Releases: Enigmatisms/cuda-pt
Parallel SBVH acceleration structure building
Credit to @AdamYuan, who coded PSBVH in his own repo (VkAdypt) and took time to migrate some of his effort to this repo.
PR #18 to #20 implements a better parallel construction procedure for SBVH. Originally, for the kitchen scene, OpenMP accelerated within-node parallelism needs 15s to build the SBVH, while the current implementation only requires less than 6s.
Contributors
Assets 2
v1.10.0: SBVH
This release implements the idea from paper 'Spatial Splits in Bounding Volume Hierarchies' by Martin Stich, et al:
- Primitive reference based BVH building (my old implementation is reference free). Also, the code is completely compatible with my GPU stackless traversal (but the design is indeed painful).
- Sutherland–Hodgman algorithm + line drawing based chopped binning.
- SAH-based spatial split and object split decisions.
- Reference unsplitting mentioned in the paper.
- A simple try for within-node parallel chopped binning.
The average performance is increased by 30%, with SAH cost significantly reduced (for example, in the Kitchen scene, 15.3-->11.14). Yet the current SBVH construction is still pretty slow. I have no intention to further accelerate it. Here are some results (BVH cost visualization):
| setting\scene | Kitchen | Rich Cars |
|---|---|---|
| BVH | ||
| SBVH |
Also, several other features / modifications have been made:
- Implements a simple
accel-onlymode, which only constructs acceleration structure and does no rendering. This is for future (and current) acceleration development purpose. - Improved BVH cost visualizing rendering mode. Now we can view the
traverse cost,intersection costandfull costin three different channels and select one of them by ImGui porting. - Adds acceleration structure evaluation metrics (see
bvh_opt.cufor more details). - Restructure BVH and SBVH by employing a more object-oriented code style instead of the old function oriented way.
v1.9.1
What's Changed
- [Code Style] Better code style via pre-commit by @Enigmatisms in #15
Full Changelog: v1.9.0...v1.9.1
Contributors
Assets 2
Volumetric Path Tracer
What's Changed
- Volumetric Path Tracer (megakernel) and Infrastructure for Online Neural Path Guiding (wavefront). See the pull request #13 by @Enigmatisms for more details:
video-med-quality.mp4
Full Changelog: v1.8.1...v1.9.0
Contributors
Assets 2
v1.8.1
What's Changed
- Accelerated WFPT implementation. by @Enigmatisms in #11
Full Changelog: v1.8...v1.8.1
Contributors
-
8000
-
Assets 2
v1.8: Python API and Distributed Parallel Rendering
What's Changed (Pull 10#)
Exported python API via nanobind:
PythonRendererclass is defined and can be imported in Python when compiled.PythonRenderersupports exporting CUDA torch.Tensor (output buffer and variance buffer), as well as the running-mean of frame time.- The structure of the CMake project is utterly changed. The core functions and classes are compiled to a static lib (RenderCore) now.
Add scripts that supports distributed parallel rendering:
- Based on PyTorch DDP module.
pyrender/ddp_render.pysupports rendering with multiple GPUs (on a same device). Multiple devices can be trivially supported, but it is not implemented yet. - The script writes output buffer (RGB image), variance buffer (single channel image), variance curves (Welford's algorithm), frame time (different processes) and average frame time curves to a tensorboard log. The variance estimation buffer can be used for adaptive sampling in the later updates.
Tested the compilation on Ubuntu 22.04, fixed several bugs that are exclusive on Linux.
| Variance Image | Variance Curves | Frame Time (ms) Curves |
|---|---|---|
Full Changelog: v1.7...v1.8
v1.7
What's Changed
- Major updates with more renderers, better visualization and higher performance by @Enigmatisms in #9
| Vader Scene | Whiskey Scene |
|---|---|
Full Changelog: v1.6...v1.7
Contributors
Assets 2
v1.6
What's Changed
- Major update: Pitched array GPU textures. by @Enigmatisms in #8
Full Changelog: v1.5...v1.6
Contributors
Assets 2
v1.5
What's Changed
- Plastic Transmission (thin) implemented. The new sports-car scene with around 3M triangles is rendered. Check the rendered image in the updated README.
- Upgrade the GUI. Now you can change the emitter / material on-the-fly. This is laborious for GPU (why would I write GPU inheritance in the first place). Check the updated README for video.
- Fixed several bugs.
v1.4
Full Changelog: v1.3.3...v1.4