TWI423680B - Design space exploration method of reconfigurable motion compensation architecture - Google Patents

Description

Design space exploration method for reconfigurable mobile compensation architecture

The present invention relates to a design space exploration method for a reconfigurable mobile compensation architecture, and more particularly to a method for developing a reconfigurable mobile compensation architecture using the concept of architecture/algorithm collaborative design.

With the development of multimedia technology, various video compression standards such as ISO/IEC video compression standards MPEG-1, MPEG-2 and MPEG-4, ITU-T video compression standards H.263 and H.264 have been Successful development has enhanced the enjoyment of video in human life.

In recent years, in response to the application of different video compression standards, it has become an inevitable trend to develop decoders capable of supporting multi-standard video. Among the above video compression standards, motion compensation is the most important and core part. For the video decoder, the motion compensation process is a motion vector obtained according to motion estimation, and a corresponding block similar to the current picture is found from the reference picture, and is interpolated. The operation obtains the motion compensation predicted value.

Because of the similarities and differences between the motion compensation processes corresponding to these video compression standards, if the motion compensation processes of different video compression standards are implemented in the decoder together, it will inevitably consume repeated hardware resources to implement these motion compensation processes. Commonality between the two. Therefore, a design methodology is required to develop an efficient mobile compensation hardware architecture for the required application specifications to support motion compensation processing for different video compression standards.

The invention provides a design space exploration side of a reconfigurable mobile compensation architecture The method uses the concept of architecture/algorithm collaborative design to develop an efficient reconfigurable mobile compensation architecture for the required application specifications.

The invention provides a design space exploration method for a reconfigurable mobile compensation architecture. First, set the predetermined application specifications. Then, the commonality between the motion compensation algorithms corresponding to the various video compression standards is extracted, and the calculation amount of each motion compensation algorithm is analyzed to determine the arithmetic components included in the processing unit. Based on the predetermined application specifications, the data flow of each mobile compensation algorithm can be analyzed in the case of the worst-case reconfigurable mobile compensation architecture with different data granularity and different number of processing units. The hardware parameters corresponding to the motion compensation algorithm. Based on predetermined design goals and hardware parameters, a predetermined data granularity and a predetermined number of processing units of the reconfigurable mobile compensation architecture are selected.

The design space exploration method of the reconfigurable mobile compensation architecture, the embodiment of the present invention further includes analyzing the partition type of each block under different data granularities according to various block segmentation types supported by each mobile compensation algorithm. The reusable data in the reference block corresponding to each block segmentation type is obtained according to the hardware parameters corresponding to each mobile compensation algorithm. When the reconfigurable mobile compensation architecture is processed at a predetermined data granularity, the reusable data in the reference block is retained in the internal memory of the reconfigurable mobile compensation architecture.

The design space exploration method of the reconfigurable mobile compensation architecture, the embodiment of the present invention further includes analyzing the reference blocks required for each pixel interpolation type according to various pixel interpolation types supported by the motion compensation algorithms. And according to the hardware parameters corresponding to each mobile compensation algorithm. When the reconfigurable motion compensation architecture processes at a predetermined data granularity, the reference block corresponding to each pixel interpolation type is retrieved from the internal memory of the reconfigurable mobile compensation architecture.

Based on the above, the present invention analyzes the computational complexity of each motion compensation algorithm based on the commonality of the extraction, and further obtains a processing unit of the reconfigurable motion compensation architecture. borrow This processing unit saves hardware resources from processing common operations between these motion compensation algorithms. In addition, the present invention analyzes the data flow of each mobile compensation algorithm by using a reconfigurable mobile compensation architecture with different data granularities and different processing unit numbers, and obtains hardware parameters corresponding to each mobile compensation algorithm. These hardware parameters are weighed against predetermined design goals to develop a reconfigurable mobile compensation architecture that meets the predetermined application specifications and is more efficient.

Since various video compression standards define different profiles and levels according to different application requirements, this embodiment first sets a predetermined application specification, and develops through the design space exploration method under the predetermined application specifications. Reconfigurable mobile compensation architecture with better efficiency.

