TW200803527A - Method for determining a motion vector describing motion realtive to a reference block and storage media thereof - Google Patents

Abstract

A method for determining a motion vector describing motion relative to a reference block, the method comprising: determining which of a plurality of prediction blocks is a good match with the reference block, according to a match criteria; performing a local area exhaustive search to produce a best match with the reference block, the search performed in an area centered around the good match prediction block, the best match having integral pixel resolution; modeling the degree of match between the best match and the reference block as a quadratic surface; analytically determining a minima of the quadratic surface, the minima corresponding to a best matching block with fractional resolution; and computing a fractional motion vector based on the best matching block with fractional resolution.

Description

200803527 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.

I [Prior Art] Personal computers and consumer electronics are used in a variety of entertainment products. These entertainment items can be roughly classified into two categories: those using computer-generated graphics, such as computer games; and those using compressed video streaifi, such as pre-recorded programs. Digital audio and video (DVD), or digital programming from cable or satellite operators to a set-top box. The second type also includes a coded analog video stream', for example, implemented 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 the computer 'computer game consoles' and some personal computers. A graphics processing unit is optimized to quickly perform three-dimensional primitive objects, such as triangles, quads, and the like. These basic objects are described by a plurality of vertices, where each vertex has an attribute (e.g., a color) and texture can be applied to the base object. 6Clienfs Docket N〇.:S3U06-0025 TT's Docket N(K〇608-A41237-TW/final/林环辉/2007/06/15 6 200803527 The result is a two-dimensional array of pixels Pixels), displayed on a computer monitor or monitor. The encoding and decoding of video streams involves different kinds of operations, such as discrete cosine transform, motion estimation, and motion compensation. : age (10) compensation), deblocking fi 1 ter. These calculations are typically handled by a general purpose central processing unit (CPU) in conjunction with a particular hardware logic circuit, such as an application specific integrated circuit (ASIC). Consumers therefore need multiple computing platforms to meet their entertainment needs. There is therefore a need for a single ten-pipe platform that can handle computer graphics and video encoding/decoding. SUMMARY OF THE INVENTION One aspect of the present invention is a method for determining a motion vector for interpolating a movement relative to a test block, the method comprising: a handle missing ^ 0 · a very consistent standard, determining a plurality of prediction blocks Which one has the -_ ρ 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 〒 The best match corresponds to a region around the environment ^ ; , , , , and the best match has an integer material solution to model the degree of matching between the best match and the reference block as X ' face; Determining a minimum value of the secondary surface, the difference f table is a best matching block having a small value corresponding to the fractional resolution; and according to the score, the number of solutions

