethe point, but theres extractors for multiple games if you want to do it yourself. Or you could even use jpcsp since it has a scene dumper of the geometry into obj models. You will be surprised how low it gets, try tekken 5 and such.

But to be honest you probably wont find anything higher than god of war psp games with the first one hovering 40k to 50k per frame and the second one slightly higher both at 30 fps. So basically below 2 million polygons a second.

The model format does not store duplicate vertices shared by multiple triangles, so it only needs to store about one vertex per two triangles. The format the GPU accepts requires some duplicate vertices to be sent, so its best case around one vertex per triangle.

I see, thats strange i also know that the makaron emulator mentioned trigger heart doing 44k triangles to 33k vertices. Heres an image, that why i was certain metallic actually meant more triangle than verts since he wouldn't be the first one.
W = verts tr = triangles.

 

[xaeroxcore]

I dont even seethe point, but theres extractors for multiple games if you want to do it yourself. Or you could even use jpcsp since it has a scene dumper of the geometry into obj models. You will be surprised how low it gets, try tekken 5 and such.

But to be honest you probably wont find anything higher than god of war psp games with the first one hovering 40k to 50k per frame and the second one slightly higher both at 30 fps. So basically below 2 million polygons a second.



I see, thats strange i also know that the makaron emulator mentioned trigger heart doing 44k triangles to 33k vertices. Heres an image, that why i was certain metallic actually meant more triangle than verts since he wouldn't be the first one.
W = verts tr = triangles.

It could mean God of War or Tekken 5 PSP could be on DC reach?

 

[milk]

But to be honest you probably wont find anything higher than god of war psp games with the first one hovering 40k to 50k per frame and the second one slightly higher both at 30 fps. So basically below 2 million polygons a second.

Since you mentioned that, do you have any idea on ho Daxter fares? Its from the same devs, and it does show they were trying their darnest to look as geometrically dense as possible to the jak games, which were among the poly kings of ps2. You can also see how they cut many corners with art and level design to aproach the look while staying performant, but the end result was still very pleasant for psp.

 

[Cloofoofoo]

Since you mentioned that, do you have any idea on ho Daxter fares? Its from the same devs, and it does show they were trying their darnest to look as geometrically dense as possible to the jak games, which were among the poly kings of ps2. You can also see how they cut many corners with art and level design to aproach the look while staying performant, but the end result was still very pleasant for psp.

Looking into it, its super dense .Close to 50k - 70k triangles per frame @ 30fps which should be like 2.1 million polygons a second peak. Not bad seems to be inline with ghost of sparta. Heres a few shots

It could mean God of War or Tekken 5 PSP could be on DC reach?

Tekken 5 seems to be very low, probably inline with virtua fighter 3 at the average. Definitely doable on dc. God of war is doable but it would look different. Dreamcast can definitely handle the geometry for chain of olympus( ghost of sparta would probably be stretching it and require some changes) but you probably have to dump the bloom and the shadow maps and probably the multiple point lights. At least the screen resolution would be much higher and probably better textures.

 

[Nesh]

It could mean God of War or Tekken 5 PSP could be on DC reach?

Surely you could. I mean.....these games got their compromises to be able to run on a PSP.
Every PSP and PS2 game could have a DC version theoretically with the right changes. We could get higher resolution textures, AA and image resolution, but we might have seen compromises in polygons, lighting and other effects.
I am sure the DC had untapped potential or other strengths where the PS2 could not match.
Back then as a fanboy I liked how the DC died and Sega games found their way on my PS2, but thinking it now in a more mature perspective, I wish the DC lived more to see more of its potential and games ported.
It is interesting that the DC felt overpowered and underpowered at the same time.

 

[Cloofoofoo]

Virtua fighter 3 tb - Dreamcast
( had akira model and jacky stage finished but crashed both and didnt save. It was similar count to these)

Wolf costume 2(in pieces but all there minus the shadow volumes) - 4,024 triangles

stage 09 - 6,887 triangles

 

[milk]

Looking into it, its super dense .Close to 50k - 70k triangles per frame @ 30fps which should be like 2.1 million polygons a second peak. Not bad seems to be inline with ghost of sparta. Heres a few shots

I appreciate your effort, but from the screenshot this looks like the critically infamous last game in the series called Jak & Daxter and the Lost Frontier, initially developed by ND themselves, but then given to High Impact games.

I was talking about "Daxter - 2006" an early PSP game by Ready at Dawn, it was a spin-off from the main jak series, and the studio then had just been created by ex-ND devs.

 

[Cloofoofoo]

²I appreciate your effort, but from the screenshot this looks like the critically infamous last game in the series called Jak & Daxter and the Lost Frontier, initially developed by ND themselves, but then given to High Impact games.

I was talking about "Daxter - 2006" an early PSP game by Ready at Dawn, it was a spin-off from the main jak series, and the studio then had just been created by ex-ND

I wouldnt expect much. This infamous last game is probably as dense as it comes. Especially judging from this wireframe i found of the 2006 one. 1300 triangles for the main. Kinda like raziel from dc port of sr.

