Brash & Plucky

Searching for a proper formulation for flyaways in wool and fabric in general.

These are some simulations of growth from the root being randomly affected. From top to bottom, using different seeds. From left to right, the resulting B-Spline being evaluated using few or many control points.

Everything happens in 3D, but this image is just the X-Z projection.

Force-field dynamics

e.g., wind, vibrations, etc…

Many sampling algorithms take as input random numbers in the [0..1) interval. On the other hand, Pseudo-Random Number Generators usually produce their output as (unsigned) integer 32-bit or 64-bit numbers.

It is trivial to turn an u:u32 integer into an f:[0..1] floating-point value, dividing u by 2^32-1. This is because u ranges in [0..2^32-1].

One might think that u/2^32 would produce a floating-point value lower than 1. But it does not, as evidenced by this loop:

The number printed out happens to be 0x100000101 which is 2^32+257.

( u / static_cast< f32_t >( 0x100000101ull ) ) is the f:[0..1) we were looking for.

( u / static_cast< f64_t >( 0x100000080ull ) ) does the same in double-precision.

One major endeavour that I undertook this past summer was the creation of a novel displacement mapping system for Maverick Render.

The main motivations behind this have been:

• Our customers were running out of GPU memory too often when using displacement.
• Our main goal is that Maverick is an exquisite-quality tool for model visualization. Hence displacement is very important to us.
• Open the door to some other geometry-generation features beyond displacement itself.

So I undusted the old virtualized geometry solution I implemented for fryrender back in 2008 and re-engineered it for the Maverick core in the GPU.

Say hi to: Micro-Patch Displacement Mapping

Non-affine MPDM in arbitrarily-curved surfaces

Affine MPDM in flat surfaces

Some good properties of Maverick’s displacement:

• No requirements whatsoever on the base mesh or on the heightmap texture.
• It costs near 0 in warm-up time.
• It costs near 0 in memory (only a small payload is appended to the height map).
• 0 information is lost; every single texel in the heightmap is turned into a displaced micro-patch.
• Implemented as a custom primitive in Maverick’s OptiX ray-tracing backend.
• Used every trick in the book to optimize it as much as I could.

MPDM vs. brute-force

MPDM is meant to address some classic problems with the subdivide-and-displace-vertices (brute-force) approach. Using brute-force:

• There is some (often significant) wait time for warm-up.
• Memory usage goes through the roof, more than often blowing up all available memory.
• Height sampling happens at the vertices and not at the heightmap texels.

Points 1-2 are very obvious. It becomes impossible to abuse displacement, or to recklessly use it for large surfaces. Some techniques such as back-face culling or camera culling can be used to work around these difficulties, but such techniques pose problems on their own. Expecting the user to be careful not to exceed his memory budget is always an ill-fated plan.

Point 3 often means that the resulting displacement is smudgy, and not as crisp as it should be. i.e., information is lost, and/or geometry is produced in excess. Some techniques (that Maverick itself has featured in the past) such as autobump can be used to work around this problem by kind-of-restoring the information that is lost in the process. But then again, optimal configuration would become the responsibility of the user.

None of these issues happen with MPDM. :-)

Are there any benefits to the old brute-force approach?

As usual, nothing really good comes without a price. This is often very true in the realm of algorithm optimization.

• MPDM unfolds virtualized geometry during ray-traversal. The displaced geometry never exists for real in memory (that is why it is said to be virtualized) and this is the reason why MPDM has a near 0-cost in memory usage. However, this virtualization comes at an algorithmic cost. Despite massively optimized, MPDM renders a bit more slowly than the same displaced geometry would if the vertices were pre-displaced during warm-up and entered the ray-tracing BVH as regular geometry.
• For MPDM, part of the optimization process relies on the fact that the input map is discretized in texels, (as in a regular bitmap). For brute-force displacement the only requirement is that the height map can be evaluated at each geometry vertex, which is true to bitmaps, procedurals, and all map types. So MPDM is a priori limited to filetex maps, but we have added to Maverick a new map type called bakemap so the user can bake any procedural map tree into a filetex to feed MPDM if necessary.

How does MPDM work?

Here’s a very broad explanation on how MPDM in Maverick works:

• A (rather classic) hierarchical partitioning scheme based in min-max mip-map levels is performed on the input heightmap. As depicted in the images below.
• Those voxels are efficiently ray-traceable… unless there is curvature in the base mesh. Hell breaks loose as soon as curvature is at play. Maverick does non-affine math here (visualized in a previous post on non-affine triangle sweeps). I’ve tried different approaches for this (non-affine, discretized, and hybrid), and the code allows to flag-switch each mode on and off.
• Because of the non-affine treatment, at the lowest hierarchy level you no longer have quads or triangles, but bilinear patches (hence the name: Micro-PATCH Displacement Mapping).

Non-affine MPDM in arbitrarily-curved surfaces

The input heightmap used

Affine MPDM in flat surfaces (special optimization branch)

The input heightmap used

But the devil lies in the details. Efficient implementation of this all was a multi-week hell working with data structures, low-level optimizations, coordinating MPDM with OptiX and the rest of the Maverick core, and dealing with plenty of non-obvious fine-grain issues such as:

• Avoiding cracks across edges.
• Dealing with UV mapping projections.
• Transfer of shading normals.
• Real-time notifications from the UI.
• …

Youtube 4K videos:

https://youtube.com/shorts/2PdRx0sALgg?feature=share

https://youtube.com/shorts/O9qUCZTxkTQ?feature=share

https://youtube.com/shorts/ZKcLPQkhFWc?feature=share

https://youtube.com/shorts/Y9DFtEo5Hgo?feature=share

I’ve been working lately on glare/bloom/fringe and other post-processing effects in Maverick Render. Some of these inherit from our lovely ArionFX Adobe Photoshop and AfterEffects plug-in.

