This is an extension of a paper I wrote which shows how to use the linear texture sampling capabilities of the GPU to calculate points on Bezier curves. You store the control points in the texture, then sample along the texture’s diagonal to get points on the curve:
GPU Texture Sampler Bezier Curve Evaluation
This extension shows how to use the technique to evaluate points on surfaces and inside of volumes, where those surfaces and volumes are defined either by Bezier curves or polynomials (Tensor products of polynomials to be more specific).
As an example of what this post will allow you to do:
- By taking a single sample of a 3d RGBA volume texture, you’ll be able to get a bicubic interpolated value (a bicubic surface).
- Alternately, taking a single sample of a 3d RGBA volume texture will allow you to get a linear interpolation between two biquadratic surfaces (a linear/biquadratic volume).
- This post also covers how to extend this to higher degree surfaces and volumes.
Here are two images generated by the WebGL2 demos I made for this post which utilize this technique for rendering surfaces, fog volumes, and solid volumes. (link to the demos at bottom of post!)
All textures are size 2 on each axis which makes it a cache friendly technique (you can grow the texture sizes for piecewise curves/surfaces/volumes though). It leverages the hardware interpolation which makes it a relatively computationally inexpensive technique, and it supports all polynomials within the limitations of floating point math, so is also very flexible and expressive. You could even extend this to rational polynomial surfaces and volumes which among other things would allow perfect representations of conic sections.
The animated Bezier curve images in this post came from wikipedia. Go have a look and drop them a few bucks if you find wikipedia useful!
Wikipedia: Bézier curve
If you’ve read my curve paper and understand the basics you can skip this section and go onto the section “Before Going Into Surfaces”.
Let’s talk about how to store curves of various degrees in textures and evaluate points on them using the GPU Texture sampler. We’ll need this info when we are working with surfaces and volumes because a higher degree curve is dual to a section of lower degree surface or an even lower degree volume.
The three ways we’ll be talking about controlling the order of curves are:
- Texture Dimensionality – 1d texture vs 2d texture vs 3d texture vs 4d texture.
- Number of Color Channels – How many color channels are used? R? RG? RGB? RGBA?
- Multiple Texture Samples – Doing multiple texture reads.
By texture dimensionality I mean how many dimensions the texture has. In all cases, the size of the texture is going to be 2 on each axis.
Starting with a 1d texture, we have a single texture coordinate (u) to sample along. As we change the u value from 0 to 1, we are just linearly interpolating between the two values. A 1d texture that has 2 pixels in it can store a degree 1 curve, also known as a linear Bezier curve. With linear texture sampling, the GPU hardware will do this linear interpolation for you.
The equation for linear interpolation between two values A and B which are at t=0 and t=1 respectively is:
Here’s the 1d texture:
Here’s a linear curve:
Going to a 2d texture it gets more interesting. We now have two texture coordinates to sample along (u,v). Using linear sampling, the hardware will do bilinear interpolation (linear interpolation across each axis) to get the value at a specific (u,v) texture coordinate.
Here is the equation for bilinear interpolation between 4 values A,B,C,D which are at texture coordinates (0,0), (1,0), (0,1), (1,1) respectively, being sampled at (u,v):
That equation interpolates from A to B by u (x axis), and from C to D by u (x axis), and then interpolates from the first result to the second by v (y axis). Note that it doesn’t actually matter which axis is interpolated by first. An equivelant equation would be one that interpolates from A to B by v (y axis) and from B to C by v (y axis) and then between those results by u (x axis).
With that equation, something interesting starts to happen if you use the same value (t) for u and v, expand and simplify, and end up at this equation:
That equation is very close to the quadratic Bezier formula, which is below:
To get to that equation, we just make B and C the same value (B), and rename D to C since that letter is unused. This tells us how we need to set up our 2d texture such that when we sample along the diagonal, we get the correct point on our quadratic Bezier curve:
Here’s a quadratic Bezier curve in action. You can see how it is a linear interpolation between two linear interpolations, just like taking a bilinearly interpolated sample on our texture is.
Taking this to a 3d texture, we now have three texture coordinates to sample along (u,v,w). Again, with linear sampling turned on, the hardware will do trilinear interpolation to get the value at a specific (u,v,w) texture coordinate.