Docket N〇.:S3U06-0025 1 s Docket N〇:0608-A41237-TW/fmal/林璟辉/2007/06/15 7 200803527 The best matching square calculates a fractional motion vector. Another aspect of the present invention is a method for determining a motion vector for interpolating relative to a movement of a parameter, the method comprising: determining, according to the ~phase, which of the plurality of prediction blocks is to be paid by the Yuan test block 1 ^ ^(go, match; perform a partial region thorough search to produce the best match with the knot|the tester's thoroughness: search for the region around the center predicted by the better match, The best match has an integer pixel solution: the square ru and the analytically determined modelling the best match with the reference block; a minimum of the quadratic surface of the degree, the minimum corresponding to the best of the score </ RTI> </ RTI> <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; </ RTI> <RTIgt; ~> ^ 弋 数 预测 预测 预测 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个 一个It’s better to thoroughly search the local area of Hai The coincident prediction block is a region around the central eve a, and the best match has an integer pixel resolution; the 私, ^, pa 乜 coincidence and the privateness paid between the reference blocks are constructed as a quadratic surface · The county I 曰%曰| at the location of the knife to determine a small value of the secondary surface, the small value corresponds to a fractional block; and according to the score of the most / degree - the best matching square motion vector * Matching box to calculate a score [Embodiment] 8Clienrs Docket N〇.:S3U06-0025 TT,s Docket N〇:〇6〇8-A4m7-TW/iinal/林环辉/2〇〇7/〇6/ I5 8 200803527 The real-time delivery period disclosed here—the graphics processing unit to improve the motion estimation system and method. 1. The operation flat color for the video scales is used for meshing and video encoding and/or decoding. A block diagram of the computing platform. The system 1GG includes a general purpose (10) 11Q (hereinafter referred to as a main processor), a graphics processor (10) 12G, a memory 13 (), and a bus bar 140. The graphics processing unit 120 includes a video acceleration unit (clear) 15 that speeds up video encoding and/or decoding, as will be described later. The video acceleration function of the graphics processing unit is an instruction that can be executed on the graphics processing unit 12A. The software decoder 160 and the video acceleration driver 17 are located in the memory 13A, and the decoder 160 is executed on the main processor 110. The decoder 16 can also issue a video acceleration command to the graphics processing unit 120 via an interface provided by the video acceleration driver 170. In this way, the system 1 performs video coding by sending a video power command to the main processor software (10) of the graphics processing unit 12 (10). In this way, the intensive computation block ((10) Putati〇nally intensive) that is often executed is offloaded to the _processing unit 120, while the more complex operation is performed by the domain || 11(). Several conventional elements that are not necessary for interpreting the video acceleration characteristics of the graphics processing unit m and are familiar to those skilled in the memory are omitted in FIG. Next, we will explain the video coding outline, and then discuss a video coding component (Mobile Estimation 9Client's Docket N〇.:S3U06~0025 TT's Docket No:0608-A41237-TW/fmal/林环辉/2007/06/15 9 200803527 ,) How to complete the video acceleration unit work provided by the graphics processing unit 12Q 2. Fig. 2 is a functional block diagram of the first _ view I encoder. Transfer to the image of the encoder (10): (205) is composed of pixels. The encoder (10) operates using the time (*temp〇ral) in the image 205 and the spatial similarity (Deng's like: Other^leS), and utilizes the decision within a frame (space) and/or between the maps (day guard) (d) differential similarity coding. Spatial coding utilizes the same or related characteristic coding of adjacent pixels within an image, so only the difference is encoded. Time Coding The use of the final multi-pixels in a series of images is usually the same value, so only the differences between the images are encoded. Encoder (10) (4) Statistical redundancy with entropy coding: some images are more likely to occur in the second picture I, so it is more often represented by shorter codes. Examples of smoke coding include *Huffy (Huff_(7) coffee), run-length encoding (Arithmetic c〇ding) and pre- and self-adaptive binary arithmetic coding ((10) binary arithmetic coding). In this exemplary embodiment, the block of wheeled image 205 is provided to a subtractor 2H) and a motion estimator 22A. The motion estimator compares the blocks in the input image 205 with the pre-stored reference images 23 to find similar blocks. The motion estimator 2 2 0 calculates a set of motion vectors 245 that represent the configuration between the matching blocks. The matching of the motion vector 245 with the reference image 2 is collectively referred to as the prediction side lOClienfs Docket N〇.: S3U06-0025 TT's Docket No: 0608-A41237-TW/fmal/林璟辉/2〇〇7/〇6yq5 10 200803527 255, representing time coding. Prediction block 255 is provided to subtractor 21A which subtracts the input block 2〇5 from prediction block 255 to produce a residual image 26〇. The remaining image is supplied to a discrete cosine converter (10), dlscrete (10), this transf〇rm) block 270 and quantizer_, which is line-space encoded. The quantizer 28〇&amp; output (for example, the grouped DCT coefficient> is encoded by the pain encoder 29〇. • For a certain dog-type image (information or 1 frame, with prediction or P frame) The spatially encoded remainder of the space from quantizer 280 (residual by spatiaUy (10) is provided to the internal decoder. The decoder uses the spatially encoded remainder in conjunction with the motion vector 245 generated by the motion estimator 220 to spatially encode the image 205. The reconstructed image is stored in a reference image buffer 295, which is provided to the motion estimator 220, as previously described. As discussed in connection with the first figure, the encoder (10) is in the main processor ιι〇 The above-mentioned implementation, however, utilizes the video acceleration instructions provided by the graphics processing unit 12. In particular, the algorithm implemented by the mobile estimator 220 uses the absolute absolute difference provided by the graphics processing unit 7L 120. The value sums up (pass, leak-〇f-abs〇lute-difference) instructions to achieve the correct motion estimation, at a relatively low computational amount. The mobile estimation flooding method will be detailed next. Estimation of deduction 1 IClienfs Dock Et N〇.:S3U06-0025 TT's Docket No:0_-A41237-TW/fmal/林璟辉/2〇〇7/〇6/15 11 200803527 L Search Window As shown in Figure 3A, B, move The estimator 220 cuts the current image 205 into non-overlapping segments, called giant tiles. The size of the giant tile is determined by the encoder (eg, 'MPEG-2, H.264, VC'). And the size of the image changes. ' 1 In the exemplary embodiment described herein, and in various coding standards, a giant tile is 16x16 pixels. A giant tile is further cut into squares, and the size of the square can be It is 4x4, 8x8, 4x8, 16x8, or 8x16. In MPEG-2, each giant tile can have only one motion vector, so the motion estimation system is based on the giant tile. H.264 allows up to 32 motion vectors (depending on the degree). Therefore, in H.264, the motion estimation is calculated based on the basis of 4x4 or 8x8 squares. The change of H·264, called AVS, is always 8x8. In U, it can be 4x4. Or 8x8. The motion estimation algorithm 220 performs a motion estimation on each macroblock in the current image 205, in accordance with a pre-encoded image 230 ( It is similar to the target of finding a square in the giant tile of the current image 205. The replacement between the giant tile in the reference image 230 and the giant tile in the current image 205 is calculated and stored as a motion vector (245). , Fig. 2). For convenience of explanation, the motion estimation program will be described in the current image 31〇-specific giant tile (320). The giant tile selected in this example is currently in the image 12Clienfs Docket N〇.:S3U06-0025 TT's Docket No:0608-A41237-TW/flnal/林璟辉/2007/06/15 12 200803527 'In the middle of 310, However, the same technology is also applied to other giant tiles. A search window (330) is in the middle of the macroblock in the reference image 230 (corresponding to the giant tile 320 of the current image 31〇). That is, if the giant tile 32 is located at (X, Y), the search window 330 in the reference image 230 is also located at (X, γ) as shown in _340. Other embodiments place the giant tile in other portions of the reference image 23, such as the upper left. The search window 33 of the example 3A, B extends horizontally through two pixels of the corresponding giant tile, one pixel in the vertical direction. Therefore, the search window 330 contains 14 different giant tiles: two giant tiles are respectively sent with 2 pixels, just to the left of the position 340; the other two giant tiles are to the left of the position; The group is above, below, top left, top right, and bottom right of position 340. 1 The coincidence block movement operation performed by the mobile estimation source 220 uses the absolute absolute difference sum as the criterion for judging the similarity (conformity) 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 pixels in a square are summed, as understood by those skilled in the art. The mobile estimator 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. b. Selecting the target giant map mobile estimator 220 uses the different search methods to generate an intra-c〇ded motion vector or an outer code (inter-coded) of the target uranium image 205 in accordance with the motion estimator 220. ) Move the vector. The mobile estimator 220 utilizes the true BClient^s Docket N〇.: S3U06-0025 TT's Docket NcK〇608-A41237-TW/fmal/林璟辉/2007/06/15 200803527 The real world of mobile know-how 320 The household, the heart/consistent giant block should be in the search for the number of target blocks in the _second t°, which is similar to the giant block in the image 205. Normally, it is fixed: speed shift: to tGPtieal The movement of objects in flQW), Wei and similar (ie, substantially continuous) are both on the work and in time. In addition, the absolute absolute difference plus the surface (i.e., the -search space m absolute difference plus the total value) is expected to be relatively moderate (i.e., a relatively small number of local minimum points). The use of this self-aware knowledge requires a command to search for the most likely to find the most consistent location of the 'News' algorithm to minimize the number of searches to be performed to find a better minimum. In this way, the algorithm is computationally efficient and can effectively mark the best match. 4 is an algorithmic flow diagram of an exemplary embodiment motion estimator 22 for computing the current motion vector of the macroblock 310 in the image 205. The motion estimation procedure begins with a step of determining that the motion vector produced by the motion estimator 22() for the current image 205 will be inter-predicted or intra-image predicted Qntra-pre (jicted). If intra-image prediction is used, then step 420 is performed, where a conjugated gradient descent search algorithm is performed to find a predicted macroblock in the search window 320, which is related to the reference giant tile ( The current giant tile 31〇 in the image 205 is a better match. The co-vehicle gradient descent (step 420) will be described in detail in conjunction with Figures 5 and 6. MClienfs Docket N〇.:S3U06-0025 TT,s Docket No:0608-A41237-TW/fmal/林璟辉/2007/06/15 200803527 Back to 410 'If you use inter-image prediction to generate motion vectors, then execute Steps, performed here, adjacent, or,, adjacent areas, search. The search contains a giant image adjacent to the current giant tile 31〇 in the current image 2〇5 and a corresponding previously encoded reference image 23. The giant tile inside. The proximity search; algorithm (step 430) will be described in detail in conjunction with Figures I and 8.卞 A common gradient descent search algorithm (step-by-step) and a neighboring search calculus. Step 430) Each of the slave-large group target prediction giant tiles recognizes a better or a 7-acceptance match. Those familiar with the art should understand the criteria for determining what is better, and the criteria can be relative or absolute. For example, 'Use in this neighborhood search algorithm—absolute criteria: The target giant tile with the lowest score is considered to be a better match. However, the common gradient gradient search algorithm used here uses a critical value, and the absolute difference plus the total value is lower than the critical value. The first-square is considered to be a better match. However, the criterion of the threshold is a design or implementation decision. After processing step 420 &lt; 430, a better candidate is recognized. Step 440 is more performed - local area Thoroughly search (1〇Cal__search) to find the best candidate. The search area is located near the better candidate giant block identified in step 42〇 or 430. In some embodiments, step 420 ' After the common gradient descent search algorithm (ie, under the condition of intra-image prediction), the local thorough search for the four pairs of the local collocations (preferred candidates) recognized by the step is recognized by the step. Angle. For example, if In the gradient step, the value used in the previous step is j, then the search is limited to the preferred 15Clienf s Docket N〇.:S3U06-0025 TT's Docket N〇..0608-A41237-TW/fmal/林璟辉/2007/ 06/15 15 200803527 Point of Candidate(). In some embodiments, after performing step 43 (ie, in the case of inter-picture prediction), the partial thorough search (step weight) is searched for among the better candidates. A candidate for a small area near the giant tile, usually the partial perpetual search of step 440 is limited from a preferred candidate macroblock to a best candidate macroblock, which is pixel aligned (pixel_aligned), ie has an integer Pixel resolution. Steps 450 and 460 find an optimal candidate giant block alignment at a fractional-Pixel b_dary. The conventional fractional motion search algorithm uses a specific codec-specific filteringalg algorithm. 〇rithm) interpolates the pixel value at the fractional position according to the surrounding integer position. In contrast, step 45 〇 establishes that the final candidate giant tile and the reference giant __ degree are secondary surfaces, and step = analytically determines the Surface The minimum value corresponds to a best-matching giant tile, which is a fraction rather than an integer resolution. (The groundbreaking is best determined by fractional resolution. • The modeling of the macroblock will be explained in the following paragraphs.) After the coincident giant block with the score "phase" is stepped, the step 450 is recognized, and then the processing step is analyzed to calculate a fractional motion vector according to the matching giant block, and the technique known to the knowledge is used. . Then the program 4 is completed. $ Those familiar with the art should understand that the above algorithm is essentially intrinsic because it uses information from nearby areas. Although hardware-accumulated designs are often used to avoid continuous algorithms, continuous design is appropriate for many reasons. First, the pixel data is in the horizontal horizontal scanning line X / 16Clienfs Docket No.: S3U06-0025 乂: TT's Docket N〇: 0608_A41237 • Ding w/fmal/林璟辉/2〇_6m 200803527 (equential raster fashion) Take, and thus can be received in advance, held in the = channel buffer. Secondly, in the case of Beiyuan containing a single-absolute difference plus a total acceleration f-element, the effectiveness is limited to whether the unit can maintain full load instead of ', and the shell is treated. The absolute difference plus total acceleration unit does not have many cache misses in the prediction block, and the load is still loaded. Since the missing rate is a function of the size of the cache, the .