Complete vertex Compression

Reprinted please indicate the source for the klayge game engine, the permanent link of this article is http://www.klayge.org /? P = 2116

In the article compressing the tangent frame, we see the method of compressing the tangent frame to four bytes. Now let's take a look at how to compress other attributes to reduce vertex data.

Vertex attributes

First, let's take a look at the attributes of the complete vertex:

Attribute	Type	Size (bytes)	Remarks
Position	Float3	12
Texcoord	Float2	8
Tangent	Float3	12
Binormal	Float3	12
Normal	Float3	12
Blend_index	Uint4	16	Only bone Animation models are available
Blend_weight	Float4	16	Only bone Animation models are available
Total		88

After tangent frame compression, generally the engine stores blend_index and blend_weight in uint32, And the vertex format becomes:

Attribute	Type	Size (bytes)	Remarks
Position	Float3	12
Texcoord	Float2	8
Tangent_quat	Uint32	4
Blend_index	Uint32	4	Only bone Animation models are available
Blend_weight	Uint32	4	Only bone Animation models are available
Total		32

This is also a common vertex layout for general engines.

Target

The above shows that all the attributes except position and texcoord are useless. Therefore, our goal is to compress Postion and texcoord to 16-bit for each component, that is:

Attribute	Type	Size (bytes)	Remarks
Position	Short4	8
Texcoord	Short2	4
Tangent_quat	Uint32	4
Blend_index	Uint32	4	Only bone Animation models are available
Blend_weight	Uint32	4	Only bone Animation models are available
Total		24

Fortunately, the current video card supports the vertex attribute formats of short4 and short2, providing the possibility for this target. Unfortunately, short1 is not supported. Otherwise, the position can be split into short2 + short1, saving one more short.

Method

The short stored in the vertex attribute reads the value in the range [-1, 1] On the GPU end. Obviously, it is impossible to directly put the position and texcoord components into the short. Here, the position and texcoord bounding boxes must be saved in the model, while the vertex attributes store the normalized parts of the bounding box. In vs, use both to reconstruct the float component and continue the subsequent calculation. Another good thing is that modern graphics cards support mad, so position * bounding_box_extent + bounding_box_center only needs one command. The computation results in almost no loss of performance, while passing properties can save almost normal bandwidth.

That is, the simple and direct compression reduces the vertex size by 1/4 compared with that of only compressing the tangent frame. In terms of precision, short can provide 65536 different values, which is more than enough in general.

Summary

	Size (bytes)	Compression rate	Remarks
Original format	88	100%
Klayge 4.1 and earlier	48	55%	Tangent frame is compressed to 8 bytes, and blend_index is compressed to 4 bytes.
Klay Ge 4.2	24	28%	Tangent frame is compressed to 4 bytes, blend_weight is compressed to 4 bytes, position is compressed to 8 bytes, and texcoord is compressed to 4 bytes.

After such compression, the new format only occupies 28% of the original format. In examples such as causticsmap and deferredrendering, the neutral performance can be improved by 10%.