Volumetric Rendering in Realtime

Most current implementations of fog in games use layered alpha images. This technique, however, does not bare significantly resemblance to how fog actually composites in real life, since density of the fog from the viewer is not modeled in any way.

A Simple Model of Fog

In order to create fog effects in a game, it is first necessary to create an analytical model that bears some resemblance to the mechanics of real fog. Fog is a cloud of water vapor consisting of millions of tiny particles floating in space. Incoming light is scattered and emitted back into the scene. This model is too complex to render in real time, and so a few assumptions and restrictions must be made. The following is a similar model to what is used in depth fog.

The first and most important assumption, common to many real time fog implementations, is that incoming light at each particle is constant. That is, one particle of fog located at one end of a fog volume and another particle are receiving the same amount of incoming light.
The next related assumption is that each particle of fog emits the same amount of light, and in all directions. This, of course, implies that a fog's density remains fixed. These two assumptions mean that, given a spherical volume of fog, equal light is emitted in all directions.

Using these assumptions, a model of fog can be defined. If a ray is cast back from a pinhole camera through the scene - the amount of fog that contributes to the color of that ray is the sum of all the light along the ray's path. In other words, the amount of contributing light is equal to the area of fog between the camera and the point in the screen. The light of the incoming ray, however, was partially absorbed by the fog itself, thus reducing its intensity.

So, the proposed model of fog is (done for each color channel):

Since the area of fog emitting light is the same for the area of fog absorbing light, and the assumption is made that the amount of light emitted is the same percentage as absorbed, then this equal simplifies to:

If this is a per pixel operation, then the incoming light is already computed by rendering the scene as it would normally appear. An analytical way of thinking of this problem is: The amount a pixel changes to the fog color is proportional to the amount of fog between the camera and the pixel. This, of course, is the same model that is used in distance fog. Thus, the problem is reduced to determining the amount of fog between the camera and the pixel being rendered.

Determining Fog Depth at each pixel location

Standard Depth fog uses the Z (or w) value as the density of fog. This works well, but limits the model to omnipresent fog. That is, the camera is always in fog, and there is (Save a definable sphere around the camera) an even amount of fog at all points in the scene.
Of course, this does not work well (or at all) for such effects as a ground fog, and this technique cannot be used for interesting volumetric lighting.

An alternative way to create fog is to model a polygonal hull that represents the fog, and to compute the area of fog for each pixel rendered on the scene. At first glance, this seams impossibly complex. Computing volume area typically involves complex integration.

However, the shaft of fog along a ray can be closely approximated subtracting the w depth a ray enters a fog volume from the w depth of the point it leaves the volume, and multiplying by some constant. (Mathematically, this is a simple application of a form of Stoke's theorem, where all but 2 of the terms cancel since the flux is constant in the interior).

A Simple Case

The first case to consider is a way of rendering a convex volumetric fog that has no object in it, including the camera. The algorithm can easily be expanded to handle objects (or parts of the scene) inside the fog, the camera inside the fog, and concave volumes.

Computing this term on a per pixel basis involves several passes. Clearly, for any view there are two distances of concern: the point the ray enters the fog, and the point the ray exists.
Finding the point a ray enters a fog volume is computed by rendering the fog volume and reading the w value. Finding the ray on the other side of the fog volume is also not difficult. Polygons not facing the camera are culled away - but since any surface not facing the camera would be the backside of the fog volume, reversing the culling order and drawing the fog again, renders the inside of the fog volume. With convex volumes - there will never be case where the ray will pass in and out of a fog volume twice.

To get the total amount of fog in the scene, the buffer containing the front side w values of the fog volume is subtracted from the buffer containing the back side w values of the fog. But, the first question is, how can w pixels operations be performed? And then, how can this value be used for anything? Using a vertex shader, the w is encoded into the alpha channel, thereby loading the w depth of every pixel into the alpha channel of the render target. After the subtraction, the remaining w value represents the amount of fog at that pixel.

So the algorithm for this simple case is:

Adding another variable: Objects in the fog
Rendering fog with no objects inside it is not that interesting, so the above algorithm needs to be expanded to allow objects to pass in and out of the fog. This task turns out to be rather straightforward. If the above fog algorithm was applied without taking into consideration that objects are in the middle, the fog would be incorrect.

