TW200821986A - Graphic processing unit and method for computing sum of absolute difference of marcoblocks - Google Patents

Abstract

A graphics processing unit (GPU) comprising: an instruction decoder configured to decode a sum-of-absolute-differences (SAD) instruction into a plurality of parameters describing an M x N (M, N are integers) pixel block and a n x n (n is an integer) pixel block in U, V coordinates; and sum-of-absolute-differences (SAD) acceleration logic configured to receive the plurality of parameters and to compute a plurality of SAD scores, each SAD score corresponding to the n x n (n is an integer) pixel block and to one of a plurality of blocks that are contained within the M x N pixel block and are horizontally offset with the n x n pixel block.

Description

200821986, IX. Description of the Invention: [Technical Field] The present disclosure relates to a graphics processing unit, and more particularly to a graphics processing unit having image compression and decompression features. [Prior Art] Personal computers and consumer electronics are used in various entertainment products. This f. entertainment items can be roughly divided into two categories: those using computer-generated graphics, such as computer games; and those using compressed video streams, such as pre-recorded programs to digital On a video disc (DVD), or from a cable or satellite industry, digital programming to a set-topbox. The second type also includes the encoded analog video stream', e.g., executed by a digital video recorder (DVR). (Computer graphics are usually generated by a graphics processing unit (GPU). A graphics processing unit is a special type of microprocessor built on computer game consoles and some personal computers. The system is optimized for fast execution of three-dimensional primitive objects, such as triangles, quadrilaterals, etc. These basic objects are described by multiple vertices, each of which has attributes (such as color) and can be Texture is applied to the basic object. The result is a two-dimensional array of 6Client5s Docket N〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/fmal/林环辉/2007/06/13 6 200821986 pi xe 1 s), displayed on a computer monitor or monitor. The encoding and decoding of video poor streams involves different kinds of operations, such as discrete cosine transform (discrete cosine transform) ), motion estimation, motion compensation, deblocking filter (deblocking filter) These calculations are usually handled by a general purpose central processing unit (CPU) in combination with special hardware logic circuits, such as application specific integrated circuits (ASICs). The nose platform meets their entertainment needs. Therefore, it needs a single computing platform that can handle computer graphics and video encoding/decoding. [Invention content]

An aspect of the present invention is a graphics processing unit, comprising: an instruction decoder configured to decode an absolute difference plus total instruction into a plurality of parameters, the plurality of parameters being described on one of u and v coordinates a block and an η X η pixel block, wherein Μ, N, n are integers; and an absolute difference summation acceleration logic circuit configured to receive the plurality of parameters and calculate a plurality of absolute differences plus total values, each The absolute difference plus the total value corresponds to the η χ η pixel block 'and corresponding to the Μ Ν Ν pixel block and has a difference from the η χ η pixel block. Another aspect of the present invention is a graphics processing unit comprising: a main & processor interface for receiving a video acceleration command; and a video acceleration unit for returning the video acceleration command, the video acceleration unit including an absolute difference The jiji speed logic circuit is set to receive the plurality of parameters and calculate a plurality of pairs of ^Client, s Docket N〇.: S3U06-0024 TT's Docket No: 0608-A41261-TW/final/林璟辉/2007/06 八3 7 200821986 • The difference sum value, each absolute difference plus total value corresponds to η χ η pixel squares, and corresponding to the plurality of squares existing in the Μ X Ν pixel square and having a difference from the η X η pixel square one. Another aspect of the present invention is a method for calculating an absolute difference plus a total value of a macroblock, wherein Μ and Ν are integers, and the method includes: performing an absolute difference plus total instruction to calculate a giant a first absolute difference value of a first η χ η portion of the block, the first portion comprising an upper left portion of the χΜ χΜ giant tile, where n is an integer; performing an absolute difference 1 summing instruction Calculating a second absolute difference sum value of the second η χ η portion of the Μ Μ Μ giant block, the second portion including an upper right portion of the ΜχΜ giant tile; accumulating the first and second portions The absolute difference sum is worth a sum; performing the absolute difference summing instruction to calculate a third absolute difference sum value of the third η χη part of the ΜχΜ giant block, the third part containing the Μ χ a lower left portion of the macroblock; adding the third absolute difference plus the total value to the sum; performing the absolute difference summing instruction to calculate a fourth η χ η portion of the one of the Μ Μ Μ giant tiles a fourth absolute difference plus a total value, the fourth part comprising the Μ Μ Μ giant tile A lower right portion; and the total value of the absolute difference plus a fourth added to the sum. [Embodiment] The embodiments disclosed herein provide for the use of a graphics processing unit to enhance the motion estimation system and method. 1. Computing platform for video coding SClienfs Docket N〇.:S3U06-0024 TT,s Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13 8 200821986 1st picture Lay code and / or decoding _ thinning operation block diagram. The system HK) includes a general purpose cpum (hereinafter referred to as a processor), a graphics processor (10) 120, and a memory 13 buffer bus 140. The graphics processing unit 12 includes a video acceleration unit (Kawasaki, which can speed up video encoding and/or decoding, and the moxibustion can be described. The video processing unit of the graphics processing unit 12 can be on the fresh element 12 () The software decoder 160 and the video acceleration driver 17 are located in the memory 13A, and the horse 160 is executed on the main processor 110. The decoder (10) can also be transmitted through an interface provided by the accelerator driver 170. The video acceleration command is sent to the graphics processing unit 120. In this way, the system 1 performs video encoding on the main processor software (graphic processor) of the graphics processing unit A 120 by issuing a video acceleration command. The intensive operation block (3) that is often executed is unloaded by the graphics processing unit 12G, and the more complicated operation is performed by the S processor 110. Several of the explanation for the graphics processing unit are omitted in FIG. 12〇 Video Acceleration is necessary for Asian and African and familiar with the familiar components of this memory. Next, the video coding summary system will be discussed next, then a video coding element How to use the video acceleration unit function provided by the graphics processing unit 120. 2. See the function block diagram of the video encoder ι6〇 of Fig. 1 of the sfl encoder. Input to 9Client5s Docket N〇 .:S3U06-0024 TT's Docket N0:〇6〇8-A41261-TW/fmal/林璟辉/2〇〇7/06/13 200821986 The image of the flat horse state 160 (205) is composed of pixels. The device ι6〇 operates using temporal and spatial similarities within the image 205 and utilizes differential similarity coding within a frame (space) and/or graph & gate. Spatial coding utilizes the same or related characteristic coding of adjacent pixels within an image, so only the difference is encoded. Time coding uses the same value of many pixels of the φ ϋ image towel, so only the difference between the images, 4 codes are encoded as 1β〇, and the statistical redundancy of entropy coding is also used: some images f are more than others Often occurs, so the more common occurrence is represented by a shorter code. Examples of entropy coding include Huffman coding, run-length encoding, arithmetic coding (Ari1: hmetic c〇ding), and two-bit arithmetic coding before and after self-adaptation (c〇ntext-adap1) Ive binary arithmetic coding) In this exemplary embodiment, the blocks of the input image 2〇5 are provided to a subtractor 210 and a motion estimator 22〇. The motion estimator 22 compares the blocks in the input image 205 with a pre-stored reference image 23 to find similar blocks. Motion estimator 220 calculates a set of motion vectors 245 that represent the configuration between the matching blocks. The moving to summer 245 coincidence block 230 with the reference image is collectively referred to as a prediction block 255, representing time encoding. Prediction block 255 is provided to subtractor 21A, which subtracts prediction block 255 from input image 2〇5 to produce a residual image 260. The remaining image 260 is provided to a (DCT ' discrete cosine transform) block 270 and a quantizer 280 that performs spatial encoding. The output of quantizer 28Q (e. g., a set of quantized DCT coefficients) is encoded by entropy encoder 290. lOClienfs Docket N〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/final/林环辉/2007/06/13 10 200821986 For some type of image (280 res1dual) is provided to the internal solution, . The decoding y e^d is combined with the movement H1 remainder, , and generated by the motion estimator 22G. Provided to the mobile estimator 22G, such as the pre-examination image buffer 295, which f is discussed in connection with the first figure, the encoder 16 哭 in the main processing cry (10) hit 'however, the use of the graphics processing unit m : Execution:: = Estimated. The absolute absolute difference provided by the object 〇12〇 is added to the early s(four)b a_ute-dl version (10)) to achieve a positive ‘ at a relatively low amount of computation. The details will be detailed in the mobile estimation, and the dynamic estimation algorithm. 3. The search window (Search Window) As shown in the 3A and B pictures, the mobile estimator 220 will present the current image 2 to 5 Cut into segments that do not overlap, called giant tiles. The size of the giant tile will vary depending on the specifications used by the editor, (2, MPEG-2, H.264, VC) and the image. The exemplary embodiments described herein are small and 16 x 16 pixels in a variety of different coding standards. A giant tile is cut into squares. The larger 11 Client's Docket No.: S3U06-0024 XT's Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13 11 200821986 The small can be 4x4, 8x8, 4x8, 16x8, or 8x16. In MPEG-2, each macroblock can have only one motion vector, so the motion estimation is based on the giant tile. H·264 allows up to 32 motion vectors (depending on the degree), so in H·264, the motion estimation is calculated based on the basis of 4x4 or 8x8 blocks. The change of H·264, called AVS, is always 8χ8. In VCM ' it can be 4x4 or 8x8. The motion estimation algorithm 220 performs a motion estimation on each macroblock in the current image 205, looking for a square in a pre-coded image 230 (which is similar to the giant tile of the current image 205). aims. The permutation between the macroblocks in the reference image "o" in the macroblock Ί4 full image 2 0 5 is calculated and stored as a motion vector (245, Fig. 2).

