实时法线贴图DXT压缩技术

需积分: 9 153 浏览量更新于2024-08-01 收藏 1.27MB PDF 举报

"这篇文档是关于实时法线贴图DXT压缩的研究，由J.M.P.van Waveren（id Software, Inc.）和Ignacio Castaño（NVIDIA Corp.）于2008年发表。主要内容涉及利用现代图形硬件进行法线贴图的实时压缩和解压缩技术，特别是针对地形渲染的优化。文档评估了多种基于现有纹理压缩格式的物体空间和切线空间法线贴图压缩技术，并提出了两种高度优化的切线空间法线贴图压缩算法，可以在CPU和GPU上实现实时性能。" 在计算机图形学中，法线贴图是一种用于增强物体表面细节的技术，无需增加几何顶点数量即可使物体显得更加立体。它通过扰动表面法线来模拟微小的高低不平，从而产生深度和纹理效果。DXT（也称为S3 Texture Compression或DirectX Texture Compression）是一种常见的纹理压缩格式，旨在减少存储和内存需求，同时在渲染时保持较高的视觉质量。文档的介绍部分提到了原始的Bump Mapping技术，由Blinn在1978年提出，它使用高度场的梯度来改变表面法线。随着时间的推移，这一技术演变为法线映射，可以使用法线贴图直接提供法线信息，而不是高度信息，这使得处理更加复杂和细腻的表面成为可能。在现代游戏和实时渲染中，实时性是一个关键因素。虽然硬件支持在渲染过程中实时解压缩DXT压缩的法线贴图，但压缩过程本身可能会消耗大量时间。文档中提到，现有的压缩器在将法线贴图压缩到这些格式时，速度可能较慢，这对实时应用是个挑战。为了应对这个问题，作者提出了两种高度优化的切线空间法线贴图压缩算法，目标是在CPU和GPU上都能实现实时性能。切线空间法线贴图特别适用于处理光照和法线变化较大的情况，因为它允许更精确地计算光照效果。文档的其余部分可能详细讨论了这些算法的工作原理、实现细节、性能比较以及可能的优化策略。这包括如何有效地编码和解码法线信息，如何减少失真以保持视觉质量，以及如何在不同硬件平台上实现高效的并行计算。这篇技术文档对于那些希望优化实时渲染性能，特别是在处理大规模地形和复杂表面细节时，提供了宝贵的理论和实践指导。通过理解和应用文中提出的压缩算法，开发者能够提升游戏或应用的性能，同时保持高质量的视觉效果。

Real-Time Normal Map DXT Compression 10 February 2008

Only storing the X and Y components is in essence an orthographic projection of the normal vectors

along the Z-axis onto the XY-plane. To reconstruct an original normal vector, a projection back onto

the unit-hemisphere is used, by deriving the Z component from the X and Y. Instead of this

orthographic projection, a stereographic projection can be used as well. For the stereographic

projection the X and Y components are divided by one plus Z as follows, where (pX, pY) is the

projection of the normal vector.

pX = X / ( 1 + Z )

pY = Y / ( 1 + Z )

The original normal vector is reconstructed by projecting the stereographically projected vector back

onto the unit-hemisphere as follows.

denom = 2 / ( 1 + pX * pX + pY * pY )

X = pX * denom

Y = pY * denom

Z = denom – 1

The advantage of using the stereographic projection is that the interpolated normal vectors behave

better under bilinear or trilinear filtering. The interpolated normal vectors are still not on the

shortest great arc, but they are closer, and have less of a tendency to go up on the unit-

hemisphere.

The stereographic projection also causes a more even distribution of the pX and pY components

with the angle on the unit-hemisphere. Although this may seem desirable, it is actually not, because

most tangent-space normal vectors are close to the tip of the unit-hemisphere. As such, there is

actually an advantage to using the orthographic projection which results in more representations of

vectors with Z close to 1. The compression techniques discussed below use the orthographic

projection because for most normal maps it results in better quality compression.

3.1 Tangent-Space DXT1

Using tangent-space normal maps only the X and Y components have to be stored in the DXT1

format, and the Z component can be derived in a fragment program. The following fragment

program shows how the Z can be derived from the X and Y.

# input.x = normal.x [0, 1]

# input.y = normal.y [0, 1]

# input.z = 0

# input.w = 0

MAD normal, input, 2.0, -1.0

DP4_SAT normal.z, normal, normal;

MAD normal, normal, { 1, 1, -1, 0 }, { 0, 0, 1, 0 };

RSQ temp, normal.z;

MUL normal.z, temp;

The following images show a XY_ DXT1 compressed normal map on the right, next to the original

normal map on the left. The DXT1 compressed normal map shows noticeable blocking and banding

artifacts.

Real-Time Normal Map DXT Compression 12 February 2008

3.2 Tangent-Space DXT5

Just like for object-space normal maps, all three components can be stored in the DXT5 format. The

best results are usually achieved when storing _YZX data. In other words the X component is

moved to the alpha channel. This technique is also known as RxGB compression, and was employed

in the computer game DOOM III. By moving the X component to the alpha channel, the X and Y

components are encoded separately. This improves the quality because the X and Y components are

most independent with the largest dynamic range. The Z is always positive and typically close to 1

and, as such, storing the Z component with the Y component in the DXT1 part of the DXT5 format

causes little distortion of the Y component. Storing all three components results in minimal

overhead in a fragment program as shown below.

# input.x = 0

# input.y = normal.y [0, 1]

# input.z = normal.z [0, 1]

# input.w = normal.x [0, 1]

MAD normal, input.wyzx, 2.0, -1.0

The following images show that, although the quality is better than DXT1 compression, there are

still noticeable banding artifacts.

DXT5 compressed normal map on the right compared

to the original normal map on the left.

Just like for object-space normal maps the quality can be improved by storing directions that are

not necessarily unit-length. The best quality is typically achieved by also moving the X component

to the DXT5 alpha channel. The following fragment program shows how the directions are re-

normalized after moving the X component back in place from the alpha channel.

# input.x = normal.x [0, 1]

# input.y = normal.y [0, 1]

# input.z = 0

# input.w = normal.z [0, 1]

MAD normal, input.wyzx, 2.0, -1.0

DP3 scale, normal, normal

RSQ scale.x, scale.x

MUL normal, normal, scale.x

The following images show that encoding directions with re-normalization in a fragment program

reduces the banding artifacts, but they are still quite noticeable.

剩余58页未读，继续阅读

aircraftabc

粉丝: 0
资源: 4

实时法线贴图DXT压缩技术

游戏客户端，UIF-DXT编辑工具

Real-Time Adaptive Scalable TextureCompression for the Web.pdf

COMPUTE_VIEW_NORMAL

UE5安卓打包ETC2和ASTC和DXT的区别

MATLAB Output argument "f1" (and possibly others) not assigned a value in the execution with "untitled>dxt" function.

OpenCV(4.6.0) /io/opencv/modules/core/src/dxt.cpp:3508: error: (-215:Assertion failed) type == CV_32FC1 || type == CV_32FC2 || type == CV_64FC1 || type == CV_64FC2 in function 'dft'什么问题

用OpenCV和C++实现两张图片的相位谱交换，并用双谱进行重构

pixi6.5可以加载什么格式的压缩纹理？

dds图片资源何如压缩

最新资源