A JPEG-like texture
compression with adaptive
quantization for 3D
graphics application
C.-H. Chen, C.-Y. Lee
Department of Electronics Engineering, National Chiao Tung University, 1001, Ta Hsueh Road, Hsinchu, 300, Taiwan, R.O.C.
E-mail: [email protected]
Published online: 2 October 2001 c
Springer-Verlag 2001
DCT-based compression is widely used in video and image compression for a high compression ratio, but it suffers from ran-dom-access problem when applied to tex-ture compression. In this paper we present a JPEG-Like DCT-based texture compres-sion technique which is suitable for 3D graphics rendering system. We apply a sim-ple adaptive quantization on an 8× 8 block-size texture, such that the length of the en-coded bit stream of one block can be approx-imated to the target. A pre-defined quantizer scale is encoded with the bit stream with a small overhead. Our technique achieved a high compression ratio, quality control of true color texture, and random access of tex-ture data.
Key words: 3D graphics – Texture
compres-sion
Correspondence to: C.-H. Chen
3D computer graphics have been widely used in various applications, for example, entertainment, medicine, and scientific research. There are few ap-plications that stress modern PC computing power more than today’s 3D-API. The continuous improve-ment of deep sub-micron semiconductor technology has brought 3D graphics processors from high-end workstations to low-end desktop PCs. These power-ful processors are now capable of supporting many 3D applications. Although current 3D graphics pro-cessors have high computing power, they still cannot render lifelike animated images in real-time. The bottleneck of 3D graphics systems includes comput-ing power, memory storage and bandwidth require-ment. Although the computing power of graphics processors doubles every six months, the memory storage and bandwidth requirement become worse due to the demand for realistic and complex 3D ap-plications. Thus the overall system performance is limited.
Among the operations in 3D graphics rendering sys-tems, texture mapping consumes the largest band-width and memory storage. The bandband-width require-ment for texture access may be up to giga-bytes per second, as shown in Table 1. Traditional texture mapping is used to increase the detail of the ge-ometry visual quality without increasing the model complexity. Now multi-texture mapping techniques are widely used in lots of applications, for example, specular light-map, environment mapping, shadow and bump mapping. Texture filtering, which includes bilinear, trilinear and anisotropic filtering, is applied to reduce texture aliasing. With multi-texture map-ping and filtering, the bandwidth requirement for texture access may jump up to 2-, 4- or 8-fold the sin-gle texture mapping (PowerVRTM1998).
Texture compression (TC) is the technique used to reduce the texture memory storage and bandwidth requirement. It increases system performance in sev-eral ways. First, it reduces accelerated graphics port (AGP) and frame buffer bus traffic, yielding supe-rior texturing performance and fill rate. Second, it re-duces the texture storage in frame buffer; this makes more detailed texture storage or performance possi-ble by using the additional memory space for doupossi-ble or triple buffering. TC is a “must have” feature in current 3D accelerators.
Textures are ordinary bit map images. Although good and optimized image compression techniques have been described in many papers during the past few decades, they are not suitable for TC
be-The Visual Computer (2002) 18:29–40
Texture 5 M polygons/s 10 M polygons/s 20 M polygons/s filtering type Without With Without With Without With cache texture cache texture cache texture
cache cache cache
Point sampling 375 37.5 750 75 1500 150 Linear 750 75 1500 150 3000 300 mip-map Bilinear 1500 150 3000 300 6000 600 mip-map Trilinear 3000 300 6000 600 12 000 1200 mip-map Anisotropic 6000 600 12 000 1200 24 000 2400
Table 1. Bandwidth estimation of texture
access (MB/s). Assume the following: 25 pixels/polygon, 24-bit RGB texture, single texture mapping, 10% texture cache miss rate
cause of several issues with regard to texture access. In Sect. 2, we review previous work and discuss these issues considering modern technology and ap-plications. In addition, we find a balance between texture quality and compression ratio (CR) for dif-ferent applications. For example, medical and sci-entific 3D images must have lossless compression, 3D games can choose lossy compression and in-ternet 3D applications can choose the highly lossy mode to shorten download time. Current results (3dfx 1999; S3TM 1999) for the CR are within the range of 4 to 8. This CR may be fine for current API, but it is not sufficient for future 3D applica-tions. Future TC should provide high CR and quality control. TREC (Torborg et al. 1996) is the first to use JPEG-like high CR and a simple quality control technique. But TREC suffers from a long texture ac-cess problem (Microsoft 1999), which may degrade performance. In this paper, we develop a JPEG-like TC technique which aims for high CR (> 8). We solve the problem of random texture access where quality control is also considered. A compar-ison of quality and CR with other methods is also shown.