For ease of illustration, the motion estimation procedure will be described in a particular macroblock in the current image 310 (320). The giant tile 32 selected in this example is in the middle of the current image 310, however the same technique is applied to other giant tiles. The technique _ (330) is in the middle of the giant tile in the McCaw image 230 (corresponding to the giant tile 320 of the current image 31〇). That is, if the macroblock 32 is located at (χ, Υ), then the search window 33〇 in the reference image 230 is also located at (χ, γ) as shown at point 34G. Other embodiments place the giant tile in the reference portion, such as the upper left. In the third example, the search window 33 in the figure extends through the two pixels of the corresponding giant block, and one pixel in the direction of the ^^. Therefore, the search brothers 330 contain 14 different giant blocks: rain 0, ^ ^ where a giant tile finds 1 and 2 pixels respectively, just to the left of position 340. Another fine is not as good as force group The two giant blocks are in position 340 HClienfs Docket No.: S3U06-0024 TT's Docket No: 0608-A41261-TW/fmal/林璟辉/2〇〇7/〇6/13 12 left of 200821986, the remaining group is at position 340 Top, bottom, top left, top right, bottom left and bottom right. The coincident block shift operation performed by the motion estimator 220 uses the absolute absolute difference sum as a criterion for judging the similarity (match) between the macroblocks. The absolute absolute difference is summed to calculate the absolute value of the difference between the two pixel values, and the absolute values of the differences of all the pixels in a square are summed, as understood by those skilled in the art. The mobile estimation method 220 combines the absolute absolute difference summation criterion with the groundbreaking method of selecting the target giant tile of the similarity to be tested, which will be explained below. L Selecting the target giant map The homesickness estimate 240 uses different search methods to generate an internal code (intra-c〇de(j) motion vector or external code (intra-c〇de(j)) according to the motion estimator 220. Inter-coded) The motion estimator 22 utilizes the knowledge of the real world of motion to predict where the coincident giant tile should be in the search window 320, reducing the number of target blocks in the search window 32, It is actually similar to the giant tile 31◦ in the current image 205. In the real world, the object is moving with the solid acceleration, which means that we can expect an object in a frame (optical, optical fl〇w). The movement is tempered and similar (ie substantially connected, ') in both the work and time. In addition, the absolute absolute difference is added to the total surface (ie, in the - search space, the absolute difference plus the total value) The system is expected to be relatively moderate (ie, a relatively small number of local minimum points). Using this knowledge requires a command to search for the most likely place to find the most consistent, and the algorithm disclosed here uses the number of searches to be performed to reduce the number of searches to be performed. Compared with 13Client5s Docket No.: S3U06-0024 TT, s Docket No: 0608-A41261-TW/flnal%_/2Q_6/i3 13 200821986 Best minimum point. So - the algorithm is efficient in calculation and effective Mark the better match. >

