Convolutional operation method using block-based distributed input output and compression
The method of block-unit distributed input/output and compression optimizes convolution operations by using sequential access for row and column vectors, improving performance and reducing data size to distribute data across multiple devices, thus speeding up convolution operations.
Patent Information
- Application Number
- KR1020240153485
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2024-11-01
- Publication Date
- 2026-11-11
- Estimated Expiration
- 2044-11-01
AI Technical Summary
Conventional convolution operations face challenges in improving input/output performance of auxiliary storage devices due to low throughput and limitations in using more than four secondary storage devices, particularly with column vector output times being lengthy and stride access methods.
A method involving block-unit distributed input/output and compression, utilizing sequential access for row and column vectors, includes steps of reading row vectors into main memory, performing Fast Fourier Transforms, transposing and distributing blocks across multiple auxiliary storage devices, and using Hadamard products to optimize data distribution and compression.
This approach enhances input/output performance by reducing data size and distributing data across more than four auxiliary storage devices, significantly decreasing the time required for convolution operations.
Smart Images

Figure 112024120014794-PAT00004_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to a convolution operation method using block-unit distributed input / output and compression, which can improve the input / output performance of an auxiliary storage device by utilizing sequential access for both row vectors and column vectors during the input / output process of a convolution operation, and can also distribute data across four or more auxiliary storage devices by reducing the data size through data compression, thereby effectively reducing the input / output time required for the convolution operation. Background Technology
[0002] As is well known, convolution is a mathematical operator that inverts and shifts a signal or function to multiply it by another signal and then integrates it. It is used for the composition of functions and signals and is utilized in a wide range of fields, such as artificial intelligence, signal processing, and holography.
[0003] To explain this convolution operation, the mathematical theorem states that the convolution of two signals is equal to the Adamdar product of their Fourier spaces, and the operation complexity of the convolution of a one-dimensional function is O(n 2 However, since the computational complexity of the Fast Fourier Transform (FFT) is O(n log n) and the computational complexity of the convolution is reduced, it can be computed faster.
[0004] Also, since the result of the convolution theorem is a circular convolution, to obtain a linear convolution using the convolution theorem, zero padding must be performed by adding zeros to the discrete signal.
[0005] Here, the Fast Fourier Transform (FFT) reduces the computational complexity of the Discrete Fourier Transform to O(n²). 2 As a fast computation algorithm that can reduce complexity from ) to O(nlog n), the computational complexity of the 2D Fast Fourier Transform is O(n 2 log n) is.
[0006] Meanwhile, wavefront propagation is a technique that utilizes convolution operations. Data can be stored in secondary storage to perform convolution operations on matrices larger than those in main memory. The matrix can perform read-1D matrix operations-write operations in three stages. By sequentially reading the row vectors of the matrix, performing a 1D Fourier transform, transposing them into column vectors, and storing them in the column direction in secondary storage, sequential access reading and stride access writing are performed. This results in improved read performance but relatively lower write performance.
[0007] In addition, convolution operations using multiple secondary storage devices are a technique for distributing input and output operations across multiple secondary storage devices. By dividing the matrix through implicit convolution and distributing the input and output operations across up to four secondary storage devices, the input and output throughput of each secondary storage device is reduced, thereby allowing for a reduction in data input / output time during the operation.
[0008] In the case of the conventional convolution operation described above, there is a problem in that outputting a column vector to a secondary storage device takes a long time due to the low output throughput caused by the stride access method, and there are limitations to performance improvement because only up to four secondary storage devices can be used. Prior art literature
[0009] 1. Korean Registered Patent No. 10-2065672 (Registered Jan. 07, 2020) The problem to be solved
[0010] The present invention aims to provide a convolution operation method using block-unit distributed input / output and compression, which can improve the input / output performance of auxiliary storage devices by utilizing sequential access for both row vectors and column vectors during the input / output process of a convolution operation, and can also reduce the data size through data compression to distribute data across four or more auxiliary storage devices, thereby effectively reducing the input / output time required for the convolution operation.
