|[ << ]||[ < ]||[ Up ]||[ > ]||[ >> ]||[Top]||[Contents]||[Index]||[ ? ]|
Texture shadows involve rendering shadow casters from the point of view of the light into a texture, which is then projected onto shadow receivers. The main advantage of texture shadows as opposed to Stencil Shadows is that the overhead of increasing the geometric detail is far lower, since there is no need to perform per-triangle calculations. Most of the work in rendering texture shadows is done by the graphics card, meaning the technique scales well when taking advantage of the latest cards, which are at present outpacing CPUs in terms of their speed of development. In addition, texture shadows are much more customisable - you can pull them into shaders to apply as you like (particularly with Integrated Texture Shadows, you can perform filtering to create softer shadows or perform other special effects on them. Basically, most modern engines use texture shadows as their primary shadow technique simply because they are more powerful, and the increasing speed of GPUs is rapidly amortizing the fillrate / texture access costs of using them.
The main disadvantage to texture shadows is that, because they are simply a texture, they have a fixed resolution which means if stretched, the pixellation of the texture can become obvious. There are ways to combat this though:
The simplest projection is just to render the shadow casters from the lights perspective using a regular camera setup. This can look bad though, so there are many other projections which can help to improve the quality from the main camera’s perspective. OGRE supports pluggable projection bases via it’s ShadowCameraSetup class, and comes with several existing options - Uniform (which is the simplest), Uniform Focussed (which is still a normal camera projection, except that the camera is focussed into the area that the main viewing camera is looking at), LiSPSM (Light Space Perspective Shadow Mapping - which both focusses and distorts the shadow frustum based on the main view camera) and Plan Optimal (which seeks to optimise the shadow fidelity for a single receiver plane).
You can also sample the shadow texture multiple times rather than once to soften the shadow edges and improve the appearance. Percentage Closest Filtering (PCF) is the most popular approach, although there are multiple variants depending on the number and pattern of the samples you take. Our shadows demo includes a 5-tap PCF example combined with depth shadow mapping.
Again as GPUs get faster and gain more memory, you can scale up to take advantage of this.
If you combine all 3 of these techniques you can get a very high quality shadow solution.
The other issue is with point lights. Because texture shadows require a render to texture in the direction of the light, omnidirectional lights (point lights) would require 6 renders to totally cover all the directions shadows might be cast. For this reason, Ogre primarily supports directional lights and spotlights for generating texture shadows; you can use point lights but they will only work if off-camera since they are essentially turned into a spotlight shining into your camera frustum for the purposes of texture shadows.
Directional lights in theory shadow the entire scene from an infinitely distant light. Now, since we only have a finite texture which will look very poor quality if stretched over the entire scene, clearly a simplification is required. Ogre places a shadow texture over the area immediately in front of the camera, and moves it as the camera moves (although it rounds this movement to multiples of texels so that the slight ’swimming shadow’ effect caused by moving the texture is minimised). The range to which this shadow extends, and the offset used to move it in front of the camera, are configurable (See Configuring Texture Shadows). At the far edge of the shadow, Ogre fades out the shadow based on other configurable parameters so that the termination of the shadow is softened.
Spotlights are much easier to represent as renderable shadow textures than directional lights, since they are naturally a frustum. Ogre represents spotlight directly by rendering the shadow from the light position, in the direction of the light cone; the field-of-view of the texture camera is adjusted based on the spotlight falloff angles. In addition, to hide the fact that the shadow texture is square and has definite edges which could show up outside the spotlight, Ogre uses a second texture unit when projecting the shadow onto the scene which fades out the shadow gradually in a projected circle around the spotlight.
As mentioned above, to support point lights properly would require multiple renders (either 6 for a cubic render or perhaps 2 for a less precise parabolic mapping), so rather than do that we approximate point lights as spotlights, where the configuration is changed on the fly to make the light shine from its position over the whole of the viewing frustum. This is not an ideal setup since it means it can only really work if the point light’s position is out of view, and in addition the changing parameterisation can cause some ’swimming’ of the texture. Generally we recommend avoiding making point lights cast texture shadows.