The reason why this is not correct is obvious; the actual volume of fog between it and the camera has been computed incorrectly. Because there is an object inside the fog - the backside of the fog is no longer the polygonal hull that was modeled, but the front side of the object. The distances of fog needs to be computed using the front side of the object as the back end of the fog.

This is accomplished by rendering the scene (defined as objects in the fog) using the W trick. If a pixel of an object lies in front of the fog's back end - it replaces the fogs backend with its own, thereby becoming the virtual back part of the fog.

The algorithm changes to:

Unfortunately, the above approach has one drawback. If an object is partially obscured by fog, then the component that is not in the fog will be rendered into the back buffer, effectively becoming the backside of the fog. Thus, then the distance from these pixels to the camera would be counted as the distance of fog - even though there is none.

Although this could be corrected by using the stencil buffer, another approach is to redraw (or frame copy) the screen in the front side of pass - thereby using the scene as the fog front as well as the back. This causes objects partially obscured by fog to render correctly - those parts not in fog result in a 0 fog depth value. This new approach looks like:

Further Optimizations and Improvements

Cleary, this is a simple foundation for fog - there are numerous improvements and enhancements that can be made. Perhaps highest on the list is a precision issue. Most hardware allows only 8 bit alpha formats. Because so much is dependent on the w depth - 8 bits can be a real constraint. Imagine a typical application of a volumetric fog - a large sheet of fog along the ground. No matter what function used to take the depth and render it into fog - there remains a virtual far and near clipping plane for the fog. Expanding these planes means either less dense, or less precise fog, while keeping them contracted means adjusting the fog clipping planes for each fog volume rendered.

On new and upcoming hardware, however, there is a trick with the pixel shaders. Why not keep some more bits of precision in one of the color channels, and use the pixel shader to perform a carry operation? At first glance it appears that 16 bit math easily be accomplished on parts designed to operate at only 8. However, there is one nasty limiting factor - on a triangle basis - the color interpolators work at only 8 bits. Texture coordinates, on the other hand, typically operate at much higher precision, usually at least 16 bits. Although texture coordinates can be loaded into color registers, the lower bits of precision are lost . An alternative is to create a 1D step function filled texture, with each texel representing a higher precision value embedded in the alpha and color channels. Unfortunately, the precision here is usually limited to the size of a texture.

Once the issue of higher precision is addressed, it is possible to render concave volumes even with limited 8-bit hardware. This must be accomplished by either rendering concave fog volumes as a collection on its convex parts, or by summing the multiple entry points of fog and subtracting away the multiple exit points. Unfortunatly, the high precision trick will not work for the latter approach since there is no way to both read and write the render target in the pixel shader. Although a system of swapping between multiple buffers carefully segmented to avoid overlap might work, this latter approach will probally not be feasible until hardware allows rendering into 16 bit formats (i.e. a 16 bit alpha format).

Finally, there are many artistic enhancements that can be made on this kind of volumetric effect. To make volumetric light, for instance, the alpha blends modes can be changed to additive rather then blend, thereby adding light to the scene. Decay constants can also be modeled in this way, to accomplish some surface variations of fog density.

Additionally, fog volumes can be fitted with textures on top of them that operate much like bump maps do - varying the height of the fog at that point without changing the actual geometry. To create an animated set of ripples in fog, for instance, one can take a ripple depth texture and move it along the surface of the fog volumes, and adding it to the w depth. Other texture tricks are possible as well - noise environment maps can be coupled to fog volumes to allow primitive dust effects.

And of course, it can be quite fun to draw the fog mask without actually drawing the object - creating an invisible object moving through the scene.

Supplement

The article, Volumetric Rendering in Real-time, covered the basis of volumetric depth rendering, but at the time of the writing, no pixel shader compliant hardware was available. This supplement describes a process designed to achieve two goals, to get more precision out of an 8 bit part, and to allow the creation of concave fog volumes.

Handling Concavity
Computing the distance of fog for the convex case was relatively simple. Recall that the front side of the fog volume was subtracted away from the backside (where the depth is measured in number of units from the camera). Unfortunately, this does not work with concave fog volumes because at any given pixel, it may have two back sides and two front sides.

The solution is intuitive and has sound mathematical backing - sum all of the front sides and subtract them from the summed front sides. As shown in diagram one - this is the mathematical equivalent of breaking the volume into convex chunks and summing the up.