[0011] The purposes of the embodiments of the present invention are not limited to those mentioned above, and other unmentioned purposes will be clearly understood by those skilled in the art from the description below. means of solving the problem
[0012] According to an embodiment of the present invention, a first step of reading each row vector of a target matrix to be operated on from one auxiliary storage device into a main memory device, performing a first zero padding and a first Fast Fourier Transform to obtain a first row vector set matrix, transposing the obtained first row vector set matrix, and dividing a compressed first compressed block into blocks that are submatrices, and distributing and storing a plurality of first divided blocks in a plurality of auxiliary storage devices; A second step of reading a column vector block among the plurality of first compression blocks from the plurality of auxiliary storage devices, decompressing the column vector block by rearranging it using a transpose matrix method, performing second zero padding and a second Fast Fourier Transform to obtain a submatrix of the same size as the row vector set matrix, multiplying the obtained submatrix by an Hadamard transfer function and then performing a first inverse Fast Fourier Transform to obtain a second row vector set matrix, unpadding and transposing the obtained second row vector set matrix, and dividing the compressed second compression block into blocks corresponding to the submatrix, thereby distributing and storing the plurality of second divided blocks in the plurality of auxiliary storage devices; A convolution operation method using block-unit distributed input / output and compression may be provided, comprising: a third step of reading a row vector block among the plurality of second compression blocks from the plurality of auxiliary storage devices, decompressing the row vector block, performing a second inverse Fast Fourier Transform to obtain a third row vector set matrix, unpadding the obtained third row vector set matrix, and storing it by sequential writing at the location of the target matrix.
[0013] In addition, according to an embodiment of the present invention, the first step may provide a convolution operation method using block-unit distributed input / output and compression, wherein each row vector is read into the main memory in a sequential access manner for the target matrix, and the plurality of first compression blocks are distributed and stored in the plurality of auxiliary memory devices in a sequential access manner.
[0014] In addition, according to an embodiment of the present invention, the first step may provide a convolution operation method using block-unit distributed input / output and compression, wherein the first step involves performing a row-direction Fourier transform on the target matrix, transposing it, dividing it into blocks, and distributing and storing them in a plurality of auxiliary memory devices.
[0015] In addition, according to an embodiment of the present invention, the second step may provide a convolution operation method using block-unit distributed input / output and compression, wherein the column vector block among the plurality of first compression blocks is read from the plurality of auxiliary storage devices in a sequential access manner, and the plurality of second compression blocks are distributed and stored in the plurality of auxiliary storage devices in a sequential access manner.
[0016] In addition, according to an embodiment of the present invention, the second step may be provided with a convolution operation method using block-unit distributed input / output and compression having the same size as the Fourier space of the zero-padded target matrix in the case of the transfer function that performs an Hadamard product on the submatrix.
[0017] In addition, according to an embodiment of the present invention, the second step may be provided as a convolution operation method using block-unit distributed input / output and compression, which crops the transfer function in accordance with the size and position of the submatrix after performing an Hadamard product.
[0018] In addition, according to an embodiment of the present invention, the third step may provide a convolution operation method using block-unit distributed input / output and compression, wherein the row vector block among the plurality of second compression blocks is read from the plurality of auxiliary storage devices in a sequential access manner and stored at the location of the target matrix in a sequential access manner. Effects of the invention
[0019] The present invention can improve the input / output performance of auxiliary storage devices by using sequential access for both row vectors and column vectors during the input / output process of convolution operations, and can also distribute data across four or more auxiliary storage devices by reducing the data size through data compression, thereby effectively reducing the input / output time required for convolution operations. Brief explanation of the drawing
[0020] FIG. 1 is a flowchart illustrating a convolution operation process using block-unit distributed input / output and compression according to an embodiment of the present invention, and FIGS. 2 to 9 are drawings for explaining the detailed configuration of a convolution operation process using block-unit distributed input / output and compression according to an embodiment of the present invention. Specific details for implementing the invention
[0021] The advantages and features of the embodiments of the present invention, and the methods for achieving them, will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims. Throughout the specification, the same reference numerals refer to the same components.
[0022] In describing the embodiments of the present invention, specific descriptions of known functions or configurations will be omitted if it is determined that such detailed descriptions could unnecessarily obscure the essence of the invention. Furthermore, the terms described below are defined in consideration of their functions in the embodiments of the present invention, and these definitions may vary depending on the intentions or practices of the user or operator. Therefore, such definitions should be based on the content throughout this specification.