To enable texture shadows, use the shadow technique SHADOWTYPE_TEXTURE_MODULATIVE or SHADOWTYPE_TEXTURE_ADDITIVE; as the name suggests this produces Modulative Shadows or Additive Light Masking respectively. The cheapest and simplest texture shadow techniques do not use depth information, they merely render casters to a texture and render this onto receivers as plain colour - this means self-shadowing is not possible using these methods. This is the default behaviour if you use the automatic, fixed-function compatible (and thus usable on lower end hardware) texture shadow techniques. You can however use shaders-based techniques through custom shadow materials for casters and receivers to perform more complex shadow algorithms, such as depth shadow mapping which does allow self-shadowing. OGRE comes with an example of this in its shadows demo, although it’s only usable on Shader Model 2 cards or better. Whilst fixed-function depth shadow mapping is available in OpenGL, it was never standardised in Direct3D so using shaders in custom caster & receiver materials is the only portable way to do it. If you use this approach, call SceneManager::setShadowTextureSelfShadow with a parameter of ’true’ to allow texture shadow casters to also be receivers.
If you’re not using depth shadow mapping, OGRE divides shadow casters and receivers into 2 disjoint groups. Simply by turning off shadow casting on an object, you automatically make it a shadow receiver (although this can be disabled by setting the ’receive_shadows’ option to ’false’ in a material script. Similarly, if an object is set as a shadow caster, it cannot receive shadows.
There are a number of settings which will help you configure your texture-based shadows so that they match your requirements.
Shadow textures take up texture memory, and to avoid stalling the rendering pipeline Ogre does not reuse the same shadow texture for multiple lights within the same frame. This means that each light which is to cast shadows must have its own shadow texture. In practice, if you have a lot of lights in your scene you would not wish to incur that sort of texture overhead.
You can adjust this manually by simply turning off shadow casting for lights you do not wish to cast shadows. In addition, you can set a maximum limit on the number of shadow textures Ogre is allowed to use by calling SceneManager::setShadowTextureCount. Each frame, Ogre determines the lights which could be affecting the frustum, and then allocates the number of shadow textures it is allowed to use to the lights on a first-come-first-served basis. Any additional lights will not cast shadows that frame.
Note that you can set the number of shadow textures and their size at the same time by using the SceneManager::setShadowTextureSettings method; this is useful because both the individual calls require the potential creation / destruction of texture resources.
The size of the textures used for rendering the shadow casters into can be altered; clearly using larger textures will give you better quality shadows, but at the expense of greater memory usage. Changing the texture size is done by calling SceneManager::setShadowTextureSize - textures are assumed to be square and you must specify a texture size that is a power of 2. Be aware that each modulative shadow texture will take size*size*3 bytes of texture memory.
Important: if you use the GL render system your shadow texture size can only be larger (in either dimension) than the size of your primary window surface if the hardware supports the Frame Buffer Object (FBO) or Pixel Buffer Object (PBO) extensions. Most modern cards support this now, but be careful of older cards - you can check the ability of the hardware to manage this through ogreRoot->getRenderSystem()->getCapabilities()->hasCapability(RSC_HWRENDER_TO_TEXTURE). If this returns false, if you create a shadow texture larger in any dimension than the primary surface, the rest of the shadow texture will be blank.
This determines the distance at which shadows are terminated; it also determines how far into the distance the texture shadows for directional lights are stretched - by reducing this value, or increasing the texture size, you can improve the quality of shadows from directional lights at the expense of closer shadow termination or increased memory usage, respectively.
As mentioned above in the directional lights section, the rendering of shadows for directional lights is an approximation that allows us to use a single render to cover a largish area with shadows. This offset parameter affects how far from the camera position the center of the shadow texture is offset, as a proportion of the shadow far distance. The greater this value, the more of the shadow texture is ’useful’ to you since it’s ahead of the camera, but also the further you offset it, the more chance there is of accidentally seeing the edge of the shadow texture at more extreme angles. You change this value by calling SceneManager::setShadowDirLightTextureOffset, the default is 0.6.
Shadows fade out before the shadow far distance so that the termination of shadow is not abrupt. You can configure the start and end points of this fade by calling the SceneManager::setShadowTextureFadeStart and SceneManager::setShadowTextureFadeEnd methods, both take distances as a proportion of the shadow far distance. Because of the inaccuracies caused by using a square texture and a radial fade distance, you cannot use 1.0 as the fade end, if you do you’ll see artifacts at the extreme edges. The default values are 0.7 and 0.9, which serve most purposes but you can change them if you like.
When rendering shadow casters into a modulative shadow texture, Ogre turns off all textures, and all lighting contributions except for ambient light, which it sets to the colour of the shadow (Shadow Colour). For additive shadows, it render the casters into a black & white texture instead. This is enough to render shadow casters for fixed-function material techniques, however where a vertex program is used Ogre doesn’t have so much control. If you use a vertex program in the first pass of your technique, then you must also tell ogre which vertex program you want it to use when rendering the shadow caster; see Shadows and Vertex Programs for full details.