4 is an algorithmic flow diagram of an exemplary embodiment motion estimator 220 for computing a motion vector of a current image 205. The motion estimation procedure begins in step 41, where it is determined that the motion vector generated by the motion estimator 22 as the current image 205 will be inter-predicted or intra-predicted. . If intra-image prediction is used, then step 42Q is performed, where a conjugated gradient descent search alg〇rithm is performed to find the intra-predictive giant tile in the search window 32G, which is related to the reference giant tile. (Currently the giant tile 310 in the image 205) is a preferred match. The conjugate gradient descent search (step 420) will be described in detail in conjunction with the fifth and beta plots. Returning to step 410, if inter-picture prediction is used to generate a motion vector, then step 430 is performed, where, adjacent, or,, adjacent, search. The search includes a giant tile adjacent to the current giant tile within the current image 2〇5, and a corresponding macroblock within the first (five) encoded reference image 230. The proximity search cool method (step 430) will be described in detail in conjunction with Figures 7 and 8. The conjugate gradient descent search algorithm (step 41A) and the neighbor search algorithm (step 430) recognize the preferred or accepted conformance in each of the large-group target prediction macroblocks. Those familiar with the art should understand that the criteria for determining how to be a "best match" can be relative or absolute as used in the proximity search algorithm described here. - Absolute criteria: lowest value The target giant tile of (score) is considered to be a better match. Bran HClienfs Docket No. :S3U06-0024 ..., 叩 TT's Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13 200821986 Two descent search algorithm utilization - critical value, absolute difference plus low total value The 篦-: block at the critical value is considered to be a better match. However, the criteria for this threshold are a design or implementation decision. After processing step 42 or 43, to recognize a more healthy match. Step 44G finds the best candidate by the local area exhaustive. The search area is located near the preferred candidate giant block identified in step 420 or r 43\. In some embodiments, after performing the step '420' common light gradient descent search algorithm (ie, in the case of intra-image prediction), the partial thorough «searched light field contains the local minimum recognized in step 420. 4 diagonals near the outside of the (preferred candidate). For example, if the value used in the previous step of the gradient descent is 1, the search is limited to a point away from the preferred candidate (±1, ±1). In some embodiments, after performing step 43 (ie, in the case of inter-picture prediction), a partial thorough search (step 440) for searching for a candidate containing a small area near the preferred candidate giant block, typically Yes (±2+2) 局部 The partial thorough search of step 440 is limited from a preferred candidate giant tile block to a best candidate giant tile block, which is pixehaligned, ie has integer pixel resolution. Steps 450 and 460 find an optimal candidate tile block alignment at a fractional-pixel boundary. The conventional fractional motion search algorithm uses a specific codec-specific filtering algorithm to interpolate the pixel values at the fractional position, depending on the surrounding integer position. In contrast, step 45 〇 establishes that the best candidate giant tile and the reference giant tile match the secondary surface, and step 46〇15Client's Docket N〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/fmal /林璟辉/2007/06八3 15 200821986 Judgment = Table: The minimum value. The minimum corresponds to the best matching giant block, the second knife and the I } number resolution. (The relationship between the scores and the giant blocks is still _ 朗.) The matching macro block of the fractional resolution is in step 45_Forbearing:: Calculated according to the matching giant block - Fractional motion vector丄::: The technology that the ▲ 者 knows. Then the program shed was completed. ★ Those who are familiar with this technique should understand that the above algorithm is essentially 4, because it uses information from nearby areas. Despite the use of hardware acceleration, the meter usually avoids continuous algorithms because many centuries, continuous designs are appropriate here. First, the 'pixel data is read in the form of a sequential raster fashi〇n, and thus can be received in advance, and held in the -circuit buffer H. Secondly, in the implementation of the single-riding difference plus acceleration unit, the implementation of the _ is in the unit is mixed to maintain full load rather than continuous processing. The absolute difference plus the total feed element does not have many cache misses in the prediction block to maintain the load rejection. Since the miss rate is a function of the cache size, the hdtv resolution image only needs 192 〇/8 = <1 Κ β motion vector in the cache, and a low cache miss rate is expected. c. Intra-image prediction within the conjugate gradient only. Figure 5 is a flow chart of the conjugate gradient step 44A of Figure 4, performed by one embodiment of the motion estimator 220. As previously mentioned, the steps are performed when it is determined that the use of intra-image prediction will be used to find that the macroblock in the search window 320 is in a better (i.e., acceptable) manner than the current block 310. The absolute difference plus the total value is a set of 5 initial candidates and is calculated. · The current giant tile, and the current giant tile 16Client's Docket No.: S3U06-0024 TT's Docket No: 0608-A41261-TW/fmal/林璟辉/2007 /06/13 200821986 Large blocks of up, down, left and right. From this initial group of 5 different kinds of magical sleep money. From these two degrees, get the steepest lying direction = domain. If the gradient is relatively shallow, or if the 5 initial candidate giant blocks have a gamma-to-value plus value, then the search extends away from the current giant tile because there is no better local minimum probability in = or The selection of the conditions. After the completion of the step 44Q of the total fortunate, the step is explained in more detail below. The default step 5G5 begins 'initialize here - candidate block ^^ and step two =, △ /. In the embodiment, the candidate giant tile q is set to the search window pull + / upper corner and the step values are all set to a small integer value, such as 8. = In step 51, the seat of the candidate giant block of the week (9) is calculated. The four candidate tiles are the upper, lower, left and right of the candidate giant U. That is, the difference 515 is This calculates the absolute value of the 5 tilted block, plus 4 (the original and the surrounding four). In step 52, calculate the gradient of the gradient and the difference between the left and right resident blocks. The heart is the difference between the absolute difference and the total value of the upper and lower giant blocks. Thus, the error value between 1 and 4 is increased or decreased, and the gradient represents x; y: : in step 525 'the gradient is - The critical value is compared. If the gradient is lower than (the gradient is relatively shallow), the shirt is in the current search for the small value of the interest rate, so the search extends to the new one: = away from the original candidate processing giant block, In a et NolLotlf ^ ^ M ^ ^ ^ ^ lg # ^ Μ ^ ^ ^ ^ ^ Claw Docket Deputy 608 catch 2 Cai w/f_林璟辉咖嶋咖17 200821986 Also "申„亥搜哥. The extension of the search for a new candidate giant block block Ί々 530 530 仃 仃 仃 仃 仃 仃 仃 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 530 Candidate giant block c" around the square corner, distance CA): Yiyi to a 〆々ς) separately perform a common gradient selection giant block (c, TL' now BL, s) = to the gradient comparison of the step coffee, if The ladder calculated at the macroblock 52〇 is again, or the (4) threshold (i.e., the gradient is relative to (4)). The absolute difference plus the total value calculated in step 515 is compared with the -threshold value. If the absolute difference plus total value is lower than the critical value, it means that the step of finding a better phase is touched back to the beer, (in step 545), the beer is provided with a lower absolute difference plus the total value. Candidate giant block. If in step 540, the total value of the pairwise difference is equal to or lower than the S-value value, the table does not find a better match, so adjust the search to 27 55°, select - the new central candidate giant block k new In the central giant (:, hole, 1^' candidate group, the lowest absolute total value of the square is calculated in step 515. Then, in step 555, the new step value △ is calculated from the ladder, and ~' The gradient of the mountain ridge represents an acceptable coincidence of the giant block system, which is far away from the current central candidate, so it increases (, △,). Conversely, the shallow ladder sound is acceptable, and the corresponding macroblock block is very close to the central candidate. Should be less than 18Client's Docket No.: S3U06-0024 TT's Docket No: 0608-A41261-TW/fmal/Lin Huanhui/2007/06/13 1 0 200821986 Different coefficients can be (Δ'Λ). Those skilled in the art will appreciate that various gradients are used to calculate (Δ, Δ,) to achieve the result. Next, the number of iterations is tested at step 560. If the number is greater than a field value, then step 440 is followed by Step 565 $成, can not find the phase: maximum:: near (four) choose a group of 瞧 ^ pick, take, winter pay, the gradient drops Step 440 returns to step 51,