τν solution=degree image is only f deleted in the cache. /8 = the side vector, and the low cache miss rate is expected. The in-image pleats of c, and the flowchart of the common gradient step 440 of Fig. 4 are performed by the embodiment of the mobile estimation gas 220. As previously mentioned, the step weighting is performed when it is determined that the intra-image prediction will be used to find that the giant block system in the search t 32G matches the current side $ _ is - preferably (acceptable). The total value of the absolute county value is calculated for the five initial candidates of the P group: the current giant tile, and the giant tile of the upper, lower, left and right of the current giant tile. From this initial group of 5 absolute differences plus the total value, count the two sets of mutual (four) silk gradient. From these two sets of gradients, the gradient of the _most (four) XI direction. The money gradient is shallow, or the five initial waiting blocks have very close absolute total values, and the job search extends far from the visual block, because there is no simple condition for practicing the local minimum probability in this area. This step will be explained in more detail after the overview of the common gradient step 44Q. 17Clienf s Docket No.:S3U06-0025 mD〇cketN〇面捕237向_林环辉/2_〇6/i5 Take this step from step 505, initialize here - candidate block ^ and step value △, and λ y. In the embodiment, the candidate block ^ is the upper left corner of the search window 17 200803527 320, and the steps are then sighed to a small integer value, such as 8. At the force 51G, the candidate giant block is calculated. These four silicon, pec W week's seat of the giant block. That is, the mouth, c-, upper, lower, left, and right four ~ ς) heart (―~+Q, C) difference force, the absolute sound of the candidate giant block, and the four surrounding. At step 52Q, the difference between the absolute value of the left and right giant blocks and the total value is calculated. The gradient a is the absolute difference between the giant block and the total value Λ » ” left, regardless of whether the error value between the matching macroblocks is ^ &quot;J&quot;b left value 疋, plus or minus, the gradient Indicates the x or v direction. At step 525, the gradient system 盥_gt is the boundary value for comparison. If the gradient is below the threshold (__目对对喊), this means that the local minimum is in the current search area' so the search extends to the new candidate giant block. These new candidates = tiles are far from the original candidate processing giant block k. In some embodiments, the search is also extended when the absolute difference plus total values calculated for the candidate giant tiles are similar in step 515. The extended search continues at step 530 where the coordinates of the four new candidate giant blocks are calculated. The original four candidate giant block blocks are located at the upper, lower, left and right distances C, A), and four new candidate giant image blocks are selected to form the square corner 'distance (△, △,) around the original candidate giant block Q: ΓΖ = (-△, +Cv, -△, C ==(△〆) △}+=(_△〆,ς) Lake=Team+匕~+Q 18Client's Docket No.:S3U06-0025 TT,s Docket No: 0608-A41237-TW7flnal/林环辉/2007/06/15 18 200803527 In step 535, a common vehicle gradient descent step 440 is performed on these new candidate giant block blocks (c, t1, t. Gradient comparison, if the gradient of the giant graph is equal to the critical value (_ do the ground to the ground (4)), the absolute difference plus the total value calculated in step 515 is compared with the second critical value. If the total value of the absolute value is below the threshold, and the job title is better, then step 440 returns to the pager, (at step 545) 'providing the candidate block of the lowest absolute difference plus the total value. ^ If the value of Wei in the step 540 is added to or below the threshold value, indicating that there is no _ record, then 驴 search _. In step ^ 550 'Select - New The candidate giant tile Q. The new central giant tile. In the candidate group, the block with the lowest absolute difference plus the total value is calculated in step 515. Then, in step 555, a new step is calculated from the ladder. The value △, and Δ, for example, the steep_gradient represents an acceptable coincidence of the giant block system, which is far away from the current central candidate, so it is increased by (10). Conversely, the shallow gradient represents an acceptable coincident giant block system. Therefore, it should be reduced (△, Α). Those skilled in the art should understand that various coefficients can be used to calculate the result from the gradients. Then 'test the number of iterations in step 560. If the number is greater than _maximum, then step 440 is completed in step 565, and an acceptable match is not found. Furthermore, an error gradient is used to select a new set of candidate giant tiles, which is expected to be more than 19Client, s Docket No:: S3U06- 0025 TT s Docket No: 0608-A41237-TW/fiiial/Lin Huanhui/2007/06/15 19 200803527 Close to the final match, the gradient descent step 44 returns to step 5ι, where a new set is generated. The common gradient descent step 44 is in the following two In this case, when an acceptable value is found (step 545), or the maximum number of iterations is reached, there is still no match (step 565). The _6 figure illustrates the use of a common gradient descent step 44 (an exemplary state of Γ. The distant giant block q is square (), and four surrounding candidate circles (61〇T, 610L, 610R, 610B). From these initial candidates, the gradients are calculated as &&;(10)X, 62GY). In this exemplary state, the gradient is too shallow, and there is no absolute difference plus the total value below the threshold. Therefore, the search is extended using 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 block blocks are from the original candidate giant block. The distance from the surrounding corner △. The giant tiles around these central candidates are shown as hexagons (64, _7;, 6 rings 2, 6 2, 6 3, 6 lions 3, 64 pts, _ Lu Fan state, two candidates 640 There is an absolute difference plus a total value below the critical value and a ''steep' gradient) (650XY, 66QXY). Another candidate is selected according to each, steep, gradient: candidate 670 is based on gradient 650XY, and then selected 680 is based on Gradient 660XY. Gradient descent search continues to use these new candidates 670, 68〇, according to the conjugate gradient descent step 440. d. Use previous neighbor inter-image predictive shift 7th image 4th neighborhood proximity search algorithm (step 430) The flow chart is executed by an embodiment of the mobile estimator 220. As described above, the search is 20Clienfs Docket N〇.: S3U06-0025 TT's Docket No: 0608-A41237-TW/fmal/林璟辉/2007 /06/15 20 200803527 The current giant tile 310 (which has been categorized in the pre-encoded reference image selection block contains the adjacent image 205 encoding) is also included as a candidate 230. a corresponding macroblock. The step of calculating the candidate giant block coordinates begins at step 710, at By using the absolute value of the current huge block 310 address (the remaining value) and the number of giant blocks per line of the ancient ten tube-flag variable ramiD. If listening to the dragon is not Q, then TQmuD is straight. In addition, TOPVALID is False. In step 720, the _flag variable leftvaud is calculated by dividing the address of the macroblock block 310 by the integer and the number of macroblocks per line. If the divisor is not 0' then the LEFTVALID is true 'in addition, the leftvalid is False. These TOPVALID and LEHVALID variables indicate that the current giant tile 31〇 has a neighboring giant tile on the top and the left, respectively, considering the upper edge and the left edge of the giant tile. In step 730, the TOPVALID and LEFTVALID variables are combined to determine the current giant. The availability or existence of four candidate macroblocks 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; and a giant tile on the upper left. TF if (T0PVALID &amp;LEFTVALID); there is a giant tile TR (TOPVALIM RIGHTVALID). Next, in step 740, the availability is determined for a previous candidate giant block ρ, which corresponds spatially to the current giant tile. One of the previously encoded reference images 230 of 310 Giant block. The relative positions of the five candidate giant blocks can be seen in Figure 8, where L system 810, system 820, TL system 830, TR system 840, and system 850. Returning to Figure 7, how many candidate giant tiles in steps 730 and 740 can be compared with 21Client's Docket N〇.:S3U06-0025 TT's Docket No:0608-A41237-TW/fmal/林璟辉/2007/06/15 200803527 From 1 to 5). Step 750 Scale - The candidate giant block is calculated to add the absolute difference. If 5 candidates are available, the total absolute value of the group is: Γ ( L Λ- Τ\ If some candidates are not available, those familiar with the art should know that the group of candidates is relatively small. Then complete (4) 43Q, shouting the candidate giant block with the lowest absolute value. ~ As previously discussed in conjunction with Figure 4, once the matching giant block is found (regardless of the use of the first-neighbor search method or the fifth gradient of the listening gradient ), then the search area is more limited, using a partial thorough search (Fig. 4, 4). After the local search, the result of the partial thorough search is calculated—the fractional motion vector. The calculation of the fractional motion vector will be detailed below. • Using the fractional motion vector of the quadratic surface model. 1 Those who are familiar with this technique should be familiar with the degree of conformity between the illustrated giant tiles and the search window to produce a “wrong surface.” Using a groundbreaking approach, swaying estimates The device 220 models the erroneous surface with a quadratic surface and analytically determines the minimum value of the surface with sub-pixel accuracy. The motion estimator 220 first gives a minimum value with the minimum value of the _square term. Estimate The device 220 then determines the minimum value of the orthogonal direction along this line. The general equation of the quadratic curve is as shown in Equation 1. 22Client,s Docket N〇.:S3U06-0025 TT's Docket No:0608-A41237-TW/fmal/ Lin Yuhui/2007/06/15 22 200803527 Equation 1 y = Cx+C2t^Cf The curve is differentiated, as in Equation 2: C2 + 2C3t => i Equation 2 ϊ -dan coefficient c&quot;.2, ς Knowing that: the solution can be solved to determine, the minimum position shift estimator 220 solves Equation 3 to determine the coefficient Ci, the heart q. C2 C3 4 31 - 27 5, -27 25 A: 5 -5 1 Σ^2 Equation 3 The motion estimator 220 uses the equation 84 for the effective efficiency of the 84 absolute difference plus total instruction provided by the field j j 用 圚 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 。 The value 'accumulates i represents the absolute difference plus the value of the adjacent macroblock in the X direction. If combined, the detailed description of the figure is as follows: 4 absolute difference plus total instruction efficient calculation of adjacent giant tiles ( x, y), (x+1, y), (x+2, yMX+3, y), the four absolute differences plus the total value, ie i = 〇 ·.. 3 and. As mentioned above, Once the coefficient has Equation 2 is used to determine 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 220 uses the 8x4 absolute difference plus the total efficiency of the instruction. The four absolute differences of the adjacent giant blocks (X, y), (x+l, y), (x+2, y), (x+3, y) plus the total value. Equation 3 solves the calculation Since these absolute differences add up to 23Clienfs Docket N〇.:S3U06-0025 TT's Docket No:0608-A41237-TW/fmal/Lin Huanhui/2007/06/15 9 . 200803527 The coefficient is as described above. The coefficients are known and the equation is solved to lead to the minimum in the t, y direction. The second erroneous surface method used by the motion estimator 22〇 is used after the preferred ride-pixel boundary—the preferred tree is used: the expensive filter goes to the better tree (10) green of the money world. improvement. In the calculation of the absolute difference total accelerator 10 (4), the motion estimator 22G judges that the giant tile in the image is better in the current image-reference giant block. The motion estimator 22 uses the absolute difference provided by the processing unit 120 to add hardware acceleration, which is a graphics acceleration unit command. The absolute difference plus total instruction is to input a 4x4 reference block and an 8x4 prediction block, and generate 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 8 χ 4 prediction block are merely examples to illustrate the present invention, and should not limit the φ size of the reference block and the prediction 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 8x4 prediction block is composed of a plurality of horizontally adjacent 4x4 blocks that overlap each other, such as blocks 910, 920, 930, 940. The absolute difference summation unit takes an input 4Χ4 reference block 950 and calculates the absolute difference plus the total value of the McCaw block and 910-940 blocks. That is, the absolute difference summation command calculates 4 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 the sum of the absolute values of the difference between block 930 and block 950; the other value is the absolute value of the difference between block 940 and block 950, 24 Client, s Docket N〇.: S3U06-0025 TT's Docket No.. 0608-A41237-TW/final/林璟辉/2007/06Λ5 24 200803527 Total. See Figure 9B for tea. The absolute difference in the graphics processing unit 12 is accelerated. The four absolute difference plus total calculation units (96〇, 97〇, _, 99〇) are used to achieve the absolute difference sum. instruction. The leftmost 4x4 box 910 is provided for the absolute difference plus the total amount of the individual 96. Then round the 4χ4 square on the right to the absolute difference feed unit (10). Then input the right square ω square (93 〇), give, % to the difference total calculation unit 98 〇. Finally, the rightmost 4χ4 squares 940々, 邑, 邑 are added to the difference total calculation unit 9 . The graphics processing unit uses independent absolute difference summation calculation units in parallel, so the absolute difference plus total instruction produces 4 absolute difference plus total values per cycle. Those skilled in the art should be aware of the algorithms used to calculate the absolute difference summation operations of two blocks of the same size and the hardware design used to perform this operation, so these details will not be described in detail. • The 4x4 reference block is listed horizontally and vertically at the edge of the pixel. However, it is not necessary to correct the 4x4 prediction block 91〇-94〇 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 block k is supplied to each of the independent absolute differences plus the total hardware acceleration unit. Each unit produces a value of 12 bits, and these values are combined into one bit round. In one embodiment, the magnitude of these values is based on the 11 texture coordinates of the prediction block (the lowest coordinate in the lowest bit position). 25Client,s Docket N〇.:S3U06-0025 TT,s Docket N〇:0608-A41237-TW/fmal/林璟辉/2007/06八5 25 200803527 The following code shows 8x8 squares, ie two adjacent 8x4 squares The absolute difference plus total value of , can be calculated using only 4 absolute difference plus total instructions. The temporary registers T, T, T, and T4 are used to temporarily store the four absolute difference plus total values. 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 in refReg. U 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 T1A refReg, U, V ; left-top of 8x8 prediction block