Computing concavity is as simple as adding the multiple front sides and subtracting them from the multiple back sides. Clearly, a meager 8 bits won't be enough for this. Every bit added would allow another summation and subtraction, and allow for more complex fog scenes.

There is an important assumption being made about the fog volume. Is must be a continuous, orientable hull. That is, it cannot have any holes in it. Every ray cast through the volume must enter through hull the same number of times it exits.

Getting Higher Precision
Although most hardware acceleration can handle 32 bits, it is really four 8-bit channels. The way most hardware works today, there is only one place where the fog depths could be summed up: The Alpha Blender.

The alpha blender is typically used to blend on alpha textures by configuring the source destination to multiply against the source alpha, and the destination to multiply against the inverse alpha. However, they can also be used to add (or subtract) the source and destination color channels. Unfortunately, there is no way to perform a carry operation here: If one channel would exceed 255 for a color value, it simply saturates to 255.

In order to perform higher bit precision additions on the Alpha Blending Unit, the incoming data has to be formatted in a way which is compatible with the way the alpha blender adds. To do this, the color channels can hold different bits of the actual result, and most importantly, be allowed some overlap in their bits.

The above will give us 12 bit precision in an 8 bit pipe. The Red channel will contain the upper 8 bits, and the blue channel will contain the lower 4 -plus 3 carry spots. The upper bit should not be used for reasons which are discussed later. So the actual value encoded is Red*16+Blue.

Now, the Alpha Blender will add multiple values in this format correctly up to 8 times before there is any possibility of a carry bit not propagating. This limits the fog hulls to ones which do not have concavity where looking on any direction a ray might pass in and out of the volume more than 8 times.

Encoding the bits in which will be added cannot be done with a pixel shader. There are two primary limitations. First, the color interpolators are 8 bit as well. Since the depth is computed on a per vertex level, this won't let higher bit values into the independent color channels. Even if the color channel had a higher precision, the pixel shader has no instruction to capture the lower bits of a higher bit value.

The alternative is to use a texture to hold the encoded depths. The advantage of this is twofold. First, texture interpolaters have much higher precision than color interpolaters, and second, no pixel shader is needed for initial step of summing the font and back sides of the fog volume.

Unfortunately, most hardware limits the dimensions of textures. 4096 is a typical limitation. This amounts to 12 bits of precision to be encoded in the texture. 12 bits, however, is vastly superior to 8 bits and can make all the difference to making fog volumes practical.

Setting it all Up
Three important details remain: The actual summing of the fog sides, compensating for objects inside the fog, and the final subtraction.

The summing is done in three steps. First, the scene needs to be rendered to set a Z buffer. This will prevent fog pixels from being drawn which are behind some totally occluding objects. In a real application, this z could be shared from the pass which draws the geometry. The Z is then write disabled - so that fog writes will not update the z buffer.

After this, the summing is exactly as expected. The app simply draws all the forward facing polygons in one buffer, adding up their results, and then draws all the backward facing polygons in another buffer. There is one potential problem, however. In order to sum the depths of the fog volume, the alpha blend constants need to be set to one for the destination and one for the source, thereby adding the incoming pixel with the one already in the buffer.

Unfortunately, this does not take into account objects inside the fog that are acting as a surrogate fog cover. In this case - the scene itself must be added to scene since the far end of the fog would have been rejected by the Z test.

At first, this looks like an easy solution. In the previous article, the buffers were setup so that they were initialized to the scene's depth value. This way, fog depth values would replace any depth value in the scene if they were in front of it (i.e. the Z test succeeds) - but if no fog was present the scene would act as the fog cover.

This cannot be done for general concavity, however. While technically correct in the convex case, in the concave case there may be pixels at which the fog volumes are rendered multiple times on the front side and multiple sides on the backside. For these pixels, if the there was part of an object in between fog layers than the front buffer would be the sum of n front sides, and the back side would be sum of n-1 back sides. But since the fog cover was replaced by the fog - there are now more entry points then exit points. The result is painfully obvious - parts of the scene suddenly loose all fog when they should have some.

The above diagram illustrates that without taking into account the object's own depth value, the depth value generated would be B1 - A1 - A2 since B2 was never drawn because it failed the Z test of the scene. This value would be negative, and no fog would get blended. In this case, C needs to be added into the equation.