If we follow the same process as the 2d texture, we will wind up with the equation for a cubic Bezier curve:
Here’s how the texture is laid out:
Here’s a cubic Bezier curve in action, where you can see 3 levels of linear interpolations, just like how trilinear interpolation works:
While I have never used a 4d texture it appears that directx supports them and there looks to be an OpenGL extension to support them as well.
If we took this to a 4d texture, we would end up with the equation for a quartic curve. If you have trouble visualizing what a 4d texture even looks like, you aren’t alone. You have four texture coordinates to sample along (u,v,w,t). When you sample it, there are two 3d volume textures that are sampled at (u,v,w), resulting in two values as a result. These values are then interpolated by t to give you the final value. A fourth dimensional texture lookup is just an interpolation between 2 three dimensional texture lookups. That is true of all dimensional texture lookups in fact. An N dimensional texture lookup is just the linear interpolation between two N-1 dimensional texture lookups. For example, a three dimensional texture lookup is just an interpolation between 2 two dimensional texture lookups. This “hierchical interpolation” is the link I noticed between texture interpolation and the De Casteljau algorithm, since that is also a hierchical interpolation algorithm, just with fewer values interpolated between.
Here’s how the 4d texture is laid out:
Here’s the quartic Bezier equation, which is what you get the answer to if you sample a 4d texture at (t,t,t,t):
Here’s a quartic Bezier curve in action, showing 4 levels of linear interpolation, just like how quadrilinear interpolation works with 4d textures:
So, the bottom line of this section is that if we sample along the diagonal of an N dimensional texture which has one color channel, we will get points on a degree N curve.
Number of Color Channels
Another way we can control the degree of a curve stored in a texture is by the number of color channels that are stored in the texture.
In the section above we showed a 1d texture that stored a linear curve. it had only one color channel:
Let’s add another color channel. A,B will be stored in the red channel, and B,C will be stored in the green channel:
When we read that texture at location (t), we will get the following values:
- R: The linear interpolation between A and B at time t.
- G: The linear interpolation between B and C at time t.
Now, if we just lerp between R and G in our shader, for time t, we will get the point at time t, on the cubic Bezier curve defined by control points A,B,C.
Pretty cool right?
What happens if we add another color channel, blue?
Well, when we sample the texture at time t, we get the following values:
- R: The linear interpolation between A and B at time t.
- G: The linear interpolation between B and C at time t.
- B: The linear interpolation between C and D at time t.
We can combine these values using the quadratic Bezier curve formula, as if these were each a control point:
The result we get is a point on the CUBIC curve defined by the four control points A,B,C,D.
In the previous section, it took a 3d volume texture to calculate a cubic curve. In this section we were able to do it with a 1d RGB texture, but it came at the cost of of having to do some calculation in the shader code after sampling the texture to combine the color channels and get the final result.
How exactly does adding a color channel affect the degree though? Each color channel added increases the degree by 1.
You can see this is true by seeing in the last section how a 3 dimensional texture can evaluate a cubic, and a 4 dimensional texture can evalaute a quartic, but the 4th dimensional texture was just two 3 dimensional textures. Adding a second color channel just doubles the size of your data (and adding two tripples, and adding three quadruples), so having a 3d volume texture that has two color channels is the same as having a 4d volume texture with a single color channel. In both cases, you are just interpolating between two 3d texture samples.
So, for every color channel we add, we add a degree.
Multiple Texture Samples
Multiple texture samples is the last way to control curve degree that we are going to talk about.
Taking extra texture samples is a lot like adding color channels.
If you have a 1d RGB texture, you get a result of 3 lerps – R,G,B – which you can use to calculate a cubic curve point (order 3). If you take a second sample, you get R0,G0,B0,R1,G1,B1 which is a result of 6 lerps, which gives you a point on a sextic curve (order 6).
If you have a 2d RGBA texture, you get the result of 4 quadratic interpolations – R,G,B,A – which gives you an order 5 curve point. Taking another texture read gives you 8 quadratic interpolation results, which you can put together to make an order 9 curve point. Taking a third texture read would get you up to order 13.