SAD T2, refReg, U+4, V ; right-top of 8x8 prediction block ADD sadS, ΤΙ, T2 SAD T3, refReg, U, V+4 ; left-bottom of 8x8 prediction block ADD sadS, sadS, T3 SAD T4 , refReg, U+4, V+4 ; right-bottom of 8x8 prediction block ADD sadS, sadS, T4 However, it is usually possible to avoid calculating and summing the values of all four sub-blocks, as long as the sum reaches the current minimum value Stop the calculation. 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 &lt; 4 II SUM &lt; MIN) SUM := SUM + SAD(refReg, U+(I%2)*4, V+ 26Clienf s Docket N〇 .:S3U06-0025 TT's Docket N(K〇608-A41237-TW/final/林璟辉/2007/06/15 26 200803527 (工&gt;&gt;1)*4); IP (SUM &lt; currMIN) currMIN = MIN ;

Go to Next Search point; The 84 absolute difference summation command in graphics processing unit 120 is used directly by the advanced search algorithm of motion estimation gas 220, such as performing a partial thorough search as illustrated in FIG. In addition, the texture cache 1 〇 6 〇 (Fig. 1) is a block correction, and the algorithm used by the motion estimator 220, as described above, is like φ prime weight. Although the multiplexer unit can be added to the graphics processing unit 120 to handle these corrections, 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 8χ4 absolute difference summing instruction provides the advantage of efficiently computing the minimum value, which involves calculating the absolute difference sum value of the neighboring 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 motion estimator 220 and the use of the 8x4 absolute difference summation instruction in the graphics processing unit 120. Next, the absolute difference summing instruction is described in detail. Graphics processing unit 120. a. Graphics Processing Unit Flow Figure 10 is a data flow diagram of the graphics processing unit 120, wherein the instruction stream 27Client's Docket No.: S3U06-0025 TT's Docket No: 0608-A41237-TW/fmal/林璟辉/2007/06/15 27 200803527 is the arrow on the left side of the picture, and the image or graphics flow is indicated by the arrow on the right. Figure 10 omits several elements familiar to those skilled in the art which are not necessary to explain the in-loop deblocking characteristics of the graphics processing unit 120. An instruction stream processor 1010 receives an instruction ΐ 020 ' from a system bus (not shown) and decodes the instruction to generate instruction data Γ 030, such as vertex data. Graphics processing unit 120 supports a conventional graphics processing instruction, as well as instructions for speeding up video encoding and/or decoding, such as the aforementioned 8 χ 4 absolute difference summing instructions. Looking at the ‘graphic processing instructions involves issues such as vertex shading, geometry shading, and pixel pixei 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 (TF1J) 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 explained later. The resulting data is then processed by a post-packer (p〇st-packer 1070) to its pressure and fine data. After post-processing (p〇st_pr〇cessing), the data generated by the video acceleration unit is supplied to the execution unit pool (execution unii: ρ0〇ι) 1〇40. 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 11 instead of the shader 28Client's Docket No.: S3U06-0025 TPs Docket No: 0608-A41237-TW/fmal/林环辉/2〇〇7/〇 6/15 28 200803527 Line 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 by the graphics commands are both 2-dimensional arrays. The graphics processing unit 120 also uses this advantage to download the image data to the video processing unit 1100 using the texture filtering unit 1050, thereby causing the texture cache 1060 to cache some of the video.

