GPU computation

GPU computation

render/computation.mojo steps float images with GLSL fragment shaders. It is three.js's GPUComputationRenderer from examples/jsm/misc/. A simulation keeps its state in the images, for example the positions and velocities of a flock, and runs one step a frame.

examples/ripples.mojo draws this picture.

var computation = GPUComputationRenderer(64, 64)
var position = computation.add_variable("position", MOVE, start^)
var velocity = computation.add_variable("velocity", STEER, computation.create_texture())
computation.set_variable_dependencies(position, [position, velocity])
computation.set_variable_dependencies(velocity, [position, velocity])
computation.init()
# For each frame:
computation.compute()
var map = assets.textures.add(computation.current_texture(position))

Variables

A variable is an image of size_x by size_y texels, four floats for each texel, and the shader that computes its next state. add_variable(name, shader, initial) adds one and returns its place. create_texture() returns an image of zeros for a start.

set_variable_dependencies(variable, dependencies) says which images the shader reads. A variable that reads its own last image lists itself. init() declares a sampler2D for each dependency, named after it, and defines resolution as the image size. Then it compiles each shader with the GLSL subset.

Set a shader's other uniforms on programs[variable] after init, for example programs[velocity].set_uniform("delta", 0.016).

A shader

A shader finds its texel as three.js's shaders do, and writes gl_FragColor:

void main() {
    vec2 uv = gl_FragCoord.xy / resolution.xy;
    vec4 p = texture2D( position, uv );
    vec4 v = texture2D( velocity, uv );
    gl_FragColor = vec4( p.xyz + v.xyz * 0.016, p.w );
}

All four numbers are kept as they are. They are not clamped and not premultiplied. A read gets the texel that its coordinate falls in, as three.js's NearestFilter does. The variable's wrap_s and wrap_t say what a coordinate past an edge reads. They are CLAMP unless set. A texel that the shader throws away with discard keeps its value.

A step

compute() runs each variable's shader once, at every texel. Each variable reads the images as they were before the step. So a variable that runs after another in the same step reads the other's old image, as in three.js.

current_image(variable) returns an image now. alternate_image(variable) returns the image before the last step. The images lie row by row from the bottom, so texel (x, y) starts at (y * size_x + x) * 4. current_texture(variable) returns a float texture for a material to read.

On the GPU

GpuComputation.compute(computation) in render/gpu.mojo runs a step on the device, one launch for each variable. The host and the device run the same function at each texel, compute_texel, so the images agree to the float. tests/test_gpu.mojo checks this.

Where this port differs

• The images are float lists. three.js keeps them in render targets.
• Only FloatType and NearestFilter are ported. setDataType and a linear filter are not.
• init raises when a shader is refused. three.js returns an error string for a missing dependency. Here a dependency that is not a variable raises in set_variable_dependencies.
• gl_FragCoord.w is one. three.js's is one over the clip-space w, which is also one on its full-screen quad.