2 Issues of texture compression and
previous work
Image and video compression technologies have been extant for a long time. Still image coding schemes have been classified into predictive, block transform, and multi-resolution approaches (Egger et al. 1999). VQ, DCT and DWT are the most widely accepted techniques among these image compression schemes. Although textures are or-dinary bitmap images, current image compression
methods may not be applicable to TC. The nature of texture access make TC differ from traditional image compression. We discuss the issues of TC below.
2.1 Issues of texture compression
2.1.1 Random access
In a standard rasterization pipeline, the texture coordinate is generated until the polygon scan-conversion stage. The texture coordinate is evaluated or interpolated in texture space. Perspective correc-tion and anisotropic and filtering texture mapping techniques result in more detail and correct tex-ture mapping. The mapping procedure introduces discontinuous texture access in texture space and sometimes may cause large displacement. Thus the TC technique must provide fast random access of texture.
2.1.2 Compression ratio and visual quality
There is a tradeoff between CR and quality. Loss-less compression, such as Lempel–Ziv compres-sion, can provide perfect reconstruction, but the CR is low. However, lossy compression introduces er-ror. For TC, existing techniques (S3TM 1999; 3dfx 1999; Ivanov and Kuzmin 2000) have average CR values of between 4 and 8. Current image com-pression techniques can easily achieve CR values higher than 8 and also provide some features for various applications, such as: quality control, SNR scaleable, multi-resolution and progressive coding and streaming. The major difference between im-ages and textures is that imim-ages are viewed on their own, while texture mapping rotates, distorts,
magnifies or minifies texture to “paste” on geom-etry surface. The visual quality is the most impor-tant for image compression, but it is not so critical for TC.
2.1.3 Decoding speed
As shown in Table 1, the texture access rate may be up to a giga-texel per second. The texture compression engine must provide high-speed de-compression and texture access. All current designs include a small on-chip texture cache to reap the benefit of texture reuse. Hakura and Gupta (1997) showed that the cache miss rate is under 10% with proper cache design. Issue texture access or pre-fetch as early as possible can solve the problem of long texture decompression latency (Igehy et al 1998; Mi-crosoft 1999). With the texture cache and a pre-fetch technique, the decoding speed may not be so critical, except in the worst-case scenario of texture cache miss. For example: currently, the fastest 3D GPU can achieve a rate of 3.2 G texel/s; assume a 10% texture cache miss rate; thus, the decoding speed is at least 320 M texel/s.
2.2 Previous work
Several TC techniques have been proposed and some are used in current products. These techniques can be classified into several categories:
2.2.1 Block truncation coding
Block truncation coding (BTC) (Knittel et al. 1996) is a simple TC technique. Two colors are selected to be the reconstruction level of a 4×4 block. Although BTC is very simple to encode and decode, its draw-backs are edge discontinuity, blocking effect and low CR. Kugler (1997) provides post-processing filtering to improve the quality. But this introduces more tex-ture access of neighboring blocks and makes decom-pression more complex. Despite its CR being high (= 12), BTC is not well suited for TC due to poor quality.