One complaint in ArionFX and also in Maverick is (was, because this post is about a successful fix) that Glare/Bloom diverge in shape and power when the input image is rendered at a different resolution, even if the Glare/Bloom parameters stay the same.

There are some relatively unobvious reasons for this. Basically, the challenges are:

• Hard challenge: Diffraction is a frequency analysis effect. For a render, this happens in the discrete realm (pixels). The size (amount of pixels) of the images involved changes what frequencies and how they show up in the Fourier Transform.
• Hard challenge: Anti-Aliasing of neighboring pixels (more prevalent at low resolution) averages their power and dims the overall Glare/Bloom overlay. This can pose a real problem for thin geometries such as lightbulb filaments.
• Easy challenge: As illustrated in some of my previous posts, the FT itself has some properties that relate its scale and power to the scale and power of the aperture/obstacle of the lens iris. These of course must be compensated for.
• Medium challenge: Changes in aspect ratio, or in padding in the image buffers (such as the padding between the IPR size in the UI vs. the canvas size) must be taken into account as well.

The upcoming release of Maverick will address these issues.

Here’s a small video with a sequence of Maverick post-processing effects, all rendered alternating landscape and portrait aspect ratios between 512 and 2048. The video is cropped as landscape to be easier on the eyes. As can be seen, at lower resolutions there’s always some power divergence, and a little bit of blur. But those are unavoidable to some extent.

Youtube 4K version: youtu.be/4K8F8vP76…

These are some technical visualizations from the all-new 0-memory geometry generation subsystem in Maverick Render which we intend to release soon.

I will be posting more related images and videos in the upcoming days.

Youtube versions of the above videos

[EDIT] I believe that micro.blog has some difficulty with (i.e., won’t play back) videos encoded at a non-standard aspect ratio. So here are the same videos in Youtube form:

youtu.be/mORIVYWTr…

youtu.be/1VBq42-cF…

youtu.be/wOYJUVqYg…

Results in production

The first beneficiary of this subsystem is Displacement Mapping (now Micro-Patch Displacement Mapping a.k.a., MPDM). This new MPDM system uses near zero memory, near zero warm-up time, and captures literally and exactly every single texel in the input texture.

Youtube 4K version: youtu.be/BfNTDcMCB…

These below are a bunch of glare patterns produced with the (much improved) bloom/glare toolchain in Maverick Render.

These tests have been conducted by @Vloba with the soon-to-be-released Maverick Studio 2022.5 and they make use of some of our procedural maps for aperture.

These images are expressive of the versatility of the new system.

Bessel functions

From Wikipedia

Iterative implementation

I came up with the above implementation, with which I plotted these.

J0-4 in [0..2pi]

2*abs(J1-4(x)/x) in [0..2pi]

Iterative methods are a terrible choice for many use cases because of error accumulation, jitter, the very presence of a loop, etc… But in this case this method also happens to be pretty weak and way too easy to break:

J0-4 in [0..8pi]

Deus ex machina

After some unsuccessful googling I was about to work on my own Taylor Series approximation or at least study the proposed implementations in Numerical Recipes. But I learnt to my surprise :o) that <math.h> provides j0/j1/jn/y0/y1/yn and even CUDA provides j0f/j1f/jnf/y0f/y1f/ynf.

I checked whether the C/CUDA implementations match my iterative implementation, and they do.

For reasons I may discuss in a future post, I am after j1 only, so… my job here is done.

One might need (as is the case for me now) to compute a FT without actually computing the FT. i.e., to find the explicit formulation for the resulting FT of a certain 2D signal.

In particular, as part of some optimizations in Maverick Render I wish to create a map that looks like the diffraction pattern of a polygonal aperture shape, but without computing the FT of said aperture.

From a previous post

Looking at this video from my previous post, intuition says that each of the streaks of a diffraction pattern is “kind of independent from” or “overlaps onto” the rest and could be modeled on its own from a 1D FT. This intuition is incorrect in the general case. But let’s focus on the one case where this reasoning is correct: the diffraction pattern of a square shape.

Separability of the FT

The Fourier Transform is separable in both directions:

• The 2D analysis formula can be written as a 1D analysis in the x direction followed by a 1D analysis in the y direction.
• The 2D synthesis formula can be written as a 1D analysis in the u direction followed by a 1D analysis in the v direction.

So it seems natural to focus on the 1D case first and then combine 2x 1D transforms for the square. The cross section of a square is (one cycle of) a 1D rectangular pulse.

Fourier Transform of a 1D unit pulse

From Wikipedia

A unit pulse of amplitude A and width T is a function g(t) that yields A for t IN [-T/2..+T/2] and 0 everywhere else. Plugging this g(t) in the definition of the Fourier Transform, we obtain this:

So the FT of a unit pulse is the real-valued function A*T*sinc(f*T). The imaginary component in this case cancels out.

The sinc function

The sin(pi*f*T)/(pi*f*T) part of the solution is such a fundamental building block in the field of signal processing that it got a name of its own: The venerable sinc function, or alternatively the sampling function.