The solution requires knowing which scenario's where the scene's w depth should be added and which scenarios the scene's w depth should be ignored. Fortunately, this is not difficult to find. The only situation where the scene's w depth should be added to the total fog depth are those pixels where the object is in between the front side of a fog volume and its corresponding backside.

The above question can be thought of asking the question: did the ray ever leave the fog volume? Since the fog hulls are required to be continuous, then if the answer is no then part of the scene must have blocked the ray. This test can be performed by a standard inside outside test.

To perform an inside/outside test - each time a fog pixel is rendered, the alpha value is incremented. If the alpha values of the far fog distances is subtracted to the corresponding point on the near fog distance, then values greater then 1 indicate the ray stopped inside the volume. Values of 0 indicate that the ray left the fog volume.

To set this test up, the alpha buffer of the near and far w depth buffers must be cleared to 0. Each time a fog pixel is rendered, the alpha will be incremented by the hex value 0x10. This value was used because the pixel shader must perform a 1 or 0 logical operation. A small positive value must be mapped to 1.0 in the pixel shader, a step which requires multiple shifts. Due to instruction count restraints - the intial value has to be at least 0x10 for the shifts to saturate a non-zero value to one.

The rest is straightforward - all the front sides and all the backsides are summed up in their independent buffers. The scene is also drawn in its own buffer. Then all three buffers are ran through the final pass where the scene's w depth is added on only if the differences of the alpha values is not 0.

This requires a lengthy pixel shader. A great deal of care must be taken to avoid potential precision pitfalls. The following pixel shader performs the required math, although it requires every instruction slot of the pixel shader and nearly every register. Unfortunately, with no carry bit, there is no way to achieve a full 8 bit value at the end of the computation, so it must settle for 8.

ps.1.1
def c1, 1.0f,0.0f,0.0f,0.0f
def c4, 0.0f,0.0f,1.0f,0.0f

tex t0 // near buffer B
tex t1 // far buffer A
tex t2 // scene buffer C

// input:
// b = low bits (a) (4 bits)
// r = high bits (b) (8 bits)
// intermediate output:
// r1.b = (a1 - a2) (can't be greater than 7 bits set )
// r1.r = (b1 - b2)

sub r1.rgb,t1,t0
+sub_4x r1.a,t0,t1 //If this value is non zero, then
mov_4x r0.a,r1.a //the were not as many backs as
mad r1.rgb,r0.a,t2,r1 //front and must add in the scene
dp3 t0.rgba,r1,c4 // move red component into alpha

// Need to shift r1.rgb 6 bits. This could saturate
// to 255 if any other bits are set, but that is fine
// because in this case, the end result of the subtract
// would have to be saturated since we can't be
// subtracting more than 127
mov_x4 r1.rgb,r1
dp3_x4 t1.rgba,r1,c1 // move into the alpha
add_x2 r0.a,t0.a,t1.a // the subtract was in 0-127
mov_d2 r0.a,r0.a // chop off last bit else banding
+mov r0.rgb,c3 // load the fog color

This pixel shader gives an alpha value which represents the density of fog, and loads the fog color constant into the color channels. The Alpha Blending stage can now be used to blend on the fog.

Finally, there is one situation which can cause serious problems - clipping. If a part of the fog volume is clipped away by the camera because the camera is partially in the fog, then part of the scene might be in the fog. Previously, it was assumed the camera was either entirely all the way in, or all the way out of the fog. This may not always be the case.

An alternative solution is to not allow polygons to get clipped. The vertex shader can detect vertices which would get clipped away and snap them to the near clip plane. The following vertex shader clips w depths to the near clip plane, and z depths to zero.

// transform position into projection space
m4x4 r0,v0,c8
max r0.z,c40.z,r0.z //clamp to 0
max r0.w,c12.x,r0.w //clamp to near clip plane
mov oPos,r0

// Subtract the Near clipping plane
add r0.w,r0.w,-c12.x

// Scale to give us the far clipping plane
mul r0.w,r0.w,c12.y

// load depth into texture, don't care about y
mov oT0.xy,r0.w

Additionally, please note that full source code will be available with the release of DX 8.1, which is imminent. The code is in the volume fog SDK sample.

________________________________________________________