2.2.2 Vector quantization
Vector quantization (VQ) was chosen for TC due to the simple decoding involved. However, VQ suf-fers from the problem of codebook selection. Code-book size determines the image quality and CR. Beers (1996) and PowerVRTM (Butler et al. 1999)
chose one codebook for the entire texture. When de-compressing, the decoder first retrieves the index of the corresponding block and then consults the code-book to find the final colors. This requires two serial memory accesses of one block unless the codebook can be stored on the chip or a universal codebook can be hard-wired on the chip. A high-quality code-book will consume a large amount of chip area and require long downloading latency. It is impossible to find a universal codebook whose size is small and has good quality. Thus the traditional VQ tech-nique is not suitable for high-performance texture systems.
2.2.3 Local palette
S3TC (S3TM1999) and 3dfx-FXT1 (3dfx 1999) are local palette techniques. They decompose texture into small 4× 4 (S3TM 1999; 3dfx 1999) or 8× 8 (3dfx 1999) blocks. Within each block, two (S3TM 1999; 3dfx 1999), three (3dfx 1999) or four (3dfx 1999) colors are encoded and linearly interpolated to four or eight colors. The drawback of this method is the lack of colors in each block. Thus Ivanov and Kuzmin (2000) proposed a “color distribution” tech-nique to provide more colors on one block. It al-lows using colors from neighboring blocks instead of simply interpolating colors stored in the current block. The visual quality may become superior to that in S3TC and FXT1. However, all of the local palette techniques have a CR of 4 to 6 for RGB format. Current local palette techniques do not pro-vide 24-bit RGB true color format, and when the alpha component is considered, the CR becomes re-duced. They cannot achieve a high CR of true color texture.
2.2.4 DCT-based coding
Texture and rendering engine compression (TREC) (Microsoft 1997) is a JPEG-like compression tech-nique, while Playstation2 (Suzuoki et al. 1999) is an MPEG2-based technique. Both of them have DCT-based coding. A CR higher than 15 can eas-ily be achieved, and good quality can also be en-sured under DCT-based coding. The challenges of DCT-based coding for TC are random access and decoding speed. Variable-length bit stream makes it unsuitable for TC. Besides, it must decode in order from the head of the bit stream to find the desired pixel. TREC solved this problem by preserving DC without DPCM. Talisman constructed an index ta-ble and link-list to address the variata-ble-length bit
Fig. 1. JPEG encoding/decoding flow
stream of each block. However, the long latency (Mi-crosoft 1999) of texture access is still a bottleneck and makes the hardware design more complex. In this paper, our task is to solve the problems of DCT-based TC.
3 The proposed JPEG-like texture
compression technique
3.1 Brief review of JPEG
DCT-based compression techniques are based on the concept of compacting energy into fewer coefficients in the transform domain and then encoding these coefficients. JPEG (Pennebaker 1993) is a standard DCT-based still image compression technique, and it provides good visual quality. JPEG partitions the image into 8×8 blocks, applies to each block the for-ward discrete cosine transform (FDCT), and scalar quantizes the DCT coefficients. The DC coefficient is encoded by DPCM, while the zig-zag-ordered AC coefficients are entropy coded using Huffman cod-ing (or variable length codcod-ing, VLC). Figure 1 shows the baseline JPEG encoding and decoding flow. In the encoding process, only the Huffman codewords
and quantization tables are left for the user to con-trol the visual quality and compression ratio. Stan-dard JPEG has provided the example VLC code-words and quantization table based on perceptual criteria.
If we want to use JPEG for TC, we need to solve the random access problem. To achieve random ac-cess of an image block in the JPEG bit stream, we need to build an index table or truncate the bit stream. Truncating the bit stream will cause a serious block effect, since the high-frequency component cannot just be discarded. The technique used to build the index table is similar to TREC. The entries in the in-dex table represent the memory address or offset of each block. The size of the index table is described as follows: Sindex= TotalBlocks × log2 TotalBlocks× (R + G + B + A) C R , (1)
where TotalBlocks= (Width/8) · (Height/8). For 512× 512 24-bit RGB format images and CR = 24, Sindex= 73 728 bits. Adding the index table, the overall C R decreases to 18.73. If we parti-tion the bit stream of each block into 32, 64, 128 or 256 segment sizes for the purpose of
mem-Fig. 2. Compression ratio decreases under different bit
stream segment sizes in JPEG
ory alignment, the overhead of index table is as follows: Sindex= TotalBlocks × log2 TotalBlocks j=1 Bj Segment Size × Segment Size , (2) where Bj is the encoded bit-stream length.x
rep-resents the minimum integer greater than or equal to
x. Note that if we down sample U and V an extra
index table should be built for the U and V streams to distinguish them from the Y stream. The overall
C R is reduced by adding an index table. Figure 2
shows the decrease in C R for different segment sizes. It indicates that the overhead of the table is high especially for a low-complexity image: Lenna and Backwall as shown in Table 3. Using an index ta-ble will introduce two memory accesses to obtain the desired bit-stream and degrade the overall perfor-mance.
