Search

CN-121979673-A - Reconfigurable hardware suite control method, equipment, medium and product

CN121979673ACN 121979673 ACN121979673 ACN 121979673ACN-121979673-A

Abstract

A reconfigurable hardware suite control method, equipment, media and products relate to the field of artificial intelligence hardware. The method comprises the steps of responding to an access event, reading and analyzing module file information in a storage unit through a golden finger interface unit to obtain hardware resource requirements and operating system adaptation information, distributing hardware resources for an extended functional module through a resource control unit based on the hardware resource requirements, responding to the module hot-plug event, initializing and driving the extended functional module based on the operating system adaptation information, responding to a task chain, decomposing the task chain into atomic operator sequences through a dynamic reconfigurable computing unit, configuring a reconfigurable interconnection network in the dynamic reconfigurable computing unit, generating system load prediction information based on computing characteristics of various atomic operators and real-time task queue depth, and dynamically configuring a reconfigurable hardware suite through a hardware management unit. The application can improve the control efficiency of the reconfigurable hardware suite.

Inventors

  • WANG XIAODONG
  • ZENG BO

Assignees

  • 上海快贴电子有限公司

Dates

Publication Date
20260505
Application Date
20251231

Claims (10)

  1. 1. The reconfigurable hardware suite control method is applied to a reconfigurable hardware suite, and is characterized in that the reconfigurable hardware suite comprises a main computing module and at least one expansion function module, the main computing module comprises a dynamic reconfigurable computing unit, a hardware management unit and a resource control unit, the expansion function module comprises a golden finger interface unit and a storage unit, and the method comprises the following steps: Responding to an access event of the extended function module, the main calculation module reads and analyzes module file information in the storage unit through the golden finger interface unit to obtain hardware resource requirements and operating system adaptation information, wherein the module file information is used for storing configuration data of the extended function module; Based on the hardware resource requirement, distributing hardware resources to the extended function module through the resource control unit; The method comprises the steps of responding to a module hot plug event, initializing and driving loading the extended function module based on the operating system adaptation information so as to enable the extended function module to enter a usable state; decomposing a task chain into an atomic operator sequence through the dynamic reconfigurable computing unit in response to the task chain issued by the artificial intelligence application, wherein the atomic operator sequence comprises a plurality of atomic operators of different types; Configuring a reconfigurable interconnection network in the dynamically reconfigurable computing unit based on the computing characteristics of the atomic operators of various types; Generating system load prediction information based on the calculation characteristics of the atomic operators and the real-time task queue depth; and dynamically configuring the reconfigurable hardware suite through the hardware management unit based on the system load prediction information.
  2. 2. The method of claim 1, wherein the golden finger interface unit includes a first configuration pin, the responding to the access event of the extended function module, the main computing module reading and analyzing module file information in the storage unit through the golden finger interface unit to obtain hardware resource requirements and operating system adaptation information, including: Responding to the access event, the main computing module accesses the module archive information of the storage unit through the first configuration pin, wherein the module archive information at least comprises a module identifier, power consumption requirement information, the number of required high-speed input and output channels, an interrupt mapping scheme and drive configuration scripts for different operating systems; and taking the power consumption requirement information and the number of the required high-speed input/output channels as the hardware resource requirement, and taking each driving configuration script as the operating system adaptation information.
  3. 3. The method of claim 2, wherein the golden finger interface unit further comprises a second configuration pin, and after the responding to the access event of the extended function module, the main computing module reads and parses the module profile information in the storage unit through the golden finger interface unit to obtain the hardware resource requirement and the operating system adaptation information, the method further comprises: the main computing module is in bidirectional communication with the expansion function module through the second configuration pin so as to acquire electric parameter sets respectively corresponding to a plurality of levels of working modes supported by the expansion function module, wherein the electric parameter sets comprise reference power supply requirements and reference signal rates; acquiring real-time power supply requirements and real-time signal rates monitored by the main computing module; Determining a target level operating mode based on the real-time power supply demand and the real-time signal rate, matching each of the electrical parameter sets, the target level operating mode being used to characterize a highest level operating mode in which the reference power supply demand does not exceed the real-time power supply demand and the reference signal rate does not exceed the real-time signal rate; and updating the power consumption requirement information of the hardware resource requirement and the required high-speed input-output channel number based on the reference power supply requirement and the reference signal rate of the target level working mode.