From Wikipedia: The normalized sinc function is the Fourier Transform of the rectangular function with no scaling. It is used in the concept of reconstructing a continuous bandlimited signal from uniformly spaced samples of that signal.

The sinc function can be expressed in unnormalized form (sin(x)/x which integral over the real line is pi) or in normalized form (sin(pi*x)/(pi*x) which integral over the real line is 1).

In both cases there is an indetermination at x=0 but you can use L’Hôpital’s rule to show that sinc(0)=1.

From Wikipedia

Fourier Transform of a rectangular shape

Since the FT is separable as mentioned above, the 2D FT of a square shape must be the 1D FT of a unit pulse in the x direction, followed by another 1D FT in the y direction.

Note that this would work for rectangles too, by using a different T value for each axis.

The top-right image is the actual FT of the square shape on the top-left. The bottom-right image is just the above piece of code. The bottom-left image is the profile along the x direction of the above code at y=0 which is proportional to abs(sinc).

The top-right and bottom-right images match exactly, except for:

• Some numerical drift.
• Unlike the above code, the true FT is periodic and the streaks wrap-around. But this happens to be a desirable property in my case of use.

So: success!

Can this method be generalized to n-sided polygons?

One might hope that the FT of n-sided polygons, or even inf-sided polygons (circles) could be conscructed similarly by composing rotating FTs of square functions somehow. But the truth is that things get a little bit more complicated than that.

In particular, the other case of interest for me in my application besides squares is circles. We’ll get into those in a future post.

The double-slit experiment

In the basic version of the double-slit experiment, a laser beam illuminates a plate pierced by two parallel slits, and the light passing through the slits is observed on a screen behind the plate. The wave nature of light causes the light waves passing through the two slits to interfere, producing bright and dark bands on the screen.

The pattern formed on the screen is the (constructive/destructive) sum of two rectangular interference patterns; one per slit. The profile of each pattern happens to be a sinc function.

In this image below, I have dimmed the slits vertically using a gaussian in order to generate an image that more closely resembles classic pictures of the experiment.

This post is an extension of a post I wrote many years ago called Playing with the Fourier Transform (I).

The Fourier Transform has become an often-used tool in my toolbox because:

• it provides a means to optimize (large) convolutions via the Convolution Theorem. i.e., image filtering, certain tonemapping operations, …
• the FT corresponds to the Fraunhofer/far-field diffraction of light through an aperture, which is a key component in ArionFX and Maverick Render.

I will not be unveiling anything new here, but wanted to post visualizations I just made of some properties of the Fourier Transform that everyone in the field of Computer Graphics (or anyone dealing with the FT) should be familiar with. The videos below come straight out of Maverick’s Unit Testing system.

In the videos:

• The input signal (top-left) is in white.
• The real part of the FT (bottom-left) is in red (-) and green (+).
• The imaginary part of the FT (bottom-right) is in red (-) and green (+).
• The magnitude of the FT (top-right) is in faux color.

Changing the amplitude of the input signal

When the amplitude of the input signal is multiplied up or down, the magnitude of the FT is equally multiplied up or down.

Shifting the input signal

Shifting the input signal causes changes in phase in the (re,im) components of the FT, but the FT magnitude stays invariant.

Note that the small fluctuations in the FT magnitude below are due to anti-aliasing in the input signal and numerical drift in the computations. If sampling sizes and numerical precision were increased, the top-right quadrant would look completely frozen.

Sampling in larger/smaller buffers

Sampling the input signal and then computing the FT in larger buffers (e.g., twice as large as in the video below) produces exactly the same FT, occupying the central part of the buffer, leaving padding space towards the edges before wrapping-around.

The opposite happens if the buffers become smaller. The FT stays the same size (which will make it wrap-around earlier on all 4 edges).

Note that these videos may appear to be the same size in your browser, because this blog’s layout will dock media content to the width of the page.

Rotating the input signal

When the input signal is rotated, the FT rotates in sync.

Scaling the input signal

When the input signal is scaled up or down, the overall shape of the FT gets scaled inversely. The overall power also changes, proportionally to the scale.

FT of a polygonal shape

The Fourier Transform of an n-sided polygon casts n streaks. Because the FT is periodic, when n is even, only n/2 streaks are apparent due to overlapping pairs. When n goes to infinity, the FT becomes the FT of a circle, which is called the Airy disk.

The periodicity of the FT is visible when the streaks wrap around the edges of the image.

FT of a gaussian

While not strikingly obvious in the video below, the Fourier Transform of a 2D gaussian function is another 2D gaussian. Note how the transform has no imaginary component in this case.

[EDIT] All throughout this post I loosely used the term FT to actually mean DFT.

Youtube versions of the above videos

www.youtube.com/shorts/Y9…

www.youtube.com/shorts/gS…

www.youtube.com/shorts/hd…

www.youtube.com/shorts/JH…

www.youtube.com/shorts/aA…

In the realm of integer hashing, the Avalanche effect is the notion that if the input of a hash function is changed slightly, its output must change significantly. i.e., flipping one bit in the input must randomly flip about half the output bits.

This criterion can be hardened. Strict (full) avalanche says that whenever a single input bit is flipped, each of the output bits must change with a 50% probability.

In simpler terms this means that a good hash function is expected to break the correlation between the input and the output as much as possible. Avalanche is one possible measure of how well this property is achieved.

