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

Headhunter is a no. Shenmue is do able just tedious. But most llikely the next 3 games are tony hawk, matt hoffman, spiderman since they use the same engine. Shenmue faaaar off.

Have you ever explained what your process for extracting those is? Do you feed the results to some on-line repository? Great work btw.
 
Have you ever explained what your process for extracting those is? Do you feed the results to some on-line repository? Great work btw.

There isnt a real process. Alot of dreamcast games use ninja chunk models, even those that at first appear to be unrelated custom format. Using existing converters is just trial and error.

Then theres some games that have dedicated converters like shenmue and some ps1 port from activision.
 
Yes i see but its not an example of gameplay character models or even the locales in the game. Not even ryu hhazukis passport model. Not sure why the objection to getting more info on shemue 1.

no no i have no objection at all, do it if you can, it'll be very appreciated ! My comment was just an informative one about some shenmue data we already had here. No offense.
 
Thanks to extractors available now its Shenmue 1 for the dreamcast turn. These are assets from the disc not emulator rips. The game seems to have multiple iterations of the same maps at different size/detail I guess as needed for cutscenes/gameplay. I didnt delve deeper because its just tedious. For some models textures are slightly warped or missing.Its very impressive for 1999/98.


Ryo Hazuki (passport model) - 14,227 tris
shenmue11.jpg


Nozomi gameplay model - 1,909 tris
shenmue12.jpg


Ryo Hazuki (gameplay model) - 1,793 tris
shenmue13.jpg


Hazuki household exterior - 24,030 tris
shenmue14.jpg


Shenhua (intro/passport model) - 9,104 tris
shenmue15.jpg


Dobuita segment - 17,265 tris
shenmue16.jpg


Dobuita + larger map segment(not sure how it would be used ingame) - 71,457 tris
shenmue17.jpg


Cat - 1,300 tris
shenmue18.jpg
 
Trigger Heart Exelica weirdness
I've been following the PS2 vs DC thread on Sega-16 and saw about it. (I have an account there, but the verification email went missing, so I've never been able to post)

I'm not really sure why the game runs like that. That's definitely a situation where culling would be worth it. Maybe since a shoot'em up isn't very taxing for the DC (mostly some simple 2D collision detection between the player and bullets) there were getting good enough performance that they didn't feel like bothering to optimize it. They might have already been getting slowdown from fillrate issues, that they never felt the need to optimize the polygon count? Does THE use Ninja? Maybe they didn't realize that they were rendering all the off screen stuff somehow because they over-estimated what engine was culling? They had a deadline and not enough time to cull correctly? Just laziness? It's really weird.

An update on fake polygon counts: I've been doing some benchmarking, and if you just flood the PVR with off-screen vertices, it looks like the best vertex count you can get is only around 7.9 million vertices per second, or 253 MB/s of vertex data (32 bytes per vertex).

I had trouble getting the benchmark to work correctly. Initially, I was getting worse results from the benchmark not drawing anything than I've gotten actually drawing stuff. (I was getting a max vert/sec of 5 million, when I've gotten a peak 6 million while doing T&L and putting stuff on screen.) Eventually I figured out that, for some reason, the GPU is slow the first frame it's initialized, and my benchmark was measuring that first frame. Rendering a dummy frame and then preforming the benchmark got much better results.

Edit: Oh, the Passport models are more than 10K. The internal strings in the Passport binary all refer to "10,000 polygon Ryo/Chao/Xiuying/etc" so seeing 50% higher numbers is suprising. Nozomi actually has two detail levels. Since that model has more polygons than Ryo, that's probably the high detail one.
 
Last edited:
I've been following the PS2 vs DC thread on Sega-16 and saw about it. (I have an account there, but the verification email went missing, so I've never been able to post)

I'm not really sure why the game runs like that. That's definitely a situation where culling would be worth it. Maybe since a shoot'em up isn't very taxing for the DC (mostly some simple 2D collision detection between the player and bullets) there were getting good enough performance that they didn't feel like bothering to optimize it. They might have already been getting slowdown from fillrate issues, that they never felt the need to optimize the polygon count? Does THE use Ninja? Maybe they didn't realize that they were rendering all the off screen stuff somehow because they over-estimated what engine was culling? They had a deadline and not enough time to cull correctly? Just laziness? It's really weird.

An update on fake polygon counts: I've been doing some benchmarking, and if you just flood the PVR with off-screen vertices, it looks like the best vertex count you can get is only around 7.9 million vertices per second, or 253 MB/s of vertex data (32 bytes per vertex).

I had trouble getting the benchmark to work correctly. Initially, I was getting worse results from the benchmark not drawing anything than I've gotten actually drawing stuff. (I was getting a max vert/sec of 5 million, when I've gotten a peak 6 million while doing T&L and putting stuff on screen.) Eventually I figured out that, for some reason, the GPU is slow the first frame it's initialized, and my benchmark was measuring that first frame. Rendering a dummy frame and then preforming the benchmark got much better results.

Edit: Oh, the Passport models are more than 10K. The internal strings in the Passport binary all refer to "10,000 polygon Ryo/Chao/Xiuying/etc" so seeing 50% higher numbers is suprising. Nozomi actually has two detail levels. Since that model has more polygons than Ryo, that's probably the high detail one.

Well , whats your sega-16 name. Maybe I can recommend you for activation or something. I had the same problem till I reached out for help from a user.