Born a new set. The common gradient descent step 440 is in the following two cases: the value of the step = (=) or the number of maximal iterations (the number of steps (6)' or the number of maximal iterations is not shown in Fig. 6 to illustrate the use of the co-Lee gradient descent step, step 44 〇.妒 Candidate giant block ~ is square (hidden), and four surrounding candidate circles (61〇T, 610L, 610R, 610B). Gradient & a620X, 62GY) are calculated from these initial candidates. In this exemplary state, the gradient is too shallow, and the absolute difference plus the total value is below the relaxation value. Therefore, the search is extended to create four new central candidate giant blocks, shown as triangles (63〇TL, 63〇TR, 63〇BL, 63〇br). These new candidate giant blocks are spaced from the original candidate giant block & The giant tiles around these central candidates are shown as hexagonal (64〇11, 64, 6, 2, 6, 2, 6, 3, 6, 3, 6, 4, 6), selected as Candidate. In this exemplary state, the two candidates 640 have an absolute difference plus a total value below the threshold and a "steep" gradient (650XY, 660XY). Another candidate is selected according to each, steep, gradient: candidate 670 is based on a gradient of 65 〇χ γ, and then 68 〇 is based on a gradient of 660 ΧΥ. The gradient descent search continues to use these new candidates 6?, 68, according to the conjugate gradient drop step 440. 19Client's Docket N〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/final/林璟辉/2007/06/13 200821986 The 7th picture is the 4th near-search calculus implemented by the mobile estimator 22() In the example of Figure 4, the macroblock contains a giant tile adjacent to the current series 2〇5 2 as described above, the search code. Also included as a candidate H current giant tile 31G (has been associated with a corresponding giant tile. & the sequence of the pre-coded reference image 1 different candidate giant tile block coordinates starting from step 7iq, in This is calculated by using the absolute value (remainder) of the current giant tile 310 address and the number of macroblocks per row - the flag Wei T〇PVALID. If the absolute value is not G, then aud is true, and in addition, T0PVALID is False. In step 72, the flag variable LEFTVAUD is calculated by dividing the current giant block 31G address by the integer and the number of macroblocks per line. If the divisor is not G' then LEFTMLID is true, in addition, LEFTVAUD is False. These T0PVALID and LEFTVALID variables indicate that the current giant block 31〇 has a neighboring giant tile on the top and the left side, respectively, considering the upper edge and the left edge of the giant tile. In step 730, the T0PVALID and LEHTALID variables are used in combination to determine the current giant. The availability or existence of the four candidate giant tile blocks adjacent to block 310. In particular, there is a giant tile L (LEFTVALID) on the left side; there is a giant tile T (TOPVALID) on the top; Block TF if (TOPVALID &LEFTVALID); The block TR is (TOPVALI D& RIGHTVALID). Next, at step 740, the availability is determined for a previous candidate giant tile p, which is spatially corresponding to one of the previously encoded reference images 230 of the current giant tile 310. Giant block. The relative positions of the five candidate giant blocks can be seen in pictures 8A and B, where L system 810, T system 820, TL system 830, TR system 840, P system 850. 20Clienfs Docket N〇 .:S3U06-0024 TT's Docket N(K〇608-A41261-TW/fmal/林璟辉/2007/06/13 200821986 How many candidate giant blocks can be added to the difference than the number of steps 73 and 740. 5 Calculate the total number of handles for each-available, and add the total difference: L + Ts 〇, ΑΓ, ρ, med(T, TL, m)\ handle is available, familiar with this The person of the skill should know the candidate block of the group. Step 430, the reply has the lowest absolute difference plus the total use ΐ ί:; , discussed, once the matching giant block is found (regardless of the use of the brother _« Search for the wrong or the 5th _ total - limit, using a partial thorough search (the first == =: using the results of the partial thorough search - score movement to = number moves to 1 The calculations will be detailed below. The score of the model in the knife Jiang Shun who is familiar with this technique should be familiar with the degree of search for the "flaw surface". _ Ingenious method, shifting estimates 22G establishes a model of the wrong surface with a secondary surface and analyzes the minimum value of the surface with soil pixel accuracy. The mobile estimation unit 220, the minimum value of the first item, the given-minimum line. The mobile detector 220 connects the minimum value of the orthogonal direction along the t lines. Think of the general equation of the quadratic curve as Equation 1. 21Client's Docket N〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/fmal/林璟辉/2007/06八3 200821986 y = q + C2t + C/ Equation 1 Differentiate the curve, as in Equation 2: Faint = + = Equation 2 Once the coefficient c" c2, c3 is known, it can be solved to determine t, the smallest position. The motion estimator 220 solves Equation 3 to determine the coefficients c15 c2, c3.

Cl C2 C3 /._V ΣΣ, X - 27% 5127 Σ Equation 3 The motion estimator 220 uses the 84 absolute difference value provided by the graphics processing unit 120 to add the total efficiency of the command calculation equation 3. Each 4 represents an absolute difference plus a total value, and i is added to represent the absolute difference plus the total value of the adjacent macroblocks in the X direction. As described in detail in FIG. 1, the 8x4 absolute difference summing command is valid (the rate is calculated by the adjacent giant block (x, y), (x+l, y), (x+2, y), (x+3,y), the four absolute differences plus the total value, ie izz()".3 and i = t=j + l. As mentioned above, once the coefficient is known, solving equation 2 yields t , the minimum value in the X direction. Equation 3 can be used to determine the minimum value t in the vertical direction. In this example, the motion estimator 2 2 0 uses the 8 X 4 absolute difference plus the total efficiency of the instruction. The total absolute difference between the four macroblocks (x, y), (x+1, y), (x+2, y), (x+3, y) plus the total value. Equation 3 solves the calculations from these The coefficient of the absolute difference plus the total value Ci, C2, C3. As mentioned above, once the coefficient is known, the solution of Equation 2 is 22Clienfs Docket No.: S3U06-0024 TT's Docket N〇: 0608-A41261-TW/fmal/林璟辉/2007/06/13 22 200821986 The minimum value in the direction of t, y. The method of moving the estimation method is more expensive than the second error table on the first-order boundary. The advancement of the familiar method of paying for it.