Implementing avalanche

Let us focus on u32->u32 hash functions.

Implementing avalanche is fairly simple. All you need to do is pass each possible u:u32 value through your hash function, and then flip each bit in u one by one, seeing what bits in hash(u) are affected. For each flip in the output you ++ the (c,r) entry in a 32x32 counters array, where (c,r) corresponds to a flip in the c-th input bit and a flip in the r-th output bit.

An exact estimation of avalanche requires that you try each of the 2^32 possible values for u once, which may take longer than reasonable. But it is generally enough to just try a large bunch of (uniformly distributed) values to get a good estimate of what avalanche will converge to.

A good hash function should accumulate to a close-to-perfectly-even counters array. i.e., if the array is normalized by the total amount of values we tested, the array will be flat 50% gray. On the other hand, poor hash functions quickly produce an evident pattern in the resulting array (see down below).

https://en.wikipedia.org/wiki/MurmurHash

Avalanche matrix for Murmur3’s finalizer

I wrote a Shadertoy for this

Over the years I have implemented plenty of stress-tests for PRNGs, QRNs, hash functions, etc… as part of Maverick Render’s MK_api Unit Testing. This time I went ahead and wrote a Shadertoy instead as an exercise:

https://www.shadertoy.com/view/NtVczW

The code allows you to try out different popular u32->u32 hash functions. You can also type in your own. In a few seconds you will see what the 32x32 avalanche array and the corresponding numerical estimation converge to.

The code can be edited to display the avalanche matrix, or the bias matrix, which is nothing but bias=abs(avalanche-.5). i.e., How much avalanche deviates from 50%. Naturally, the bias is meant to converge to 0 for a high-quality hash function.

Some results with other hash functions:

Avalanche matrix for iquint1

Avalanche matrix for Schechter and Bridson hash

Avalanche matrix for Wang’s hash

Valuable links

There is a methodology to create your own hash functions by combining certain reversible operations. I won’t get into the details here, but I will give a couple of fantastic links instead:

• The pretty amazing Hash Function prospector project, by Chris Wellons (a.k.a., skeeto):

https://github.com/skeeto/hash-prospector

skeeto found some extraordinary 32-bit and 64-bit direct/reverse hash functions himself. In particular the lowbias32 one, which comes implemented in my Shadertoy above.

• A dissertion on the search for perfect hashing, by Bob Jenkins:

http://burtleburtle.net/bob/hash/integer.html

Takeaway

Many CG use cases where what matters the most is performance can get away with cheap hash/RNG functions. On the other hand, industrial-strength rendering applications such as Maverick Render just can’t.

But there is a sweet spot between high-performance and low-bias (as proven by, for example, the Murmur3 finalizer or skeeto’s low-bias hash) where you can have a negligible bias with just a handful of arithmetic instructions.

I am super-stoked about the new tech we’ve been creating for Maverick Render recently.

Test render courtesy of Studio Podrini

Our Product Manager wouldn’t like me to spoil any surprises by revealing what new features are involved in this image, so I will leave that to the viewer’s imagination. But here’s a hint: new/updated BSDF stuff + new heavy raytracing-related stuff + even more new heavy raytracing-related stuff.

All these will see the light in Maverick Render 2022.5 (as soon as it is ready).

A “triangle sweep” can be defined as 2 triangles (A,B,C) and (D,E,F) and the 3 bilinear patches at the sides of both triangles. It’s kind of a “twisted” triangular prism.

We can define some sort of (u,v,w) barycentric coordinates here, such that each point P interior to the sweep could be written as P=(A+AD*w)*u+(B+BE*w)*v+(C+CF*w)*(1-u-v).

As soon as the sweep is “twisted” the behavior of those barycentric coords is no longer linear. i.e., linear motion in the (u,v,w) space does not translate to linear motion in 3D space and viceversa. However these coordinates can be useful for certain applications.

One way to find (u,v,w) for a given P is to solve for w first (the red plane below) and then find u and v as regular barycentrics in the triangular section. After massaging the numbers a bit, solving for w turns out to be a cubic equation, which is the wiggly orange curve below.

We are only interested in solutions where w IN [0..1].

There are many situations in Computer Graphics where you may need to store massive amounts of unit-length vectors (e.g., directions, surface normals, …). The normals AOV in a G-Buffer, incoming directions in a Photon map, or uploading a 3D model to the GPU for rendering are situations that greatly benefit from a bit-efficient encoding for normals.

Spherical coordinates encoding

The straightforward way to encode a normal is to store its (x,y,z) coordinates, which takes 3x32=96 bits in single-precision floating-point. This naive approach doesn’t take advantage of the fact that in normals or directions the vector length is irrelevant. So assuming unit length, you may use spherical coordinates (theta, phi) and encode each normal with 2 numbers instead of 3.

Since (theta, phi) have a [0..2pi] and [0..pi] range, it sounds reasonable to quantize its values using fixed-point math. This way both angles become a couple of integers of, say, 16-bit each, with their ranges turned into [0..65535]. Note that this gives twice as much resolution to phi but that’s fine. We have achieved a reasonably fine-grained 16+16=32-bit encoding. Depending on how tolerable precision losses are in your application, you may go even lower (e.g., 12+12=24-bit, or even 8+8=16-bit).

The problem with this simple approach is that spherical coordinates are not an equal-area transform, as they concentrate points near the poles.

Spherical coordinates encoding: phi gets more precision than theta, and density between the equator/poles is highly uneven.