Table 1 is a table of predetermined application specifications supported by a reconfigurable mobile compensation architecture in accordance with an embodiment of the present invention. Referring to Table 1, the reconfigurable mobile compensation architecture developed in this embodiment is intended to support the main profile of the video compression standard MPEG-2 with the high level and the video compression standard MPEG-4. The advanced coding efficiency profile is the same as the L4 level and the main profile in the video compression standard H.264. It supports P-picture and B-picture processing, and can process 30 color format YCrCb per second. The ratio is a high resolution (1920×1088) picture of 4:2:0.

In order to support the video compression standard described above, the present embodiment analyzes the similarities and differences between the motion compensation algorithms corresponding to the video compression standards, and extracts at least one commonality between the motion compensation algorithms to enable reconfigurable movement. The compensation architecture can operate more efficiently and save hardware resources.

For example, the motion compensation algorithm corresponding to the video compression standard H.264 supports 1/4 pixel precision luminance interpolation and 1/8 pixel precision chrominance interpolation, and video compression standard MPEG. The motion compensation algorithm corresponding to -4 supports interpolation of 1/4 pixel precision and 1/2 pixel precision in luminance prediction and chrominance prediction, respectively. For 1/4 pixel accuracy, although the H.264 and MPEG-4 motion compensation algorithms use 6th and 8th order finite impulse response (FIR) filters for interpolation, respectively, H. 264 motion compensation algorithm uses a sequence of order coefficients [1, 5, 20, 20, -5, 1] and a sequence of order coefficients used by the MPEG-4 motion compensation algorithm [-8, 24, -48, 160, -48, 24 There is a sequence of common factors between , -8]. The sequence of the common factor coefficients is obtained by a simple addition operation and a displacement operation to obtain one of the above-mentioned sequence of order coefficients.

The MPEG-4 motion compensation algorithm uses a sequence of order coefficients [-8, 24, -48, 160, -48, 24, -8] divided by a divisor of 8 to obtain a sequence of order coefficients [-1, 3, -6, 20, 20, -6, 3, -1], this order coefficient sequence and the sequence coefficient sequence [1, 5, 20, 20, -5, 1] adopted by the H.264 motion compensation algorithm can be composed of the common factor sequence [1, 2] , 4, 16] can be obtained by the addition and displacement operations shown in Table 2 below. The order factor 3 is the coefficient 1 shifted to the left by 1 bit (that is, the coefficient 2 is obtained) and the coefficient 1 is added. The order coefficients -5 and -6 are the coefficient 1 shifted to the left by 2 bits (that is, the coefficient 4 is obtained) and added. The coefficient 2 or the coefficient 1, and the order coefficient 20 is the coefficient 1 shifted to the left by 4 bits and the coefficient 4 is added. In addition, video compression standards MPEG-2 and MPEG-4 support interpolation of 1/2 pixel precision in luminance prediction and chrominance prediction. Here, a simple bilinear filter can be used to generate the predicted value of motion compensation.

Then, based on the commonality described above, the embodiment analyzes the operation amount of the motion compensation algorithm corresponding to each video compression standard, and determines the operation included in the processing unit (PU) in the reconfigurable motion compensation architecture. element. FIG. 1 is a schematic diagram of non-integer pixel interpolation supported by the video compression standard H.264. Referring to FIG. 1, in the H.264 motion compensation algorithm, the 1/2 pixel position is interpolated by using 6 integer point pixels adjacent to the left and right, for example, 1/2 pixel position b=round ((E-5F) +20G+20H-5I+J)/32), where round() indicates rounding. Based on the commonality of the above extractions, in the H.264 motion compensation algorithm, the 1/2 pixel position b = [(G + J) (16 + 4) + (F + J) ((-4) + (- 1))+(E+J)+16]>>5, so that 1/2 pixel position b requires 8 addition operations and 1 displacement operation. Here, only the sequence of common factor coefficients can be analyzed, and H can be analyzed. The calculation amount required for each pixel position in the .264 motion compensation algorithm is not added to the addition and displacement operations obtained by the sequence of the common factor coefficients. Based on the above-described series of common factor coefficients, the operations required for each pixel position in the MPEG-4 and MPEG-2 motion compensation algorithms can also be analogized.