[0023] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.
[0024] FIG. 1 is a flowchart showing a convolution operation process using block-unit distributed input / output and compression according to an embodiment of the present invention, and FIG. 2 to 9 are drawings for explaining the detailed configuration of a convolution operation process using block-unit distributed input / output and compression according to an embodiment of the present invention.
[0025] First, the convolution operation method using block-unit distributed input / output and compression according to an embodiment of the present invention can be implemented in a device including a processor (10), a main memory (20), and a plurality of auxiliary memory devices (30). The processor (10) can handle the overall process for performing the convolution operation method using block-unit distributed input / output and compression. It can read data from the plurality of auxiliary memory devices (30), temporarily store it in the main memory (20), process the convolution operation, and then store the result back in the plurality of auxiliary memory devices (30).
[0026] And, the main memory (20) is a space that the processor (10) directly accesses to read and write data and stores currently running programs and data, and may include, for example, RAM, DRAM, etc.
[0027] Additionally, a plurality of auxiliary storage devices (30) are used to store and keep data for a long period of time and may include, for example, a hard disk drive (HDD), a solid-state drive (SSD), a USB drive, an external hard drive, etc.
[0028] These multiple auxiliary storage devices (30) refer to non-volatile data storage devices or memory devices used to preserve data of a computer device. Since they are primarily created for the purpose of permanently preserving data, they have a larger storage capacity and slower speed than the main memory (20), and in an embodiment of the present invention, they can be used to store data larger than that of the main memory (20).
[0029] Here, a Solid State Drive (SSD) is a type of secondary storage device that can store data using semiconductor memory. Compared to a Hard Disk Drive (HDD), it has faster data input / output speeds but smaller storage capacity. Internally, it can be composed of DDR cache memory, NAND flash memory, and controllers. It is known to have the characteristic of having higher random write performance than random read performance, and is primarily known to have a problem where performance degrades when reading and writing are performed simultaneously.
[0030] In the embodiments of the present invention, it is explained under the assumption that all data layouts are processed in a row-major manner.
[0031] To briefly explain this row-major method, when multidimensional data is represented in one-dimensional memory, it indicates the adjacency of data of any dimension. For example, in a two-dimensional matrix, row-major means that row elements are adjacent in memory, while column elements are separated by the length of the row. In this case, accessing row data is sequential access, and accessing column data is stride access.
[0032] Here, sequential access refers to sequentially accessing (I / O) contiguous data in the memory area; generally, because cache efficiency is high and data is adjacent, data I / O can be performed quickly from the storage device.
[0033] In addition, stride access refers to accessing data located at intervals in a memory area. Although the data is not randomly scattered like in random access (input / output) where data in a memory area is accessed randomly, data input / output performance is low because data cannot be accessed sequentially, and it generally occurs when accessing multidimensional data.
[0034] The process of performing a convolution operation using the processor (10), main memory (20), and multiple auxiliary memory devices (30) as described above is explained below.
[0035] Referring to FIGS. 1 to 10, a first step can be performed in which a processor (10) reads each row vector of a target matrix to be processed from one auxiliary storage device (30) into a main memory device (20), performs a first zero padding and a first fast Fourier transform to obtain a first row vector set matrix, transposes the obtained first row vector set matrix, divides the compressed first compressed block into blocks that are submatrices, and distributes and stores a plurality of first divided blocks in a plurality of auxiliary storage devices (30) (110).
[0036] In the first step (110) above, each row vector is read into the main memory (20) by sequential access to the target matrix from the processor (10), and a plurality of first compressed blocks can be distributed and stored in a plurality of auxiliary memory devices (30) by sequential access.
[0037] In addition, in the first step (110), the processor (10) can perform a row-direction Fourier transform on the target matrix, transpose it, divide it into blocks, and distribute and store it in multiple auxiliary memory devices (30).
[0038] For example, the processor (10) can perform an X-axis Fast Fourier Transform (FFT) on the target matrix, and as shown in FIGS. 2 and 3, the processor (10) can read each row vector of the target matrix (M0) to be operated on from one of the auxiliary memory devices (30, d0) into the main memory (20) and perform a first zero padding and a first Fast Fourier Transform to obtain a first row vector set matrix (2N×B).