Octahedron encoding

A smarter solution which produces a much more even distribution, while being strikingly simple and more efficient computationally (no trigonometry whatsoever) is octahedron encoding.

Octahedron encoding: The distribution is pretty even.

The idea is to project positions in the sphere (unit vectors) onto a sphere-centered octahedron. And then piece together the 8 faces of the octahedron in the unit cube. This diagram should be self-explanatory.

Reference: Octahedron environment maps

Fibonacci encoding

Recently I have been using the Fibonacci lattice here and there and had the idea to use its equal-area spherical flavor for normal encoding. After googling a bit I came across this paper, so somebody had the idea first:

Unorganized Unit Vectors Sets Quantization

Actually, pulling from this thread I also came across these Shadertoy samples by Íñigo Quilez. They visualize what happens with either method as more bits are budgeted for normals encoding:

Fibonacci: https://www.shadertoy.com/view/4t2XWK

Octahedron: https://www.shadertoy.com/view/Mtfyzl

Fibonacci encoding: Looks great, but comes at a price.

This looks elegant (pretty even!). Unfortunately the implementation to encode/decode unit-vectors with n bits is rather convoluted and involves N=2^n and 1/N. As soon as n>24 or so numerical degradation makes this solution unusable in single-precision. Also, this method is significantly slower than octahedron encoding, and while the distribution looks prettier, the difference in practice is negligible.

Basic concepts in this post, but worth refreshing.

Very often in Computer Graphics (in Monte Carlo integration in particular) you need to sample N points uniformly over the surface of a square, a disk, or a sphere.

Depending on many (many!) factors such as how critical the quality of the random distribution needs to be, performance or code simplicity considerations, availability of a high-quality randomization seed, etc… one may choose one method or another. There are plenty of options, and each comes with their own strengths and weaknesses.

In this post I will talk about:

• The Equal-Area transform to turn a pair of uniformly distributed random numbers in the unit square into uniformly distributed points over the surface of the the unit disk/sphere.
• A super-simple Fibonacci-based method to generate N points that uniformly cover the unit square (and hence the unit disk/sphere).

Equal-Area transform

Given u a uniform distribution of 2D points in the surface of the unit square, the transforms to produce uniform distributions of 2D points in the unit disk, or 3D points in the unit sphere are:

There are very clean-cut explanations for these on the always fantastic wolfram.com website.

Disk Point Picking

Sphere Point Picking

These are based on the area (disk) and solid angle (sphere) differential elements.

Fibonacci lattice

The algorithm to generate the first N points in the Fibonacci lattice goes like this:

This produces a stratified distribution of points in the unit square.

The beauty of this method is that it’s a strikingly simple couple of divs and a mod. Unlike most series generators, however, it requires that you know N beforehand. Other usual methods like the ones depicted later in this post are incremental, so each new sample is spawned next to the previous ones. The Fibonacci lattice, on the other hand, “squeezes” previous samples as new ones come in, as illustrated by the videos below.

Fibonacci lattice (square/disk/sphere)

This videos simply plug the Fibonacci lattice formula to the Equal-Area transforms above.

As a bonus, the next videos are generated like the ones above, but using a uniformly distributed QRN generator (the popular and very strong Mersenne Twister) and the van Der Corput stratified series (in particular, the Halton sequence using primes 2 and 3).

Mersenne Twister (Random)

Halton (Stratified)

Quiz: Does this loop end, and if it does, what is the value of k after the loop? (assuming IEEE-754 and usual 32-bit semantics for int/float)

Running this code in VS you get k=16777217 which equals 2^24+1.

In other words, in line with the true intent of the quiz, float can encode -exactly- (without any precission loss) all natural numbers up to 2^24 (included). Because float encodes sign in a dedicated bit, this property holds true for negative values as well. So you could hypothetically encode any [-2^24..+2^24] int in a float with a static/implicit cast.

I said hypothetically because generally speaking I would never do such thing. Carelessly casting between int and float (without an explicit floor, ceil, trunc, round…) is often a sign of lousy coding, unless performance is at stake and you know well what you are doing.

However, I came across the situation in Unity recently, where the Vertex ID node in their Shader Graph hands over SV_VertexID as a float (disappointed face :-P). I would’ve expected an uint/int output or a float that you could reinterpret-cast to retrieve the raw SV_VertexID bits with asuint(). But nope. You are handed over a float which seemingly carries (float)SV_VertexID.

One may recklessly use this float to index a StructuredBuffer ignoring the way static-casting works. This is one case where ignorance is bliss, because the float received exactly matches the original int as long as the original value is <=2^24. That is, as long as you are dealing with fewer than (roughly) 16.7 million vertices, which is usually the case.

I believe that Unity’s ShaderGraph doesn’t support int/uint as In/Out values between nodes, so I guess that the Vertex ID and Instance ID nodes just static-cast the corresponding SV_... value to a float. But it would be better (in the pedantic sense) to reinterpret the bit pattern of the raw values with asfloat and then let the user retrieve them with asint()/asuint().

reinterpret-cast between int/float in HLSL

asint()

asuint()

asfloat()

A (loose) explanation of the 2^24 limit

This is the IEEE-754 standard for floating-point numbers, as described in Wikipedia:

The value of a float-encoded number is reconstructed as (-1)^S * M * 2^E.

