[TA frost wolf \u may - "hundred people plan"] Figure 3.6 texture compression - inclusion slimming

3.6.1 Texture compression

Texture compression is to solve the problem of memory 、 Bandwidth issues , Image compression technology specially used for storing textures in computer graphics rendering system ;

Image format

Picture format is the storage format of picture file , Usually on disk 、 Used when storing and transferring files in memory ;

Texture format

Texture format is the texture data format that the graphics card can sample directly , Usually used when loading textures into a graphics card ;

Texture pipeline

Difference ：

• The texture compression format is based on block compression , It can read the byte block of the pixel faster and decompress it to support random access ;
• The image compression format is based on the whole image , The resolution of a single pixel cannot be realized directly ; also , The image compression format cannot be GPU distinguish , Also need to CPU Decompress it into an uncompressed liberal arts format to be recognized ;

3.6.2 Common texture formats

Uncompressed format

DXTC

DXTC Texture compression format comes from S3 The company put forward S3TC Algorithm , The basic idea is to put 4x4 The pixel blocks are compressed into a 64 or 128 Bit data block , The advantage is that a fixed size and independent coding fragment is created , No shared lookup tables or other dependencies , The decoding process is simplified ;

DXT 1

Each block has 2 individual 16 position RGB Color value （RGB565 565 The number of digits representing three channels ）, Represents this 4x4 Extreme value of color in pixel block , Then two intermediate color values are calculated by linear interpolation ,16 individual 2 The bit index value represents the color value index of each pixel ;

For having Alpha Mapping of information , Then the interpolation only generates an intermediate color , When two bits are indexed to the fourth color, it means that the pixel is transparent ;

The reference object is ：RGB24
16 Pixels share 64 Bit data , Then a single pixel is used 4 Bit data ;

DXT2/3

And DXT1 be similar , In the original 64 Bit basis , newly added 64 Bit data is used to represent Alpha value , The whole data block becomes 128 position ; Thus, the occupation of each pixel 8 Bit data ,0-3 Bits represent transparent information ,4-7 Bits represent color information ;

DXT4/5

And DXT2/3 The difference is that Alpha The information is obtained by linear interpolation , Representing color information 64 The bit data block remains unchanged , and Alpha Information is provided by 2 individual 8 position Alpha Extreme values and 16 individual 3 Bit index values ;

expand

If in Unity The inner map type is normal , Will use DXTnm Compressed format , This format will map normals R Channel deposit A passageway , then RB The channel is cleared to 1, In this way, we can find XY The information is stored in RGB/A Respectively compress , So as to obtain higher accuracy , And then based on XY build Z Channel data ;

Compression ratio reference object RGBA 32 position
DXT2/3/4/5 The number of digits is 128 position , Can correspond to 16 Pixel , So the number of bits occupied by each pixel is 8 position , So the compression ratio is 32/8 = 4

ATI1/2

ATI1 Also known as BC4, Each data block stores a single color data channel , Coding and DXT5 Medium Alpha Same data , It is often used to store height map 、 Smoothness map , The effect is basically the same as the original image ;
If the comparison object is a single channel 8 position , use 64 Who said 16 Pixel , Then the compression ratio is 8/ (64/16) = 2:1

ATI2 Also known as BC5, Each block stores the data of two color channels , Ditto with DXT5 in Alpha Data is encoded in the same way , It's equivalent to storing two BC4 block ;
If yes, store the discovery in XY ... is used in the dual channel BC5 Format compression , Because each channel has its own index , So normal mapping XY Information can be compared to BC1 Keep more fidelity in , The disadvantage is that you need to use twice as much memory , More bandwidth is also required to pass textures to shaders ;
If the comparison object is two channels 16 position , use 128 Who said 16 Pixel , Then the compression ratio is 16/(128/16) = 2:1

BC6/7

Only in D3D11 And above are supported in graphics hardware , Each block occupies 16 byte ,BC7 in the light of 8 position RGB or RGBA data , and BC6 in the light of RGB Half precision floating point data , therefore BC6 Is the only one that can be stored natively HDR Of BC Format ;
BC6 It's specifically for HDR（ High dynamic range ） Compression algorithm of image design , The compression ratio is 6:1;
BC7 It's specifically for LDR（ Low dynamic range ） Compression algorithm of image design , The compression ratio is 3:1, This format is used for high quality RGBA Compress , It can significantly reduce the error effect caused by compressing normals ;

ETC

ETC Widely used on Android platform ：


First select the column horizontally through the brightness index , Then use the pixel index to select which row in the corresponding column , Finally, adjust on the basis of the original color , obtain 16 A different color ;