3.2 The proposed algorithm
Figure 3 shows the compression and decompression flow of our JPEG-like TC algorithm for an RGBA,
32-bit texture. The major distinctions of our algo-rithm from JPEG are:
• It operates on one 8 × 8 block individually. Each 8× 8 Y or U or V or Alpha block is encoded, de-coded and accessed independently.
• It compresses every block toward a fixed length bit stream, thus random access can be achieved and memory space can be utilized efficiently. The bit stream length of every encoded block can be any size, as controlled by user definition or qual-ity control.
• Every 8 × 8 block has its corresponding quantizer scale, which is pre-defined to control quality and
C R.
• The “Alpha” component, which represents the transparency of every texel, is compressed in ad-dition to the R, G and B color components. Both DCT and VLC are the same as those for JPEG; however, we change the quantization step in our al-gorithm. In the JPEG quantization stage, Hung and Meng (1991) found the optimized quantizer of JPEG using a complex search. But if the same quantizer is applied to every image block as in JPEG, it will lack local adaptation and also result in a variable-length bit stream. In our algorithm, we introduce a localized quantizer, an adaptive quantization for every block. One of the pre-defined quantizer scale or quantiza-tion tables is chosen to quantize each 8× 8 block to the desired bit stream length.
In JPEG or MPEG, the quantizer scale can be speci-fied to control the rate of the encoded bit stream. Rate control is introduced to provide consistent visual quality in video transmission. The problem found is, given a target bit rate, how to encode the video sequence for minimized distortion and consistent vi-sual quality. In JPEG and MPEG, the quantization table or quantizer scale is applied to all blocks. We can describe the quantization for each 8× 8 block as follows: Cn= Round Cn Qn· Qp n= 1, 2, . . . , 64, (3) where Cn represents the DCT coefficients in one block, Qn is the corresponding quantization step and Qp is the quantizer scale, whose default is 1.
Round(x) represents the closest integer to x.
Although an optimized quantizer can be found to achieve minimum distortion of one block, the de-coder has to know the quantizer of each block. As a consequence the C R will strongly decrease. Thus
Fig. 3. Our JPEG-Like texture compression/decompression flow
we have to define some quantizer scale Qp for TC and encode 4 bit data into each corresponding block to indicate which quantizer scale or table is selected.
3.3 Adaptive quantization for
8
× 8
blocksConsidering hardware implementation, we define sixteen Qp for the quantizer scale. Thus a 4-bit in-dex is encoded in each 8× 8 block to indicate which
Qphas been selected. Therefore (3) can be written as follows: Cn= Round 16· Cn Qn· Qp Qp= 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30, 32, 40, 48. (4) The range and resolution of Qp can be increased or decreased to control the rate more precisely or coarsely. Assume the desired compression ratio is
C R. The allowed bit count for one 8× 8 block j is B= 64 · 24/CR for RGB true color texture. The bit
count after Huffman coding is Bj. The encoding
tar-get is to let Bj come as close as possible to (but
remain less than) B. The selection of Qp could be made by searching all of the Qp. However, we intro-duce the activity-based bit estimation model (Cheng and Hang 1995) used in the rate control of video se-quences to increase the encoding speed.