Tomb _min ^~ As mentioned earlier, the motion estimator 22 is in the current image

The giant figure in the predicted image has a relatively simple payout. The motion estimator 220 uses the 7-shaped metric unit 12 (the absolute difference provided by 3 plus the total hardware acceleration, which is the graphics plus the 7 〇 command. The absolute difference plus the total command is to be input - 4 χ 4 reference blocks and one 8x4 predicts the block and produces 4 absolute difference plus total values. The size of the reference block and the prediction block can be changed as needed. The 4 χ 4 reference block and the ( 4) prediction block are merely examples to illustrate the present invention, and should not limit the reference block and Predicting the size of the block. Figure 9 is a block diagram showing the operation of the absolute difference plus total instruction for the reference and prediction blocks. As shown in Figure 9, the 8χ4 prediction block consists of multiple horizontally adjacent 4x4 blocks that overlap each other. The composition is as in blocks 910, 920, 930, 940. The absolute difference summing unit takes an input 4Χ4 reference block 950 and calculates the absolute difference plus the total value of the reference block and 910-940 squares. The value summation instruction calculates four values: one value is the sum of the absolute values of the difference between block 910 and block 950; the other value is the sum of the absolute values of the difference between block 920 and block 950; the other value is block 930. versus The sum of the absolute values of the differences of block 950; the other value is the sum of the absolute values of the differences between block 940 and block 950. Referring to Figure 9, the absolute difference in the graphics processing unit 120 is summed up by 23Clienfs Docket N〇 .:S3U06-0024 TT's Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13 23 200821986 Single 兀 uses 4 absolute difference plus total calculation unit (960, 970, 980, 990) The absolute difference is added to the total command. The leftmost 4χ4 box 91〇 is supplied to the absolute difference plus the total unit 96〇. Then the 4χ4 square (92〇) on the right is input to the absolute difference total calculation unit 97〇. Enter the 4χ4 square (93〇) on the right to add the total difference calculation unit 98〇 to the absolute difference. Finally, provide the rightmost 4χ4 square 940 to the absolute difference total calculation unit 9(10). The graphics processing unit 12〇 uses independent absolutes in parallel. The difference sums up the calculation unit, so the absolute difference plus total instruction produces 4 absolute differences plus the total value per ρ cycles. Those skilled in the art should be aware of the absolute difference used to count two identically sized pixel blocks. Value plus total operation The method, and the hardware design used to perform this operation, will not be described in detail. The 4x4 reference block is horizontally and vertically listed at the edge of the pixel. However, the 4x4 prediction blocks 910-940 need not be corrected vertically. In one embodiment, the data is corrected by rotating (logic circuit 995) the reference block. Rotating the reference block instead of rotating the 4 prediction blocks separately saves the number of logic gates. The rotated reference I block is provided for each individual The absolute difference adds up the hardware acceleration unit. Each unit produces a value of 12 bits, and these values are combined into a 48-bit output. In one embodiment, the magnitude of these values is based on the U texture coordinates of the predicted block (the lowest coordinate in the lowest bit position). The following code shows that 8x8 squares, that is, two adjacent 8x4 squares, can be calculated using only 4 absolute difference plus total instructions. Temporary Τ, Τ, Τ, Τ 4 series are used to temporarily store the four absolute differences plus the total value. The variable sadS is used to accumulate these absolute differences plus the total value. The address of the 8x4 reference block is assumed to be 24Clienfs Docket N〇.: S3U06-0024 TT’s Docket No: 0608-A41261-TW/fmal/林璟辉/2007/06A3 24 200821986 refReg.纹理 and V are 8x8 prediction block texture coordinates. The following code produces the total absolute difference plus the total value of the entire 8x8 block, stored in sadS. SAD Tl, refReg, U, V ; left-top of 8x8 prediction block SAD T2, refReg, U+4, V ; right-top of 8x8 prediction block

ADD sadS, Tl, T2 SAD T3, refReg, U, prediction block ADD sadS, sadS, T3 SAD T4, refReg, U+4 8x8 prediction block ADD sadS, sadS, T4 V+4 / left-bottom of 8x8 , V+ 4 ;right-bottom of However, it is usually possible to avoid calculating and summing the values of all four sub-blocks, as the calculation can be stopped as long as the sum reaches the current minimum. The following virtual code shows how to use the absolute difference plus total instruction in a loop, which stops when the sum reaches a minimum. WORK:=〇; SUM := 0; MIN = currentMIN; WHILE (work < 4 II SUM < MIN) SUM := SUM + SAD(refReg, U+(I%2)*4, V+ (I>〉1 *4); IF (SUM < currMIN) currMIN = MIN; 25Client's Docket N〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13 200821986 G〇to Next Search The 84 absolute difference summation command in the graphics processing unit 120 is directly used by the advanced search algorithm of the sway estimation state 220, for example, the partial partial search in the $FIG. In addition, the texture cache (筮Ί, 罘iU map) is block corrected, and the algorithm used by the motion estimator 220, as described above, is pixel corrected. Although multiplexer units can be added to the graphics processing unit 12 to handle these correction errors, doing so increases the number of logic gates and power consumption. Instead, the graphics processing unit 120 uses these extra budgets to four absolute difference summing units instead of just one. In some embodiments, the absolute difference summation instruction provides the advantage of efficiently computing the minimum value, which involves calculating the absolute difference sum value of adjacent blocks. In some embodiments, the 8χ4 absolute difference summation instruction provides another advantage of a thorough search (block 440), which calculates the absolute difference plus the total value for each diagonal when the step value is one. 4. The graphics processor has discussed the implementation of the soft algorithm of the mobile estimator 220 and the use of the 8x4 absolute difference summing instruction in the graphics processing unit 120, followed by a detailed description of the absolute difference summing instructions and The graphics processing unit 12 is. a. Graphics Processing Unit 湳 Figure 10 is a data flow diagram of the graphics processing unit 120, wherein the instruction stream is indicated by the arrow to the left of Figure 10, and the image or graphics stream is represented by the arrow to the right. Figure 10 omits several components familiar to those skilled in the art. These pairs are 26Client's Docket N〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/fmal/林环辉/2007/06/13 200821986 It is not necessary to interpret the deblocking feature in the loop of the graphics processing unit 12〇. An instruction stream processor 1010 receives an instruction 1020 from a system bus (not shown) and decodes the instruction to generate instruction material 1030, such as vertex data. Graphics processing unit 12 supports a conventional graphics processing instruction and instructions for speeding up video encoding and/or decoding, such as the aforementioned 8x4 absolute difference summing instructions. ‘Knowledge graphics processing instructions involve issues such as vertex shading, geometry shading, and pixel shading. Therefore, the instruction data 1〇3 is applied to the pool 740 of shader execution units. The shading execution unit necessarily uses a texture filter unit (TFU) 750 to apply a texture to a pixel. The texture data is taken from texture cache 1〇6〇, which is behind the main memory (not shown). Some instructions are sent to the video processing unit 1100, the operation of which will be described later. The resulting data is then processed by a post-wrapper (p〇st-packer low) which compresses the data. After post-processing (p〇st_pr〇cessing), the data generated by the video acceleration unit is provided to an execution unit pool 1040. The execution of the video encoding/decoding acceleration instructions, such as the aforementioned absolute difference summing instructions, is different in many respects from the conventional graphics instructions described above. First, the video acceleration command is executed by the video processing unit 1100 instead of the shader execution unit. Second, the video acceleration instructions do not use their texture data. However, the image data used by the video acceleration command and the texture data used in the graphics instructions are both 2-dimensional arrays. The graphics processing unit 120 also utilizes this advantage, and uses the texture filtering unit 1〇5〇 to download to the video processing unit 27Clienfs Docket N〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/fmal/林环辉/2007/06 /13 27 200821986 φ 1100 image data, thus making the texture cache 1G6Q cache - some image data operated by the video processing unit 1100. Therefore, as shown in the first diagram, the video processing unit 1100 is located between the texture filtering unit 1〇5〇 and the post-packer 1070. The texture filtering unit 1050 checks the instruction data l〇3 extracted from the instruction 1〇2 (Whether the data 1G3G further provides the coordinates of the image data desired by the texture filtering unit 1〇5Q main memory (not) β. In one embodiment, these squat marks are U and V pairs. Those who are familiar with the art should be familiar with this. When the command 1020 is a video acceleration command, the command data is commanded to command the texture filter and the wave unit 1050. Any texture filter (not shown) in the texture filtering unit 1〇5〇 is skipped. Therefore, the texture filtering unit 1〇5〇 is controlled by the video acceleration command to download the image data to the video processing unit 11〇〇. The texture filtering and wave unit 1050 is manipulated as a video acceleration command to download the image I data to the video acceleration unit A 11. The texture filtering unit 1〇5〇 on the data path of the processing unit 1100 receives the image data. (; and the command data 1030 on the command path, and perform - operation on the image data according to the command data 1030. The image data output by the video processing unit n (10) is fed back to the execution single S pool 1_, in the post wrapper m After processing b. The instruction parameters now say that the processing unit of the dragon processing unit (recording) (4) the difference sums up the operation of the video acceleration instruction. As previously explained, each graphics processing unit instruction is decoded and parsed into the instruction material 1030. It can be regarded as the specific reference of each instruction 28Client, s Docket N〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/final/林璟辉/2007/06/13 28 200821986 Number set. Absolute difference plus total instruction parameters Shown in the first table. Table 1: Absolute difference of the graphics processing unit plus total instruction input / output name size _____________^——- Narrative input FieldFlag 1-bit if FieldFlag 22 1 Field Picture, the rest Frame Picture Input TopFieldFlag 1-bit if TopFieldFlag 1 1J Top-Field-Picture, Other Bottom-Field_Picture If FieldFlag is input Input PictureWidth 16-bit For example: 1920 for HDTV Input PictureHeigh t 16-bit For example: 1080 Input BaseAddress 32-bit unsigned predictive picture base address input BlockAddres s U: 16-bit signed V: 16-bit Signed Predicted Picture Texture Coordinates (Related to Basic Address) In SRC1 Opcode SRC1[0:15] = U? SRC1[31:16] = VU, V is 13.3 format, ignore fractional part 29Client's Docket No.: S3U06- 0024 TT's Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13 29 200821986 Enter RefBlock 128-bit reference picture data"-------Destination 4x16-bit in SRC2 Opcode output-- The least important 32-bit Operand in the 128-bit scratchpad is in DST Opcode

