RTSDF Real-time Signed Distance Fields for Soft Shadow Approximation in Games Yu Wei Tan

2025-04-26 0 0 5.46MB 8 页 10玖币

侵权投诉

RTSDF: Real-time Signed Distance Fields for Soft Shadow

Approximation in Games

Yu Wei Tan a, Nicholas Chua, Clarence Koh and Anand Bhojan b

School of Computing, National University of Singapore, Singapore

{yuwei, nicholaschuayunzhi, clarence.koh}@u.nus.edu, banand@comp.nus.edu.sg

Keywords: Real-time, Signed Distance Field, Jump Flooding, Ray Tracing, Soft Shadow, Rendering, Games.

Abstract: Signed distance ﬁelds (SDFs) are a form of surface representation widely used in computer graphics, hav-

ing applications in rendering, collision detection and modelling. In interactive media such as games, high-

resolution SDFs are commonly produced ofﬂine and subsequently loaded into the application, representing

rigid meshes only. This work develops a novel technique that combines jump ﬂooding and ray tracing to gen-

erate approximate SDFs in real-time. Our approach can produce relatively accurate scene representation for

rendering soft shadows while maintaining interactive frame rates. We extend our previous work with details

on the design and implementation as well as visual quality and performance evaluation of the technique.

1 INTRODUCTION

Signed distance ﬁelds (SDFs) are scalar ﬁelds that

store the shortest distance between a point in space to

a model. Their sign indicates if that point is inside or

outside the bounds of said model. In interactive me-

dia, models are most commonly represented by tri-

angle meshes. SDFs are typically generated ofﬂine

through ray tracing and scan conversion etc., limiting

their use to rigid meshes. While existing real-time

GPU-based methods can update SDFs per frame, they

are unable to handle high resolutions efﬁciently.

We present a novel SDF method that integrates

jump ﬂooding and ray tracing to generate an approx-

imate real-time SDF (RTSDF) of reasonably high

quality for a ﬁxed small scene. Additionally, we eval-

uate the technique by applying it to raymarched soft

shadow approximation, offering trade-offs between

speed and quality for real-time application require-

ments. This paper extends our previous work (Tan

et al., 2020) with a detailed analysis of the design,

implementation and evaluation of the technique.

2 DESIGN

As shown in Figure 1, jump ﬂooding produces a

fast approximation of the SDF which allows for real-

ahttps://orcid.org/0000-0002-7972-2828

bhttps://orcid.org/0000-0001-8105-1739

time calculation. Conversely, ray tracing gives a

more accurate scene representation as it queries the

hardware-generated triangle mesh and slowly con-

verges. Hence, we propose a real-time SDF that com-

bines the speed of jump ﬂooding with the precision of

ray tracing. We ﬁrst perform an initial jump ﬂooding

which creates an SDF of the voxelized scene. Next,

we use the voxelized SDF as a mask to decide where

to attempt ray tracing on the triangle mesh for more

accurate scene representation. Naturally, we choose

locations closer to surfaces to ray trace and fall back

on the distances generated by jump ﬂooding in empty

regions. Our technique is built on NVIDIA’s Falcor

library (Benty et al., 2020) which provides an abstrac-

tion over the graphics API for our implementation.

(a) Jump ﬂooded SDF (b) Ray-traced SDF

Figure 1: Soft shadows from SDF raymarching.

2.1 Ray Tracing

The basic implementation of an SDF is a uniform grid

where the discrete points in the scalar ﬁeld are stored

as a 3D texture (Wright, 2015). Uniform grids are

arXiv:2210.06160v1 [cs.GR] 11 Oct 2022

easy to implement and allow us to perform hardware

interpolation of neighbouring points, making sam-

pling efﬁcient. A brute force approach to generate

uniform SDF is ray sampling (Wright, 2015).

For each discrete point in the distance ﬁeld, we

shoot rays in random directions to query the distance

to the closest mesh. To calculate the direction, we ran-

domly generate a point on the surface of a sphere with

a uniform distribution (Weisstein, 2019). The mini-

mum distance traced for each point is stored in the

red channel of a 400 ×200 ×400 3D texture. Rays

traced in future frames will overwrite the value if the

newer value is smaller. To determine the sign, we ﬁrst

check if it is a front or back face hit via the dot prod-

uct of the ray direction and normal of the primitive

intersected. We then accumulate the number of front

and back face hits in the green and blue channels of

the texture. Finally, we set the sign to negative if the

majority of hits are back face hits (Wright, 2015).