Just like adding color channels, taking extra texture samples requires you to combine the multiple results in your shader, which increases computational cost.
Besides that, you are also doing more texture reads, which can be another source of performance loss. The textures are small (up to 2x2x2x2) so are texture cache friendly, but if you have multiple textures, it could start to add up I’m sure.
IMO this option should be avoided in favor of the others, when possible.
Before Going Into Surfaces
Before we start on surfaces, I want to mention a few things.
Even though we’ve been talking about Bezier curves specifically, a previous post explained how to convert any polynomial from power basis form into Bernstein basis form (aka you can turn any polynomial into a Bezier curve that is exactly equivelant). So, this generalizes to polynomials, and even rational polynomials if you do division in your shader code, but I’ll point you towards that post for more information on that: Evaluating Polynomials with the GPU Texture Sampler.
You can also extend the above for piecewise curves easily enough. You just set up a different curve (or surface or volume, as we describe below) for different ranges of your parameter space values. From time 0 to 1, you may use one texture, and from time 1 to 2, you may use another. Better yet, you would store both curves in a single texture, and just make the texture be a little larger, instead of having two separate textures.
Also, many other types of curves – B-splines, nurbs – can be broken down exactly into piecewise Bezier curves (rational, if the source curve is rational). Check these links for more info:
Algorithms for B-Spline Curves
Wikipedia: De Boor’s Algorithm.
Finally onto surfaces!
I’m going to show how to extend the curve calculation technique to calculating points on Bezier rectangles. A Bezier rectangle is a rectangular surface which has one or more bezier curves across the X axis and one or more bezier curves across the Y axis. The degree of the curve on each axis doesn’t need to match so it could be quadratic on one axis and cubic on the other as an example.
To actually evaluate a point on the surface at location (u,v), you evaluate a point on each x axis curve for time u, and then you use those resulting values as control points in another curve that you evaluate at time v.
Just like linear interpolation, it doesn’t matter which axis you evaluate first for a Bezier rectangle surface so you could switch the order of the axis evaluation if you want to.
The image above shows a bicubic surface, the blue lines show the x axis cubic bezier curves, while the yellow lines show the y axis cubic bezier curves. Those lines are called “isolines” or “isocurves”. The 16 control points are shown in magenta.
Another name for a Bezier rectangle is a tensor product surface. This is a more generalized term as it isn’t limited to Bezier curves.
Note: there is another type of Bezier surface called a Bezier Triangle but I haven’t worked much with them so can’t say if any of these techniques work with them or not. It would be interesting to explore how these techniques apply to Bezier triangles, if at all.
Hopefully it should come as no surprise that a 2d texture using regular bilinear interpolation is in fact a Bezier rectangle which is linear on the x axis and linear on the y axis. It has a degree of (1,1) and is stored in a 2d texture (2×2 pixels), where the four control points are just stored in the four pixels. You just sample the texture at (u,v) to get that point on the surface. Pretty simple stuff.
Order (1,1) Bezier Rectangle:
Something interesting to note is that while the isolines (edges) of the rectangle are linear, the surface itself is curved. In fact, we know that the diagonal of this surface is in fact a quadratic Bezier curve because we calculate curves by sampling along the diagonal! (if the middle corners are different, it’s the same as if they were both replaced with the average of their values).
There are other ways to store this Degree (1,1) surface in a texture besides how i described. You could also have a 1 dimensional texture with two color channels, where you sample it along the u axis, and then interpolate your R and G values, using the v axis value. This would come at the cost of doing a lerp in the shader code, instead of having the texture sampler hardware do it for you.
Now that the simplest case is out of the way, how about the next simplest? What if we want a surface where we linearly interpolate between two quadratic curves? That is, what if we want to make a degree (2,1) Bezier rectangle?
Order (2,1) Bezier Rectangle:
Well if you think about it geometrically, we can store a quadratic curve in a 2d texture (2×2) with a single color channel. To linearly interpolate between two of those, we need two of those to interpolate between. So, we need a 3d texture, since that is just an interpolation between two 2d textures.
When we sample that texture, we use the coordinates (u,u,v). That will make it quadratic in u, but linear in v.
Stepping up the complexity again, what if we wanted to make a biquadratic surface – aka degree (2,2)?
