TW395135B - A high throughput and regular architecture of 2-D 8x8 DCT/IDCT using direct form - Google Patents

Abstract

The invention describes a novel 8x8 2-D DCT/IDCT architecture based on the direct 2-D approach and the rotation technique. The computational complexity is reduced by taking advantage of the special attribute of complex number. Two architectures are proposed: the parallel architecture is suitable for low speed and low power design; the folded architecture is suitable for higher speed and applications of various standards. Unlike other approach, the proposed architecture is regular, hence, it is suitable for VLSI implementation. Compared to the row-column method, the throughput of the proposed folded architecture can achieve two times that of the row-column method with little hardware increased.

Description

Printed by the Consumers' Cooperative of the Central Bureau of Standards of the Ministry of Economic Affairs A7 B7 The present invention is a fast method derived from the direct 2_D met plus 5 method and the rotation characteristics of complex numbers. It is derived from the fast χ W two-transform cosine transform / inverse discrete cosine transform. BACKGROUND OF THE INVENTION Video and video compression have become a major issue in communication technology. Nowadays, many people's livelihood products such as video phones, video conferences, high-quality digital television (HDTV), etc. All need to use data compression technology to reduce the amount of data they transmit in order to meet the channel ’s bandwidth and rain to obtain higher quality. 15 Because of the current standards for many video and video compression, such as JPEG, Η · Z61, MPEG, etc. , Both have adopted Discrete Cosine Transform (DCT) algorithm as the second ring in its compression step, so a fast and efficient discrete cosine transform (DCT) hardware will be necessary to capture future images and The development of video compression, especially when the amount of data such as high-quality digital television (HDTV) is increasing in the future, discrete cosine transformation with high computing power DCT) to meet their needs. In addition, considering the technical aspects of hardware fabrication and very large integrated circuits (VLSi), a regular and modularized hardware and architecture can reduce the required hardware area, and can also shorten the design of the hardware. Time spent with detectives. Comprehensive the above (please read the precautions on the back before filling this page) Cantonese / I -------- Order -------- ϋ This paper size is applicable to China National Standard (CNS) Α4 specification (210X297 A7 B7 Five 'Invention Note' (words' A discrete cosine transform (DCT) algorithm with high output (thtputput > high regularity) will be suitable for its requirements .... Row-column inetHod>-A two-dimensional version of the W discrete cosine transform (DCT) formula can be expressed as ^ 丄 1, t · 4! ____ < χέ =:. = 1. ....— Jill. Ni, n2rkuk2t〇, l, ... n, 1 (!) (Please read the notes on the back before filling out this page) 15 Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs 25 C (0) = 古 与 c ( n) = 1 When n is off, where Km, μ is the input of the discrete cosine transform (DCT), but the output of DCt. Although many algorithms have been proposed, the most commonly used method is to convert two-dimensional data The one-dimensional operation is called the row-column method. A discrete cosine transform is only needed for the row-column method. The operation of the dimensional DCT is complicated. The structure of the columnar method is shown in Figure 1. Most of the discrete cosine transform (DCT) hardware in the past has adopted the row-column method (r0w-c〇1_ thief because of its regularity). ^) Realize its hardware, but its complexity is far greater than the direct method 〇4 This paper jumps to the national standard of China (c is called A4 silk (2 offering 297 meals) —r pack ----- -Order • 1 ^ __ Φ—! 1 ^ ___ I II 』__! 15 Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs 20 (b) Even number 5. Description of the invention (J): 4 弭 Overview 5 The primary objective of the present invention is To reduce the complexity of the conventional technology, we have developed 5-a discrete cosine transform (dct) algorithm using the direct method, and use this algorithm to implement the hardware of the real DCT. The present invention also aims at parallel architecture and folding. Two types of architectures, such as architecture, are suitable for sluggish and low-material design, while the folding 10 architecture is suitable for shouting design and suitable for various video standards. Due to the reduced complexity, relatively, the same hardware The output volume of the data under the body area will be doubled. The algorithm also aims at direct = inherent material contact hiding. Pure analysis and money make the county algorithm know the rules and modularization, very old in the production of the narrative type Lai circuit (coffee worker) hardware. Table 1 shows the age diagram of ki, k2 and a, b Illustrate the structure of the row-column method. Figure 2 shows the age mapping method from ^ 2 to ~, £ when the figure is displayed. Figure 3 Step-One Structure (a) n; L = Odd Number 25 Figure 4 Step 2 Structure Paper size applies Chinese National Standard (CNS) A4 specification (Η0 × 297 mm) (Please read the precautions on the back before filling this page)