A plurality of rounding parameters are used in combination to determine the 4x4 block address captured by the texture filtering unit 1〇5〇. The BaseAddress parameter indicates the starting point of the texture metric in the texture cache. The upper left square of this area is given to (4) this smart parameter. Pic-tureHeighl; and PictureWidth input parameters are used to determine the range of the square, the lower left coordinate. Finally, the video graphics can be progressive progessive or interlace. For interlaced scanning, it consists of two directions (upper and lower). Texture filtering unit use

FieldFlag and TopFieldFlag to properly interlace the scanned image. c. The image data is converted into an absolute difference summing command, and the video processing unit 11 操 manipulates the input pixel block from the texture filtering unit 105 and converts the block into a suitable format for absolute use. The difference plus 'total acceleration 960-990 processing. The pixel square is then provided to 0 Ί 力 矣 $ $ unit 960-990, which replies the absolute difference plus the total value. ..., 1 minute 对 对 佶Λ έέ ι 值 值 值 值 值 值 值 值 值 值 值 值 值 值 值These functions will be received in the first 1, _ video processing unit Π00 to define two input parameters that calculate the absolute difference plus 8x4 squares. Reference block data 仫 SRC2 operation code is directly defined: 8Χ4Χ8 bit squares are treated as 128-bit straight:: material. In contrast, the SRC1 operation code defines the address of the prediction block instead of = poor material 30Clienfs Docket N〇.: S3U06-0024 TT's Docket No: 0608-A41261-TW/final/林璟辉/2007/06/13 30 200821986 Video Processing Unit The 1100 provides these addresses to the texture filtering unit 1〇5〇, which fetches 128-bit prediction block data from the texture cache 1060. Although the image data contains luminance (Y) and chroma (Cb, Cr) planes, motion estimation typically uses only the gamma component. Therefore, when the absolute difference summing instruction is executed, the pixel block operated by the video processing unit 1100 contains only components. In one embodiment, the video processing unit 11 generates a disable flag that directs the texture filtering unit 1050 not to fetch the Cr/Cb pixel data from the texture cache iQ6 class. The figure is a block diagram of the texture filtering unit 1050 and the texture cache ι〇6〇. The texture filtering unit 1050 is designed to extract the texel boundry from the texture cache logo and download the 4x4 texture image block from the texture cache to the filter input buffer mo. When the captured data represents the video processing unit 1100, the texture image 112 is considered to be 4 channels (ARGB) of each punctured bit, for a 128-bit texture image size. When the data is extracted for the absolute difference plus total instruction, the texture filtering unit 1〇5〇 ^ : downloads the 8x4x8 bit block. To handle the correction problem, the 8x4 block is downloaded to two 4X4 pixel input buffers (1110A and 1110B). The image data that is not used by the video processing unit 1100 may be corrected by the byte. However, the texture filtering unit 1050 is designed to extract texture image boundaries from the outside. Therefore, when the data is captured by the video processing unit 1100, the texture filtering unit 1〇5〇 may need to capture up to 4 texture image correction 4×4 blocks around a specific byte correction 8×4 block of each 4×4 half block. . 6 Hai program can be seen in the brother 11 picture, in which the left half 4X4 block (target 31Client5s Docket N〇.: S3U06-0024 TT's Docket No: 0608-A41261-TW/final/林璟辉/2007/06/13 31 200821986 - Block 1130) is aligned on the texture image boundary, either in the vertical direction or in the horizontal direction. In other words, the block 113 0 extends two pixel images. The U and V addresses of the target block 1130 define the top left corner of the 4x4-8 bit, the byte correction block. In this example, the texture filtering unit 1〇5〇 determines that the images 1140, 1150, 1160, 1170 should be fetched to obtain the target block η3〇. After the determination, the texture filtering unit 1050 retrieves the block 丨 〇 17 17 and then combines the rows and columns selected by the bits from the blocks 1140-1 170 to the target side (the leftmost 4x4 bit of the block 1130 is written to the filter). Buffer m〇B. Those familiar with the art should know how to use multiplexers, shifters, mask bits to achieve this result, not to get from the texture cache 1〇6〇 The 4 χ 4 target correction is taken. In the embodiment shown in Fig. 11, when the target block 113 〇 contains a vertical texel boundary, the data is not rearranged vertically. In this case, the birthday ^ is downloaded to the filter buffer 111 The order of the data of 〇A and ill(10) is different from the original order in the cache. In this embodiment, the t-staff 11QG must vertically rearrange (rotate) 128 bits> Predicting the order of the blocks. In another embodiment, before the -filter buffer 111G, the texture filtering unit 1〇5〇: directly rearranges the cache texture image data to conform to the original cache group, =1 The block in the clear or flow chart should It is understood to mean the modulo # or two: ° " & code" which contains one or more executable instructions for implementing a specific logic circuit function. Those skilled in the software door should be #_, other implementations The method is also included in the scope of the disclosure of 2008 2008. In the other implementation methods, the functions may be performed in the order in which they are performed or disclosed, including substantially synchronous or reverse, depending on the function involved. The disclosed system and method can be implemented in software, hardware or a combination thereof. In some implementations, the system and/or method is implemented in software stored in a memory and is suitably processor located in a computing device. Executed (including but not limited to a microprocessor, microcontroller, network processor, re-assemblable processor, scalable processor). In other embodiments, the system, system, and/or method is logical Circuit implementation, including but not limited to a PLD 'programmable logic device, a programmable gate array (PGA), a field programmable logic gate array (FPGA, f Ield pr〇grammable卯忭array) or an application specific circuit (ASIC). In other embodiments, these logic statements are performed in a graphics processor or graphics processing unit (GPU). The systems and methods disclosed herein can be embedded Any computer readable medium (using or linking to an instruction execution system, apparatus, or device. The instruction execution system includes any computer-based system, a processor-containing system, or other system that can be retrieved from the instruction execution system A system for executing these instructions (c(10)puter-readable medium) can be any tool that can hold, store, communicate, transfer or transfer the program for use as or in connection with the execution system of the instruction. The computer can be a system or transmission medium such as (unrestricted) electronically based, magnetic, optical, electromagnetic, infrared or semiconductor technology. 