I Processing unit Π00 operation of the image data. Thus, as shown in FIG. 10, video processing unit 1100 is located between texture filtering unit 1050 and back wrapper 1070. The texture filtering early 10 5 0 checks the instruction tribute 1030 from the instruction 10 2 0. The command material 1030 further provides coordinates of desired image data in the main memory (not shown) of the texture filtering unit 1050. In one embodiment, these coordinates are designated U and V pairs and should be familiar to those skilled in the art. When the instruction 1020 is a video acceleration command, the retrieved instruction data 1030 further commands the texture filtering unit 1050 to skip any texture filters (not shown) within the texture filtering unit 1050. Therefore, the texture filtering unit 1050 is controlled by the video acceleration command to download the image data to the video processing unit Π00. According to this method, the texture filtering unit 105Q is manipulated as a video acceleration command to download image data to the video acceleration unit 1100. The video processing unit Π00 receives the image data from the texture filtering unit 1050 on the data path, and the command material 1030 on the command path, and performs an operation on the image data according to the command data 1030. The image data output by the video processing unit 1100 is fed back to the execution unit pool 1040 after being processed by the post wrapper 1070. 29Clienfs Docket N〇.:S3U06-0025 TT's Docket No:0608-A41237-TW/fmal/林璟辉/2007/06/15 29 200803527 b. Instruction parameters Now the video processing unit 11GG is performing the absolute difference summation video acceleration command Operation. As previously stated, the _ processing _ processing instructions are decoded and parsed into the instruction material 1030, which can be considered as a specific parameter set for each instruction. The parameters of the absolute difference plus total instruction are shown in the first table. Table 1: Absolute difference plus total instruction of graphics processing unit