A7 B7 V. Description of the invention (4 '). — ~ (A) When owing 2 = 2,1,5 (or 6,7,3) (b) When; c2 = 0,4 Figure 5 8x8 two-dimensional discrete cosine Parallel architecture diagram of DCT transformation Six-shaped 8 Folding architecture of two-dimensional inverse discrete cosine transformation (2-D IDCT) 5 Detailed description of the invention The direct dilemma "y_j, w Diverse cosine transformation (dct) fast algorithm 10 For the sake of convenience and accuracy, we ignore the two-dimensional Shi # Discrete Cosine Transform (DCT) formula (1), the term in the formula, so formula (1) can become "ΛΓ-1 ΛΓ-1 Ykl, k2 = Σ Σ X Βι-0 η2 = 0 β2 cos 2π (2ΰι + 1) ~ 4N cos 1π {2ηζ +1) 1, 4N (2) 15 Next, we assume W =, _〇 = 〇, ι, 2 A 1, η 2 can be converted into κ through proper sequence transformation.

Yn lr n 2 = X2nl, 2n2 Πχ = 〇, ··· N / 2 ^ 1 „n ^ = X2N-2nl-l? N J /. N2 = 〇, ... N / 2-1 x ^ 2n2 ^ i = W2r .... n ~ i 〇2 = L1 n1 == 0, ... N / 2 ^ x ^ :: 'N / 2_ 1 ~ ^ 2N " 2ni'lf 2N-2n0-i ηι = Ν / 2 KT 1 2… · Ν 1 So the formula can be written as 2 1 /2'.....Kη2=N/2,..··N-1

Ykl, k2 Λ ^ 'Ι N ~ l Σ Σ rric iJ2 = 〇al' «2 ㈣--c〇s Ml £ 2 + 1) ^ 2 gate. In order to calculate (3), we can use (4) L ^ J n ^ ~ l M ^ (5) is quickly obtained: U, i, k2 = V1 ^ 1, 1 〇 Knife 2 = 0 and Vfl (4)

A7 B7

D kl, k2 Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs V. Invention Description 1 (5) W 4hIin (t ^ > Rea /… 2)] Because of the symmetrical service system in (5), we do not We need to calculate all the 42 values of the mind. 'Actually, we only need to calculate all the hearts and some allowable 2', as long as any value of owing 2 can be expressed in the mind or the Chinese.

To further simplify the formula (4), we replace ynl, η2 with formula (6) to ynl, t 4n2 +1 = (4t + l) (4ni + l) ia〇d 4N and '〇 $ Hi, Ti2 In the formula SN ~] _), if we fix the value, the relationship with t map will be 1 to 1 'However, if it is a different value, the way of map will be different. Figure 2 illustrates the mapping from Xnl, n2 to yni, t when W 岣. I replaced ^ 2 in formula (4) with ynl, t, and rewritten formula (4) into 15 (7a, 7 called: s \ ttrr AN \ f d.) T Σ: 〇 ^ β1ι / ^ 4 (; ί1 + 1) [" 1 + (4, +,) /: 2ΐ3 (7b) In the formula (7b), the calculation of the sum of the inner layers is very similar, except that Ul + (4t + 1) i2: The term i is not limited to 0 to.; L is within the range of 20. In order to solve this difference, consider the following relation: [^ 1 + (4 t + 1) ^ 2] = a N + bi, and 〇S bs N-1 is an integer, and the office is 〇❹ 一 i Integer. Bring this relation into (7b >), that is to obtain (8): This paper ruler is used for phase_scale (CNS) Α4 specifications (^ 97 mm (please read the precautions on the back before filling in this page)