Order (2,2) Bezier Rectangle:
Well, to make a quadratic curve we need 3 control points, so for a biquadratic surface we need 3 quadratic curves to quadratically interpolate between.
One way to do this would be with a 4d texture, sampling along (u,u,v,v) to make it quadratic in both u and v.
But, because 4d textures are kind of exotic and may not be supported, we can achieve this by instead having a 3d texture with two color channels: R,G.
When we sample that texture, we sample at (u,u,v) to get two values: R,G. Next we linearly interpolate from R to G using v. This makes us quadratic in both u and v.
There are other ways to encode this surface as well, but i’ll leave that to you to think about if you want to (:
Lastly, what if we wanted a bicubic surface? A cubic curve has 4 control points, so we need 4 cubic curves to cubically interpolate between to make our final surface.
Order (3,3) Bezier Rectangle:
Thinking back to the first section, a 3d texture can evaluate a cubic curve. Since we need four cubic curves, let’s just use all four color channels RGBA. We would sample our texture at (u,u,u) to get four cubic curves in RGBA and then would use the cubic Bezier formula to combine those four values using v into our final result.
Generalizing surface calculations a bit, there are basically two steps.
First is you need to figure out what your requirements for the x axis is as far as texture storage for the desired degree you want. From there, you figure out what degree you want on your y axis, and that degree is what you multiply the x axis texture storage requirements for.
It can be a little bit like tetris trying to figure out how to fit various degree surfaces into various texture sizes and layouts, but it gets easier with a little practice.
It’s also important to remember that the x axis being the first axis is by convention only. It could easily be the y axis that defines the texture storage requirements, and is multiplied by the degree of the x axis.
Volumes aren’t a whole lot more complex than surfaces, but they are a lot hungrier for texture space and linear interpolations!
Extending the generalization of surfaces, you once again figure out requirements for the x axis, multiply those by the degree of the y axis, and then multiply that result by the degree of the z axis.
The simplest case for volumes is the trilinear case, aka the Degree (1,1,1) Bezier rectangle.
Order (1,1,1) Bezier Cube:
It’s a bit difficult to understand what’s going on in that picture by seeing the data as just fog density, so the demos let you specify a surface threshold such that if the fog is denser than that amount, it shows it as a surface. Here is the same trilinear Bezier volume with a surface threshold.
Order (1,1,1) Bezier Cube:
You just store your 8 values in the 8 corners of the 2x2x2 texture cube, and sample at (u,v,w) to get your trilinear result.
The next simplest case is that you want to quadratically interpolate between two linear surfaces – a Degree (1,1,2) Bezier rectangle.
Order (1,1,2) Bezier Cube:
To do this, you need 3 bilinear surfaces to interpolate between.
One way to do this would be to have a 2d Texture with R,G,B color channels. Sample the texture at (u,v), then quadratically interpolate R,G,B using w.
Another way to do this would be to have a 3d texture with R and G channels. When sampling, you sample the 3d texture at (u,v,w) to get your R and G results. You then linearly interpolate from R to G by w to get the final value.
Yet another way to do this would be to use a 4d texture if you have support for it, and sample along (u,v,w,w) to get your curve point using only hardware interpolation.
The next simplest volume type is a linear interpolation between two biquadratic surfaces – a Degree (2,2,1) Bezier rectangle.
Order (2,2,1) Bezier Cube:
From the surfaces section, we saw we could store a biquadratic surface in a 3d texture using two color channels R,G. After sampling at (u,u,v) you interpolate from R to G by v.
To make a volume that linearly interpolates between two biquadratic surfaces, we need two biquadratic surfaces, so need to double the storage we had before.
We can use a 3d texture with 4 color channels to make this happen by storing the first biquadratic in R,G and the second in B,A, sampling this texture at (u,u,v). Next, we interpolate between R and G by v, and also interpolate between B and A by v. Lastly, we linearly interpolate between those two results using w.
