I've been thinking about this for a while and I am kinda getting to the "code something and see if it bloody works" stage, but I want to have a chat about it with youguys first.

The problem I am trying to get my head around is the shear scale of things. I've written planetary terrain systems in the past, but they don't require that Buckingham Palace is recognisable and in the correct place.

Say I want to be able to have a theatre of operation that covers western Europe in the 1940's. Your are looking at about 1000 miles by a 1000 miles.

That's a big area. Using a 4K texture gives me a resolution of about 0.25 miles per pixel. Not very nice.

So I started thinking about using the texture pixel as the base height and adding in pseudo random noise for detail. In effect turning every pixel into a CLOD (well I'm leaning towards ROAM) patch.

Then I can have the resolution I want at the cost of a lot more processing, but at first glance it looks possible.

The next thing I thought about was visibility. If I was at ground level, then it's not too bad. Maybe an 8 by 8 set of patches would suffice.

However at 35,000 feet. The visible range is huge and I'm having to handle huge numbers of patches. Admittedly all of them would be reduced to a single quad, or at least most of them, but it's still a lot of patches.

So I started thinking about organising sets of patches into region patches, and groups of region patches into sector patches, and sector patches into country patches, and about that point I decided to go for a beer instead.

Has anyone had any expierience with a system like this?

Anyone got a better idea (Please have a better idea, pretty please with sugar on)

Create a quad tree for your terrain. Make each leaf contain a mesh fixed size such as 33x33 vertices (you can even use tesselation hardware for this). Subdive the quadtree nodes based on the distance from camera, for example : if camera distance to node bound is smaller than the bound size, subdivide (this may not be optimal algorithm).

WIth few hundred quad tree nodes you'll be able to simulate pretty huge terrains at variable resolution. Increasing the visible range gets cheaper and cheaper since bigger nodes cover bigger terrain areas. Of course you should try to keep the terrain triangles small enough so that the screen space error is small. This can be handled with a proper node subdivision algorithm.

In the best case you'll have a height texture or texture array which contains 33x33 blocks.With a 1K x 1K height texture, you can have around 900 nodes in the GPU memory in the same time. When the camera moves, you'll need to update the height texture from system ram, or hard disk, or use any random terrain algorithm you find suitable.

Just some ideas,

Cheers!

35,000 feet ... in the 1940's. Your are looking at about 1000 miles by a 1000 miles

35,000 feet? Spacecraft in the 40s? You might consider revising your parameters before you design everything. A P15 Mustang had a maximum altitude ~10,000 feet. So maybe more like 125x125 miles. You can limit that even further by adding haze (fog/clouds) as, at high altitudes there wouldn't be a real use for a pilot to see more than (maybe) 10 miles.

The Me-262 had a service ceiling of 37,565 ft, and the P51 Mustang had a service ceiling of 41,900 ft.... what are you talking about?

Something to think about, is that other flight sim games actually use a curved surface to represent the terrain, I'm spotty on the details but MS Flight did that.

Oh, this might help? http://www.microsoft.com/Products/Games/FSInsider/developers/Pages/GlobalTerrain.aspx

Thanks ferrous, that doc really has given me a few ideas.

I don't want the whole Earth. I want to be able to have theatre's of operation and expend more resources to make that much smaller area as high detail as I can.

Despite that , the techniques in the doc are very close to what I have bouncing around in my skull.

It's food for thought.

When I'm doing a whole planet I use six quadtrees arranged as a cube them deform them to a sphere in the vertex shader, I don't want to do that here.

I suggest using quad tree for geometry management and procedural virtual texturing for texturing.

Lots of good information is here:

http://dice.se/publications/terrain-in-battlefield-3-a-modern-complete-and-scalable-system/

My implementation results:

I tested this approach to terrain sizes of 1000 km x 1000 km (with 10m elevations sampling span) and texture sizes

up to 256M x 256M pixels (this gives you denisty about 200 pixels per meter).

If you want to play with my implementation in realtime, you can grab it here (there is readme file, which

should help you get it running):

https://drive.google.com/file/d/0B1nsNgFzDF54SGx3alN5WEZHWjQ/edit?usp=sharing

This is slightly older build which does not support async streaming, but still should represent the potential of used

techniques :)

Cheers

P.

That looks really nice, I'll grab a copy and have a play. Thanks.

I will have to think about handling buildings etc as well.