Input/output name size Description Input FieldFlag 1-bit if FieldFlag == 1 Bay ij Field Picture, and the rest Frame Picture Input TopFieldFlag 1-bit if TopFieldFlag 1 Bay ij Top-Field-Picture, Other Bottom-Field-Picture FieldFlag is input. InputWidth is 16 bits. For example: 1920 is used for HDTV input PictureHeight 16-bit. For example: 1080 for 30P HDTV input BaseAddress 32-bit unsigned predicted picture Basic address input BlockAddres s U: 16 - Bit-signed signed picture texture coordinates (relative to the base address) at SRC1 Opcode SRC1[0:151 = U, SRC1"31:161=V 30Ciient's Docket No: :S3U06-0025 TT,s Docket No: 0608-A41237-TW/fmal/林璟辉/2007/06/15 30 200803527 V: 16-bit signed U, V is 13.3 format, ignore fractional input RefBlock 128-bit reference picture data ~ SRC2 Opcode output · Destination 4x16-bit I28-bit scratchpad among the least significant 32-bit f Operand elements in DST Opcode

A plurality of input 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 coordinates of the top left square in this area are given to BaseAddass. The PictureHeight and PictureWidth input parameters are used to determine the square of the box, ie the lower left coordinate. Finally, the video shape can be a progressive progessive or an intentce. For interlaced scanning, it consists of two directions (upper and lower). The texture filtering unit 75 uses FieldFlag and TopFieldFlag to properly process the interlaced scanned image. c. The image data is converted into an absolute difference summing instruction, and the video processing unit extracts the input pixel blocks from the texture filtering unit 1050 and converts the squares into a suitable format to facilitate the absolute difference plus acceleration: Said processing. The pixel block is then provided to the absolute difference summation acceleration 兀 960- ‘ ’ replies its absolute difference plus the total value. Each absolute difference plus total value is accumulated to the target register. These features will be detailed later. Video processing unit H00 receives the definition calculation. The absolute = 3 IClienfs DocketNo.: S3U06-0025 宁 difference plus,,, heart value TT s Docket N〇: 0608_A41237-丁 W/flnal/林璟辉/2〇〇7/〇6/15 31 200803527 Two input parameters for the 8x4 block. The data in the reference block is directly defined by the SRC2 operating code. The 8x4x8 bit box is treated as 128-bit data. In contrast, the SRC1 operational code defines the address of the prediction block rather than the data. The video processing unit 1100 provides these addresses to the texture filtering unit lQ5, which takes the prediction block data of the 128-bit from the texture cache 1060_.

I Although the image data contains the luminance (Y) and chroma (Cb, Cr) planes, the motion estimation usually uses only the Y component. Therefore, when the absolute difference summing instruction is executed, the pixel block operated by the video processing unit 1100 contains only the gamma component. In one embodiment, video processing unit 11 generates a disable signal 'which directs texture filtering unit 1050 not to extract Cr/Cb pixel data from texture cache (9). Figure 11 is a block diagram of texture filtering unit 1050 and texture cache 1 〇 6 。. Texture filtering unit 1050 is designed to extract texture boundary (texel boundry) from texture cache (9) and download 4x4 texture image block from texture cache iQgQ to filter input buffer 111. When the captured data represents the video processing unit 1100, the texture image 1120 is treated as a texture image size of four channels (ARGB) each having 32 bits for 128 bits. When the data is fetched for the absolute difference summing instruction, the texture filtering unit 1〇5〇 downloads the 8x4x8 bit block, which is stored in the 2 pixel input buffers (1110A, 1110B). The 8X4 image block used for the absolute difference plus total command is as previously described in connection with Figure 9. The image data used by the video processing unit 1100 may be corrected by a byte. However, the texture filtering unit 1050 is designed to extract the pattern from the outside 32Qient's Docket N〇.: S3U06-0025 TT’s Docket No: 0608-A41237-TW/fmal/林璟辉/2007/06/15 32 200803527 Image boundary. Therefore, when subtracting the data retrieved by the processing unit n(10), the texture filtering unit 1 〇 5 〇 may need to capture up to 4 texture image correction 4 blocks around a particular byte correction 8x4 block. *Hai private order can be seen in the 11th towel, the squares captured by the towel (target block 1130) are on the boundary of the texture image, whether in the vertical direction or in the horizontal direction. The u and ν addresses of the target block 丨13G define the top left corner of the 8 χ 4 χ 8 bits, and the byte correction block. In this example, texture cache unit #1050 fetches texture images 1140, 115〇, 116〇, 117〇 to obtain target block 1130. The texture filtering unit 1〇5() then combines the rows and columns selected from the bits of the block li4〇_ii7〇, and the leftmost 4χ4 bits of the target block (10) are written to the filter buffer 1110Β. Those familiar with the art should know how to use the multiplexer, shifter, and mask bits to achieve this result, regardless of the 4χ4 target correction taken from texture cache 1Q6(). In the embodiment illustrated in Figure 11, when the target block 113G contains a vertical texel boundary, the material is not rearranged vertically. When this happens, the order of the data downloaded to the filter buffers 1 ΠΟΑ and Π 10 在 is different from the original order in the cache. In this embodiment, the video processing crew 11 must vertically rearrange (rotate) the 128-bit reference block data to match the order of the predicted blocks. In another embodiment, before writing to one of the filter buffers 1110, the texture filtering unit vertically rearranges the cached texture image data to conform to the original cache order. 0 33Client's Docket N〇.: S3U06-0025 TT's Docket No ..0608-A41237-TW/fmal/林珠辉/2007/06/15 33 200803527 The blocks in the description or the flowchart are to be understood as representations. A second program code that contains one or more executable instructions for implementing a special logic circuit work ::= step. Those who are familiar with software β should understand other implementations. In other implementation methods, the townships are executed in the order of the stipulations, and the functions involved in the f are performed or reversed.