The next higher surface would be a triquadratic volume, which is degree (2,2,2). Since you can store a biquadratic surface in a 3d 2x2x2 texture with two color channels, and a triquadratic volume needs 3 of those, we need a 3d texture 6 color channels. Since that doesn’t exist, we could do something like store 2 of the quadratic surfaces in a 2x2x2 RGBA texture, and the other quadratic surface in a 2x2x2 RG texture. We would take two texture samples and combine the 6 results into our final value.
Tricubic is actually pretty simple to conceptualize luckily. We know that we can store a bicubic surface in a 3d 2x2x2 RGBA texture. We also know that we would need 4 of those if we want to make a tricubic volume. So, we could do 4 texture reads (one for each of our bicubic surfaces) and then combine those 4 samples across w to get our final volume value.
Hopefully you were able to follow along and see that this stuff is potentially pretty powerful.
Some profiling needs to be done to better understand the performance characteristics of using the texture sampler in this way, versus other methods of curve, surface and volume calculation. I have heard that even when your texture samples are in the texture cache, that it can still take like ~100 cycles to get the information back on a texture read. That means that this is probably not going to be as fast as using shader instructions to calculate the points on the curve. However, if you are compute bound and can offload some work to the texture sampler, or if you are already using a texture to store 1d/2d/3d data (or beyond) that you can aproximate with this technique, that you will have a net win.
One thing I really like about this is that it makes use of non programmable hardware to do useful work. It feels like if you were compute bound, that you could offload some work to the texture sampler if you had some polynomials to evaluate (or surfaces/volumes to sample), and get some perf back.
I also think this could possibly be an interesting way to make concise representations (and evaluation) of non polygonal models. I imagine it would have to be piecewise to make things that look like real world objects, but you do have quite a bit of control with Bezier curves, surfaces and volumes, especially if you use rational ones by doing a divide in your shader.
Here’s a few specifica areas I think this technique could help out with:
- Higher order texture interpolation with fewer samples – You’d have to preprocess textures and would spend more memory on them, but it may be worth while in some situations for higher quality results with a single texture read.
- 2D signed distance field rendering – SDF textures are great for making pseudo vector art. They do break down in some cases and at some magnification levels though. It would be interesting to see if using this technique could improve things either with higher order interpolation, or maybe by encoding (signed) distances differently. Possibly also just useful for describing 2d vector art in a polynomial form?
- 3d signed distance field rendering – Ray marching can make use of signed distance fields to render 3d objects. It can also make use of functions which can only give you inside or outside status based on a point. It would be interesting to explore encoding and decoding both of these types of functions within textures using this technique, to sample shapes during ray marching.
If you are interested in the above, or curious to learn more, here are some good links!
2D Catmull-Rom in 4 samples
Distance Field Textures
Inigo Quilez: raymarching distance fields
If you have any questions, corrections, feedback, ideas for extensions, etc please let me know! You can leave a comment below, or contact me on twitter at @Atrix256.
Feedback / Ideas
@anders_breakin had some ideas that could possibly pan out:
- The derivative of a Bezier curve is another Bezier curve (Derivatives of a Bézier Curve). You could encode the derivative curve(s) in a texture and use that to get the normal instead of using the central differences method. That might give higher quality normals, but should also decrease the number of texture reads needed to get the normal.
- If you want more accuracy, you may subdivide the curve into more numerous piecewise curves. The texture interpolator only has 8 bits of decimal precision (X.8 fixed point) when interpolating, but if you give it less of the curve/surface/volume to interpolate over at a time, it seems like that would result in more effective precision.
@Vector_GL suggested reading the values in the vertex shader and using the results in the pixel shader. I think something like this could work where you read the control points in the VS, and pass them to the PS, which would then be able to ray march the tensor product surface by evaluating it without texture reads. So long as you have fewer VS instances than PS instances (the triangles are not subpixel!) that this could be an interesting thing to try. It doesn’t take advantage of the texture interpolator, but maybe there would be a way to combine the techniques. If not, this still seems very pragmatic.
I was thinking maybe this could be done via “rasterization” by drawing a bunch of unit cubes and having the PS do the ray marching. With some careful planning, you could probably use Z-testing on this too, to quickly cull hidden pixels without having to ray march them.
Here are the WebGL2 demos:
Analytical Surfaces Evaluated by the GPU Texture Sampler
Analytical Volume Evaluated by the GPU Texture Sampler