  4. 4. The method of claim 1, wherein the dynamically reconfigurable computing unit includes a neural network processing component NPU and a hardware task chain parsing component, and wherein the decomposing, by the dynamically reconfigurable computing unit, the task chain into an atomic operator sequence in response to a task chain issued by an artificial intelligence application includes: Analyzing the task chain through the hardware task chain analysis component to obtain a dependency analysis result of the task chain, wherein the dependency analysis result is used for representing data dependency in the task chain, and the data dependency at least comprises a parallel branch relationship and a serial dependency; and decomposing the task chain into a plurality of atomic operators comprising an execution order based on the dependency analysis result to obtain the atomic operator sequence.
  5. 5. The method of claim 4, wherein the NPU comprises a plurality of scalar processing elements SPUs, tensor processing elements TPU, and vector processing elements VPUs, the configuring a reconfigurable interconnect network in the dynamically reconfigurable computing unit based on the computing characteristics of the various types of atomic operators, comprising: Determining a target processing element set of the atomic operator sequence based on the computing characteristics, wherein the target processing element set comprises a plurality of target processing elements, one type of the atomic operators corresponds to one or more target processing elements, and the target processing elements are used for representing any one processing element of the SPU, the TPU and the VPU; Establishing a data path for each target processing element through the reconfigurable interconnect network based on the set of target processing elements; And connecting atomic operators with the data dependency relationship based on the data path, and configuring the reconfigurable interconnection network.
  6. 6. The method of claim 1, wherein generating system load prediction information based on the computed features of the various atomic operators and real-time task queue depth comprises: Acquiring reference power consumption values of the atomic operators under standard voltage frequency and historical memory access frequencies of the atomic operators in the historical execution process based on the calculation characteristics of the atomic operators; Identifying the type and the number of atomic operators to be executed based on the real-time task queue depth, and calculating the total reference power consumption value of the current queue based on each reference power consumption value and a preset weight set; Predicting the bandwidth configuration and the interface data throughput of the memory controller in a preset future time window through a preset linear regression algorithm based on the historical memory access frequency and the type and the number of atomic operators to be executed; Calculating a power consumption increment value through a preset power consumption model based on the memory controller bandwidth configuration and the interface data throughput; calculating the sum of the total reference power consumption value and the power consumption increment value to obtain a predicted power consumption value in the preset future time window; And taking the bandwidth configuration of the memory controller, the data throughput of the interface and the predicted power consumption value together as the system load prediction information.
  7. 7. The method of claim 6, wherein dynamically configuring the reconfigurable hardware suite by the hardware management unit based on the system load prediction information comprises: Based on the memory controller bandwidth configuration, adjusting the bandwidth, the working mode and the clock frequency of the memory controller of the main computing module; Based on the interface data throughput, adjusting the link rate and the power management state of a high-speed input/output interface between the main computing module and the extended function module through the resource control unit; If the predicted power consumption value is larger than a preset power consumption threshold value, searching a first frequency voltage combination corresponding to the power consumption value not exceeding the preset power consumption threshold value in a preset frequency voltage configuration table, and reducing the working voltage and the working frequency of the dynamic reconfigurable computing unit to values corresponding to the first frequency voltage combination; If the predicted power consumption value is smaller than or equal to the preset power consumption threshold, calculating a difference value between the preset power consumption threshold and the predicted power consumption value to obtain a power consumption allowance, searching a second frequency voltage combination corresponding to the power consumption allowance in the frequency voltage configuration table, and improving the working voltage and the working frequency of the dynamic reconfigurable computing unit to values corresponding to the second frequency voltage combination.
  8. 8. A reconfigurable hardware suite control device comprising one or more processors and memory coupled with the one or more processors, the memory to store computer program code comprising computer instructions that the one or more processors invoke to cause the reconfigurable hardware suite control device to perform the method of any one of claims 1-7.
  9. 9. A computer readable storage medium comprising instructions which, when run on a reconfigurable hardware suite control device, cause the reconfigurable hardware suite control device to perform the method of any one of claims 1-7.
  10. 10. A computer program product, characterized in that the computer program product, when run on a reconfigurable hardware suite control device, causes the reconfigurable hardware suite control device to perform the method of any one of claims 1-7.