• S is the 1-bit sign.
• M is the 23-bit mantissa, interpreted as 1.xxxxx (in binary).
• E is the 8-bit exponent, used as E-127 where 127 is often called bias.

e.g., a power-of-2 integer number like 4 would be encoded as:

• M=1.00 which is the binary representation of 4, with the point at the left-most 1.
• E=2 (+bias).

The restored number is 1.00 shl 2=4.

We should be able to do the same for all power-of-2 integers until we max out E.

For non-power-of-2 integers the process is similar. e.g., number 5:

• M=1.01...00.
• E=2 (+bias).

The restored number is now 1.01 shl 2=5.

This works the same for all integer numbers as long as M can hold the raw binary representation of the number. The tallest binary number that M can hold is 23 consecutive 1s. That is: 1.11...11 (24x1s in total). With E=23 (+bias) this equals 2^24-1.

The next integer 2^24 would be encoded as 1.00...00 (24x0s clamped to 23, but the trailing 0s are meaningless here). With E=24 (+bias) this equals 2^24 (the answer provided above!!).

But the next integer 1.{23x0s.1} can’t be encoded in a 23-bit M. So from 2^24 onwards, there is necessarily a loss. Some integers beyond 2^24 (such as powers-of-2) may be luckily encoded exactly by float. But not all consecutive numbers will. Actually, 2^24+1 is the first integer that won’t.

Whoa! As always, apologies for any mistakes in any of my posts.

[EDIT] The same reasoning can be applied to double, where the mantisa M is 52-bit long.

Hence double can encode exactly all integers in the range [-2^53..+2^53].

The technique described here is very basic, but useful nonetheless.

Let’s make the following broad assumptions:

1. No interactions between surfaces (i.e., no GI, no occlusion, etc…).
2. The only light source is an IBL (environment) map placed at infinity which orientation is anchored to the camera.
3. Each surface point P is made of a perfectly rough white material which simply “collects” the light that arrives to P from above the hemiplane defined by (P,N).

Under such assumptions a ball floating in the void looks like this image rendered in Maverick Studio with a procedural sphere primitive and a standard material:

The illumination collected by each point P and radiated back to the camera (+ tonemapping) happens to be, exactly, one particular pixel in the rendered ball as depicted in the hand-drawn diagram above. Note that due to the constraints set, this color depends exclusively on the normal N at P but not on P itself. In other words, the rendered ball effectively precomputes what any surface point P sees through its normal N.

The goal in this post is to benefit from this knowledge in a rasterization viewport. So let’s jump now to the world of pixel shaders. In our vertex shader, let’s transform the object’s normals N to camera space Nc and let’s drop the resulting Z coordinate. (Nc.x,Nc.y) are the geometry normals projected to screen space. Let’s normalize these: Nc'=normalize(Nc.x,Nc.y)

Let’s upload the ball render as a regular 2D texture and let’s define the following UV space:

Nc' is normalized, so the UVs generated will cover the interior of the ball only. Normals at grazing angles will be near the perimeter of the texture’s ball.

In our fragment shader, let’s simply output the color found in the ball image at the UVs computed as described. Here’s the result:

Of course, physically-based materials are more complicated than this. In general, the amount of light received by the surface from the (P,N) hemiplane through each possible incoming direction and then bounced back through each possible outgoing direction depends on both directions, and also on the properties of the material. This is basically the definition of a BRDF, BTW.

But because we set the constraint that the lights are not moving (i.e., the IBL stays fixed with respect to the camera) the incoming and outgoing directions and hence the behavior of the BRDF ends up being a function of N and nothing else. So this method still works no matter what material we render the ball with (well… excluding refractions).

Here goes another example with the same IBL but adding some specularity to the material:

Of course this could be pushed further with some SSAO or any other technique that approximates self-shadowing. But this is a very compact and cheap (memory/code/effort) way to display models in an easy-to-author appealing way.

Some final notes:

• If you want to be accurate, the ball should be rendered with an orthographic camera (i.e., not with a perspective camera).
• For blurry/glossy or diffuse reflections, you can get away with a low-res ball render. However, make sure that the UVs that you generate don’t step on the antialias pixels at the boundary of the rendered ball. Otherwise shading will look b0rked at grazing angles.
• This technique can be extended by using 2 ball images: one with the diffuse component of the material, and another with the specular component. This way the fragment shader could modulate the diffuse part and the specular part by separate colors that could be passed as inputs to the shader. Keep in mind that this would require that the ball images and the colors are passed in linear (instead of sRGB) space and that the resulting color is sent back to sRGB.

I’m currently swimming in an ocean of NURBS with their trimming loops. This is experimental work for NURBS rendering in Maverick Render.

[EDIT] Throwing some GLSL vert/frag shaders in to the mix for normal/tangency inspection. 2 days in and this is still WIP.

A couple of weekends ago I was tinkering with fluid dynamics and implemented a few compute shaders in Unity. Among the ones I implemented, my favorite is the Lattice-Boltzmann Method featured here.

This early implementation allows to draw SDF entities in the simulation space behaving as blockers, inflows or outflows. I intend to expand on the features, but most likely at a slow pace, because this is a no-rush side-project. But I definitely wish to post now and then about my progress.

I also hope to find some spare time soon to write a quick article on how to implement an LBM cellular automata from a light-on-maths perspective.

Some pet project I am making some progress on. More info soon. Definitely. Maybe.

[EDIT] I had to install a micro.blog plug-in so Twitter cards display the main post image when cross-posting to Twitter. Apologies for the multiple test posts.