[0039] Here, the first zero padding means adding zero values to the data. This process can be used to match the input / output size of the data or to remove circularity in the data, and as a result of zero padding, the length of the vector increases from N to 2N.
[0040] In addition, the first Fast Fourier Transform is an algorithm that transforms a signal in the time domain into the frequency domain, and the computational complexity of the Discrete Fourier Transform is O(n 2 It can be reduced from ) to O(nlog n).
[0041] Next, the first row vector set matrix obtained from the processor (10) is transposed (B×2N), and the compressed first compressed block is stored in a buffer. Then, the block is divided into submatrix blocks (2N / B B×B) to form a plurality of first divided blocks in a plurality of auxiliary storage devices (30, d0,..., d 2N / B It can be distributed and stored in each of the ).
[0042] Next, a second step can be performed in which a column vector block among a plurality of first compressed blocks is read from a plurality of auxiliary storage devices (30) in the processor (10), the column vector block is rearranged using a transpose matrix method to decompress, a second zero padding and a second Fast Fourier Transform are performed to obtain a submatrix of the same size as the row vector set matrix, the transfer function is multiplied by the obtained submatrix and a first inverse Fast Fourier Transform is performed to obtain the second row vector set matrix, the obtained second row vector set matrix is unpadding and transposed, and the compressed second compressed block is divided into blocks that are submatrices, and the plurality of second divided blocks are distributed and stored in a plurality of auxiliary storage devices (30) (step 120).
[0043] In the second step (120) above, the processor (10) can read column vector blocks from a plurality of first compression blocks in a sequential access manner from a plurality of auxiliary storage devices (20), and distribute and store a plurality of second compression blocks in a plurality of auxiliary storage devices (30) in a sequential access manner.
[0044] And, in the second step (120) above, the transfer function that multiplies the submatrix by the Hadamard function can have the same size as the Fourier space of the zero-padded target matrix.
[0045] In addition, in the second step (120) above, the processor (10) can crop the transfer function after performing an Hadamard product, corresponding to the size and position of the submatrix.
[0046] For example, the processor (10) can perform a Y-axis Fast Fourier Transform (FFT), an Amadara product, and a Y-axis Inverse Fast Fourier Transform (IFFT) on the target matrix. As shown in FIGS. 4 and 5, the processor (10) reads a column vector block from a plurality of auxiliary storage devices (30) among a plurality of first compression blocks, rearranges the column vector block stored in the buffer using a transpose matrix method (N×B) to decompress it, and then performs a second zero padding and a second Fast Fourier Transform to obtain a submatrix of the same size as the row vector set matrix.
[0047] Here, the second zero padding is similar to the treatment of the first zero padding as described above, so it will be omitted here.
[0048] In addition, the second Fast Fourier Transform (ID-FFT) performs a one-dimensional Fast Fourier Transform, reducing the computational complexity of the Discrete Fourier Transform to O(n 2 The time domain signal can be rapidly converted to the frequency domain by reducing it to O(nlog n), and as a result of performing a one-dimensional Fast Fourier Transform, a submatrix of size 2N×B may exist in the memory area, which is equal to the submatrix of the Fast Fourier Transform of the target matrix (M0).
[0049] Next, the transfer function (H(fx, fy)) is applied to the Haramard product of the submatrix obtained from the processor (10), and then the first inverse Fast Fourier Transform is performed to obtain a second row vector set matrix; the obtained second row vector set matrix is un-padding and transposed, and the compressed second compressed block is stored in a buffer; then, the submatrix is divided into blocks (2N / B) and a plurality of second divided blocks are stored in a plurality of auxiliary storage devices (30, d0,..., d 2N / B It can be distributed and stored in ).
[0050] Here, the transfer function (H(fx, fy)) has a size of 2N×2N, which is the same as the Fourier space of the zero-padded target matrix (M0), and can be cropped to a size of 2N×B corresponding to the size and position of the corresponding submatrix.
[0051] And, the first inverse Fast Fourier Transform (1D-IFFT) performs a one-dimensional inverse Fast Fourier Transform, reducing the computational complexity of the Discrete Fourier Transform to O(n 2 It is possible to quickly convert (restore) a signal in the frequency domain to the time domain by reducing it from ) to O(nlog n).