33Client5s Docket N〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13 33 200821986 Specific examples of computer-readable media using electronic technology (non-restrictive) may include one or more electrical (electronic) connections Line; a random access memory (RAM); read-only memory (ROM, read-only memory); a wipeable programmable read-only memory (EPROM or flash memory) . Specific examples of computer readable media that use magnetic technology (without limitation) may include: a portable computer diskette. Specific examples of non-limiting media that use optical technology for fine-readable media may include: a fiber optic and a portable CD-ROM. The invention is illustrated and described herein by way of example only, and is not intended to Many different modifications and structural changes are made within the scope and scope of equalization. Therefore, it is preferable to interpret the scope of the appended patent application broadly and in a manner consistent with the field of the invention, and to make this statement before the scope of the subsequent patent application. BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a block diagram of an exemplary computing platform for graphics and video coding and/or decoding. Figure 2 is a functional block diagram of the video encoder 16A of Figure 1. The brothers 3 A and B illustrate the division of the witness image into giant tiles of non-overlapping segments. Figure 4 is a flow diagram of an exemplary embodiment of an algorithm used by the motion estimator of Figure 2. 34Client5s Docket N〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/fmal/Lin Huanhui/2007/06/13 200821986 Figure 5 Figure 4 Conjugate Gradient Step Figure 6 illustrates the use of Figure 5 A total of the flow chart of the embodiment. A flowchart of one embodiment of a common light gradient descent step 440. Figure 7 is the adjacent position of the search algorithm in Figure 4, the relative position of the selected giant block.

The 5th order of the absolute difference plus the total refers to the block diagram of the reference and prediction block operations. Figure 10 is a data flow diagram of the graphics processing unit of Figure 1. Figure 11 is a block diagram of the texture filtering unit and texture cache of the 1Gth image. [Main component symbol description] 1〇〇~ system, 110~ main processor, 12〇~ graphics processor (Gpu), . Fe, body, 140 to bus, 150 to video acceleration unit (Vpu), ι6〇 ~ software decoder, 170 to video acceleration driver. 205~image, 210~subtractor, 220~moving estimator, 230~reference image, 245~moving vector, 255~prediction block, 260~remaining image, 27〇~ discrete cosine converter, 280~ Quantizer, 290~entropy decoder, 21〇〇~ decoder. ^Client^ Docket N〇.:S3U06-0024 s Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13 310~The current giant block, 320~ giant block, 330~ search window, 340~ point. 35 200821986 纛, 400~ program, 410~ determine the motion vector will be inter-image prediction or image inside, measure, 420~ perform conjugate gradient descent search algorithm, 430~ execute adjacent search brother 440~ perform a partial area thoroughly Search, 'π~ establish the best candidate giant block and the reference giant block to match the degree of secondary surface, 460~ find a best candidate giant block alignment at a fractional pixel boundary, 47〇~ according to the match giant The tile calculates a fractional motion vector. 505~ initialize a candidate block, 51〇~ calculate the candidate giant block^, the coordinates of the surrounding candidate giant block, 515~ calculate the absolute difference sum of the five candidate giant blocks, respectively, 520~ calculate the gradient meaning and heart 525~The gradient is lower than a critical value, 530~ calculates the coordinates of the four new candidate giant block, 535~ performs the conjugate gradient lowering step 44〇, 54〇~ compares the absolute difference plus for each candidate giant block respectively Whether the total value is lower than a critical value, 545~return the candidate giant block with the lowest absolute difference plus the total value, 550~select a new central candidate giant block, 555~from gradient & The step value Δλ· and ~, 560~ test whether the number of iterations of the loop is greater than a maximum value, and the 565~back pass does not match. 610C~candidate giant block, 610L-610R-610Τ-610Β~four surrounding candidates, 620Χ-620Υ~ initial candidate calculation gradient, 630TL-630TR-630BL-630BR~ four new central candidate giant blocks, 640L-640R -640Τ-640Β~candidate, 670-680~candidate 710~ Calculate a flag variable TOPVALID using the absolute value of the current huge block 310 address and the number of macroblocks per line. If the absolute value is not 〇, then TOPVALID is true. In addition, TOPVALID is false 36Client's Docket N〇.:S3U06-0024 TT,s Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13 36 200821986 720 The ~flag variable LEFTvaLID is calculated by dividing the current giant tile address by the integer and the number of macroblocks per line. If the divisor is not 〇, LEFTVALID is true, and LEFTVALID is false. 730~ combines the T0PVALID and LEFTVALID variables to determine the availability of the four candidate giant tiles adjacent to the current giant tile. 740~ determines the availability of a previous candidate giant block p. 750~ Calculate the absolute difference sum for each available candidate block. 810-850~ Candidate giant block. 910-940~4x4 blocks, 950~4x4 reference blocks. 234~rotation logic, 950~prediction block, 960-990~absolute difference sum total calculation unit, 1010~ instruction stream processor, 1020~ instruction, 1030~ instruction data, 1040~execution unit pool, 1050~texture filter unit, 1060~ texture cache, 1〇7〇~ post wrapper, 1100~video processing unit. 1120 ~ texture image, Π 30 ~ target block, 1140-1170 ~ texture image, 1110A-B ~ buffer. 37Client’s Docket No.:S3U06-0024 TT’s Docket No:0608-A41261-TW/final/林璟辉/2007/06/13 37