Referring to FIG. 1, among the various block partition types supported by the H.264 motion compensation algorithm, for example, block partition types such as 4×4, 8×4, 8×8, and 16×8, H.264 When the motion compensation algorithm performs pixel position i, f, k or q processing on the 4×4 block partition type, the interpolation value obtained by the operation under the 4×4 block partition type is low in reusability. And the interpolation pixel position i, f, k or q is computationally large and is the worst case. Similarly, the MPEG-4 motion compensation algorithm is the worst case for the pixel position a, b, c or d processing of the 8×8 block partition type, and the MPEG-2 motion compensation algorithm is for 16×16 block partitioning. The worst case is when the type is processed in pixel position a.

Therefore, based on the commonality described above, the present embodiment can estimate the amount of computation required for each of the different motion compression algorithms to process the luminance and chrominance interpolation in the P picture in the worst case, and the processing B. The brightness and chrominance interpolation in the picture are the operations required respectively. In terms of addition, since the MPEG-4 motion compensation algorithm is more complicated than the other two motion compensation algorithms, it takes a lot of addition operations to obtain the first pixel position in the worst case interpolation. Therefore, the processing unit is assumed. Contains enough additions to complete the MPEG-4 motion compensation algorithm for the worst case processing in one clock cycle. Thereby, the embodiment can determine the arithmetic components included in the processing unit.

In order to develop a reconfigurable mobile compensation architecture, this embodiment refers to various abstraction layers divided by the top-down design methodology, such as application specifications, algorithm layers, and architecture layers, and discusses the mobile compensation algorithms from the uppermost layer. The data flow, and the various design conditions are explored one by one to weigh the hardware that meets the intended design goals.

For the reconfigurable mobile compensation architecture, the top-level data flow is to perform the acquisition of the required data and the operation of the data in the unit clock cycle. Herein, the present embodiment analyzes the reconfigurable mobile compensation architecture based on the predetermined application specifications, and performs the mobile compensation calculation with different data granularity and different number of processing units in the case of peak computing and data configuration worst case. The data flow corresponding to the legal time, and according to the multiple hardware parameters corresponding to each mobile compensation algorithm, such as: peak bandwidth, number of bus rows, required memory capacity and operating frequency.

For example, the H.264 motion compensation algorithm supports a minimum of 4×4 block points. Since the type is cut, the data flow corresponding to each mobile compensation algorithm is calculated by using different numbers of processing units under the condition that each partition type is processed by 4×4 data granularity. Tables 3 to 5 are tables of the operating frequency estimated by the reconfigurable mobile compensation architecture in different numbers of processing units in the video compression standards MPEG-2, MPEG-4, and H.264, respectively. , B and P represent I picture, B picture, and P picture, respectively.

As shown in Tables 3~5, when 1 or 2 processing units are used, the reconfigurable mobile compensation architecture supports the video compression standard MPEG-4 and H.264, which requires a higher operating frequency, so it may consume too much hardware. Resources to achieve this working frequency. When three processing units are used, the reconfigurable mobile compensation architecture supports the above video compression standard, and the data flow is low in regularity and high in complexity. Although the use of five processing units reduces the operating frequency of the reconfigurable mobile compensation architecture, it also faces the same problems as using three processing units. Therefore, in the embodiment, based on the predetermined design target and the hardware parameters corresponding to the respective motion compensation algorithms, the predetermined data granularity (for example, 4×4) adopted by the reconfigurable mobile compensation architecture may be selected, and the predetermined a processing unit of a quantity (for example, four), wherein the predetermined design target is, for example, a hardware parameter corresponding to each mobile compensation algorithm, so that the most appropriate operating frequency can be obtained under a predetermined data granularity and a predetermined number of processing units, Bandwidth and the amount of memory required.