[0052] In addition, the second row vector set matrix can be unpaved to form an N×B matrix.
[0053] Meanwhile, the processor (10) can perform a third step of reading row vector blocks from a plurality of auxiliary storage devices (30) among a plurality of second compressed blocks, decompressing the row vector blocks, performing a second inverse fast Fourier transform to obtain a third row vector set matrix, unpadding the obtained third row vector set matrix, and storing it in the location of the target matrix by sequential writing (step 130).
[0054] In the above third step (130), the processor (10) can read row vector blocks from a plurality of secondary storage devices (30) in a sequential access manner and store them in the location of the target matrix in a sequential access manner.
[0055] For example, the processor (10) may perform an inverse Fast Fourier Transform (IFFT) on the target matrix, and as illustrated in FIGS. 6 and 7, the processor (10) reads row vector blocks from a plurality of secondary storage devices (30) among a plurality of second compressed blocks, rearranges and stores them in a buffer, decompresses the row vector blocks stored in the buffer, performs a second inverse Fast Fourier Transform (1D-IFFT) to obtain a third row vector set matrix, and after unpadding the obtained third row vector set matrix, one of the secondary storage devices (30, d a It can be saved by sequential writing to the target matrix location of ).
[0056] Here, the second inverse Fast Fourier Transform (1D-IFFT), unpadding, etc. have been described above, so their explanation is omitted here.
[0057] To further explain the allocation to auxiliary storage devices by block in the convolution operation method using block-unit distributed input / output and compression according to the embodiment of the present invention as described above, since the number of available auxiliary storage devices is generally relatively smaller than the number of blocks, one auxiliary storage device must input / output multiple blocks. As shown in FIG. 8, in order to evenly distribute the input / output load of the auxiliary storage devices, the index of the block can be allocated as shown in Equation 1 below.
[0058]
[0059] Here, ID device represents the index of the secondary storage, and b i,j represents the block of the 1st row and jth column, and M represents the total number of secondary storage devices.
[0060] In the convolution operation method using block-unit distributed input / output and compression according to the embodiment of the present invention as described above, a matrix larger than the main memory (20) is stored in a plurality of auxiliary memory devices (30), and input / output is performed efficiently to perform the operation, and the convolution theorem, which performs the convolution through a Fourier transform, can be utilized. A two-dimensional Fourier transform can be expressed as a convolution of one-dimensional Fourier transforms with respect to rows and columns, and a two-dimensional convolution can be performed by performing the one-dimensional Fourier transform multiple times using this.
[0061] In addition, in the convolution operation method using block-unit distributed input / output and compression according to an embodiment of the present invention, the data can be compressed before storing the data in a plurality of auxiliary storage devices (30) to reduce the data size, and various compression techniques presented in the prior art can be used, and the input unit of compression can be compressed into a unit relatively smaller than the block (i.e., input / output size), and the compression throughput must be relatively higher than the sum of the input / output and operation throughput of the convolution operation process.
[0062] In addition, in the convolution operation method using block-unit distributed input / output and compression according to the embodiment of the present invention, when the result of performing a one-dimensional Fourier transform is output to a plurality of auxiliary storage devices (30), the column vector is written using a sequential access method rather than a stride access method. To this end, the set of column vectors can be divided into blocks (square matrices) to perform sequential writing for each block, and at this time, the number of the plurality of auxiliary storage devices (30) can be set to be equal to the number of blocks.
[0063] Here, if the number of available auxiliary storage devices (30) is relatively smaller than the number of blocks, the blocks and the multiple auxiliary storage devices (30) can be mapped so that they can be evenly distributed among the multiple auxiliary storage devices (30).
[0064] Meanwhile, in the convolution operation method using block-unit distributed input / output and compression according to an embodiment of the present invention, as illustrated in FIG. 9, by using a GPU in addition to a CPU, data can be divided once more to fit the GPU memory size. In order to perform operations even on a GPU with a small memory size without reducing the block size, data in the main memory buffer can be divided and transmitted to the GPU. In this case, data in the main memory can be divided by preserving rows or columns so that all operations (rearrangement to compression) can be performed on the GPU.