The bit count Bj depends on the complexity in the
block. For example, a texture block with higher ac-tivity indicates that the content of the block is more complex than one with low activity and that more bits will be needed to encode it. Activity here means an absolute-value summation of the 63 DCT AC coefficients in one 8× 8 block. Therefore we can construct a linear equation of block activity and bit counts for different Qp. The following is an empiri-cal first-order bits model:
Bj = K1
actj
Qp j+ K2, (5)
where K1 and K2 are constants derived from
train-ing data to minimize the error and actj is the ac-tivity of block j. By using this model, we can train lots of various blocks under different Qp
and then obtain 16 approximation lines. When en-coding block j, we calculate actj first and then solve (6):
Qp j = K1
actj
B− K2
. (6)
Thus we obtain the approximate Qp j. Figure 4 illustrates the concept of Qp j selection. By con-trolling Qp j, Bj can be specified to be as close as possible to B. Since the content of neighbor-ing blocks usually has little variation, the previous blocks Qp j can be used as a reference for Qp j+1.
Fig. 4. Qpselection under different activities
The pseudo-code to encode one 8× 8 block is as follows:
1. Calculate activity actjof block j.
2. Solve (6) to obtain the value of Qpclosest to or
equal to Qp j−1.
3. Quantize j with Qpand Huffman coding to
ob-tain the bit count Bj; if Bj> B go to step 4, else if Bj< B go to step 5, else go to last step 6. 4. If Bj> B, decrease the bit count:
while ( Bj> B) {
increase Qpone level to Qp;
quantize with Qp;
Huffman encoding to obtain new Bj; };
Qp j = Qp.
5. If Bj< B, increase the bit count but do not allow it to exceed the budget B:
while ( Bj< B) {
decrease Qpone level to Qp;
quantize with Qp;
Huffman encoding to obtain new Bj; };
Qp j = Qp.
6. Encode corresponding index of Qp jinto block. As shown above, complex blocks are assigned large
Qp, and low-activity blocks smaller ones. Although
higher Qp means higher loss and lower Qp means
smaller loss, all of the blocks are encoded towards the target length and stored in a fixed-size mem-ory block with efficient memmem-ory utilization. Thus the texture in any block in the memory can be accessed. This fixed-length technique also provides efficient memory allocation and consistent decoding speed.
3.4 Bit budgets
The texture formats in 3D graphics include gray-scale, 16-bit RGB, 24-bit RGB, 32-bit RGBA and others. How to choose the bit target for each compo-nent in each block is another problem in our adaptive quantization technique. Here, we profile the bit al-location of each component in JPEG to see the bit counts of each component. When arranging the bit budget, the memory alignment problem should also be considered. Bit counts of 32, 64, 128 or 256 are suitable for memory access and hardware design. Other size may cause two or three memory accesses in one texture request and decrease the performance of texture mapping.
We down-sample U and V by two in our algorithm. This refers to the YUV 4: 2 : 0 format used in JPEG and MPEG. Figure 5 is the bit allocation profile of three 24-bit true color JPEG images, where the quan-tizer and Huffman coding are set to default. Ob-viously, we can find that almost all of the U and
V allocated bit counts are within 32 bits in one
block. But Y may be distributed widely from 16 to 160 bits. Thus we can conclude that a 32-bit bud-get is sufficient for U and V for most cases in TC. For Y , it depends on the complexity of the texture. A 64-, 96-, 128- or 256-bit budget can be chosen for Y .
In addition to the RGB format, the texture format may include the “Alpha” component, which is used to represent the transparency of the texture. The Al-pha component can be treated as a gray-level image; thus we encode it individually. The Alpha resolution may be 4, 5, 6 or 8 bit, depending on the application. When the resolution is 5 or 6 bit, we can combine the bit stream of Alpha with Y for a total bit length of 128 or 256, which is a good memory alignment size. If an 8-bit Alpha is used, the bit stream is sep-arated from Y . Table 2 shows the performance of different resolution Alpha under various C R in our algorithm. Here, we treat the gray-level image as the Alpha component, since the detail of Alpha can be any shape.