[EDIT] Here’s a hint. It’s a cutsie CSG-based Solid modeling experiment in Unity Editor.

Recently I have been revisiting in my spare time one of my favorite subjects: topology. I will probably be posting some stuff about topological operators, subdivision surfaces, DCEL structures, irregular tilings and such things in the near future.

Below: A little demo I’ve written that supports an extended flavor of Conway’s notation for polyhedra: adotaocD.

[EDIT] I paused this pet project for the time being… because the 2022.2 release of Maverick has virtually blown all my productivity hours away. But I plan to write some entries on topological operators at some point further into this year… Hopefully…

These are some very old MLT tests that I found on my hard drive today.

Metropolis Light Transport (MLT) is a global illumination variant of the Metropolis-Hastings algorithm for Markov-Chain Monte Carlo (MCMC).

In a nutshell, MCMC algorithms explore the neighborhood of a Monte Carlo sample by jittering (mutating) said sample. A metric is then used to define how good a sample is, so the next mutation is accepted or rejected based on how good it is compared to the current sample. This leads to a random-walk in the space of samples which tends to crawl up in the increasing direction of the metric used.

This method biases sampling towards successful samples. So a bit of statistics juggling is required to counter that bias. The link below is a seminal paper that beautifully describes the nuts and bolts of this process. The idea presented there is to jitter samples, not in sample space directly, but in quasi-random number tuples space instead:

Simple and Robust Mutation Strategy for Metropolis Light Transport Algorithm

In the particular case of MLT, samples are random light paths, and the metric used is the amount of light energy transported by a path. This can be implemented on top of regular Path Tracing. Arion was the first (?) commercial render engine that did this on the GPU back in 2012 or so.

MLT has fallen out of favor over the years (path guiding is all the rage these days). But MLT proved to be a competent universal solution to global illumination, being capable of boosting path tracing so it could efficiently solve even the hardest cases, such as refractive caustics.

The beauty of MLT is that it requires 0 configuration from the user and does not need any additional memory for data structures. The main drawbacks are the ugly splotchy and non-uniform look of noise, as compared to regular path tracing, and the inability to guarantee noise stability across animation frames.

Experiments in 2D

Before I implemented MLT in Arion, I did some MCMC simulations with 2D sampling. In the videos below I used the (x,y) coords of pixels as samples, and the grayscale luminance Y of the image as the metric. The goal of the mutations here is to reconstruct the image by sampling harder in the directions of higher luminance.

MCMC is prone to getting stuck for long amounts of time in local maxima. The practical solution proposed in the above paper is to introduce a plarge probability that kicks the sampling scheme and sends the next sample to a purely random position in the sample space. The videos below visualize this very well I think.

Mutating in QRN tuple space without plarge - Watch on Youtube

Mutating in QRN tuple space with plarge - Watch on Youtube

MLT reconstructing a grayscale image - Watch on Youtube

Implementing MLT successfully in our render engine did require some specialization in our sampling routines. I won’t get into the details, but basically, since mutations happen in the random tuples that generate the paths, you expect continuous mutations in the tuples to produce continuous changes in the generated paths. So all samplers involved must avoid discontinuities in the [QRN<->path] bijection.

Back in time

Some MLT-heavy images rendered in Arion back then:

[EDIT] This post was migrated from my blog from 2011…

This is a classic subject in numerical (Monte Carlo) integration.

Uniform 2D distribution vs. Halton series for the first 2 dimensions

To the left: 32768 points in a 512×512 image using a uniform random number generator (Mersenne Twister). To the right, the first 32768 pairs in the Halton series, using dimensions #0 and #1. Click to enlarge!

[EDIT] This post was migrated from my blog from 2011…

[EDIT] This post was extended on 2022.09.

The beauty and power of the Fourier Transform never cease to amaze me. And since several effects in ArionFX are based on it, I have had to play with it a lot in recent times.

As explained in a previous post, diffraction patterns (e.g., the glare simulated in ArionFX for Photoshop) come from the Fourier Transform of the lens aperture. I will use the FT of an aperture mask for visualization in this post.

I will use pow-2 square sizes (my FT implementation is an FFT). Let’s start by this Aperture x Obstacle map output directly from ArionFX for Photoshop v3.5.0.

Aperture x Obstacle mask, rasterized @ 512×512

The Fourier Transform of this image, rasterized at that size in said 512px buffer, is the following.

FT of the mask above

The faux-color is done on the magnitude of each complex number in the FT. All the FT images in this post are normalized equally, and offset to look “centered” around the mid-pixel.

Such diffraction patterns and some heavy convolution magic are what ArionFX uses to compute glare on HDR images:

Resulting glare in ArionFX

Now, let’s focus on what happens to the FT (frequency space) when one does certain operations on the source data (image space). Or, in this exemplification: what happens to the diffraction pattern when one plays with the rasterized aperture mask.

Note that we’re speaking of the Discrete Fourier Transform, so sampling (rasterization, pixelization, anti-aliasing, …) issues are mostly ignored.

Rotation about the center

A rotation of the source buffer about its center doesn’t change the frequencies present in the data; only their orientation. So a rotation in the source data rotates the FT in the exact same way.

As we will see next, this property holds true regardless of the center of rotation, because the magnitude of the FT is invariant with respect to translations.

Rotation about the center

Translation (with wrap-around)

Frequencies arise from the relative position of values in the data field, and not from their absolute position in said field. For this reason, shifting (wrap-around included) the source data does not affect the magnitude of the corresponding Fourier Transform in any way.

