Jump to content
  • Advertisement
Sign in to follow this  
3TATUK2

Deferred Shadow Mapping?

This topic is 2141 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

So I've implemented multiple point-lights with shadowmapping using forward rendering..

 

The performance is horrible though. I *do* use a cubemap to reduce texture() fetches, but even then I can only get about 6 lights before I start running around 100-150 FPS (down from 200-300).. Even if the lights aren't on screen, like if I'm in a corner staring at a wall.

 

I also know that when doing multiple *lights* like this, you basically need to do deferred rendering for decent performance, although I don't know much specific details about deferred rendering.

 

Now, what I'm really curious about is specifically whether the *shadow-mapped* part of the point lights will be faster with deferred rather than forward?

 

I mean, from my understanding - the slowdown is mainly due to all the texture() fetches for each shadow-mapped light.... But - will deferred rendering still basically do the same thing?

 

If not, and it *is* considerably faster, what are some of the details?

 

For example, I've seen it's possible to do basically hundreds of lights with deferred rendering... But how many *shadow-mapped* lights can you get away with?

 

If it *is* worth it and deferred shadow-mapped lighting is *fast* - then, then please provide any decent links you may have to guides about both simple deferred rendering and shadow-mapped deferred lighting.

 

Thanks

Share this post


Link to post
Share on other sites
Advertisement

I'm in the process of doing similar (but with a LOT more lights than you have) and here are some observations from my work so far.

 

Frustum cull the lights!  Each light is bounded by a sphere, so you get to do a nice fast sphere/frustum intersection test, and if the light volume is outside of the view volume, then don't draw it.

 

If you're drawing fullscreen quads for each light, then stop it; this isn't specific to shadow mapping but applies to deferred in general.  With a lot of lights you end up doing a lot of fullscreen passes, where only a relatively small portion of the screen may actually need to be updated for each pass.  Define a light volume (i.e. some 3D geometric shape; there's a tradeoff between number of verts and fillrate overhead here, I use icosahedrons but it may be easier to use boxes) and draw it, transforming by your standard MVP.

 

Rather than having a single shadow map texture that's continually updated, I use a separate shadow map per-light, with shadow maps moving to and from a free pool as they go in and out of the view.  That's not a performance consideration, but it is important for the next bit...

 

The fastest work is the work you don't do.  Your main bottleneck is updating shadow map textures, and if you can skip a lot of that updating then you're going to get a huge performance win.  By having a separate shadow map per-light you get to be able to cache shadowmaps, and then for each light, if it doesn't have a moving object in it, you can just reuse it's shadow map from the previous frame.  Transient lights (e.g. explosions) or lights which move are always updated, but for other static scene lights you can easily lop off a large amount of your updating overhead.

 

Yes, that means extra memory, but it's a tradeoff - memory versus performance - and on any reasonable GPU you really don't have memory problems.  Moving them to and from a free pool means that in general you only ever have to deal with memory overhead for lights in the current scene anyway, rather than for every single light in your map.

 

For updating shadow maps I use a geometry shader so I only need to submit one batch of draw calls rather than 6; you can frustum cull and backface cull per-cube face in a geometry shader too - there's some nice sample code for it here: http://diaryofagraphicsprogrammer.blogspot.ie/2011/02/shadows-thoughts-on-ellipsoid-light.html (D3D/HLSL but you should be able to see the principle) but that's a more minor performance gain.

 

Deferred isn't the magic solution to shadowing that it is to lighting, and much of this applies to a forward renderer too.

 

Finally, and it does need to be said - 100-150fps compared to 200-300 is not "horrible".

Share this post


Link to post
Share on other sites

I understand that regardless of whether you use deferred rendering or not... Dynamic shadowmap generation is the same expense. But what I don't fully get is the *application* of the shadowmap data... Like, in my non-deferred renderer what I do in my fragment shader is loop each light and do a texture2D() lookup on each shadowmap... But this ends up being very slow with a very small number of shadowed lights. Is the *application of the shadowmap* faster in deferred? Or still the same? (i know the lighting itself is faster - so is the "application of shadowmaps" also called "lighting" )

 

Like if you use static shadowmaps generated only once for unchanging geometry

Share this post


Link to post
Share on other sites

The performance is horrible though. I *do* use a cubemap to reduce texture() fetches, but even then I can only get about 6 lights before I start running around 100-150 FPS (down from 200-300).


I can't comment on the techniques you are using (outside my knowledge domain), but I don't consider 150 FPS "horrible". Are you taking into account that there are initial upfront costs for setting up everything? Also, are you measuring your performance in microseconds?

Share this post


Link to post
Share on other sites
Sign in to follow this  

  • Advertisement
×

Important Information

By using GameDev.net, you agree to our community Guidelines, Terms of Use, and Privacy Policy.

GameDev.net is your game development community. Create an account for your GameDev Portfolio and participate in the largest developer community in the games industry.

Sign me up!