3.5 Compression ratio and visual quality comparison
As discussed in Sect. 3.4, we can use the adaptive quantization in Sect. 3.3 for each block to efficiently use the bit budget. Table 3 shows the CR and a visual-quality comparison with other techniques. The CR,
Fig. 5. Bit allocation profile of YUV components in JPEG
Table 2. PSNR performance of different resolution Alpha; CR
is 8
Baboon Lenna Backwall Banner Texture 4 bits/pixel 45.84 51.05 49.68 46.92 49.57 5 bits/pixel 41.12 47.80 46.93 42.78 45.71 6 bits/pixel 36.01 44.01 43.08 38.11 41.07 8 bits/pixel 24.76 34.04 33.58 27.57 31.61
when U and V are down-sampled by 2, is defined as follows: C R= (R + G + B + A) × 64 BitY+ BitU+BitV 4 + BitA, (7)
Where R, G, B, A represent the format of the tex-ture and BitY, BitU, BitV, BitA represent the bit counts of one block. For a 24-bit RGB format,
R= G = B = 8 when Y = 64, U = V = 32 and C R= 19.2. For U = V = 32 and Y = 128 or 64,
sizes suitable for memory alignment, C R= 10.67 and 19.2. As shown in Table 3, our TC technique can achieve a C R about 2-fold greater than the S3TC local palette technique and still hold the vi-sual quality of JPEG. The random access problem of TREC is also solved. When the Alpha precision is low, for example 4 or 5 bit, combining the Al-pha bit stream with Y to 128 bits results in good quality, and CR becomes 14.2, which is more than three times that in S3TC. For a low-complexity im-age (Lenna and Backwall), the quality is still good at C R= 19.2. The rendering results in Figs. 6 and 7 show the overall visual quality is good even under high C R. It is hard to find the difference between compressed and uncompressed texture-mapped 3D scenes. A feature comparison with other TC tech-niques is given in Table 4. DCT-based coding is good for high C R and good quality, while others cannot preserve good quality under high C R. With a fixed length bit stream technique, our technique is better than TREC, where both a high C R and good quality can be achieved.
3.6 Decoding speed
The decoding speed is an important issue in TC. Previous works (S3TM 1999; 3dfx 1999; Ivanov
and Kuzmin 2000) challenge the decoding speed and hardware cost of JPEG. But from the a system view, it is not critical in current 3D graphics sys-tems. Current 3D graphics accelerators all include
Table 3. Compression ratio and visual quality (average RGB PSNR) comparison. The number in the parenthesis is C R
Baboon Lenna Backwalla
Bannera Texture Venusa
FXT1 S3TC JPEGb TRECc Proposition 1, Proposition 2, Proposition 3, Proposition 4,
C R= 6 C R= 6 C R= 8 C R= 10.67 C R= 12 C R= 19.2 Baboon 28.30 27.82 24.94 (15.34) 26.20 (12.16) 25.92 25.03 24.64 23.08 Lenna 34.91 35.03 31.71 (30.27) 31.32 (36.03) 33.16 32.89 32.72 31.51 Backwall 33.14 32.6 29.44 (21.10) 28.87 (24.46) 29.71 29.56 29.45 28.47 Banner 30.26 30.01 25.88 (14.39) 26.89 (12.08) 26.80 26.10 25.81 24.31 Texture 37.97 36.51 44.78 (21.94) 34.93 (24.5) 43.65 42.49 41.44 31.85 Venus 33.44 33.05 29.90 (19.17) 30.36 (17.49) 32.10 31.35 30.90 28.83
aTextures from 3D WinbenchTM2000, 3D WinbenchTMis a trademark of ZD Inc. bJPEG [13]: UV down-sample, and use default quantizer scale and default Huffman table. cTREC [8]: UV down-sample, and use uniform quantization mode.
the MPEG video decoder. The transistor count of current 3D accelerators may be up to twenty mil-lion, and the MPEG video decoder is under 5% of the total chip area. Our JPEG-like TC algorithm can share hardware resources with the MPEG de-coder, for example, the IDCT, inverse quantizer and VLD. High throughput IDCT and VLD tech-niques are available to support high-speed decom-pression, for example, 600 Mpixels/s ISDCT (Lin and Lee 2000) and group-based VLD (Shieh et al. 2001). Even under sequential decoding, this decod-ing speed is sufficient to support DCT-based texture decompression.