Description

Reconfigurable hardware suite control method, equipment, medium and product Technical Field The application relates to the technical field of artificial intelligence hardware, in particular to a reconfigurable hardware suite control method, equipment, medium and product. Background Along with the wide application of artificial intelligence technology in the edge computing scene, the reconfigurable hardware suite becomes the key for realizing high-efficiency end-side intelligent computing because of being capable of adapting to different computing tasks through dynamic configuration. When the traditional control method processes dynamic and changeable task flows and heterogeneous computing resources, the coordination among the computing units is not smooth, so that the effective computing time of the hardware resources is greatly compressed, and the throughput and the energy efficiency of the whole system are far lower than theoretical peaks. The existing dynamic voltage frequency adjustment (DVFS) technology can adjust the working voltage and frequency of a processor according to a load, but the adjustment granularity of the dynamic voltage frequency adjustment (DVFS) technology usually takes the whole processor or a fixed computing cluster as a unit, cannot perform fine-granularity differential control on heterogeneous computing units in a neural network processor, cannot perform predictive allocation of computing resources based on task type characteristics and task queue depth before task execution, and the existing dynamic adjustment scheme only performs power consumption management on a single hardware module, lacks prediction of the whole load and coordination capability of cross-module resources, and reduces the control efficiency of a reconfigurable hardware suite. Disclosure of Invention The embodiment of the application provides a reconfigurable hardware suite control method, equipment, medium and product, which are used for solving the technical problem of how to improve the control efficiency of the reconfigurable hardware suite. In a first aspect, an embodiment of the present application provides a reconfigurable hardware set control method, which is applied to a reconfigurable hardware set, where the reconfigurable hardware set includes a main computing module and at least one extended functional module, the main computing module includes a dynamic reconfigurable computing unit, a hardware management unit, and a resource control unit, and the extended functional module includes a golden finger interface unit and a storage unit, and the method includes: Responding to an access event of the extended function module, the main calculation module reads and analyzes module file information in the storage unit through the golden finger interface unit to obtain hardware resource requirements and operating system adaptation information, wherein the module file information is used for storing configuration data of the extended function module; Based on the hardware resource requirement, distributing hardware resources to the extended function module through the resource control unit; The method comprises the steps of responding to a module hot plug event, initializing and driving loading the extended function module based on the operating system adaptation information so as to enable the extended function module to enter a usable state; decomposing a task chain into an atomic operator sequence through the dynamic reconfigurable computing unit in response to the task chain issued by the artificial intelligence application, wherein the atomic operator sequence comprises a plurality of atomic operators of different types; Configuring a reconfigurable interconnection network in the dynamically reconfigurable computing unit based on the computing characteristics of the atomic operators of various types; Generating system load prediction information based on the calculation characteristics of the atomic operators and the real-time task queue depth; and dynamically configuring the reconfigurable hardware suite through the hardware management unit based on the system load prediction information. Optionally, the golden finger interface unit comprises a first configuration pin, the response to the access event of the extended function module, the main calculation module reads and analyzes module file information in the storage unit through the golden finger interface unit to obtain hardware resource requirement and operating system adaptation information, the golden finger interface unit comprises the response to the access event, the main calculation module accesses the module file information of the storage unit through the first configuration pin, the module file information at least comprises a module identifier, power consumption requirement information, the number of required high-speed input and output channels, an interrupt mapping scheme and driving configuration scripts for different operating systems,