As previously mentioned, one of the downsides of texture shadows is that the texture resolution is finite, and it’s possible to get aliasing when the size of the shadow texel is larger than a screen pixel, due to the projection of the texture. In order to address this, you can specify alternative projection bases by using or creating subclasses of the ShadowCameraSetup class. The default version is called DefaultShadowCameraSetup and this sets up a simple regular frustum for point and spotlights, and an orthographic frustum for directional lights. There is also a PlaneOptimalShadowCameraSetup class which specialises the projection to a plane, thus giving you much better definition provided your shadow receivers exist mostly in a single plane. Other setup classes (e.g. you might create a perspective or trapezoid shadow mapping version) can be created and plugged in at runtime, either on individual lights or on the SceneManager as a whole.
Shadow textures need a depth buffer like many other RTs (Render Textures). Prior to Ogre 1.8, the depth buffer behavior was left undefined leaving a very small possibility of causing inconsistencies across different window resolutions and render systems. Depending on the render window’s resolutions and/or rendersystem being used, the depth buffer might been shared with the render window or a new one could get created to suite the shadow textures. If the application was depending on the depth buffer contents from the previous scene render (that is, no clear was performed) where a shadow texture render pass was in the middle; then the depth buffer would’ve contained garbage (but not consistent on all machines) causing graphical glitches hard to spot.
From Ogre 1.8 onwards the depth buffer usage & sharing can be flexible controlled through the use of depth pool IDs. These pool IDs are not specifically part of shadow textures, but rather anything involving RTs. All RTs with the same pool ID share the same depth buffers when possible (following RenderSystem API rules, check RenderSystemCapabilities flags to find out what the behavior will be). The default ID for shadow textures is 1; which is the same default value for render windows, and RTTs; thus maintaining the same behavior with older applications while achieving maximum memory saving and performance efficiency because the number of created depth buffers is as lowest as possible.
However there are some reasons to put shadow textures in a separate pool. This holds specially true if the application depends on the previous contents from the depth buffer before the shadow pass, instead of doing a clear:
For example, the DeferredShading sample suffers from this problem. If this is a problem for a particular effect you’re trying to achieve, you can specify a custom pool ID so that shadow textures get their own depth buffer(s), ensuring they aren’t shared with other RTs. You can set the poolId parameter from either SceneManager::setShadowTextureSettings or setShadowTextureConfig
mSceneMgr->setShadowTextureSettings( size, count, format, PoolId ); mSceneMgr->setShadowTextureConfig( 0, 512, 512, PF_FLOAT16_R, 50 );
Note a poolId of 0 will make the shadow textures not to use a depth buffer, which isn’t usually a desired behavior.
Texture shadows have one major advantage over stencil shadows - the data used to represent them can be referenced in regular shaders. Whilst the default texture shadow modes (SHADOWTYPE_TEXTURE_MODULATIVE and SHADOWTYPE_TEXTURE_ADDITIVE) automatically render shadows for you, their disadvantage is that because they are generalised add-ons to your own materials, they tend to take more passes of the scene to use. In addition, you don’t have a lot of control over the composition of the shadows.
Here is where ’integrated’ texture shadows step in. Both of the texture shadow types above have alternative versions called SHADOWTYPE_TEXTURE_MODULATIVE_INTEGRATED and SHADOWTYPE_TEXTURE_ADDITIVE_INTEGRATED, where instead of rendering the shadows for you, it just creates the texture shadow and then expects you to use that shadow texture as you see fit when rendering receiver objects in the scene. The downside is that you have to take into account shadow receipt in every one of your materials if you use this option - the upside is that you have total control over how the shadow textures are used. The big advantage here is that you can can perform more complex shading, taking into account shadowing, than is possible using the generalised bolt-on approaches, AND you can probably write them in a smaller number of passes, since you know precisely what you need and can combine passes where possible. When you use one of these shadowing approaches, the only difference between additive and modulative is the colour of the casters in the shadow texture (the shadow colour for modulative, black for additive) - the actual calculation of how the texture affects the receivers is of course up to you. No separate modulative pass will be performed, and no splitting of your materials into ambient / per-light / decal etc will occur - absolutely everything is determined by your original material (which may have modulative passes or per-light iteration if you want of course, but it’s not required).
You reference a shadow texture in a material which implements this approach by using the ’content_type shadow’ directive in your texture_unit. It implicitly references a shadow texture based on the number of times you’ve used this directive in the same pass, and the light_start option or light-based pass iteration, which might start the light index higher than 0.
|[ << ]||[ < ]||[ Up ]||[ > ]||[ >> ]||[Top]||[Contents]||[Index]||[ ? ]|
This document was generated on August 20, 2012 using texi2html 5.0.