4 Conclusion
In this paper, we review the texture compression techniques of previous works and address the re-quirements of texture compression. The compres-sion ratio of current standard S3TC and FXT1 is under 8: 1. A high-compression-ratio technique such as TREC suffers from the texture random ac-cess problem. By analyzing the properties of JPEG, we have proposed a JPEG-like DCT-based texture compression technique to obtain higher compres-sion ratios (> 12). We have solved the random ac-cess problem in TREC. In the proposed technique,
6
7
Fig. 6. 3D rendering result 1 (CR= 16 RGB format) Fig. 7. 3D rendering result 2 (CR= 16 RGB format)
Table 4. Features comparison of texture compression techniques
Kugler [11] Beers [1] PowerVR [5] S3TC [2] FXT1 [3] TREC [8] Proposed
TC type BTC VQ VQ Local palette Local palette DCT DCT
Encoding speed Fast Slow Slow Medium Medium Fast Fast
Decoding complexity Low Low Low Medium Medium Highest High
CR 12 > 8 2−16 4∼ 6 4∼ 8 > 10 > 10
Quality Bad Medium Medium Very good Very good Good Good
we apply localized block adaptive quantization to each 8× 8 block and encode it to target bit stream length. A 4-bit quantizer-scale is encoded with the bit stream in each block. And a bit-count estimation model is proposed to increase the encoding speed for real-time texture encoding. The quality of our technique is at an acceptable level and the compres-sion ratio can be defined by the user. Thus quality control, which is used in video applications, can also be achieved. Applications can choose the com-pression ratio versus quality trade-off depending on the hardware computing power or network band-width. This is important for internet and wireless 3D applications.
Bandwidth and storage are always the issues in var-ious applications. Texture compression is also an issue. High compression ratio and quality control are the trends for future complex 3D applications. Our technique provides both features and brings a video technique into graphics applications. It provides an alternative approach to meet the ever-increasing memory storage and bandwidth requirements.
Acknowledgements. The authors would like to thank their colleagues
within the SI2 group of NCTU and the Multimedia Division of SiS (Silicon Integrated Systems Corp.) for many fruitful discussions. The support from SiS and the NSC is also acknowledged. This work was supported by the National Science Council of Taiwan, ROC, under Grant NSC89-2218-E-009-080.
References
1. 3dfx Interactive (1999) FXT1 texture compression technol-ogy white paper. Available at
http://www-dev.3dfx.com/fxt1/fxt1whitepaper.pdf 2. Hung AC, Meng TH-Y (1991) Optimal quantizer step sizes
for transform coders. In: Chan YT, Venetsanopoulos AN (eds) IEEE Int. Conf. Acoustics Speech Signal Process., Toronto, Canada, IEEE Press, New York, pp 2621–2624 3. Beers AC, Agrawala M, Chaddha N (1996) Rendering from
compressed textures. Proc. SIGGRAPH ’96, pp 373–378 4. Kugler A (1997) High performance texture decompression
hardware. Visual Comput 13(2):51–63
5. Shieh B-J, Lee Y-S, Lee C-Y (2001) A new approach of group-based VLC codec system with fully table pro-grammability. IEEE Trans Circuits Syst. Video Technol. 11(2):210–221
6. Ivanov D, Kuzmin Ye (2000) Color distribution – a new approach to texture compression. Comput Graph Forum 19(3):283–289
7. Knittel G, Schilling A, Kugler A, Straber W (1996) Hard-ware for superior texture performance. Comput Graph 20(4):475–481
8. Igehy H, Eldridge M, Proudfoot K (1998) Prefetching in a texture cache architecture. Proc. 1998 EUROGRAPHICS/ SIGGRAPH Workshop Gr. Hardware, pp 133–142 9. Cheng J-B, Hang H-M (1995) Adaptive piecewise linear