Invariance to translation

Let’s recall that the idea behind the FT is that “any periodic function can be rewritten as a weighted sum of sines and cosines of different frequencies”. Periodic being the keyword there.

Repetition (tiling)

Tiling the data buffer NxM times (e.g., 2×2 in the example below) produces the same FT, but with frequencies “exploded” every NxM cells, canceling out everywhere else.

This is because no new frequencies are introduced, since we are transforming the same source data. However, the source data is NxM times smaller proportional to the data buffer size (i.e., the frequencies become NxM times higher).

Exploded frequencies on tiling

Data scaling

Normalization and sampling issues aside, scaling the data within the source buffer scales the FT inversely.

This is because encoding smaller data produces higher fundamental frequencies, while encoding a larger version of the same data produces lower fundamental frequencies.

Inverse effect on scaling

In the particular case of glare (e.g., ArionFX) this means that the diffraction pattern becomes blurrier if the iris is sampled small. Or, in other words, for a given iris, the sharpest diffraction pattern possible is achieved when the iris is sampled as large as the data buffer itself.

[EDIT] This post was migrated from my blog from 2011…

Here’s what we could call a mean pyramid of the 512×512 Lena image. i.e., a sequence of progressive 1:2 downscalings, where each pixel in the i-th downscaled level is the average of the corresponding 4 pixels in the previous level. For people familiar with OpenGL and such, this is what happens when the mipmaps for a texture map are computed:

The 9+1 mipmap levels in the 512×512 Lena image

There are many smart, efficient, and relatively simple algorithms based on multi-resolution image pyramids. Modern GPUs can deal with upscaling and downscaling of RGB/A images at the speed of light, so usually such algorithms can be implemented at interactive or even real-time framerates.

Here is the result of upscaling all the mip levels of the Lena image by doing progressive 2:1 bi-linear upscalings until the original resolution is reached:

Upscaled mipmap levels

Note the characteristic “blocky” (bi-linear) appearance, specially evident in the images of the second row.

Lately, I have been doing some experimentation with pixel extrapolation algorithms that “restore” the missing pixels in an incomplete image for a 2D DOF filter based on the Depth AOV.

My pixel extrapolation algorithm works in 2 stages:

1. The first stage (analysis) prepares the mean pyramid of the source image by doing progressive 1:2 downscalings. Only the meaningful (not-a-hole) pixels in each 2x2 packet are averaged down. If a 2x2 packet does not have any meaningful pixels, a hole is passed to the next (lower) level.
2. The second stage (synthesis) starts at the smallest level and goes up, leaving meaningful pixels intact, while replacing holes by upscaled data from the previous (lower) level.

Mean pyramid (with holes)

Note that the analysis stage can stop as soon as a mip level doesn’t have any holes.

Here is the full algorithm, successfully filling the missing pixels in the image.

Mean pyramid (filled)

Conclusions

This algorithm can be implemented in an extremely efficient fashion on the GPU, and allows for fantastic parallelization on the CPU as well. The locality of color/intensity is preserved reasonably well, although holes become smears of color “too easily”.

A small (classic) inconvenience is that the source image must be pre-padded to a power-of-2 size for the progressive 1:2 downscalings to be well defined. I picked an image that is a power-of-2 in size already in the above examples.

TL;DR: Given the ease of implementation and the extreme potential in terms of efficiency, the results are quite decent for many applications.

[EDIT] This post was migrated from my blog from 2011…

The Halton sequence, which is one of my favourite algorithms ever, can be used for efficient stratified multi-dimensional sampling. Some references:

• Van der Corput sequence
• Halton sequence

It is possible to do stratified sampling of hyper-points in the s-dimensional unit hyper-cube by picking one consecutive dimension of the Halton series for each component. A convenient way to do so is to use the first s prime numbers as the basis for each Halton sequence.

It is well-known, however, that while this approach works great for low dimensions, high dimensions often exhibit a great degree of undesired correlation. The following image displays a grid where each cell combines two components of the 32-dimensional Halton hyper-cube.

Raw 32-dimensional Halton hyper-cube

One can easily spot how some pairs exhibit an obvious pattern, while others fill their corresponding 2D area very densely. This happens more aggressively for higher dimensions (i.e., to the right/bottom in the image) and for pairs formed with close components (i.e., near the diagonal in the image). Click to enlarge!

A successful approach to dissolve this problem without losing the good properties of stratification is to do “random digit scrambling” (a.k.a., rds). During the construction of a Halton number, digits in the range [0..base[ are combined. Given a Pseudo-Random Permutation of length=base, all that one must do is use PRP(digit) instead of digit directly. This somewhat shuffles Halton pairs in rows and columns in a strong way so the correlation disappears. However, since the PRP is a bijection, the good properties of stratification are generally preserved.

[EDIT] While this is true, it has been proven that Owen-Scrambling is a superior method.

How to build a strong and efficient randomised PRP of an arbitrary length is a very interesting subject which details I won’t get into here. You can build your own carefully combining a number of reversible hashing operations.

Here’s the scrambling strategy in action:

Scrambled 32-dimensional Halton hyper-cube

Now all the blocks in the grid look equally dense. Magic!

As long as one picks a good seedable PRP, it is possible to generate any number of different samplings, all with the same good properties.

[EDIT] These are the guidelines of the QRN system that Arion used for many years.