Access method of matrix data and storage device of the matrix data

Abstract

A matrix data access method is disclosed that includes following steps: the data storage space containing the presetting high-dimensional matrix is processed by dimensionality reduction to convert into two-dimensional matrix data storage space, and the two-dimensional matrix data storage space is divided into the separated two-dimensional matrix data storage subspace; the storage type of data in the high-dimensional matrix data storage subspace: the high-dimensional matrix is reduced to low-dimensional matrix, the storage type in every two-dimensional matrix data storage subspace: from left to right, from upper to lower, stored according to the modulus; the two-dimensional sub matrix data storage space is mapped to several physical memories in the data storage model, the data in data block of the two-dimensional sub matrix data storage space is read from the corresponding address of every memory in the several memories according to the matching requirements. The invention also provides a matrix data storage device. The invention is high storage effect and can satisfy the requirements for storing the various types of data.

Description

(1) Technical field

[0001] The invention relates to a matrix data access method and a matrix data storage device thereof. (2) Background technology

[0002] The vector processor is widely used in the fields of multimedia processing and scientific computing. The data objects it operates are mostly matrix data with different dimensions and data types. Therefore, it is very important to realize efficient matrix data access.

[0003] For example, Chinese invention patent 95103825.7, the patent name is a parallel storage device for image processing using linear transformation, which discloses a parallel storage device for image processing using linear transformation, including multiple address calculation units, multiple storage units and omega A network and a shift circuit unit, the address calculation unit is adapted to generate an address each corresponding to a module in the storage module according to the access mode, reference coordinates and the number of storage modules o...

Examples

Embodiment 1

[0098] As shown in Figure 1, a matrix data storage method includes a predetermined high-dimensional matrix data storage space 1, which can be transformed into multiple two-dimensional matrix data storage spaces 2 after multi-level dimensionality reduction processing, and each two-dimensional matrix data storage space 2 The two-dimensional matrix data storage space 2 can be further decomposed into a series of logically independent two-dimensional sub-matrix data storage spaces 3 . The data is stored in the high-dimensional matrix data storage space 1 from high-dimensional to low-dimensional, and stored in each two-dimensional sub-matrix data storage space in the order from left to right and from top to bottom , according to the mode of storage;

[0099] The two-dimensional sub-matrix data storage space 3 is mapped into a plurality of physical memories through the configuration of the data access mode, as shown in Figure 2, the content of the configuration includes: base address...

Embodiment 2

[0137] A matrix data storage device composed of multiple physical memory units connected in parallel, the logical storage space of the matrix data storage device includes a predetermined high-dimensional matrix data storage space, and the high-dimensional matrix data storage space is finally processed through multi-level dimensionality reduction It can be transformed into multiple two-dimensional matrix data storage spaces, and each two-dimensional matrix data storage space can be decomposed into a series of logically independent two-dimensional sub-matrix data storage spaces. As shown in Figure 5, the matrix data storage device includes:

[0138] Configuration registers, including base address registers, data element byte width registers, row span registers, column span registers, data storage mode registers, and interleave quota registers.

[0139] The address calculation unit group is composed of the same number of address calculation units as the physical memory unit, each...

Embodiment 3

[0148] In this embodiment, the matrix data access method is applied to video signal processing. Usually, a frame of video data is divided into several macroblocks, and each independent macroblock is composed of a luminance signal Y and two chrominance signals U and V according to 4 :2:0 ratio composition. Wherein, the luminance signal is a 16×16 matrix, and the two chrominance signals are 8×8 matrices respectively. Using the matrix storage device of the present invention can effectively realize the storage of such macroblock data.

[0149] Consider each independent macroblock as a four-dimensional matrix data storage space, where one dimension is time, as shown in Figure 7, a group of mutually independent matrix data storage spaces 1 are arranged in sequence along the time t axis, each The three-dimensional data space 1 is composed of three mutually independent two-dimensional matrix data storage spaces 2, and each two-dimensional matrix data storage space 2 stores a kind of ...

(1) Technical field

[0001] The invention relates to a matrix data access method and a matrix data storage device thereof. (2) Background technology

[0002] The vector processor is widely used in the fields of multimedia processing and scientific computing. The data objects it operates are mostly matrix data with different dimensions and data types. Therefore, it is very important to realize efficient matrix data access.

[0003] For example, Chinese invention patent 95103825.7, the patent name is a parallel storage device for image processing using linear transformation, which discloses a parallel storage device for image processing using linear transformation, including multiple address calculation units, multiple storage units and omega A network and a shift circuit unit, the address calculation unit is adapted to generate an address each corresponding to a module in the storage module according to the access mode, reference coordinates and the number of storage modules o...

Examples

Embodiment 1

[0098] As shown in Figure 1, a matrix data storage method includes a predetermined high-dimensional matrix data storage space 1, which can be transformed into multiple two-dimensional matrix data storage spaces 2 after multi-level dimensionality reduction processing, and each two-dimensional matrix data storage space 2 The two-dimensional matrix data storage space 2 can be further decomposed into a series of logically independent two-dimensional sub-matrix data storage spaces 3 . The data is stored in the high-dimensional matrix data storage space 1 from high-dimensional to low-dimensional, and stored in each two-dimensional sub-matrix data storage space in the order from left to right and from top to bottom , according to the mode of storage;

[0099] The two-dimensional sub-matrix data storage space 3 is mapped into a plurality of physical memories through the configuration of the data access mode, as shown in Figure 2, the content of the configuration includes: base address...

Embodiment 2

[0137] A matrix data storage device composed of multiple physical memory units connected in parallel, the logical storage space of the matrix data storage device includes a predetermined high-dimensional matrix data storage space, and the high-dimensional matrix data storage space is finally processed through multi-level dimensionality reduction It can be transformed into multiple two-dimensional matrix data storage spaces, and each two-dimensional matrix data storage space can be decomposed into a series of logically independent two-dimensional sub-matrix data storage spaces. As shown in Figure 5, the matrix data storage device includes:

[0138] Configuration registers, including base address registers, data element byte width registers, row span registers, column span registers, data storage mode registers, and interleave quota registers.

[0139] The address calculation unit group is composed of the same number of address calculation units as the physical memory unit, each...

Embodiment 3

[0148] In this embodiment, the matrix data access method is applied to video signal processing. Usually, a frame of video data is divided into several macroblocks, and each independent macroblock is composed of a luminance signal Y and two chrominance signals U and V according to 4 :2:0 ratio composition. Wherein, the luminance signal is a 16×16 matrix, and the two chrominance signals are 8×8 matrices respectively. Using the matrix storage device of the present invention can effectively realize the storage of such macroblock data.

[0149] Consider each independent macroblock as a four-dimensional matrix data storage space, where one dimension is time, as shown in Figure 7, a group of mutually independent matrix data storage spaces 1 are arranged in sequence along the time t axis, each The three-dimensional data space 1 is composed of three mutually independent two-dimensional matrix data storage spaces 2, and each two-dimensional matrix data storage space 2 stores a kind of ...

Images