It should be noted that Tables 3 to 5 of the above embodiment are based on the commonality of the extraction and the application specifications of the reservation, and explore the data flow of the reconfigurable mobile compensation architecture to support different motion compensation algorithms under different design conditions, thereby obtaining each The hardware parameters corresponding to the motion compensation algorithm. However, the commonality of extractables supports different video compressions. The standards vary, and the data granularity and number of processing units used to meet the design goals will also vary with different application specifications. The above embodiments provide a design space exploration method for reconfigurable mobile compensation architecture. Efficiently develop hardware that meets the specifications of the intended application.

As described above, it can be summarized here as the following method flow. 2 is a flow chart of a method for exploring a design space of a reconfigurable mobile compensation architecture according to an embodiment of the present invention. Referring to FIG. 2, first, a predetermined application specification is set (step S201). The commonality is extracted by a motion compensation algorithm from a plurality of video compression standards (step S202), and the calculation amount of each motion compensation algorithm is analyzed to determine an arithmetic element included in the processing unit (step S203). Based on the predetermined application specifications, the data flow of the reconfigurable mobile compensation architecture for each mobile compensation algorithm with different data granularities and different numbers of processing units is analyzed and obtained under the worst case of peak computing and data configuration. The hardware parameters corresponding to the respective motion compensation algorithms (step S204). The predetermined data granularity and the predetermined number of processing units of the reconfigurable mobile compensation architecture are selected based on the predetermined design target and the hardware parameters (step S205).

Furthermore, in the design space exploration method of the reconfigurable mobile compensation architecture, other design strategies can be adopted to improve the work efficiency of the reconfigurable mobile compensation architecture. According to another embodiment of the present invention, according to the block segmentation type supported by each mobile compensation algorithm, when the segmentation type of each block is processed under different data granularities, the reference block corresponding to each block segmentation type can be reused. Information. FIG. 3 is a schematic diagram of a reference block required for a H.264 motion compensation algorithm based on a 16×16 block partition type according to an embodiment of the present invention. Referring to FIG. 3, when the H.264 motion compensation algorithm is processed by the 16×16 block partition type 310, the size of the 21×21 reference block 320 is required to calculate the pixels in the 16×16 block partition type 310. Interpolation.

If the reconfigurable motion compensation architecture completes pixel interpolation of 16 4×4 blocks included in the 16×16 block partition type 310 one by one with 4×4 data granularity, then 21×21 parameters Test block 320 has reusable material as indicated by shaded area 330. Under different data granularities, the reusable data in the reference block corresponding to each block segmentation type will be different. The design strategy adopted in this embodiment is to keep the reusable data in the internal memory for processing in the next block to reduce the bandwidth load on the external memory access data. In this embodiment, the hardware parameters corresponding to the motion compensation algorithms (for example, internal memory capacity and bandwidth) can be obtained, and the reconfigurable motion compensation is selected based on the predetermined design target and the hardware parameters. The architecture is best suited to the granularity of the intended data.

In addition, another embodiment of the present invention further analyzes the reference block required for the pixel interpolation type supported by each motion compensation algorithm. For example, in the video compression standard MPEG-2, the luminance and chrominance interpolation operations of integer dot pixels require M×N reference block size data and horizontal 1/2 pixel luminance and chrominance interpolation operations are required ( M+1)×N reference block size data, where M×N is the block partition type size. Similarly, the video compression standards MPEG-4 and H.264 need to capture reference blocks of different sizes for different pixel interpolation types. When the design strategy adopted in this embodiment is processed by the reconfigurable mobile compensation architecture, the corresponding reference block is extracted in the internal memory according to the pixel interpolation type to reduce the access to the external memory. Bandwidth load. Through this analysis, the hardware parameters (for example, bandwidth, etc.) corresponding to each mobile compensation algorithm can be obtained, and based on the predetermined design target and the hardware parameters, the most suitable reservation for the reconfigurable mobile compensation architecture is selected. Data granularity.

