cryvosh/temporalsetinversion

Reference implementation for "Temporal Set Inversion for Animated Implicits" (SIGGRAPH 2023)

Temporal Set Inversion for Animated Implicits

Project page with paper pdf and supplemental video can be found here.

Note that this is all research-quality code. The interesting bits are contained entirely within /src/compute.wgsl.

Setup instructions

1. Download and install Rust
2. Clone this repository
3. Run cargo run --release -- --Awarnings in the root directory

Controls

1. Press r to hot-reload /src/compute.wgsl
2. Camera controls are w, a, s, d + mouse to look
3. Scroll up or down to change movement speed
4. If the camera controls feel wrong, disable the scene-dependant view matrix modifications in fn rt_main in /src/compute.wgsl

Shader Parameters

1. Ctrl+f for "PARAMETERS" in /src/compute.wgsl, here you will find the following:
2. localsubdiv controls the octree resolution, e.g., 8 -> (2^8, or 256)^3 and 10 -> (2^10, or 1024)^3
3. delta controls the temporal error tolerance, e.g., 0.01 (depends on the Lipschitz bounds of f)
4. pointsample enables world-space point-sampling within indeterminate leaves
5. fastenabled enables the fast-evaluation optimization
6. lipschitz_corrected enables Lipschitz-correction on delta
7. swept_volume enables swept-volume rendering
8. normalvis sets the surface color to the world-space normals
9. cinematic enables naive pathtracing with totalsamples samples per pixel
10. fast_normals enables per-pixel normals via finite-difference
11. scene selects which scene's normals (if fast_normals is true) / lighting / camera settings to use

Other parameters

1. To enable fullscreen, set shouldbefullscreen in /src/main.rs to true
2. To save frames to disk, set save_frames_to_disk in /src/main.rs to true
3. To save timing data to disk, set save_timing_to_disk in /src/main.rs to true

To switch scenes

1. Make file_contents in /src/IR.rs read from scenes/something_else.ll
2. Set scene in /src/compute.wgsl according to the comment next to it

To add a new scene

1. Create new file /scenes/example.cpp and paste in your GLSL code
2. The field function must be named f, mapping 4 floats x,y,z,t to a single float
3. Paste this at the top of /scenes/example.cpp

   #include <math.h>
   #include <cstring>
   #include "include/swizzle/setup.h"
   #include "overloads.h"
   

4. Install Clang and run ./compile.sh example to generate example.ll
5. Fiddle with the code until example.ll contains no branches or casts in the field function
6. In /src/IR.rs point the constructor's file contents to scenes/example.ll
7. Set scene in /src/compute.wgsl to some new value
8. Make sure fast_normals in /src/compute.wgsl is set to false to get per-voxel normals. If you want per-pixel normals, keep reading

To enable fast per-pixel normals on the new scene

1. Paste your GLSL code into a temp file called example.frag

2. Paste this at the top of example.frag

   layout(location = 0) out vec4 color;
   void main() {
   	color = vec4(0.0);
   }
   

3. Run naga example.frag example.wgsl in the CLI

4. Paste the contents of example.wgsl at the top of /src/compute.wgsl, ignoring the FragmentOutput and color lines (top of the file) as well as fn main() (bottom of file)

5. Rename fn f to something else, e.g., fn example

6. Add a new case to fn finite_diff_normal in /src/compute.wgsl to call fn example if scene is set appropriately

7. Make sure scene in /src/compute.wgsl is set to your new value

8. Set fastenabled in /src/compute.wgsl to true so that fn finite_diff_normals is used

I've tested this code on Windows 10, Vulkan backend, with an RTX 3080 / i7 3930k. It should in theory work on other operating systems / GPUs and even the web thanks to wgpu. If you need help setting it up, please open an issue.