The system and method can be software, hardware or a combination of real money, (4) inspection / or system (4) in the memory (packaged and not executed by a suitable processor located in a computing device = a micro-processing , microcontroller, network processor, re-provisioning processing boundary, ok, right _ σ extended domain n in other embodiments, the method is implemented by a logic circuit, including but not limited to - programmable Logic device (PLD 'progr_able 1〇gic —), programmable logic closed array (PGA, (ra) ra_ble gate array), field programmable logic gate array (Held pr〇grammable gate array) or 4-inch application circuit (ASIC) In other embodiments, the logic is implemented by a graphics processor or a graphics processing unit (GPU). The system and method disclosed herein can be embedded in any computer readable medium, or a link-instruction. Execution system, device, device. The instruction execution system includes any computer-based system, a system containing a processor, or other system that can retrieve and execute these instructions from the instruction execution system. The text is fine-readable media (c(10)puter-readable 34Client's Docket N〇.:S3U06-0025 TT's Docket No:0608-A41237-TW/fmal/林璟辉/2007/06/15 34 200803527 medium)" can be any, Storing, communicating, transmitting or transmitting the program as a means of using or interfacing with the instruction execution system. The computer readable medium can be, for example, (non-limiting) electronically based, magnetic, optical, electromagnetic, infrared. Or a system or medium of semiconductor technology. A specific example of a computer readable medium using electronic technology (non-limiting) may include: a line having one or more electrical (electronic) connections; a random access memory (RAM, random access memory); - read-only memory (ROM, read-only memory); - wipeable programmable read-only memory (EPROM or flash memory). Computer-readable media using magnetic technology Specific examples (non-limiting) may include: portable computer disks. Specific examples of computer readable media using optical technology (non-limiting) may include: an optical fiber and a portable CD-ROM (CD-ROM) While the invention has been illustrated and described herein by way of the embodiments of the invention Many different modifications and structural changes are possible within the scope and scope of the scope of the invention. Therefore, it is preferred that the scope of the appended claims be broadly and construed in a manner consistent with the field of the invention, This statement. [Simple diagram of the diagram] Exemplary operation Figure 1 is a block diagram of one of the platforms for graphics and video coding and/or decoding. 35Client,s Docket No.:S3U06-0025 TT's Docket No:0608-A41237-TW/fmal/林环辉/2007/06/15 35 200803527 Figure 2 is a functional block diagram of the video encoder 16〇 of Figure 1. . Figures 3A and B illustrate giant tiles that divide the current image into segments that do not overlap. Figure 4 is a flow diagram of an exemplary embodiment of an algorithm used by the motion estimator of Figure 2. r Figure 5 is a flow diagram of an embodiment of a conjugate gradient step 440 of Figure 4. Figure 6 illustrates an exemplary state using the conjugate gradient descent step 44A of Figure 5. Figure 7 is a flow diagram of an embodiment of a proximity search algorithm of Figure 4. Figures 8A and B illustrate the relative positions of the five candidate giant tiles used in the proximity search algorithm of Figure 7. Figures 9A and B are block diagrams showing the operation of the absolute difference summing instructions for the reference and prediction blocks. Figure 10 is a data flow diagram of the graphics processing unit of Figure 1. Figure 11 is a block diagram of texture filtering unit and texture cache in Figure 10. [Main component symbol description] 100~ system, 110~ main processor, 120~ graphics processing unit (GPU), 130~memory, 140~bus, 150~video acceleration unit (VPU), 160~36Clienfs Docket N〇 .:S3U06-0025 TT's Docket No:0608-A41237-TW/final/林璟辉/2007/06/15 36 200803527 Do the tr decoding mode, 17〇~video acceleration driver. 3 image, 210~subtractor, 220~moving estimator, mo~to test image 245~moving vector, coffee~predicting block, 26〇~remaining image,moving,~discrete cosine converter,~quantizer , 29G ~ entropy solution, measuring. Code. , 1 〇 萷 giant block, 320 ~ giant block, 330 ~ search window, 34 〇 ~ point. , 400 ~ program, ~ determine the motion vector will be inter-image prediction or image within = 彳, 42G ~ implementation of the total (four) degree drop search algorithm, 43G ~ execution of the proximity search 440 ~ perform a partial area thorough search, 45 〇 ~ Establish the best candidate giant block and the reference £ block to match the secondary surface, ~ at the - fractional pixel boundary - the best candidate giant block alignment, 470 ~ calculate a fractional movement based on the matching giant block vector. • 505~ initialize a candidate block, 510~ calculate the coordinates of the candidate giant block around the candidate giant block, 515~ calculate the absolute difference sum of the five candidate giant blocks, respectively, 52〇~ calculate the gradient &amp; And &amp;, 525~ gradient is lower than a critical value, 530~ calculate coordinates of four new candidate giant block, 5 evil~ perform conjugate gradient descending steps 440, 540 to absolute difference of each candidate giant block respectively Value plus value 疋 No 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~ Calculate from the gradient g and A The new step value ~ and 丨, 560 ~ test the number of iterations of the loop is greater than a maximum value, 565 ~ back pass does not match. 37Client5s Docket N〇.:S3U06-0025 TT's Docket No:0608-A41237-TW/fmal/林璟辉/2007/06/15 37 200803527 610C~Candidate giant block, 61OL-61 OR-61OT-610B~ four surrounding candidates 620X-62GY~ initial candidate calculation gradient, 630TL-630TR-630BL-630BR~ four new central candidate giant block blocks, 640L-640R-640T-640B~candidate, 670-680~candidate 710~ utilize current giant tile The absolute value of the address is different from the number of macroblocks per line by the flag variable T0PVALID. If the absolute value is not 〇, the opvalid is true. In addition, the T0PVALID is false. The 720~flag variable LEFTVALID is calculated by dividing the current giant tile address by the integer and the number of the parent macroblock. 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~Rotating logic, 950~predictive block, 960-990~absolute difference plus total calculation unit, 38Clienfs Docket N〇.:S3U06-0025 TT's Docket No:0608-A41237-TW/iinal/林璟辉/2007/06/15 38 200803527 1010~ instruction stream processor, 1020~ instruction, 1030~ instruction data, 1040~ execution unit pool, 1050~ texture filtering unit, 1060~ texture cache, 1070~ post wrapper, 1100~ video processing unit. 1120 ~ texture image, 113 〇 ~ target square, 1140-1170 ~ texture image

1 I , 1Π0Α-Β~ buffer. : ;

39Clienfs Docket No.:S3U06-0025 TT’s Docket No:0608-A41237-TW/fmal/林璟辉/2007/06/15 39

Claims (1)

