A Voxel Rendering Pipeline in CUDA for Real-time Indirect Illumination

See our results on YouTube

Dave Kotfis and Jiawei Wang CIS 565 Final Project – Fall 2014

Goal - Implement a voxel rendering pipeline in CUDA capable of real-time global illumination.

Approach

• Utilize open source VoxelPipe library from NVidia to get up and running quickly.
• Reuse our CUDA Rasterizer to visualize the result and check our progress.
• Build a Sparse Voxel Octree representation on the GPU.
• Implement octree cone tracing for indirect illumination by Crassin et. al.
• Compare performance against our CUDA Pathtracer.

![Project Plan] (images/project_plan.png)

##Phase I Progress:

• Completed OpenGL Equivalent Rasterization Pipeline with quick swap to CUDA Rasterizer.
• Integrated NVidia VoxelPipe
• Added conversion kernels that can create a cube-ized mesh to render through the standard rasterizers. Works with both OpenGL and CUDA rasterization.

OpenGL Stanford Dragon: ![Voxel Dragon GL] (images/stanford_dragon_voxelized_gl.png)

[OpenGL Pipeline] (https://www.youtube.com/watch?v=ynhujtzyh6s&list=UUpiGNsmZZFAvKHIT0y5Q32Q)

[CUDA Pipeline] (https://www.youtube.com/watch?v=cD8gOQchpoM&feature=youtu.be)

##Phase II Progress:

• Port to Linux
• Texture Mapping (in progress)
• Increased efficiency for 1024^3 voxels
• GPU SVO (in progress)

1024^3 Voxels Stanford Dragon

##Phase III Progress

• Texture Mapping in VoxelPipe is Complete
• SVO Cube Extraction is Complete
• Filling values to interior SVO nodes is Complete
• Added the ability to extract cubes from SVO at arbitrary resolution

This shows the stanford dragon with textures mapped into the voxel grid, then filled into the SVO and extracted.

![Textured Dragon Octree] (images/textured_dragon_svo_gl.png)

This is the same as before, though the cubes rendered are extracted from a higher level in the octree than the original voxelization, so the values are mip-mapped.

![Textured Dragon Octree Mip-Mapped] (images/textured_dragon_svo_mipmapped_gl.png)

##Performance Analysis

• Data Structure Sizes:

Model	Resolution	# Occupied Voxels	# SVO Nodes
Bunny	128	3,234	9,656
	256	11,603	38,408
	512	43,126	151,552
Dragon	128	6,327	14,696
	256	20,604	58,536
	512	69,865	225,360

• The following timing statistics were produced using an NVidia GTX 770. Data Construction / Cube Extraction Timing:

Model	Resolution	Voxelization	Voxel Grid to Cubes	SVO from Voxel Grid	SVO to Cubes
Bunny	128	16.2 ms	8.46 ms	1.85 ms	7.78 ms
	256	54.2 ms	19.7 ms	2.07 ms	22.6 ms
	512	203.9 ms	48.6 ms	2.3 ms	39.9 ms
Dragon	128	26.8 ms	12.1 ms	1.06 ms	12.06 ms
	256	49 ms	30.5 ms	2.79 ms	33.6 ms
	512	221 ms	90.9 ms	2.7 ms	71.8 ms

• CUDA vs. OpenGL. Voxel rendering pipeline performance comparison (*). (With NVidia GT750M)

Voxelize	Model	CUDA Render Time	CUDA FPS	OpenGL Render Time	OpenGL FPS
Without Voxelization	Dragon	47 ms	18	15 ms	60
	Two-Cows	16 ms	35	0~5 ms	60
	All Three**	55 ms	17	15 ms	58
	Bunny	31 ms	29	~0 ms	60
With Voxelization	Dragon	156 ms	5	31 ms	30
	Two-Cows	23 ms	27	15 ms	60
	All Three	141 ms	7	16 ms	34
	Bunny	109 ms	10	15 ms	47

(*) All scenes have resolution set as 512.

(**) All Three is a model scene contains dragon, bunny and buddha objects.

##References

[Interactive Indirect Illumination Using Voxel Cone Tracing] (https://hal.inria.fr/hal-00650196)

Voxel Cone Tracing and Sparse Voxel Octree for Real-time Global Illumination

[OpenGL Insights - Octree-Based Sparse Voxelization Using the GPU Hardware Rasterizer] (http://www.seas.upenn.edu/~pcozzi/OpenGLInsights/OpenGLInsights-SparseVoxelization.pdf)

[GigaVoxels] (http://maverick.inria.fr/Membres/Cyril.Crassin/thesis/CCrassinThesis_EN_Web.pdf)

[VoxelPipe] (https://research.nvidia.com/sites/default/files/publications/voxel-pipe.pdf)

##Bloopers

This is what happens when not enough memory is allocated for the SVO.

![2 Cows No Memory SVO] (images/2_cows_octree_out_of_memory.png)

![Dragon with Texture in GPU Memory] (images/dragon_octree_texture_map_512.png)