University of Pennsylvania, CIS 565: GPU Programming and Architecture, Project 4 CUDA Rasterizer

• Ziyu Li
• Tested on: Windows 7, Intel Core i7-3840QM @2.80GHz 16GB, Nvidia Quadro K4000M 4GB

Features

All Features

• All Basic Graphics Pipeline
  • Vertex Shading
  • Primitive assembly with support for triangles read from buffers of index and vertex data.
  • Rasterization
  • Fragment shading
  • Depth Buffer and Depth Test
  • Lambert and Blinn shading Model
• Other Features
  • Screen Space Ambient Occlusion (SSAO)
  • Super-Sampling Anti Aliasing (SSAA)
  • Tile-Based Pipeline
  • Back-face Culling (optimized using stream compaction)
  • Display Wire-frame / Points with line width and point size adjustment
  • UV Texture Mapping with Perspective Correction
  • Bilinear Texture Filter
  • Toon Shading

For better result, all the images demonstrate below are using SSAA 2x

Shading Models

Lambert	Blinn

To switch the shading model to Blinn, change the BLINN to 1

 #define BLINN 1

Tile-Based Pipeline

(Image from OpenGLInsights, TileBasedArchitectures)

Break framebuffer into several tiles, and render one tile per thread instead of one primitive per thread.

Primitive-Based	Tile-Based

By using the Tile-Based Pipeline, the interaction is much smoother and GPU Load is siginificant reduced.

To switch the pipeline to tile-based, change the TILE_BASED_RENDER to 1

 #define TILE_BASED_RENDER 1

For details performance comparison, please check Performance Section.

Back-face Culling

Remove the triangles which are not facing to the camera(eye) before rasterization.

Only Front Face	Only Back Face

Here is the Back-face Culling performance comparision in Primitive-Based Pipeline.

No Culling	Back-face Culling

To toggle the backface culling, change the BACKFACE_CULLING to 1

 #define BACKFACE_CULLING 1

For details performance comparison, please check Performance Section.

Display Wire-frame / Points

Solid	Wire-frame	Points

To switch different mode of display, change the SHADING to 1 for rendering wireframe and 2 for rendering vertices

 #define SHADING 1 // 0: Solid, 1: Wireframe, 2: Vertices

UV Texture Mapping with Perspective Correction

Wrong	Correct

Bilinear Texture Filter

No Filter	Bilinear Filter

To toggle the bilinear texture filter, change the BILINEAR_INTERP to 1

 #define BILINEAR_INTERP 1

Screen Space Ambient Occlusion (SSAO)

No AO	SSAO	SSAO Pass

Given any pixel in the scene, it is possible to calculate its ambient occlusion by treating all neighboring pixels as small spheres, and adding together their contributions. The occlusion can be contributed by distance to occludee and angle between occludee's normal and occluder.

(Image from A Simple and Practical Approach to SSAO https://www.gamedev.net/articles/programming/graphics/a-simple-and-practical-approach-to-ssao-r2753/)

Because we need a random direction of occlusion ray. A cached or per-generated noise texture is necessary. If the SSAO feature is turn on a 16x16 noise texture will be generate rasterizer after the primitive has loaded.

This screenshot below, shows the AO Pass of the model.

To toggle the SSAO, change the SSAO to 1

 #define SSAO 1

Super-Sampling Anti Aliasing (SSAA)

No AA	SSAA x2	SSAA x4

To toggle the SSAA, change the SSAA to 2 for 2x or 4 for 4x...

 #define SSAA 2 // A value greater than 1

Toon Shading

Orignal	Toon Shading

To toggle the Toon Shading, change the TOON to 1

 #define TOON 1

Bloom (BUGGE......)

There are still something not corrected yet. But the basic idea is to blend the framebuffer with different amount of blur.

Orignal	Bloom

To toggle the Bloom Effect, change the BLOOM to 1

 #define BLOOM 1

##Performance

Tile-Based Pipeline

The benchmark is based on test model CesiumMilkTruck and use SSAA 2x feature. And the results of graphics card gpu load and memory usage is recorded on GPU-Z.

Based on the result, the Tile-Based Pipeline reduces the GPU load and increases the framerate. It is much more efficient than Primitive-Based Pipeline, especially when less triangle occupy the the most of screen.

BackFace Culling

The backface culling slightly increase the preformance under some scenarios, but it does not affect the performance when a triangle is almost occupy the screen.

Runtime Analysis

Based on the CUDA Device Functions, rasterization is the bottleneck and most intense function in the whole program. For the primitive-based pipeline, the reason of that could caused by too much calculation redundancy and extremely unbalance the job schedualing to rasterize triangles. However, for the tile-based pipeline, There are some empty tiles which within the bounding box but outside the triangle still occupy the runtime of the kernel function. So I believe one of the optimization can be do a triangle-tile intersection test and use stream compaction to remove the empty tiles.

CUDA Device Functions	Primitive-Based (%)	Tile-Based (%)
initDepth	0.2	0.4
vertexTransformAndAssembly	0.7	1.7
primitiveAssembly	0.9	2.2
decompose	N/A	0.5
rasterization	68.2	33.1
render	5.7	14.2
sendImageToPBO	0.6	1.6
others	23.7	46.3

CUDA Runtime API	Primitive-Based (%)	Tile-Based (%)
cudaFree	5.87	6.96
cudaMalloc	0.44	0.52
cudaMemset	0.11	0.34
cudaMemcpy	0.27	0.31
cudaLaunch	1.21	3.47
cudaDeviceSynchronize	83.85	70.4
others	8.25	18

References

• Blinn Shading
• SSAO
• Back-Face Culling
• Kernel Filter

Credits

• factomicMin by @hyqneuron
• tinygltfloader by @soyoyo
• glTF Sample Models
• gpu-z by @TechPowerUp