Through the above-described design space exploration method and design strategy, it is possible to develop the most suitable reconfigurable mobile compensation architecture in accordance with this predetermined application specification. 4 is a schematic diagram of a reconfigurable mobile compensation architecture according to an embodiment of the present invention. Referring to FIG. 4, based on the data reusable design strategy and/or the design strategy of the required reference block according to the pixel interpolation type, the present embodiment provides a data communication module in the reconfigurable mobile compensation architecture 400. Group 410, by which each block segmentation class is extracted from external memory The reference block corresponding to the type is in the internal memory 420, and the desired reference block can be retrieved in the internal memory 420 even depending on the type of pixel interpolation.

Then, when the reconfigurable motion compensation architecture 400 performs the pixel interpolation operation, the data communication module 410 obtains the required data from the internal memory 420 according to the predetermined data granularity, and stores the reusable data back into the internal data. The memory 420 is used for the next block operation. Since the pixel interpolation operation may refer to pixels in the horizontal direction or pixels in the vertical direction, the present embodiment provides the data supply module 430 in the reconfigurable motion compensation architecture 400 to arrange the data communication module 410 to access the data communication module 410. data. Through the above-mentioned design space exploration method, the number of processing units of the plurality of video compression standards can be determined by the reconfigurable mobile compensation architecture 400. In this embodiment, an interpolation module 440 including a predetermined number of processing units is provided to perform each The operation of the motion compensation algorithm. In order to control the cooperation between the modules, the embodiment needs to provide a parameter control module 450 in the reconfigurable motion compensation architecture 400, which receives parameters (for example, motion vectors) required to perform each motion compensation algorithm. And according to the operation of each module, the buffer module 460 is used for temporarily storing data.

In summary, the foregoing embodiment is based on the commonality between the motion compensation algorithms of different video compression standards, analyzes the computational complexity of each motion compensation algorithm, and thereby obtains a processing unit of the reconfigurable motion compensation architecture, where The processing unit can save the common operation between the hardware compensation processing and the motion compensation algorithm by utilizing the commonality. In addition, the above embodiment analyzes the data flow of each mobile compensation algorithm by the reconfigurable mobile compensation architecture by using different data granularities and different processing unit numbers, and the hardware parameters corresponding to each mobile compensation algorithm can be obtained from the data flow. Based on the predetermined design goals and the hardware parameters, the predetermined data granularity and the predetermined number of processing units used in the reconfigurable mobile compensation architecture may be selected, thereby developing a reconfigurable mobile compensation architecture with better efficiency.

E~N, P, Q‧‧‧ integer point pixel position

A~k, m, n, p~s, cc, dd, e, ff‧‧‧ non-integer pixel locations

310‧‧‧16×16 block partition type

320‧‧‧Reference block

330‧‧‧Reusable data

400‧‧‧Reconfigurable mobile compensation architecture

410‧‧‧ Data Communication Module

420‧‧‧ internal memory

430‧‧‧data supply module

440‧‧‧Interpolation module

450‧‧‧Parameter Control Module

460‧‧‧buffer module

S201~S205‧‧‧ steps of a design space exploration method for a reconfigurable mobile compensation architecture according to an embodiment of the present invention

FIG. 1 is a schematic diagram of non-integer pixel interpolation supported by the video compression standard H.264.

2 is a flow chart of a method for exploring a design space of a reconfigurable mobile compensation architecture according to an embodiment of the present invention.

FIG. 3 is a schematic diagram of a reference block required for a H.264 motion compensation algorithm based on a 16×16 block partition type according to an embodiment of the present invention.

4 is a schematic diagram of a reconfigurable mobile compensation architecture according to an embodiment of the present invention.