·· 1 JV- Σ Σ y, χ, w 7 ----, A7 B7 V. Description of the invention (J & J .Ukifk2 = Σ) a [Σ N-: 4ΛΓ 1) 6] (8) If we will The calculation result of the sum of m is recorded as π trb then / ?! = 0 N ~ l

4N 15 Printed by the Consumer Cooperatives of the Central Bureau of Standards of the Ministry of Economic Affairs 20 Σ (c0s ^ l ^ H21_ y-C0S2 £ ll £ Llw ^ = 〇 Poles J tos ^ Uft, b is a complex number, however, we can see from the above relation , Its real part (actual part) is actually a one-dimensional DCT of a point, and its imaginary part has the following relationship with the real part. · = -Re {UVt, b} So, U ' t, b can be realized by one-dimensional DCT with one IV point. The sum operation of another t in formula (3) only needs the addition (subtraction) operation, 'no multiplication is needed. The process of the algorithm is as follows: v 1 * 5 DCT of ten different W groups of W points to go to U't b: yni, t: {y 〇, t 'yi, t, ... y ^ -i, t) * 〇 ^ t ^ Ni 2 * Use the formula to calculate some necessary Umm values. 3. Calculate the value using equation (5). In total, this algorithm only needs to calculate the discrete cosine transform (DCT) of the W group ΑΓ points and then cooperate with the one-touch < addition (subtraction) method; compared with the traditional row-column method, it needs to calculate the DCT of the 2ΑΓ group crying point, which is much less computationally expensive. . This paper size applies Chinese National Standard (CNS) A4 specification (210 X 297 mm) (please read the precautions on the back before filling out this page) -1, · · 1T · 1 A7 B7 V. Description of the invention (7 ). Although the direct algorithm of DCT (DCT) beer algorithm mentioned above requires less computation, it has a more irregular structure and is not suitable for hardware production. Circuit. In the Tao, we aim at two-dimensional discrete discrete cosine transformation (DCT). First, we propose a regular parallel cosine that looks like e). This trunk architecture is very modular and requires a low operating frequency, so it is very suitable for use. For low power design (10w ρ ·· design). In addition, we have appropriately folded the parallel architecture to reduce the area. This folded architecture is suitable for application in various video standards. I line frame when W = 8 _ 'We only need to calculate the value of w2 when allow 2 = 〇, 1,2,4,5', so according to different; the value 'I will expand the formula (7a) as follows: (please (Please read the notes on the back before filling out this page) Order. 15

, 0 = Σ ^, 0 ^ 324 Qiao = 0 (9a) and i / h, y-(472! +1) (^ + 4) = 1… (9b) and ~ printed by the staff consumer cooperative of the Central Standards Bureau of the Ministry of Economic Affairs System Σ (, i); / = 〇-ΦΓ. ^ Il = 0 and ~ = (¾) + · ^ Πΐ, 2 —Λΐ, 6 >-· / (Λΐ4 + * heart, 5-%, 3 (9d ) ^ l ~ 〇 and one · ^, 6 > ^ Ε ^, 1—, Cantonese '7 sheets_ ^

Q The size of this paper is applicable to the Chinese National Code of Practice (CNS) Α4 (21 × 297mm A7 V. Description of the invention (8 B7 l, 5 = Zs, 5 ^ 1 + 1) a '+ 5) (9e) 5 15 Printed by the Consumer Standards Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs 20 According to formulas (9) and (5), we divide the calculation into two steps: Step 1: Calculate Uni, 0, Uju, 4, 2 & gt according to formula (9) & gt ϋη1,: L, and ϋη1,5 values. The architecture of this step is shown in Figure 3. Because the mouth and unl, 5 have a (-ί) spam relationship, when Ώΐ = odd number and, the architecture is slightly the same. Step 2: Calculate u, M, k2, corpse 0, 4, 2, L 5 according to the formula; then calculate 1 to 2 according to the formula (5). When owing 2 = 2 1 Γ (or 6, is 3), the structure is shown in Figure 4 (a); 'When the large 2 = 0, $, the structure is shown in Figure (b). Figure 4 (a) differs from Figure 4 (i) in that Figure 4 (b) requires multiple levels of addition. Figure 5 is the proposed parallel architecture of discrete cosine transform (CT). It can be seen that the entire architecture is quite regular and modular. Only a large area of data connection (gl〇bai interconnection) is required between the two stages. This winding problem still has to be overcome.6 Even so, this part has regularity. Compared with the 2-D DCT architecture proposed by P. Duhame in 1990, N · Cho et al. In 1991, this parallel architecture obviously has a more regular architecture and fewer winding problems. m This paper size applies to China National Standard (CNS) A4 (210X297 mm) (Please read the precautions on the back before filling in this page.