bits estimation model for MPEG based video coding. In: Liu B, Chellappa R (eds) Int Conf Image Process, Wash-ington DC, IEEE Computer Society Press, Los Alamitos, 2:551–554
10. Torborg J, Kajiya JT (1996) Talisman: commodity realtime 3D graphics for the PC. Proc. SIGGRAPH ’96, pp 353–363 11. Ribas-Corbera J, Lei S (1999) Rate control in DCT video coding for low-delay communications. IEEE Trans Circuits Syst Video Technol 9(1):172–185
12. Ramchandran K, Vetterli M (1994) Rate-distortion opti-mal fast thresholding with complete JPEG/MPEG decoder compatibility. IEEE Trans Image Process 3(5):700–704 13. Crouse M, Ramchandran K (1997) Joint thresholding and
quantizer selection for transform image coding: entropy-constrained analysis and applications to baseline JPEG. IEEE Trans Image Process 6(2):285–297
14. Butler M, Pinter-Krainer M, VideoLogic Ltd (1999) PowerVRTMsecond generation white paper of vector
quan-tization texture compression, hardware bump mapping and generalized modifier volumes. Available at
http://www.powervr.com
15. Suzuoki M, Kutaragi K, Hiroi T, Magoshi H, Okamoto S, Oka M, Ohba A, Yamamoto Y, Furuhashi M, Tanak M, Yatak T, Okada T, Nagamatsu M, Urakawa Y, Funyu M, Kunimatsu A, Goto H, Hashimoto K, Ide N, Murakami H, Ohtagu (1999) A microprocessor with a 128-Bit CPU, ten floating-point MAC’s, four floating-point dividers and an MPEG-2 decoder. IEEE J Solid State Circuit 34(11):1608– 1618
16. Microsoft (1997) Escalante hardware overview. Talisman Graph Multimedia Syst, pp 89–106
17. Microsoft (1999) Method and system for accessing texture data in environments with high latency in a graphics render-ing system. United States Patent, Patent Number: 5880737
18. Egger O, Fleury P, Ebrahimi T, Kunt M (1999) High-performance compression of visual information – a tutorial review – part I: still pictures. Proc. IEEE 87(6):974–1011 19. PowerVRTM(1998) White paper: the future of 3D graphics
technology. Available at http://www.powervr.com
20. S3TMInc (1999) White paper of S3TC. Available at http://www.s3.com/savage3d/s3tc.pdf
21. Lin S-T, Lee C-Y (2000) Analysis and design of a high-throughput two dimension inverse scan discrete cosine transform processor. Master’s thesis, NCTU EE, Taiwan 22. Pennebaker WB (1993) JPEG still image data compression
standard. Van Nostrand Reinhold, New York
23. Hakura ZS, Gupta A (1997) The design and analysis of a cache architecture for texture mapping. In: Pleszkum A, Mudge T (eds) Proc. 24th Int Symp Comput Archit, Denver, Colorado, ACM, New York, pp 108–120
CHEN-YI LEE received his B.S. degree from the National Chiao Tung University, Hsinchu, Taiwan, in 1982 and his M.S. and Ph.D. degrees from Katho-lieke University Leuven (KUL), Belgium, in 1986 and 1990, re-spectively, all in electrical engi-neering. From 1986 to 1990 he was with IMEC/VSDM, work-ing in the area of architecture synthesis for DSP. In February 1991, he joined the faculty of the Electronics Engineering De-partment, National Chiao Tung University, where he is currently a Professor. His research in-terests mainly include VLSI algorithms and architectures for high-throughput DSP applications. He is also active in vari-ous aspects of high-speed networking, system-on-chip design technology, very low bit rate coding, and multimedia signal pro-cessing.
CHENG-HSIEN CHEN re-ceived his B.S. degree from the Department of Electron-ics Engineering, National Chiao Tung University, in 1997. Since September 1997, he has been working toward a Ph.D. de-gree in electronics engineering at the same university. His re-search interests include VLSI algorithms and architectures (in-clude 3D graphics and video systems) and memory optimiza-tion for system-on-chip design.
