feature: 3D gaussian splatting viewer

[altunenes]

cargo run --release --example gaussian3d
I tried it with Apple's mlsharp ply export.
30fps: MacBook M3 16GB.

gpu radix sort ported from: wgpu_sort (Copyright (c) 2024, Simon Niedermayr, Josef Stumpfegger: BSD 2-Clause License). I must admit that I spent some time researching whether there are more modern sorting approaches for 3dgs. However, I think this is still the best approach for now.

ply reader: a new dep but very very tiny: https://github.com/bourumir-wyngs/ply-rs so

note that for now, its compute + frag approach. When I have time, I want to try to convert this entirely into a compute pipeline. However, I realized that I need to work harder on this because I had some performance issues.

As I did previously in gaussian.wgsl, I tried a purely tiled approach for creating entire compute pipeline. However, given the immense size of the 3D gaussians, this approach proved to be too naive and ineffective result: 8-10 fps and lower quality. but I'll still push for it. I'll take that as a “starting point” for now.

NOTE: I was able to get my ply file in minutes using the following simple script via Google Colab:
An even crazier idea would be to export this model to ONNX, make the necessary adjustments, and serve it in real time :-P
but considering the complexity, this would be a huge work.

import os
import shutil
from google.colab import drive

# cfg
drive_folder = '/content/drive/My Drive/your_folder/'
image_filename = 'your_image.png'

# env
if not os.path.exists('/content/drive'):
    drive.mount('/content/drive')

repo_dir = '/content/ml-sharp'
if not os.path.exists(repo_dir):
    !git clone https://github.com/apple/ml-sharp.git {repo_dir}

os.chdir(repo_dir)
!pip install -q -r requirements.txt
!pip install -q .

# Run Prediction
input_path = os.path.join(drive_folder, image_filename)
output_dir = '/content/inference_output'

!sharp predict -i "{input_path}" -o "{output_dir}" --device cuda

# Save Result to Drive
ply_filename = os.path.splitext(image_filename)[0] + '.ply'
shutil.copy(os.path.join(output_dir, ply_filename), os.path.join(drive_folder, ply_filename))

[altunenes]

note: With WGPU 28, we can also try this with mesh shaders. However, this is just an idea for now...

#115