Claims (1)

200821986 - X. Patent application scope: 1. A graphics processing unit, comprising: an instruction decoder, configured to decode an absolute difference plus total instruction into a plurality of parameters, the plurality of parameters being described on U and V coordinates a pixel block and a η X η pixel block; and an absolute difference summation acceleration logic circuit configured to receive the plurality of parameters and calculate a plurality of absolute difference plus total values, each absolute difference plus a total value pair f It should be η X η pixel squares, and corresponding to the Μ X Ν pixel square and have a difference from the η X η pixel square. 2. The graphics processing unit of claim 1, wherein the absolute difference summation acceleration logic circuit further comprises: a plurality of absolute difference sum total calculation units, each absolute difference total calculation unit is configured to receive the ηχη a pixel block, and receiving one of the plurality of blocks included in the 像素 pixel block, and calculating a corresponding one of the plurality of absolute difference sum values. 3. The graphics processing unit of claim 1, wherein the parameter of the Μ X Ν pixel block defines an address of the Μ X 像素 pixel block in a texture cache. 4. The graphics processing unit of claim 1, wherein the parameter describing the pixel block is defined in one of the texture caches as one of the relative pixel addresses. 5. The graphics processing unit of claim 1, wherein the difference is a horizontal difference. 38Client's Docket N〇.:S3U06-0024 TT,s Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13 38 200821986 ' 6. The graphic processing unit of the application scope of the patent application, The % X Ν pixel square represents a motion estimation prediction block, and the η χ η pixel square represents a motion estimation reference block. 7. The graphics processing unit of claim 2, wherein the number of absolute difference plus total calculation units can process data in parallel. 8. The graphics processing unit of claim 2, further comprising a first logic circuit for accumulating the plurality of absolute difference plus total values to the _ _ _ _ _ _ _ _ $9. The graphics processing unit of claim 8, wherein the first logic circuit stores the plurality of absolute difference plus total values in an order into the target register, the order being The square of the ΜχΝ: and the decision. $10· The graphic processing unit of claim 2, further comprising: a texture cache memory, configured to store the pixel data in a texture image format having a predetermined number of bits; and, a texture filtering unit, configured to Determining whether the MxN pixel block extends a texture image boundary, and correspondingly extracting one or more texture image correction η χ squares around the χ 像素 N pixel block from the texture cache memory, and combining and correcting from the texture image The η χ η block is selected by the bit 疒 and column such that the leftmost bit is written to the first filter buffer and the rightmost bit is written to the second filter buffer. 11· A graphics processing unit, comprising: a main processor interface, receiving a video acceleration command; and 39Client5s Docket N〇.: S3U06-0024 TT, s Docket No: 0608-A41261-TW/fmal/林璟辉/2007/06/ 13 39 200821986 A video acceleration unit, which responds to a video acceleration command, the video acceleration unit includes an absolute difference summation acceleration logic circuit configured to receive the plurality of parameters and calculate a plurality of absolute difference plus total values, each absolute difference The value plus value corresponds to the η X η pixel block, and one of a plurality of blocks corresponding to the Μ XN pixel block and having a difference from the η χ n pixel block. 12. The graphics processing unit of claim 11, wherein the absolute difference summation acceleration logic circuit further comprises: a plurality of absolute difference sum total calculation units, and each absolute difference sum total calculation unit is configured to receive the ηχη a pixel block, and receiving one of the plurality of blocks included in the 像素 pixel block, and calculating a corresponding one of the plurality of absolute difference sum values. 13. The graphics processing unit of claim 12, wherein the plurality of absolute difference sum total calculation units can process the data in parallel. 14. The graphics processing unit of claim 11 further includes a first logic circuit for accumulating the plurality of absolute difference plus total values to a target register. 15. If the first processing circuit of the graphic processing unit of the nth item of the patent scope is applied, the plurality of absolute difference plus the total value is stored in the target register in the order, and the order, the order It is determined by the 1; coordinates of the squares in each of the ridges. 16·If the graphic processing unit of the application scope patent item u is more packaged, the texture memory format is set to the pixel data storage item and the number of bits; and 40Client5s Docket N 〇.:S3U06-0024 TT's Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13 40 200821986 A texture filtering unit is set to determine whether the MxN pixel block is applied, the text of the figure is corresponding, and corresponds to Extracting one or more texture image correction η χ n squares around the M χ N pixel square from the texture cache memory, and combining the rows and columns selected by the # yuan from the texture image correction ηχn square, so that the most The bit on the left is the first filter buffer and the rightmost bit is written to the second filter buffer. 1 is the graphics processing unit of claim U, wherein the parameter describing the ΜXN pixel block is defined as one of the 像素N pixel blocks in a texture cache and a relative address. I8. The graphics processing unit of claim 11, wherein the parameter describing the NX N pixel block directly defines the pixel data. 19. The graphics processing unit of claim U, wherein the ΜχΝ pixel square represents a motion estimation prediction block, and the “n pixel square represents a motion estimation reference block., 2〇· a calculation Μ X A method of swaying a giant block to a gentleman, and adding a total value to the difference, wherein M and N are integers, the method includes: executing an absolute difference plus total instruction to calculate a χ χ M The -nxn portion of the megablock block - the first - absolute difference plus the total value, the first containing the upper left portion of the MXM giant block, where the integer is; performing an absolute difference plus total instruction to calculate the M x a second absolute difference sum of the second nxn portion of the M giant block, the second: an upper right portion including the MxM giant block; the file 0 accumulating the first and second absolute difference plus Total Worth - Total 41 Client's Docket N〇.:S3U06-0024 TT's Docket N〇:0608-A41261-TW/flna17 林璟辉/2〇〇7/〇6/13 41 200821986 ^ Execute the absolute difference plus total instruction to calculate a third absolute difference of the third η X η portion of the Μ Μ Μ giant block plus a value, the third portion includes a lower left portion of the Μ Μ giant tile; adding the third absolute difference plus the total value to the sum; performing the absolute difference summing instruction to calculate the Μ Μ Μ giant tile a fourth absolute difference value of a fourth η χ η portion, the fourth portion including a lower right portion of the Μ Μ Μ giant tile; and f adding the fourth absolute difference value to the total value To the sum. 42Client's Docket No.: S3U06-0024 TT's Docket No:0608-A41261-TW/fmal/林璟辉/2007/06/13