-、 1T ρ · A7 Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs. 5. Description of the Invention (9). I The stacking structure mentioned above has the same design as Xu Xiang, so it can only be built; These same designs use time sl ^ ring 5 to use this module. This is the leap year p Duhame et al. ICASSP ^ 0 Vol. 4 151: 5-1518 and 1991 NI Cho et al. In Gongwaguan Tr as the name · syst • 苐 CAS_ 38 pp. 397-305 published about the two-dimensional discrete cosine transform (24 DCT) can not be done. Next, we will use the 8,8 two-dimensional inverse discrete cosine 10 transform (2D iDCT) as an example to obtain a folded structure. 9 The data of the 8'8 two-dimensional discrete cosine transform (2-D DCT) architecture in Figure 5 The flow direction is reversed, and the phase can get 8'8 two-dimensional inverse discrete cosine transform (2_d IDCT). Now the original parallel frame 4 冓 is disassembled into the original quarter. At this time, there are three points that must be traded: 〇 and L, 4, there must be one level of addition operation, it is not necessary to input other slopes, so you must have a multi-X (multiplexer) to choose. 2. In Figure 5, the operations from $ 1 to 1 ^ 1,1 ^ can be expressed as (1〇) in Λ Un 1 _ η Ιύ (4λι + Π (^ ι + 12). Li Yi ~, 32 and k2 = 1, 2, 5α〇) ^ 1 = 0 Different large 1 and magic, the corresponding a and the value are also different, the relationship between h, large 2 and accounted for, as shown in Table 1. This (1〇) type can be thrown as follows 11 \ '.j r----------- clothing--Μ ^. (Please read the precautions on the back before filling this page} 、 1Τ f. This paper size applies the Chinese National Standard (CNS) A4 specification (21〇 × 297 public A7 B7 V. Description of the invention (10) Implementation: First, calculate the following formula: 0 = 2X ^ 3, + ... Then, according to ^ 1 = 0 a block in Table 1 multiplies part of the output by (-)); finally, according to the block in Table 1, multiply the 8 output values by a qirCuiar shift, which is 5 ° and 仏 1 to U. 々L, k2 can be implemented by inverse operation—multiply by (-j) —to calculate the sum. 3. The data link between step 1 and step 2. According to its regularity, a conversion memory can be used ( tranSp〇se mem〇ry). Figure 6 illustrates the folding structure of this 8'8 two-dimensional inverse-scatter cosine transform (2-P IDCT). The data is divided into four groups of inputs, and after the operation, it is also divided into four groups of outputs The Substage 1 and the nterconnection in the picture are equivalent to the architecture required for the row-column method. The other parts are some butterfly adders and constant multipliers (co nstant 15 multiplier). Compared to the row-column method, the increased hardware area is not much, but the throughput of the entire revolution has doubled. ----------- 7 " -(Please read the notes on the back before filling this page)

、 1T Ρ-. Printed by the Consumer Cooperatives of the Central Bureau of Standards of the Ministry of Economic Affairs 20 t5 ___ 1? This paper size is applicable to the Chinese National Standard (CNS) A4 specification (210X297 mm) A7 B7 V. Description of the invention (11) Table 1 ^ 2 = 〇k2 = l moxibustion 2 = 2 first 2 — 5 left 1 left 1 + no 2 ab ki 4- ab + ab hi + A: 2 ab 0 0 0 0 1 0 1 2 0 2 5 0 5 1 1 0 1 2 0 2 3 0 3 6 0 6 2 2 0 2 3 0 3 4 0 4 7 0 7 3 3 0 3 4 0 4 5 0 5 8 1 0 4 4 0 4 5 0 5 6 0 6 9 1 1 5 5 0 5 6 0 6 7 0 7 10 1 2 6 6 0 6 7 0 7 8 1 0 11 1 3 7 7 0 7 8 1 0 9 1 1 12 1 4 (Please read the notes on the back before filling this page ) Printed by the Consumer Cooperatives of the Central Standards Bureau of the Ministry of Economic Affairs13