[0065] Accordingly, according to an embodiment of the present invention, in the input / output process of a convolution operation, the input / output performance of an auxiliary storage device can be improved by using sequential access for both row vectors and column vectors, and the data can be distributed to four or more auxiliary storage devices by reducing the data size through data compression, thereby effectively reducing the input / output time required for the convolution operation.
[0066] In conclusion, in the convolution operation method using block-unit distributed I / O and compression according to an embodiment of the present invention, the I / O cost per auxiliary storage device is 5N 2 / (N d ×CR), where Nd represents the number of secondary storage devices and CR is the compression ratio.
[0067] On the other hand, the input / output cost in the prior art is 5N2, so according to an embodiment of the present invention, not only can the input / output performance be theoretically improved by Nd×CR times, but by performing block-unit input / output, both reading and writing can be processed in a sequential access manner, thereby improving the input / output performance.
[0068] Meanwhile, as a result of testing the convolution operation method using block-unit distributed input / output and compression according to the embodiment of the present invention as described above in the experimental environment of Table 1 below, it can be confirmed that the convolution operation can be processed up to 3.12 times faster than the prior art as shown in Table 2 below, the operation can be processed up to 3.10 times faster when the number of secondary storage devices is the same, and the convolution operation can be processed up to 2.09 times faster than the prior art using multiple secondary storage devices.
[0069]
[0070]
[0071] Although various embodiments of the present invention have been presented and described in the above description, the present invention is not necessarily limited thereto, and those skilled in the art will readily understand that various substitutions, modifications, and changes are possible within the scope of the technical concept of the present invention. Explanation of the symbols
[0072] 10 : Processor 20 : Main memory 30 : Multiple auxiliary storage devices
Claims
Claim 1 A first step of reading each row vector of a target matrix to be processed from one auxiliary storage device into main memory, performing first zero padding and a first Fast Fourier Transform to obtain a first row vector set matrix, transposing the obtained first row vector set matrix, dividing the compressed first compressed block into blocks that are submatrices, and distributing and storing a plurality of first divided blocks in a plurality of auxiliary storage devices; reading a column vector block among the plurality of first compressed blocks from the plurality of auxiliary storage devices, rearranging the column vector block using a transpose matrix method to decompress it, performing second zero padding and a second Fast Fourier Transform to obtain a submatrix of the same size as the row vector set matrix, multiplying the obtained submatrix by an Hadamard transfer function and performing a first inverse Fast Fourier Transform to obtain a second row vector set matrix, and unpadding the obtained second row vector set matrix A second step of transposing and dividing the compressed second compressed block into blocks that are submatrices, and distributing and storing a plurality of second divided blocks in the plurality of auxiliary storage devices; and a third step of reading a row vector block among the plurality of second compressed blocks from the plurality of auxiliary storage devices, decompressing the row vector block, performing a second inverse Fast Fourier Transform to obtain a third row vector set matrix, unpadding the obtained third row vector set matrix, and storing it by sequential writing at the location of the target matrix;A convolution operation method using block-unit distributed I / O and compression, comprising: a first step of reading each row vector of the target matrix into the main memory using a sequential access method for the target matrix, distributing and storing the plurality of first compression blocks in the plurality of auxiliary memory devices using a sequential access method, performing a row-direction Fourier transform on the target matrix, transposing it, dividing it into blocks, and distributing and storing them in the plurality of auxiliary memory devices; a second step of reading the column vector block among the plurality of first compression blocks from the plurality of auxiliary memory devices using a sequential access method, distributing and storing the plurality of second compression blocks in the plurality of auxiliary memory devices using a sequential access method, wherein the transfer function that performs an Hadamard multiplication on the submatrix has the same size as the Fourier space of the zero-padded target matrix, and crops the result corresponding to the size and position of the submatrix after performing the Hadamard multiplication of the transfer function. Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 delete Claim 6 delete Claim 7 In claim 1, the third step is a convolution operation method using block-unit distributed input / output and compression, wherein the row vector block among the plurality of second compression blocks is read from the plurality of auxiliary storage devices in a sequential access manner and stored at the location of the target matrix in a sequential access manner.
Citation Information
Patent Citations
Multiple auxiliary storage media used convolution operating method
KR1020230144193A