S201~S205‧‧‧ steps of a design space exploration method for a reconfigurable mobile compensation architecture according to an embodiment of the present invention

Claims (9)

A design space exploration method for a reconfigurable mobile compensation architecture includes: setting a predetermined application specification; extracting at least one commonality between a plurality of motion compensation algorithms respectively corresponding to the plurality of video compression standards; and analyzing each of the plurality of motion compensation algorithms based on the commonality The operation amount of the motion compensation algorithm is determined, and the arithmetic component included in the processing unit is determined; based on the predetermined application specification, the reconfigurable mobile compensation architecture is analyzed to be different in the case of the peak calculation amount and the worst case of the data configuration. The data granularity and the number of different processing units perform the data flow corresponding to each of the motion compensation algorithms, and obtain a plurality of hardware parameters corresponding to each of the motion compensation algorithms; and based on a predetermined design target and each The hardware parameters corresponding to the motion compensation algorithm select one of the predetermined data granularities used by the reconfigurable mobile compensation architecture, and a predetermined number of the processing units. For example, the design space exploration method of the reconfigurable mobile compensation architecture described in claim 1 further includes: processing each of the different data granularities according to various types of block segmentation supported by the mobile compensation algorithm. When the block segmentation type is used, one of the block segmentation types corresponds to the reusable data in the reference block, and the hardware parameters corresponding to each of the motion compensation algorithms are obtained, wherein the reconfigurable When the mobile compensation architecture processes the predetermined data granularity, the reusable data in the reference block is retained in the internal memory of the reconfigurable mobile compensation architecture. The method for exploring a design space of the reconfigurable mobile compensation architecture described in claim 2, further comprising: providing a data communication module in the reconfigurable mobile compensation architecture, and extracting each partition type Corresponding to the reference block in the internal memory, and Accessing the required data from the internal memory; providing a data supply module in the reconfigurable mobile compensation architecture to arrange the data accessed by the data communication module; and providing an interpolation module The reconfigurable mobile compensation architecture is configured to receive and process the data arranged by the data supply module, wherein the interpolation module is composed of the predetermined number of the processing units to perform each of the motion compensation Algorithm. The method for exploring a design space of the reconfigurable mobile compensation architecture described in claim 3, further comprising: providing a parameter control module in the reconfigurable mobile compensation architecture to receive and perform each of the motion compensation calculations The parameters required by the method, and accordingly, the internal memory, the data communication module, the data supply module, and the interpolation module are controlled. For example, the method for exploring the design space of the reconfigurable mobile compensation architecture described in claim 1 further includes: analyzing each pixel interpolation type according to various pixel interpolation types supported by the motion compensation algorithm. Dereferencing the block and obtaining the hardware parameters corresponding to each of the motion compensation algorithms, wherein when the reconfigurable motion compensation architecture processes the predetermined data granularity, each pixel interpolation is captured The reference block corresponds to the internal memory of the reconfigurable mobile compensation architecture. The method for exploring a design space of the reconfigurable mobile compensation architecture described in claim 5, further comprising: providing a data communication module in the reconfigurable mobile compensation architecture, and extracting each of the pixel interpolation types Corresponding the data block is in the internal memory, and accessing the required data from the internal memory; providing a data supply module in the reconfigurable mobile compensation architecture to arrange the data communication The data accessed by the module; and providing an interpolation module in the reconfigurable mobile compensation architecture to receive the capital The material supply module processes and processes the data, wherein the interpolation module is composed of the predetermined number of processing units to perform each of the motion compensation algorithms. The method for exploring a design space of the reconfigurable mobile compensation architecture described in claim 6 further includes: providing a parameter control module in the reconfigurable mobile compensation architecture to receive and perform each of the motion compensation calculations The parameters required by the method, and accordingly, the internal memory, the data communication module, the data supply module, and the interpolation module are controlled. The design space exploration method of the reconfigurable mobile compensation architecture described in claim 1, wherein the commonity comprises: using the addition operation and the displacement operation to reduce the plurality of stages of the motion compensation algorithms respectively. A sequence of common factor coefficients obtained from the sequence of coefficients. The design space exploration method of the reconfigurable mobile compensation architecture as described in claim 1, wherein the hardware parameters include a peak bandwidth, a number of bus ranks, a required memory capacity, and an operating frequency.