This paper size applies to Chinese National Standard (CNS) A4 (210X297 mm)

Claims (1)

Fan Li is requested to print by ABCD 15 20 25 Employees' cooperation of the Central Bureau of Standards of the Ministry of Economic Affairs 30 1 A conversion method of video data, including ... (a) A series of 8X8 original digital signal data is displayed in cosine The reeled format is converted into 8X8 cosine-transformed digital data, which is &transmitted; (b) converts two-dimensional raw digital signal data into one-dimensional data, and rotates in the way of word-parallel bit-serial (C) Complex DCT can be transformed into two 8-p0int one-dimensional discrete cosine transforms (1-D DCT) and eight butterfly additions / (⑴cosine transforms the required rearrangement of digital data output (re_ ordering ) The circuit can be combined with the subsequent Zig_zag circuit. 0 2 · As described in item 1 of the scope of patent application, where the data is input from the original output terminal and the DCT is changed to IDCT, a series of cosine-transformed digital data can be converted. 'Conversion by inverse cosine transform' into the original digital signal data. 3. The method as described in item 1 of the scope of patent application, in which the computational architecture is reduced to a quarter 'from & 1, wee to ynl, antipodal way of t, Middle: (in) Digital signal data is divided into four groups to be entered sequentially, and cosine-transformed digital assets are also divided into four groups to be sequentially output; (b) Structured tribute connection. (Global interconnection) Four 4X4 conversion memories 4. The method described in item 3 of the scope of patent application, where the data is input from the original round and the DCT is changed to IDCT, a series of cosine-transformed digital data can be converted by inverse cosine transform. Into the original digital signal data. -J --- 1 ^ —11 II? I m · nt-I a ... i, (Please read the notes on the back first ^^ Write this page}-* 1T- # 线 · Sheet ") A4M (210X297 / &Jt; J ^ iT Wai Fanli, please make-· ABCD 15 20 25 Printed by the consumer cooperation of the Central Bureau of Standards of the Ministry of Economic Affairs 30 1 A method of converting video data, including ... ( a) Convert a line of 8X8 original digital signal data into cosine-rolled format to 8X8 cosine converted digital data for transmission &; (b) Convert two-dimensional raw digital signal data into one-dimensional Data, rotated in the way of word-parallel bit-serial; month (c) Complex DCT can be transformed into two 8-p0int one-dimensional discrete cosine transforms (1-D DCT) and eight butterfly addition / (⑴ cosine transform digital data output required for reordering (re_ ordering) circuit can be followed by Zig_zag circuit combined with 0 2 · The method described in item 1 of the scope of patent application, where the data is input from the original output and the DCT is changed to IDCT, a series of cosine-transformed digital data can be converted by inverse cosine Way to convert it into raw digital signal data. 3. The method as described in item 1 of the scope of patent application, in which the computing architecture is reduced to a quarter of the mapping from & 1, to ynl, t, where: (in) digital signal data is divided into The four groups are entered in sequence, and the cosine-transformed digital assets are also divided into four groups and output sequentially; (b) The architecture's global interconnection. (Global interconnection) is replaced by four 4X4 conversion memories. 4. The method described in item 3 of the scope of patent application, where the data is input from the original round and the DCT is changed to IDCT, a series of cosine-transformed digital data can be converted into the original by inverse cosine transform. Digital signal data. -J --- 1 ^ —11 II? I m · nt-I a ... i, (Please read the precautions on the back ^^ write this page first)-* 1T- # 线 · 表 纸 ") A4M (210X297 / > J ^ iT