A systolic array that simultaneously supports fine-grained deformation and multi-data stream switching

By designing a pulsating array system that supports fine-grained deformation and multi-data stream switching, the problem of low utilization of computing units was solved, and computational efficiency was improved to efficiently adapt to various neural network models.

CN116644263BActive Publication Date: 2026-11-17BEIHANG UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202310080631.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-01-19
Publication Date
2026-11-17
Estimated Expiration
2043-01-19

AI Technical Summary

Technical Problem

Existing systolic arrays suffer from low utilization of computing units when dealing with matrix multiplication operations of special shapes. They are unable to dynamically change the logic structure and support multiple data stream switching, resulting in limited computing efficiency.

Method used

Design a pulsating array system comprising a computing unit array supporting four-way full-duplex data transfer and multi-data stream switching, a fine-grained multi-mode on-chip data cache, and a controller, capable of dynamically changing logic shape and data flow to improve computing unit utilization.

Benefits of technology

It improves the utilization rate and acceleration effect of systolic arrays when calculating matrix multiplications of various shapes and sizes, and is suitable for various deep neural network models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116644263B_ABST
    Figure CN116644263B_ABST
Patent Text Reader

Abstract

This invention discloses a systolic array capable of supporting fine-grained deformation and data stream switching. The systolic array includes a two-dimensional computing unit array, a set of fine-grained, reconfigurable on-chip data caches, and a controller. The systolic array can be divided into four subarrays connected end-to-end to achieve different logical shapes and can arbitrarily switch between three data streams. In this systolic array, each computing unit can perform full-duplex data transfer with adjacent computing units in four directions (up, down, left, and right), and can simultaneously perform computational operations and data transfer operations across adjacent subarrays. In this systolic array, the on-chip data cache can be finely divided into several independently read and write cache blocks and dynamically configured according to different array shapes. This invention significantly improves the computational efficiency of the systolic array when performing matrix multiplications of different shapes and sizes, thereby improving the acceleration effect of neural network models.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence accelerator design technology, specifically to a pulsating array that can be efficiently applied to matrix multiplication of different shapes while supporting fine-grained deformation and multi-data stream switching. Background Technology

[0002] With the rise of artificial intelligence technology, deep neural networks are increasingly being applied in various fields. Although different deep neural networks differ significantly in structure, scale, and the operators that make up the network, they all use matrix multiplication as the most common basic operation. Matrix multiplication accounts for a very high proportion of the network's computational needs; therefore, the computational efficiency of matrix multiplication largely determines the execution efficiency of the entire neural network model. The systolic array is the most commonly used architecture for matrix multiplication computation. With a two-dimensional array of computational units at its core, it can achieve matrix multiplication computations with high data reuse and high computational parallelism.

[0003] Current systolic arrays still suffer from low computational unit utilization in matrix multiplication, especially when dealing with matrix multiplication operations of special shapes (e.g., when one of the matrices is elongated or even degenerates into a vector). In these cases, the computational unit utilization of systolic arrays is severely reduced, failing to achieve significant parallel acceleration. However, similar matrix multiplications inevitably occur extensively in various neural network operators, greatly limiting the computational efficiency of systolic arrays on these neural network models.

[0004] A major reason for the low utilization of computing units is that current systolic arrays either cannot support dynamic changes in their own logical structure, thus failing to adapt to changes in matrix shape, or they cannot support switching between multiple data streams, resulting in significantly limited flexibility. Therefore, by designing a systolic array that can simultaneously support fine-grained deformation and multi-data stream switching, its computing unit utilization for various neural network models can be greatly improved, thereby significantly enhancing the acceleration effect. Summary of the Invention

[0005] This invention provides a systolic array system that simultaneously supports fine-grained deformation and multi-data-stream switching, comprising three parts: a two-dimensional computing unit array consisting of computing units supporting four-way full-duplex data transfer and multi-data-stream switching; a fine-grained, multi-mode on-chip data cache to meet the different storage requirements of the systolic array under different shapes and data flows; and a controller for controlling the systolic array and on-chip data cache to perform matrix multiplication calculations under different shapes and data flows. The purpose of this invention is to enable the systolic array to dynamically and finely change its logical shape and data flow, improving the utilization and acceleration effect of computing units when calculating matrix multiplications of various shapes and sizes, thereby efficiently adapting to various deep neural network models. Overall, this invention can support an R... p Line C p The column of pulsating array in min{R p C p It allows for arbitrary switching and combination between 1 shape and 3 data streams.

[0006] The basic process of the systolic array system to accelerate deep neural network models is as follows: In the compilation stage, a neural network model is transformed into a sequence of matrix multiplication operations. For each matrix multiplication in the sequence, the optimal array logic shape and data flow are calculated using a greedy algorithm. In the execution stage, the systolic array calculates each matrix multiplication sequentially according to the sequence order. Each matrix multiplication operation includes a configuration stage and a calculation stage. In the configuration stage, the configuration information of the logic shape and data flow is first sent into the array. At the same time, the resident data of the previous round is optionally removed from the array, and the resident data of the current round is loaded into the array. In the execution stage, the non-resident operation data is sent into the array for calculation according to a certain format. At the same time, the non-resident operation results are optionally written back to the cache.

[0007] The pulsating array is a two-dimensional array composed of special computing units, and its physical shape is represented as (R p C p ), R p and C p These are the number of rows and columns of the array, respectively. Each computing unit is connected to only its four adjacent computing units (top, bottom, left, and right), and supports full-duplex data transmission in all four directions. When no deformation occurs, the systolic array maintains its physical shape (R... p C p ) Calculate matrix multiplication; when fine-grained deformation occurs, the systolic array is in logical shape (R l C l ) Calculate matrix multiplication. For a square two-dimensional array (R... p =C pThe basic steps of fine-grained deformation are: First, the pulsating array is rotate-symmetrically divided into four adjacent but non-overlapping subarrays of the same shape, the shape of which is represented as (R... s C s ), satisfying R s +C s =R p And R s No more than R p Half of the array; the second step is to connect the four subarrays end-to-end to form a logic array (R). l C l Depending on the different data cache allocation and subarray connection configurations, the logical shape after connection could be (4R). s C s ), or possibly (R) s 4C s For a non-square two-dimensional array (R) p ≠C p The basic steps of fine-grained deformation are similar, but the four subarrays may contain two different shapes. According to the above deformation scheme, the pulsating array can support a total of min{R} p C p}+1 different logical shapes.

[0008] In the systolic array, data transfer between subarrays utilizes internal paths. These internal paths exist between the tail boundary of the preceding subarray and the head boundary of the following subarray, and are formed by short connections between registers within several adjacent computational units. The function of the internal path is to transfer data from the tail boundary of the preceding subarray to the head boundary of the following subarray after a delay of several clock cycles. The computational units on the internal path simultaneously handle computational tasks within the subarray and the task of transferring data across subarrays. By utilizing adjacent computational units for data transfer, the internal path avoids introducing long connections across computational units and a large number of additional multiplexers, ensuring timing constraints and reducing design complexity.

[0009] The systolic array supports dynamic switching between three data streams: input dwell, weight dwell, and output dwell. For a matrix multiplication operation C = AB (A is the input feature matrix, B is the weight matrix, and C is the output feature matrix), the three data streams represent the temporary storage of each element of the corresponding matrix in the dwell registers within each computation unit, while the elements of the other two matrices are passed along the horizontal and vertical axes of the transformed logic array, respectively. For example, for output dwell, the partial sum generated by the matrix multiplication (i.e., the intermediate value in the process of forming the output matrix) is temporarily stored in the dwell registers within each computation unit, while the input feature matrix and weight matrix are passed along the horizontal and vertical axes of the transformed logic array, respectively.

[0010] The computing unit described is a computing unit that supports four-way full-duplex data transmission and multi-data stream switching. It includes a multiply-accumulator, a resident register, four pipeline registers, a data exchange network, and a control unit. It has an output data port and an input data port in each of the four directions (up, down, left, and right). The multiply-accumulator performs a multiplication and addition operation on three operands sequentially, which is the most basic atomic operation constituting matrix multiplication. The resident register temporarily stores elements of a matrix in the matrix multiplication process. The pipeline register temporarily stores data between adjacent computing units, constraining the critical path within a single computing unit. The data exchange network selects data from adjacent units and data generated within the unit and accurately sends it to the appropriate components within the unit. The control unit stores the configuration information of the computing unit and generates control signals to guide the data flow of the exchange network.

[0011] In the computing unit, the data exchange network selects data from adjacent units and data generated within the unit, and accurately sends it to the various components within the unit. Specifically, for data from adjacent units in any direction, the exchange network can selectively send it to a pipeline register in one of the other three directions, or it can send it to a multiply-accumulator; in particular, for data from above, the exchange network can also selectively send it to a resident register or a control unit; for the output data of the resident register, the exchange network can selectively send it to a multiply-accumulator or the pipeline register above; for the output data of the accumulator, the exchange network can selectively send it to a pipeline register in one of the four directions, or it can send it to a resident register.

[0012] In the computing unit, the control unit stores the configuration information of its own unit and generates control signals to guide the data flow of the switching network. The configuration information changes depending on the logic shape of the systolic array, the data flow, and the unit's position within the systolic array. Once a specific logic shape and data flow are selected, the configuration information for each computing unit can be determined. During the configuration phase, this configuration information is pipelined from top to bottom to the control unit of the corresponding unit for temporary storage. Once all control registers are reassigned, it signifies that the systolic array has completed a fine-grained deformation and data flow switch.

[0013] The on-chip data cache is a multi-mode data cache that supports fine-grained reallocation, comprising one resident data cache and four non-resident data caches. The resident data cache stores the resident data corresponding to each of the three data streams; the non-resident data caches store the other two types of non-resident data. The four non-resident data caches are located outside the four boundaries of the systolic array. Each non-resident data cache is divided into several cache blocks equal to the number of computing units on the corresponding boundary, and each cache block supports independent read and write operations. When the systolic array operates in different logical shapes and data streams, each cache block is reallocated according to its location to three operating modes: storing the first non-resident data, storing the second non-resident data, or idle. Attached Figure Description

[0014] Figure 1 This is a complete structural schematic diagram of one embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the computing unit structure according to an embodiment of the present invention.

[0016] Figure 3 This is a schematic diagram of the physical and logical shape of an embodiment of the present invention under a certain fine-grained deformation condition.

[0017] Figure 4 This is an embodiment of the present invention in a way that is different from... Figure 3 A schematic diagram of the physical and logical shapes under fine-grained deformation.

[0018] Figure 5 yes Figure 4 The illustrated embodiment, after fine-grained deformation, demonstrates six different configuration schemes by selecting different data streams and different data caching allocation strategies.

[0019] Figure 6 This is a schematic diagram of a typical matrix multiplication with large differences between dimensions, which is processed by the present invention.

[0020] Figure 7 yes Figure 3The illustrated embodiment, after undergoing fine-grained deformation, [is affected by / is subject to ... Figure 6 The diagram shows the mapping of the three matrices in matrix multiplication.

[0021] Figure 8 yes Figure 7 The illustrated embodiment is a schematic diagram of the data element transfer in the upper part of the computing unit during the first 8 clock cycles. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. It should be noted that the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0023] Figure 1 This is a complete structural schematic diagram of one embodiment of the present invention. The core of this systolic array system is a two-dimensional array of computing units composed of the computing units described in this invention. Each computing unit establishes bidirectional connections with its four adjacent units (top, bottom, left, and right). Outside each boundary of the array, there exists an adder array and a non-resident data buffer. The number of adders in the adder array is equal to the number of computing units on the corresponding boundary of the systolic array, and the number of buffer blocks in the non-resident data buffer is equal to the number of adders. The input data of the boundary computing units comes from the corresponding buffer blocks, and the output data can be selectively accumulated by adders or not before entering the corresponding buffer block to achieve accumulation or overwrite functions. For each computing unit at the upper boundary of the array, it is additionally connected to a resident data buffer to load resident data and array configuration information from the resident buffer to the systolic array, or to move resident data from the systolic array to the resident buffer. The systolic array system also includes a global controller that controls the operation of each buffer and computing unit array with a specified logical shape and data flow.

[0024] Figure 2 This is a schematic diagram of a computing unit structure according to an embodiment of the present invention. The computing unit includes a multiply-accumulator, a resident register, four pipelined registers, a data exchange network, and a control unit. The data exchange network serves as the center for data exchange and forwarding, connecting to the other components, while the other components are not directly connected to each other. The computing unit has a pair of input / output ports at each of the four locations. The four input ports are directly connected to the data exchange network, while the four output ports are respectively connected to the four pipelined registers.

[0025] Figure 2The control unit is used to store the configuration information of this computing unit and control the switching behavior of the data exchange network. The configuration information it inputs comes from the upper input port, flows into the control unit through the exchange network, and is stored in the control unit. The control signals it outputs are directly sent into the exchange network and play the role of controlling the data flow.

[0026] Figure 2 The multiply-accumulator in the circuit performs one multiplication and one addition operation. It has three input data points: one from a resident register, and the other two from input ports located at two different positions. All three inputs flow into the multiply-accumulator through a switching network. The output data then passes through the switching network, potentially flowing to a resident register or a pipelined register, and finally exiting from the corresponding output port.

[0027] Figure 2 The resident register in the network is used to temporarily store data. Its input data may come from the input port above or the multiplier and flow into the resident register through the switching network. Its output data may flow to the pipeline register above through the switching network and then flow out from the output port above, or it may flow into the multiplier.

[0028] Figure 2 In addition to completing the data flow described above, the switching network can also selectively send input data from one location directly to a pipeline register in one of the other three locations, and then send it out from the corresponding output port.

[0029] Figure 3 and Figure 4 This is a schematic diagram of two fine-grained deformations according to an embodiment of the present invention, illustrating how a pulsating array with a physical shape of (4, 4) can be transformed into different logical shapes through fine-grained deformation.

[0030] like Figure 3 As shown in (a), the pulsating array is divided into four subarrays of shape (2, 2), which are connected end to end to achieve... Figure 3 The logical shape shown in (b). For ease of understanding, Figure 3Only the data transfer across subarrays is shown; the transfer of other data is omitted. Solid lines with arrows indicate data flow within each subarray, similar to the data flow in a typical systolic array. Dashed lines with arrows indicate internal paths used to implement data flow between subarrays. Specifically, the upper left and upper right subarrays are connected by two internal paths, passing through computing units 11-12-02 and 01-02-03 respectively. Data from computing unit 11 is first passed to 12 (where it is only forwarded and not used in calculations), then after one clock cycle, it is passed to 02 (where it is only forwarded and not used in calculations), and after another clock cycle, it becomes the actual input data for 02, thus completing the data transfer between the two subarrays. It should be noted that... Figure 3 The implementation can be either (2, 8) or (8, 2), depending on the allocation of the non-resident data cache.

[0031] like Figure 4 As shown in (a), the pulsating array is divided into four subarrays of shape (1, 3), which are connected end to end to achieve... Figure 4 The logical shape shown in (b). For ease of understanding, Figure 4 Only the data transfer across subarrays is shown; the transfer of other data is omitted. Solid lines with arrows represent data flow within each subarray, similar to the data flow in a typical systolic array; dashed lines with arrows represent internal paths used to implement data flow between subarrays. Specifically, the upper left and upper right subarrays are connected by a short internal path, passing through computing units 02-03. Data from computing unit 02 is first passed to 03 (where it is only forwarded and not used in computation). After one clock cycle, it becomes the actual input data for 03 and participates in computation, thus completing the data transfer between the two subarrays. It should be noted that... Figure 4 The implemented logical shape can be either (1, 12) or (12, 1), depending on the allocation of the non-resident data cache.

[0032] Figure 5 yes Figure 4 The illustrated embodiment, after fine-grained deformation, demonstrates six different configuration schemes achieved by selecting different data streams and different data caching allocation strategies. For example... Figure 5 As shown in the diagram, for ease of understanding, four different markers are used to distinguish between buffer blocks configured as input feature data, weight data, output feature data, or idle data. The corresponding black solid line, gray solid line, and dashed line represent the transmission of input feature data, weight data, and output feature data in the pulsating array, respectively.

[0033] Figure 5(a) The logical shape of the configuration shown is (1, 12). The data flow is output dwell, which allocates the dwell data cache to the output dwell, allocates the first cache block of the left non-dwell data cache to the input feature data, allocates the other three cache blocks of each of the four non-dwell data caches to the weight data, and the remaining cache blocks are idle.

[0034] Figure 5 (b) The logical shape of the configuration shown is (1, 12). The data flow is weight-resident, which allocates the resident data cache to the weight data, allocates the first cache block of the left non-resident data cache to the input feature data, allocates the other three cache blocks of each of the four non-resident data caches to the output feature data, and the remaining cache blocks are idle.

[0035] Figure 5 (d) The logical shape of the configuration shown is (12, 1). The data flow is output resident, which allocates the resident data cache to the output resident, allocates the first cache block of the left non-resident data cache to the weight data, allocates the other three cache blocks of each of the four non-resident data caches to the input feature data, and the remaining cache blocks are free.

[0036] Figure 5 The configurations shown in (b)(e)(f) are similar and will not be elaborated further.

[0037] Figure 6 This is a schematic diagram of a typical matrix multiplication with significant differences between dimensions, processed by this invention. Matrix A represents the input features, matrix B represents the weights, and matrix C represents the output features obtained by multiplying matrices A and B.

[0038] Figure 7 yes Figure 3 The illustrated embodiment, after undergoing fine-grained deformation, [is affected by / is subject to ... Figure 6 The diagram illustrates the mapping of the three matrices in the matrix multiplication. The logic shape after fine-grained deformation in this embodiment is (2, 8), and the data flow selects the output feature to reside in. Around the systolic array, the elements of matrices A and B are arranged along the clock axis in different queues, indicating the clock and position of each element flowing into the systolic array. Elements in matrix A flow into the systolic array from the left side of the upper left subarray during clock cycles T0-T2. The first part of the elements in matrix B flow into the systolic array from the top side of the upper left subarray during clock cycles T0-T2; the second part of the elements in matrix B flow into the systolic array from the right side of the upper right subarray during clock cycles T4-T6; the third and fourth parts are similar. There is a 2-clock-cycle delay in the input of B elements between adjacent subarrays because the transfer of A matrix elements between subarrays requires an additional 2 clock cycles (in...). Figure 3(Detailed explanation provided). The C matrix elements marked in each cell of the pulsating array indicate the positions where the actual computation results are generated. After all computation results have been generated, the C matrix elements are moved out of the pulsating array as a whole from above.

[0039] Figure 8 yes Figure 7 The illustrated embodiment is a schematic diagram showing the data element transfer process of the upper-half computing unit during the first 8 clock cycles. (See diagram below.) Figure 8 As shown in the diagram, a computational unit can contain a maximum of four elements. Given a logical shape of (2, 8) and a data flow where the output feature resides, these four elements are the input feature value, the weight value, the output feature value, and the input feature value being passed between subarrays. For example... Figure 8 As shown in (a), (b), and (c), the input feature values ​​(elements of matrix A) are passed from left to right within the first subarray, while the weight values ​​(elements of matrix B) are passed from top to bottom within the first subarray; Figure 8 As shown in (e), (f), and (g), the input feature values ​​(elements of matrix A) are passed from top to bottom within the second subarray, while the weight values ​​(elements of matrix B) are passed from right to left within the second subarray. Furthermore, as... Figure 8 As shown in (c)(d)(e)(f), the input feature (A matrix element) is also passed between subarrays. Specifically, a00 is passed between subarrays starting in clock cycle T2, and arrives at the target cell in the neighboring subarray in clock cycle T4 and participates in the calculation; a10 is passed between subarrays starting in clock cycle T3, and arrives at the target cell in the neighboring subarray in clock cycle T5 and participates in the calculation.

[0040] The present invention has been described in detail above using specific embodiments. It should be noted that the above content is merely a specific embodiment of the present invention and should not be used to limit the scope of the invention. Any modifications, substitutions, or improvements within the design concept of the present invention should be within the protection scope of the present invention.

Claims

1. A pulsating array that simultaneously supports fine-grained deformation and multi-data stream switching, characterized in that, It includes a two-dimensional computing unit array, an on-chip data cache that supports fine-grained reallocation, and a controller; the controller is used to enable the two-dimensional computing unit array to perform matrix multiplication according to a specified logical shape and data flow, and to control the on-chip data cache to meet the corresponding storage requirements; In the two-dimensional computing unit array, the connection between computing units is limited to adjacent computing units in the up, down, left, and right directions. Each computing unit supports four-way full-duplex data transmission. Each computing unit includes a multiply-accumulator, a resident register, four pipeline registers, a data exchange network, and a control unit, and has input and output data ports in the four directions. The four input data ports are directly connected to the data exchange network, and the multiply-accumulator, resident register, control unit, and four pipeline registers are respectively connected to the data exchange network, but there is no direct connection between them. The four pipeline registers are connected to the output data ports in four directions, respectively, to temporarily store data once between adjacent computing units and constrain the critical path within a computing unit; The control unit temporarily stores the configuration information of the computing unit and generates control signals accordingly to control the data flow of the data exchange network. Under the control of the control signals, the data exchange network sends the data input from two different directions selected from the four input data ports and the output data of the resident register to the three input terminals of the multiply-adder, respectively. The multiply-adder performs a multiplication and addition calculation on the three operands in sequence, and the output data can be selectively sent to the resident register via the data exchange network, or output via the pipeline register and output data port selected from the four directions. The two-dimensional computing unit array can be divided into four non-overlapping subarrays, each containing computing units at one of the four corners of the array, and adjacent to each other around the perimeter of the array. Under the control of the control signal, the four subarrays are connected end to end in sequence through internal paths according to the data transmission order, forming a logical array corresponding to the configuration information, so that the same two-dimensional computing unit array can perform matrix multiplication with different logical shapes. The internal path is located between two adjacent subarrays in the data transmission order, connecting the tail boundary of the previous subarray to the head boundary of the next subarray, and is composed of registers inside several adjacent computing units and short connections between adjacent computing units. Data across subarrays is forwarded step by step through the data exchange network and pipeline registers along the way, and is transmitted from the tail boundary to the head boundary after several clock cycles. The computing unit located on the internal path simultaneously undertakes the computing tasks within its own subarray and the data transfer tasks across subarrays. During cross-subarray data transfer, its data exchange network, under the control of the control signal, directly sends the data from the input data port in one direction to the pipeline register of one of the other three directions, and outputs it through the corresponding output data port without passing through the multiplier-accumulator and resident register. The pulsating array supports three data streams: input dwell, weight dwell, and output dwell. For matrix multiplication C=AB, A, B, and C are the input, weight, and output matrices, respectively. The matrix corresponding to the selected data stream is the dwell matrix, whose data is temporarily stored in the dwell registers of each computing unit. During output dwell, the partial sum generated by matrix multiplication is temporarily stored. The other two matrices are the first non-dwelling matrix and the second non-dwelling matrix. The elements of one of them are passed horizontally along the logic array, and the elements of the other are passed vertically. The pulse array rewrites the configuration information temporarily stored in the control unit of each computing unit, causing the control unit to generate corresponding control signals to control the data flow of the data exchange network, thereby realizing fine-grained deformation and switching of the three data flows. The configuration information enters the array through the upper input data port of the computing unit at the upper boundary of the array, and is pipelined from top to bottom to the corresponding control unit for temporary storage through the data exchange network of each computing unit.

2. The pulsating array that simultaneously supports fine-grained deformation and multi-data stream switching according to claim 1, characterized in that, The two-dimensional computing unit array is square, and its physical shape is as follows: ,in and These represent the number of rows and columns of the array, respectively. When no deformation occurs, the array performs matrix multiplication according to the physical shape; when fine-grained deformation occurs, the logic array performs matrix multiplication according to the logical shape. Perform matrix multiplication, where and These represent the number of rows and columns of the logical array, respectively. When fine-grained deformation occurs, the layout of the four sub-arrays is rotationally symmetrical about the array center at 90° and has the same shape; the shape of each sub-array, represented by the same orientation, is as follows: ,in and Let these represent the number of rows and columns of the downward-facing subarray, respectively, satisfying... , The logical shape is or .

3. The pulsating array that simultaneously supports fine-grained deformation and multi-data stream switching according to claim 1, characterized in that, The on-chip data cache includes one resident data cache and four non-resident data caches controlled by the controller; The resident data cache connects to each computing unit at the upper boundary of the array, loading the corresponding resident matrix data into the array during input resident or weight resident, and removing the output matrix data from the array during output resident; The four non-resident data caches are located outside the four boundaries of the array, and each is divided into a cache block that supports independent reading and writing. The number of blocks is equal to the number of computing units at the corresponding boundary. The non-resident input data of the boundary computing units comes from the corresponding cache block. Under different logical shapes and data flows, each cache block is reallocated according to its location to three working modes: storing the first non-resident matrix data, storing the second non-resident matrix data, or idle.

Citation Information

Patent Citations

  • Method and computing system for reconfigurable architecture based on GPGPU

    CN113076521A

  • Array arithmetic logic unit structure

    WO2009155761A1