 

[xaeroxcore]

Looking into it, its super dense .Close to 50k - 70k triangles per frame @ 30fps which should be like 2.1 million polygons a second peak. Not bad seems to be inline with ghost of sparta. Heres a few shots

Tekken 5 seems to be very low, probably inline with virtua fighter 3 at the average. Definitely doable on dc. God of war is doable but it would look different. Dreamcast can definitely handle the geometry for chain of olympus( ghost of sparta would probably be stretching it and require some changes) but you probably have to dump the bloom and the shadow maps and probably the multiple point lights. At least the screen resolution would be much higher and probably better textures.

I´d dare to say DOA2 looks superior. I really hope source CODE of this Tekken 5 game sometime leaks to see if someone can make the trick and bring it to DC!

 

[Nesh]

DOA2 does look superior. It holds up very well

 

[Cloofoofoo]

DOA2 does look superior. It holds up very well

Doa2 seems to be beyond superior, i did extract some models using naomi lib extractor( not emulation) and the character models range up to 9k triangles and the stages( the biggest ive seen) up to 11k. Basically one character pushes as almost much as the entire tekken 5 psp scene surprisingly. Comparatively virtua fighter 4 ps2 character model i seen is like 7600 triangles. Doa 2 really isnt that bad at all. Probably dc at its best.

 

[Nesh]

Doa2 seems to be beyond superior, i did extract some models using naomi lib extractor( not emulation) and the character models range up to 9k triangles and the stages( the biggest ive seen) up to 11k. Basically one character pushes as almost much as the entire tekken 5 psp scene surprisingly. Comparatively virtua fighter 4 ps2 character model i seen is like 7600 triangles. Doa 2 really isnt that bad at all. Probably dc at its best.

Which makes me wonder, how many polygons the same Tekken stage is on PS2. I am trying to remember how many polygons TTT and Soul Calibur 2 or 3 push btw. The character detail in TTT especially was insanely high. The stages also were amazingly realistic and detailed but since they are non interactive, it is one area I wonder what's going on in terms of polygons. After that game, character geometry in Tekken games looked like they were less and less.

DOA2 was just crazy. It is one of those launch games were devs overhyped themselves about next gen and pushed to the limit to reach the theoretical results of the next gen hardware.
We tend to praise Soul Calibur a lot on the DC, but DOA2 was beyond anything technically. The stages in SC were barren of geometry in comparison. Whereas in DOA2 stages were much more complex, interactive and exhibited a larger variety of materials. I think the most impressive stage especially was that with the Opera. I dont think there was any other fighting game series during the generation that had such large interactive environments.

DOA2 also benefited a lot from DC's capabilities, looked cleaner, more defined and properly saturated on DC compared to the PS2 version. Geometry wise appear exactly the same:

Here is Soul Calibur for reference:

Tekken 5 PSP:

Tekken 5 PS2:

Tekken Tag Tournament PS2:

Characters in TTT feel so much bigger most likely due to the geometry

 

[Cloofoofoo]

Which makes me wonder, how many polygons the same Tekken stage is on PS2. I am trying to remember how many polygons TTT and Soul Calibur 2 or 3 push btw. The character detail in TTT especially was insanely high. The stages also were amazingly realistic and detailed but since they are non interactive, it is one area I wonder what's going on in terms of polygons. After that game, character geometry in Tekken games looked like they were less and less.

DOA2 was just crazy. It is one of those launch games were devs overhyped themselves about next gen and pushed to the limit to reach the theoretical results of the next gen hardware.
We tend to praise Soul Calibur a lot on the DC, but DOA2 was beyond anything technically. The stages in SC were barren of geometry in comparison. Whereas in DOA2 stages were much more complex, interactive and exhibited a larger variety of materials. I think the most impressive stage especially was that with the Opera. I dont think there was any other fighting game series during the generation that had such large interactive environments.

DOA2 also benefited a lot from DC's capabilities, looked cleaner, more defined and properly saturated on DC compared to the PS2 version. Geometry wise appear exactly the same:

Here is Soul Calibur for reference:

Tekken 5 PSP:

Tekken 5 PS2:

Tekken Tag Tournament PS2:

Characters in TTT feel so much bigger most likely due to the geometry

Namco were really good artist and it shows, look at naomi mazan and you will pretty much see all tell tale signs of namcos style. Characters in tekken tag tournament are actually 2,000 to 3,000 lower triangles compared to dead or alive 2. I cant speak for the stages of TTT but the characters are much lower. The polygon allocation is mostly on the face and less on the hands, probably what gives the illusion of more detail than doa 2 who were crazy enough to model their finger nails and other nonsense. TTT probably does look better due to all the environments mapsthey did on the characters to imitate different types of materials. They took a similar approach in Sc2 and expanded that alot in sc3.

On the subject of soul calibur 2 , I spoke to someone who has a book on japanese developers art that included interviews with dev teams of soul calibur 2. So basically each soul calibur 2 character is around 4,000 polygons and the stage is around 20,000 they said. So its characters are actually lower than doa2 and ttt, if you ever extracted silent hill 2 model its similar to the main character which around 4.5k triangles. Overall in the scene detail is a similar count to doa2 due to the fact the stage being 20k. And all these is much higher than psp tekken5 . Funny enough they considred soul calibur 2 for dc as lead platform but decided against and used ps2 instead, they seem to complain dc used yuv textures.

jin from ttt 6,148 triangles: