I'm Morgan McGuire (@CasualEffects). I've been working on computer graphics and games for 20 years at great places including NVIDIA, Williams College, Brown University, and Activision.

See my home page for a full index of my blog posts, books, research, and projects.

Saturday, August 3, 2013

Starfield Shader

I'm preparing assets for an upcoming mini-game jam/hackathon.  My current plan is to make a spaceship video game inspired by Galaxy Trucker, in which you build and then fly a ship.  This post is about the starfield shader that I created by modifying Kali's amazing Star Nest shader on Shadertoy to fit into the game framework.

The starfield
The starfield renders in 3.7ms at 1080p on GeForce 650M (the Kepler GPU in a 2012 Macbook Pro) on Windows 7 64-bit, including upsample time.  It is procedural, so it has effectively infinite extent and fine detail down to the pixel level, and has parallax as it scrolls.

I was only willing to spend 25% of the frame time on the background. To keep the performance high, I render downsampled by 2.5x in each direction to an RGB8 texture and then bilinearly interpolate up to full screen resolution.

Here's the GLSL shader:

// -*- c++ -*-
// \file starfield.pix
// \author Morgan McGuire
// \cite Based on Star Nest by Kali 
// https://www.shadertoy.com/view/4dfGDM
// That shader and this one are open source under the MIT license
// Assumes an sRGB target (i.e., the output is already encoded to gamma 2.1)
#version 120 or 150 compatibility or 420 compatibility
#include <compatibility.glsl>

// viewport resolution (in pixels)
uniform float2    resolution;
uniform float2    invResolution;

// In the noise-function space. xy corresponds to screen-space XY
uniform float3    origin;

uniform mat2      rotate;

uniform sampler2D oldImage;

#define iterations 17

#define volsteps 8

#define sparsity 0.5  // .4 to .5 (sparse)
#define stepsize 0.2

#expect zoom 
#define frequencyVariation   1.3 // 0.5 to 2.0

#define brightness 0.0018
#define distfading 0.6800

void main(void) {    
    float2 uv = gl_FragCoord.xy * invResolution - 0.5;
    uv.y *= resolution.y * invResolution.x;

    float3 dir = float3(uv * zoom, 1.0);
    dir.xz *= rotate;

    float s = 0.1, fade = 0.01;
    gl_FragColor.rgb = float3(0.0);
    for (int r = 0; r < volsteps; ++r) {
        float3 p = origin + dir * (s * 0.5);
        p = abs(float3(frequencyVariation) - mod(p, float3(frequencyVariation * 2.0)));

        float prevlen = 0.0, a = 0.0;
        for (int i = 0; i < iterations; ++i) {
            p = abs(p);
            p = p * (1.0 / dot(p, p)) + (-sparsity); // the magic formula            
            float len = length(p);
            a += abs(len - prevlen); // absolute sum of average change
            prevlen = len;
        a *= a * a; // add contrast
        // coloring based on distance        
        gl_FragColor.rgb += (float3(s, s*s, s*s*s) * a * brightness + 1.0) * fade;
        fade *= distfading; // distance fading
        s += stepsize;
    gl_FragColor.rgb = min(gl_FragColor.rgb, float3(1.2));

    // Detect and suppress flickering single pixels (ignoring the huge gradients that we encounter inside bright areas)
    float intensity = min(gl_FragColor.r + gl_FragColor.g + gl_FragColor.b, 0.7);

    int2 sgn = (int2(gl_FragCoord.xy) & 1) * 2 - 1;
    float2 gradient = float2(dFdx(intensity) * sgn.x, dFdy(intensity) * sgn.y);
    float cutoff = max(max(gradient.x, gradient.y) - 0.1, 0.0);
    gl_FragColor.rgb *= max(1.0 - cutoff * 6.0, 0.3);

    // Motion blur; increases temporal coherence of undersampled flickering stars
    // and provides temporal filtering under true motion.  
    float3 oldValue = texelFetch(oldImage, int2(gl_FragCoord.xy), 0).rgb;
    gl_FragColor.rgb = lerp(oldValue - vec3(0.004), gl_FragColor.rgb, 0.5);
    gl_FragColor.a = 1.0;

The preamble uses some G3D-specific shader preprocessing. #expect declares a macro argument, G3D allows multiple version numbers and promotes the highest one to the first line, and implements #include.  The compatibility.glsl file smoothes over differences between GLSL versions and between GLSL and HLSL.

At its heart is the "KaliSet" iterated function p := | | / | |2 - k.  I tuned the constants for the specific look I was targeting and for performance.  It is of course very performance sensitive to the number of iterations of the inner loop and the number of volumetric steps.  I also extracted some of the expressions into uniforms in a way that the Shadertoy framework wouldn't have allowed for the original.  I tried using the max component or the average component in place of the expensive length() call, but this creates a 3-fold symmetry on some "suns" that is undesirable.

In my game, the background will move slowly as the ship flies. Very bright single pixels flicker in and out of existence as they are undersampled with the unmodified shader, so I suppressed these cheaply using derivative instructions and blending in against the previous frame.  The previous frame blending is primarily there to create motion blur and soften the background under motion to keep attention on the shops. I also wiggle the screen by a random offset of up to 1/3 pixel when the camera is still.  This makes stars appear to twinkle and closes the visual difference between the slightly-incoherent rendering under motion and when still.


Morgan McGuire is a professor of Computer Science at Williams College and a professional game developer. He is the author of The Graphics Codex, an essential reference for computer graphics that runs on iPhone, iPad, and iPod Touch.