KABOOM! in 110 lines of Futhark

This is a port/theft of tinykaboom from C++ to Futhark. It was done for the following purposes:

• To see how easy it is to port such straightforward graphics-ish C++ to Futhark (very easy).

• To see how fast it would run (see below).

• If fast enough, to render the explosion in real time.

It looks like this (the actual rendering is much faster; I think browsers don't like high-framerate GIFs):

module tk = import "tinykaboom"

let step = tk.main

> :video (step 640i64 480i64 4.1e-2f32, 0.0f32, 300i64)

Details

I tried to match the coding style closely, even in the naming and order of functions. One of the notable differences is that the C++ version is able to use an ad-hoc polymorphic template function, lerp, for linear interpolation of both vectors and scalars. For the Futhark version, two functions are needed. This could have been written as a single parameterised higher-order function, but I don't think the complexity is worth it here.

I have added a time parameter to some functions, which is used to vary the radius of the sphere, in order to give it an expanding feel. This is used in the signed_distance function.

I think I do the colour computation slightly differently - my darks are not really very dark.

I depend on two small Futhark packages (vector) and (matte) for vectors and ARGB colouring respectively. If I were enough of a purist, I'd inline the necessary parts. The 110 lines mentioned in the title do not include these packages.

Performance

The runtime when rendering a 640x480 frame is as follows:

• tinykaboom.cpp compiled with -O3 -fopenmp and running on a Ryzen 1700X: 1.3s

• tinykaboom.fut compiled with futhark opencl and running on a Vega 64: 0.012s

• tinykaboom.fut compiled with futhark multicore and running on a Ryzen 1700X: 0.38s

• tinykaboom.fut compiled with futhark cuda and running on an RTX 2080 Ti: 0.003s

It runs slightly slower than I would like, especially on the Vega 64, but that's probably just the nature of the beast. I inspected the generated code and could not find any obvious smoking guns, except for the kernel being fairly large. It's good enough to be real-time in a small window.

The future

It would be cool if the explosion could be rotated a bit, or maybe not be the same for every run.