200803527 &lt;10. Applying for the patent garden: The quantity is &quot;The movement of the test box&quot; The direction of the website is thoroughly searched to generate the ::::: area to thoroughly search for the better phase =: will ^ area 'The best match has an integer pixel resolution; the degree of conformity with the reference block is modeled as a long-term surface; the two counties _-minimum, the fractional resolution-best match block are determined analytically; ...~有向向=The best matching block with fractional resolution calculates a fractional shift. 2. The method of determining the relative motion of the motion vector according to the claim of the patent scope, wherein the decision is made in a plurality of predictors Which: a step that corresponds to the reference block preferably includes: determining whether the current frame is an intra-image prediction, and if the frame of the money is an image_test, then the co-technical brother determines the plurality of Which of the prediction blocks is the best match. Down~3=Apply the subjective (four) item's decision to describe the method of changing the mobile vector relative to the temple, wherein the step of determining the plurality of predictors and the reference block preferably includes . Client s Docket N〇.:S3U06-0025 匕3 · °c et No-〇608-A41237-TW/fmal/#if^/2〇〇7/〇6/15 40 200803527 Is the current frame Inter-image prediction; and if the x-ray frame is inter-image prediction, then search for _fu &amp; +: adjacent blocks • the plural predictions two =: 4. If the square shifts two: =:=, Refer to / for the face-to-face pre-blocks for the secret position - group 4 blocks, perform a partial search for this partial area. ▲ 5. The method of claim 1, wherein the step of determining the minimum value of the two-person surface comprises: determining a first The minimum value of the direction; the minimum value of the second direction perpendicular to the first direction. 6. If the determination of the scope of the patent application section describes the method of moving the movement vector relative to the reference block, the step of analyzing the minimum value of the secondary surface by the &amp; towel comprises: calculating the optimum Matches one of the adjacent blocks of the prediction block to the absolute difference plus the total value. The method of claim 1, wherein the step of determining the minimum value of the secondary surface relative to the movement of the reference block further comprises: counting one of the plurality of squares The sum of the difference values, the first of the plurality of blocks being adjacent to the best matching prediction block in a first direction, 41Clienfs Docket N〇.: S3U06-0025 TT'SD〇cketN0: _8-A41237-TW/ Fmal_f4/2〇〇7/〇6/15 4 200803527 1 The remaining squares of the plurality of squares are respectively adjacent to the other of the plurality of squares. 8. The method of claim 1, wherein the step of determining the recorded small value of the secondary surface comprises: ^ calculating a plurality of squares. An absolute difference plus a total value: and wherein the absolute difference plus total value of the plurality of calculated blocks is performed using an absolute difference summing instruction executed by a graphics processing unit 。. 9. A method of determining a motion vector describing movement relative to a reference block, the method comprising: determining, according to a matching criterion, which of the plurality of prediction blocks has a better match with the reference block; performing a partial region thoroughly Searching to produce a best match with the reference block, the local area is thoroughly searched for a region around the center of the preferred coincident prediction block, the best match having integer pixel resolution; and analytically determining the modeling The best match corresponds to a minimum of a quadratic surface of the degree of coincidence between the reference blocks, the minimum corresponding to a best matching block of fractional resolution. 10. The method of claim 9, wherein the step of determining the movement vector relative to the movement of the reference block, wherein the step of determining which of the plurality of prediction blocks has a better match with the reference block comprises: The current frame is an intra-image prediction, using a conjugate gradient to search 42Client's Docket N〇.:S3U06-0025 TT,s Docket No:0608-A41237-TW/fmal/林璟辉/2007/06/15 42 200803527 Method, searching for the plurality of prediction blocks to determine the preferred match. 11. The method of claim 1, wherein the determination of the movement vector relative to the movement of the reference block, wherein the step of searching for the plurality of prediction blocks to determine the preferred match further comprises: a prediction block selection, a candidate block; calculating a first absolute difference sum value of a first block to the left of a fixed distance of the candidate block and a second block to the right of the fixed distance of the candidate block a second absolute difference plus a horizontal gradient between the total values; calculating a third absolute difference sum value of a third-party block above the fixed distance of the candidate block and the fixed distance at the candidate block a fourth vertical difference between a fourth square and a vertical gradient between the total values; if the horizontal and vertical gradients are lower than a gradient threshold, the fixed distance is adjusted according to the horizontal and vertical gradients; a plurality of new candidate blocks located at a fixed distance from the block having the lowest absolute difference plus the total value in the first, second, third, and fourth blocks; and a plurality of new candidate blocks The new candidate block repeats the steps following the selection of a candidate block step. 12. The method of claim 11, wherein the method for describing the movement vector relative to the movement of the reference block further comprises: if the horizontal and vertical gradient system is greater than or equal to the gradient threshold, then the first The second, third, and fourth absolute difference plus total values are compared with an absolute difference plus the total threshold; 43Client's Docket N〇.: S3U06-0025 TT's Docket No: 0608-A41237-TW/fmal/林璟辉/2007/ 06/15 43 200803527 Any one of the first, second, third, and fourth absolute difference plus total values of the right 低于 is lower than the absolute difference plus the total threshold, then the first, second, second, The fourth block has the lowest difference plus the total value of the block to determine that the preferred match is 0 ^ 13. - a computer-readable medium having a program for determining a motion vector. The program contains the set wire to perform the following steps. Logic: &amp; Determine which of a plurality of prediction blocks and one according to the - match criterion The reference block has a better match; the field performs a partial search of the local area to produce a better match with the reference block. The local consumption domain search is in the (four) preferred matching prediction block as the central friend - (4), the domain can be used (4) Digital pixel resolution 1 Modeling the degree of matching between the best match and the reference block - the subsurface; ,, determining the minimum value of the secondary surface by the knife to determine the value of the singular t a best matching block having a knife angle resolution; and calculating a motion vector based on the best matching block having the fractional resolution. The method of claim 13 of the patent application scope is used to determine that the movement is readable by the computer, wherein the step of determining which of the plurality of prediction blocks has a preferred alignment with the reference block further comprises: Judging - the current frame is inter-image prediction or if the image in the image is in-image prediction, then a conjugate gradient is used to determine which of the plurality of pre-funds is the best match. 44Client5s Docket No.:S3U06-0025 ' TT s Docket No.0608-A41237-TW/fmal/林环辉/2007/06/15 44 200803527 轻参若°海目筋框框为 inter-image prediction, then Search for the neighboring box of 今夫# _ 扪 了 ° ° ° 苓 苓 。 。 。 。 。 。乂 该 该 犹 犹 犹 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测 预测Marriage system, miscellaneous search for the test of the surrounding area of the plural _ square ... =: heart: =: pre-two? Bureau, the better match around the prediction side - the region performs the local area thorough search. To the computer readable medium having the program for determining the movement σ in the 13th item of the patent scope, the method further comprises: determining whether the current frame is an intra-image prediction; In-image prediction, using a plurality of prediction blocks to determine which one of the plurality of prediction blocks is the preferred match and/or the preferred match_block (four) position - the group of 4 blocks, performing a partial area of 5 Search thoroughly. π· ”Please refer to the computer readable medium of claim 13 for determining the movement direction of the program, wherein the step of determining the minimum value of the secondary surface by the analysis further includes: The minimum value in one direction; 45Client5s Docket N〇.:S3U06-0025 TT's Docket No:0608-A41237-TW/fmal/林璟辉/2007/06/15 The determination is perpendicular to the first direction—the minimum of the second direction 45. The method of claim 13 is the computer readable medium having a program for determining a motion vector, wherein the step of determining a minimum value of the secondary surface analytically comprises: calculating the most The absolute difference between one of the adjacent blocks of the prediction block plus the total value of r I. : : 19. The computer readable medium having the program for determining the motion vector according to claim 13 of the patent application, wherein the determination is analytically determined The step of the minimum value of the secondary surface further comprises: calculating an absolute difference plus a total value of the plurality of blocks, wherein a first one of the plurality of blocks is adjacent to the best matching predictor in the first direction Block, the remaining squares of the plurality of squares are respectively adjacent to the other one of the plurality of squares 46Clienfs Docket N〇,:S3U06-0025 TT’s Docket No:0608-A41237-TW/fmal/林璟辉/2007/06/15