2.2 Jump Flooding

The jump ﬂooding algorithm (JFA) (Rong and Tan,

2006) which can be run on parallel on the GPU al-

lows us to generate an approximate SDF in real-time.

We ﬁrst initialize a 3D SDF texture where each texel

represents a 3D point or a grid point. JFA gives us in-

formation about the closest seed at any point in space.

Setting every point on each triangle in the mesh as a

seed, we can obtain the closest distance to a surface

for our distance ﬁeld. To determine if a grid point

contains a triangle efﬁciently, we voxelize the scene

with voxel resolution equal to our ﬁnal distance ﬁeld.

After voxelization, we simply check if a grid point

contains a model voxel to determine if it is a seed.

We now have a 3D texture containing grid points

that are either empty or are seeds. Without loss of

generality, we assume that the dimensions of the 3D

texture are equal (i.e. 3D cube) and that its length n

is a power of 2. For each grid point, we query a con-

stant number of neighbouring grid points a predeﬁned

offset away. For each query, if the queried grid point

is a seed or contains seed information, we check if

that seed is closer to its currently stored seed and up-

date its seed information if so. We start with an offset

of length n

2and halve it for each subsequent iteration

until it reaches 1 for our ﬁnal iteration.

With current GPUs that can write to 3D textures,

we adapt the 2D JFA algorithm (Rong and Tan, 2006)

for 3D space. During a single iteration, we run the

querying in parallel, allowing us to use the GPU to

accelerate the calculations. However, the algorithm

produces an unsigned distance ﬁeld. To determine the

sign, we subtract a small βfrom the distance ﬁeld,

causing surface points to be of negative value and ef-

fectively thickening the surface. The surfaces gener-

ated are also hollow as they contain positive values in

their interior. With jump ﬂooding, we can generate

an approximate SDF in real-time for a decently large

resolution of 2563at 30ms and 1283at 2.34ms.

2.3 Ray Mask

We obtain a rough approximation of the SDF or

coarse SDF via jump ﬂooding to locate regions in the

scene to apply ray tracing. In raymarching, regions

far from the surface act as a way to accelerate the pro-

cess but closer regions require a more accurate surface

representation. Hence, we can detect regions closer to

surfaces based on a distance dfrom the coarse SDF

and only ray trace in these regions at a higher resolu-

tion for better surface representation. Essentially, the

coarse SDF is a ray mask that determines which ar-

eas in the SDF should be ray-traced to generate a ﬁne

SDF as shown in Figure 2. dcan be used to trade-off

performance for accuracy where a larger dresults in

more rays traced as texels further from surfaces would

be within ddistance from a surface point.

(a) Coarse SDF (b) Fine SDF

Figure 2: Slice of SDF.

Unlike adaptively sampled ﬁelds which increase

the resolution at regions with ﬁner details, we limit

the number of levels of detail to two as generation

of hierarchical SDF is difﬁcult to parallelize (Liu and

Kim, 2014). Additionally, real-time traversal of the

SDF may require multiple texture lookups to sample

until the leaf node despite saving space with a sparse

voxel texture (Aaltonen, 2018).

With the combination of the techniques, we turn

a blocky representation of the scene into a more re-

ﬁned triangle mesh representation without ray tracing

every texel in the SDF as shown in Figure 3. Addi-

tionally, we can resolve issues identiﬁed in the jump

ﬂooding of voxelization of thin surfaces. As seen in

Figure 4, while the coarse SDF gets a disconnected

representation of the plant, we ﬁll in the holes through

reﬁnement with the ray trace pass.

2.4 Raymarching

While SDFs are surface representations, they only

provide proximity information without a direction. To

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

RTSDF:Real-timeSignedDistanceFieldsforSoftShadowApproximationinGamesYuWeiTana,NicholasChua,ClarenceKohandAnandBhojanbSchoolofComputing,NationalUniversityofSingapore,Singaporefyuwei,nicholaschuayunzhi,clarence.kohg@u.nus.edu,banand@comp.nus.edu.sgKeywords:Real-time,SignedDistanceField,JumpFlooding,Ra...

展开>> 收起<<

RTSDF Real-time Signed Distance Fields for Soft Shadow Approximation in Games Yu Wei Tan.pdf

共8页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

RTSDF Real-time Signed Distance Fields for Soft Shadow Approximation in Games Yu Wei Tan

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: