Reconfigurable many-core computing device and method for matrix operation

TWI946244BActive Publication Date: 2026-11-11IND TECH RES INST
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
TW114150940
Authority / Receiving Office
TW · TW
Patent Type
Patents
Current Assignee / Owner
Filing Date
2025-12-24
Publication Date
2026-11-11
Estimated Expiration
2045-12-23

Smart Images

  • Figure TWG2TB001920996_001
    Figure TWG2TB001920996_001
  • Figure TWG2TB001920996_002
    Figure TWG2TB001920996_002
  • Figure TWG2TB001920996_003
    Figure TWG2TB001920996_003
Patent Text Reader

Abstract

This disclosure provides a reconfigurable multi-core computing device and a method for matrix operations. The device includes multiple processing elements coupled in a two-dimensional chip network architecture, a power manager, and a main controller. The main controller receives a matrix operation task including a first input matrix and a second input matrix, and decomposes the first input matrix into multiple first-type submatrices and the second input matrix into multiple second-type submatrices. The main controller selects a target operation mode from multiple operation modes based on the number of first-type submatrices, the number of second-type submatrices, and the dimension of the operation array. Based on the selected operation mode, a target processing element is determined from the processing elements, and the first-type and second-type submatrices are allocated to the target processing element for operation.
Need to check novelty before this filing date? Find Prior Art

Claims

1. A reconfigurable multi-core computing device, comprising: An computing array comprising multiple processing elements (PEs) coupled in a two-dimensional chip network (NoC) architecture; A power manager, coupled to the computing array, is used to independently control the power state of each processing element; The system also includes a master controller electrically coupled to the computing array and the power manager, wherein the master controller is configured to: receive a matrix operation task including a first input matrix and a second input matrix; decompose the first input matrix into a plurality of first-type submatrices and the second input matrix into a plurality of second-type submatrices; select one of a plurality of operation modes as a target operation mode based on a first number of the plurality of first-type submatrices, a second number of the plurality of second-type submatrices, and the dimension of the computing array, wherein each of the plurality of operation modes corresponds to a different processing element enabling configuration strategy; and, based on the selected operation mode, determine one or more target processing elements from the plurality of processing elements and allocate the plurality of first-type submatrices and the plurality of second-type submatrices to the one or more target processing elements to perform the operation.

2. The reconfigurable multi-core computing device as claimed in claim 1, wherein in the operation of determining one or more target processing elements from among the plurality of processing elements based on the selected computing mode, and allocating the plurality of first type sub-matrices and the plurality of second type sub-matrices to the one or more target processing elements to perform the computing operation: the master controller generates a power distribution map based on the selected computing mode, and commands the power manager to set the power state of the one or more target processing elements according to the power distribution map; and the master controller schedules the plurality of first type sub-matrices and the plurality of second type sub-matrices to allocate the plurality of first type sub-matrices and the plurality of second type sub-matrices to the corresponding target processing elements to perform the computing operation.

3. The reconfigurable multi-core computing device as described in claim 2, wherein the plurality of computing modes include: A full-speed computing mode, wherein the corresponding processing element enables the configuration policy to enable all of the plurality of processing elements in the computing array; An energy-saving computing mode, wherein the corresponding processing element enables a configuration strategy that enables only a first subarray in the computing array; and a performance-enhancing mode, wherein the corresponding processing element enables a configuration strategy that enables a second subarray in the computing array, and the dimension of the second subarray is different from that of the first subarray.

4. The reconfigurable multi-core computing device as claimed in claim 3, wherein when both the first number and the second number are greater than or equal to the dimension size of the computing array, the master controller selects the full-speed computing mode, wherein in the full-speed computing mode, the power distribution diagram generated by the master controller enables all of the plurality of processing elements in the computing array.

5. The reconfigurable multi-core computing device as claimed in claim 3, wherein when at least one of the first number and the second number is less than the dimension of the computing array, the master controller selects the power-saving computing mode, wherein in the power-saving computing mode, the power distribution map generated by the master controller enables only the plurality of target processing elements of a subarray, the dimension of which is determined according to the first number and the second number, and shuts down the power state of non-target processing elements.

6. The reconfigurable multi-core computing device as described in claim 3, wherein when both the first number and the second number are less than a preset ratio of the dimension size of the computing array, the main controller selects the performance enhancement mode, and the performance enhancement mode further includes dividing a common dimension of the plurality of first type sub-matrices and the plurality of second type sub-matrices to allocate a single computing task to at least two processing elements for parallel computing.

7. The reconfigurable multi-core computing device as described in claim 2, wherein the plurality of computing modes further includes an asymmetric computing mode, and the master controller selects the asymmetric computing mode when one of the first quantity and the second quantity is less than the dimension size of the computing array, while the other is greater than or equal to the dimension size.

8. The reconfigurable multi-core computing device as described in claim 2, wherein after completing a round of scheduling, the master controller is further configured to perform operation again, selecting one of the plurality of computing modes as the target computing mode, based on the first remaining quantity of the plurality of first type sub-matrices and the second remaining quantity of the plurality of second type sub-matrices.

9. The reconfigurable multi-core computing device as claimed in claim 2, wherein the master controller, configured to utilize a multi-point broadcast function of the two-dimensional chip network, simultaneously transmits first data of the plurality of first-type sub-matrices to an entire column of processing elements in the computing array and simultaneously transmits second data of the plurality of second-type sub-matrices to an entire row of processing elements when scheduling the plurality of first-type sub-matrices and the plurality of second-type sub-matrices.

10. The reconfigurable multi-core computing device as claimed in claim 1 further includes a Scatter-Gather DMA (SG DMA) unit, wherein the master controller is configured to generate a data transfer list and commands the SG DMA unit to perform batch data transfers of the plurality of sub-matrixes from a main memory to the plurality of target processing elements according to the list, wherein the data transfer list includes a plurality of entries, and each entry defines at least a source address, a destination address, and a transfer length.

11. The reconfigurable multi-core computing device as claimed in claim 1, wherein each processing element includes a local memory that includes at least two independent data buffers to allow one data buffer to provide data for computation while the other data buffer receives the next batch of computation data.

12. The reconfigurable multi-core computing device as claimed in claim 11, wherein each processing element includes a local direct memory access unit for transferring data directly from the local memory of the processing element to the local memory of another processing element without going through the host controller.

13. The reconfigurable multi-core computing device as claimed in claim 11, wherein each processing element is configured to retain a computation result and accumulate the computation result with a subsequent computation result to produce a final partial computation result.

14. The reconfigurable multi-core computing device as claimed in claim 11, wherein when a target submatrix stored in the local memory needs to be reused in the next round of computing, a microcontroller in the processing element is configured to set a full flag for the data buffer corresponding to the submatrix to prevent the main controller from repeatedly transferring the target submatrix from a main memory.

15. A matrix operation method applied to a reconfigurable multi-core computing device, the device comprising a main controller, a power manager, and a computing array consisting of multiple processing elements, the method comprising: The main controller receives a matrix operation task, which includes a first input matrix and a second input matrix. The main controller decomposes the first input matrix into multiple first-type submatrices and the second input matrix into multiple second-type submatrices. The main controller selects one of multiple operation modes as a target operation mode based on the number of the multiple first-type submatrices, the number of the multiple second-type submatrices, and the dimension of the operation array. Each of the multiple operation modes corresponds to a different processing element enabling configuration strategy. Based on the selected operation mode, the main controller determines one or more target processing elements and allocates the multiple first-type submatrices and the multiple second-type submatrices to the one or more target processing elements for operation.

16. The method of performing matrix operations as described in claim 15, wherein the step of determining one or more target processing elements among the plurality of processing elements based on the selected operation mode and allocating the plurality of first type submatrices and the plurality of second type submatrices to the one or more target processing elements for performing the operation includes: The main controller generates a power distribution diagram based on the selected operation mode, and commands the power manager to set a power state for the one or more target processing elements according to the power distribution diagram. The master controller schedules the plurality of first-type submatrices and the plurality of second-type submatrices to allocate the plurality of first-type submatrices and the plurality of second-type submatrices to the corresponding target processing elements for the operation.

17. The method of matrix operation as described in claim 16, wherein the plurality of operation modes include: A full-speed computing mode, wherein the corresponding processing element enables the configuration policy to enable all of the plurality of processing elements in the computing array; An energy-saving computing mode, wherein the corresponding processing element enables a configuration strategy that enables only a first subarray in the computing array; and a performance-enhancing mode, wherein the corresponding processing element enables a configuration strategy that enables a second subarray in the computing array, and the dimension of the second subarray is different from that of the first subarray.

18. The matrix operation method as described in claim 17, wherein when at least one of a first quantity corresponding to the plurality of first-type submatrices and a second quantity corresponding to the plurality of second-type submatrices is less than the dimension size of the operation array, the master controller selects the power-saving operation mode, wherein in the power-saving operation mode, the power distribution map generated by the master controller will only enable the plurality of target processing elements of a subarray, the dimension of the subarray being determined according to the first quantity and the second quantity, and shut down the power state of non-target processing elements.

19. The matrix operation method as described in claim 17, wherein when a first number corresponding to the plurality of first type sub-matrices and a second number corresponding to the plurality of second type sub-matrices are both less than a preset ratio of the dimension size of the operation array, the main controller selects the performance enhancement mode, and the performance enhancement mode further includes dividing a common dimension of the plurality of first type sub-matrices and the plurality of second type sub-matrices to allocate a single operation task to at least two processing elements for parallel operation.

20. A method for performing matrix operations as described in claim 16, wherein the scheduling step includes: Using the multi-point broadcast function of a two-dimensional chip network, the first data of the multiple first-type sub-matrices are simultaneously transmitted to a whole column of processing elements in the computing array, and the second data of the multiple second-type sub-matrices are simultaneously transmitted to a whole row of processing elements.

21. The matrix operation method as described in claim 15 further includes: The main controller generates a data transfer list, which includes multiple entries, and each entry defines at least a source address, a destination address, and a transfer length; and commands a distributed direct memory access unit to perform batch data transfers from a main memory to the multiple target processing elements according to the list.

22. The method of matrix operation as described in claim 16, wherein the step of selecting one as the target operation mode further includes: Receive an external policy instruction; when a first quantity corresponding to the plurality of first type submatrices and a second quantity corresponding to the plurality of second type submatrices both meet a geometric trigger condition for a performance enhancement mode, determine whether to enable the performance enhancement mode according to the external policy instruction; if the external policy instruction indicates a power saving priority state, force the selection of the power saving operation mode to reduce the number of enabled processing elements; and if the external policy instruction indicates a performance priority state, select the performance enhancement mode to utilize additional processing elements to accelerate the operation.

Citation Information

Patent Citations

  • Dynamic reconfigurable zero-knowledge proof accelerator and acceleration method

    CN120602098A

  • Convolution engine and methods of operating and forming thereof

    WO2023140778A1