ETC1

The comparison object is RGB, common 24 position
Compression ratio ：24 / （64 /16） = 6：1

ETC2

Yes ETC1 An extension of , Support Alpha Channel compression , Hardware requirements OpenGL ES 3.0 and OpenGL4.3 above ;
ETC1 The required length and width is 2 The power of
ETC2 It is required that the length and width can be 4 to be divisible by

ASTC

from AMD and ARM Jointly developed texture compression format ,ASTC All the indicators are very good , The advantage is that algorithms with different compression rates can be selected according to different pictures , The picture does not need to be 2 The power of , Support at the same time LDR and HDR, The disadvantage is that the compatibility is not perfect and the decoding time is long ;

ASTC It is also a block based compression algorithm , And BC7 similar , Its data block size is fixed to 128 position , But the number of pixels is variable , from 4x4 To 12x12 Pixels have ;

Two interpolation endpoints are stored in each data block , But it doesn't necessarily store color information , It could be Layer Information , This can be used to Normal or Alpha Better compression ; Different forms of compression can be performed according to the map type ;

For each texture element in the block , Store the weight of its corresponding interpolation endpoint , The number of stored weights can be less than the number of texture elements , The weight value of each texture element can be obtained by interpolation , Then calculate the color ;

PVRTC

Support only Iphone、Ipad And some Android phones ;
PVRTC Divide the image into low frequency signal and high frequency signal , The low frequency signal consists of two low resolution images AB form , High frequency signals are low precision modulated images , The weight of each pixel blend is recorded , When decoding AB The image is magnified by bilinear interpolation , Then the mixture is carried out according to the weight of the modulated image ;

PVRTC 4-bpp（4 bit per pixel） Put one 4x4 The pixel unit is pressed into a 64 Bit data block , One... Is stored in each block 32 Bit modulation data （ Then each pixel is divided into 2 Bit modulation data , Expressed as 00,01,10,11）, One 1 Bit modulation flag （ Used to control modulation data ）,15 Bit color A,1 Bit color A Opaque logo ,14 Bit color B,1 Bit color B Opaque logo （ Used to decide according to RGBA The storage is still RGB For storage , If different , Then judge according to the figure with high digits ）;

If the result corresponds to RGB Then the compression ratio is 6:1, If the corresponding RGBA Then the compression ratio is 8:1;

3.6.3 summary

The picture quality is better ：RGBA>ASTC4x4 > ASTC 6x6 > ETC2 ≈ ETC1

Compression ratio ：

DXT1 6:1
DXT2/3 4:1
DXT4/5 4:1
ATI1 4:1
ATI2 4:1
BC6 6:1
BC7 3:1
PVRTC 6:1
ASTC 4:1 ~ 35.95:1

Format summary

PC：

1. Low quality use DXT1 Format （ I won't support it A passageway ）, Use DXT5 Format support A passageway ;
2. High quality use BC7 Format , Support A passageway ;

Android ：

1. Low quality use ETC1 Format , But does not support A passageway ;
2. Low quality use ETC2 Format , Support A passageway , Need to be in OpenGL ES 3.0 / OpenGL 4.3 Above version ;
3. High quality use ASTC Format , Need to be in Android 5.0/ OpenGL ES 3.1 Above version ;

IOS:

1. High quality use ASTC Format , need IPhone6 above ;
2. Low quality use PVRTC2 Format , Support IPhone6 The following versions ;

Homework （ Mistake , See the supplement ）

Experimental scenario ：

The texture format is compressed to ASTC, The size of a single texture is 0.6MB, The memory usage is as follows ：


The texture format is compressed to ETC2, You can see that the size of the compressed texture is 0.7MB The memory usage is as follows ：


The texture is compressed to ETC Format , The compressed texture size is 0.7MB, The memory usage is as follows ：

It was found that , When the picture of the whole scene is reduced to a picture , Memory usage has not changed ; The following two images use different texture compression methods to get a more obvious contrast in texture size ;

Add

The previous modification of the image compression format in the system settings has no effect , After searching, it is found that it needs to be modified in a single texture image , After modification, there is a difference , Or use URP Perform performance analysis , give the result as follows ：（ And closed here Mipmap Compare )

You can see ASTC Compressed picture ratio RGBA The original picture of is in Texture2D The sampling process consumes less memory space .

Another experiment , Turn on MIPMAP It will slightly increase the memory consumption , But it is still less than uncompressed , It should still be related to the test scenario .

In addition, there are many compressed formats ,ASTC6x6 The compression result of may be less than ETC2 Of .