Yeah its crazy to think it runs like that but it isnt the only one, as you saw another example was radirgy. Trigger Heart does not appear to use ninja(though I didnt check thoroughly ). Radirgy does(as most milestone titles), why? Triggerheart is a curious case because the developer of it, warashi, had worked on the dc before on macross m3. Abysmal performance , probably no more than a few thousand triangles on screen(looked through the models) with a framerate below 30/20 fps.

"only" 7.9 million. 2million verts per second seems like a huge jump(well even if its not on screen like your other test.) I wonder how the performance is if you used a model which either long triangles that cut across the screen + very very tiny ones(kinda like a real game situation). Would you still be hitting anywhere 6 million verts/polygons per second? I imagine not right?

I figured that was the case but the way shenmue I just found it tedious to dig through the files. Yes I remember seeing that 10K somewhere years ago when inspecting the files. The ryu hazuki model seems unique , while I didnt extract the others I suspect they werent modeled past the shoulders or they were like shenhuas model. She has low definition hands, very low despite the rest of the model being high definition(i guess since it isnt visible in the passport or intro). Ryo passport model has the high polygon hands.
 
Thanks to extractors available now its Shenmue 1 for the dreamcast turn. These are assets from the disc not emulator rips. The game seems to have multiple iterations of the same maps at different size/detail I guess as needed for cutscenes/gameplay. I didnt delve deeper because its just tedious. For some models textures are slightly warped or missing.Its very impressive for 1999/98.


Ryo Hazuki (passport model) - 14,227 tris
shenmue11.jpg


Nozomi gameplay model - 1,909 tris
shenmue12.jpg


Ryo Hazuki (gameplay model) - 1,793 tris
shenmue13.jpg


Hazuki household exterior - 24,030 tris
shenmue14.jpg


Shenhua (intro/passport model) - 9,104 tris
shenmue15.jpg


Dobuita segment - 17,265 tris
shenmue16.jpg


Dobuita + larger map segment(not sure how it would be used ingame) - 71,457 tris
shenmue17.jpg


Cat - 1,300 tris
shenmue18.jpg
Now, THATS some good use of polygons!
Achieving that level of character quality with just 500 quads is an artistic and technical mastery.
BTW 14,227 tris is around as much as the character models in FF13 I think?
 
Last edited:
What is up with those quads for her dress using 4 tris each?
I suspect its because they wanted flexibility in the way the dress is animated. If the triangles were shaped with only one diagonal line, especially with such large quads, it would have potentially created some problems with how the cloth behaves. The "folding" would have been subject to those single quad diagonal breaks to form the triangles. Would they all be formed from left to right or right to left? Would they have some right to left and some left to right? Randomly? or should they do it by hand? If they do it by hand what determines how the quads should be best cut to triangles?
 
I suspect its because they wanted flexibility in the way the dress is animated. If the triangles were shaped with only one diagonal line, especially with such large quads, it would have potentially created some problems with how the cloth behaves. The "folding" would have been subject to those single quad diagonal breaks to form the triangles. Would they all be formed from left to right or right to left? Would they have some right to left and some left to right? Randomly? or should they do it by hand? If they do it by hand what determines how the quads should be best cut to triangles?

humm. Makes sense.
 
I may be wrong though. Thats just my guess. It looks like there is probably an overlapping of surfaces there too. @Cloofoofoo is it possible to send me the model?
 
Well , whats your sega-16 name. Maybe I can recommend you for activation or something. I had the same problem till I reached out for help from a user.
It's also TapamN there. The account still works for things like tracking viewed threads, I just can't post.

I wonder how the performance is if you used a model which either long triangles that cut across the screen + very very tiny ones(kinda like a real game situation). Would you still be hitting anywhere 6 million verts/polygons per second? I imagine not right?
The 6 million was drawing a bunch of tori whose models fit entirely in cache. I couldn't actually draw that much to the screen, because there's not enough video RAM to fit the display list buffer for that much at once, and the rasterization part couldn't keep up. I was getting something like 4.15 million per second actually drawn, 6 million potential.

I benchmarked full screen quads as well. Performance dropped to about 156,000 quads per second.
 
Last edited:
Headhunter is a no. Shenmue is do able just tedious. But most llikely the next 3 games are tony hawk, matt hoffman, spiderman since they use the same engine. Shenmue faaaar off.

amazing! Learning about specially Pro Skater could give us some clues about how possible it would have been the PS2 version of THPS3 on DC.

In other hand, what about Maken -X or Outrigger?
 
Thanks to extractors available now its Shenmue 1 for the dreamcast turn. These are assets from the disc not emulator rips. The game seems to have multiple iterations of the same maps at different size/detail I guess as needed for cutscenes/gameplay. I didnt delve deeper because its just tedious. For some models textures are slightly warped or missing.Its very impressive for 1999/98.


Ryo Hazuki (passport model) - 14,227 tris
shenmue11.jpg


Nozomi gameplay model - 1,909 tris
shenmue12.jpg


Ryo Hazuki (gameplay model) - 1,793 tris
shenmue13.jpg


Hazuki household exterior - 24,030 tris
shenmue14.jpg


Shenhua (intro/passport model) - 9,104 tris
shenmue15.jpg


Dobuita segment - 17,265 tris
shenmue16.jpg


Dobuita + larger map segment(not sure how it would be used ingame) - 71,457 tris
shenmue17.jpg


Cat - 1,300 tris
shenmue18.jpg


Haven’t watched this. Awesome! I cant believe ryo and nozomi models spot less than 2k triangles. They look more more detailed than that. AM2 were gods!
 
Back
Top