WO2024143656A1

WO2024143656A1 - Electronic device for performing prefetcher algorithm appropriate for neural network model by using phase change memory, and operation method thereof

Info

Publication number: WO2024143656A1
Application number: PCT/KR2023/000841
Authority: WO
Inventors: 김현; 장지훈
Original assignee: 서울과학기술대학교 산학협력단
Priority date: 2022-12-26
Filing date: 2023-01-18
Publication date: 2024-07-04
Also published as: KR20240102611A

Abstract

The present invention relates to an electronic device for performing a prefetcher algorithm appropriate for a neural network model by using a phase change memory, wherein a heterogenous memory structure is provided through additional utilization of a phase case memory (PCM) as a memory for replacing a dynamic random access memory (DRAM), whereby universal application to a CNN model is possible and power saving can be achieved while high performance is maintained.

Description

Electronic device and operating method for performing a prefetcher algorithm suitable for a neural network model using phase change memory

Various embodiments of the present invention relate to an electronic device that performs a prefetcher algorithm suitable for a neural network model using a phase change memory and a method of operating the same.

A prefetcher is a technology that predicts the data access address requested by the processor and pre-allocates data at the address the processor wishes to allocate. Predicted data is allocated to cache memory to maximize the efficiency of cache memory by processing data requested by the processor from main memory in cache memory and to hide the time required to read and retrieve data from main memory.

The present invention provides a heterogeneous memory structure that additionally utilizes phase change memory (PCM) as a memory to replace dynamic random access memory (DRAM) for an efficient low-power system. The present invention provides a PCM-based system that is very easy to use in mobile/edge devices through power savings compared to systems using DRAM. The present invention provides a prefetcher algorithm that can respond to data locality, can be universally used in CNN models, and can achieve excellent performance without requiring many additional hardware resources.

According to various embodiments, an electronic device that performs a prefetcher algorithm suitable for a neural network model using a phase change memory includes a processor, and when executed by the processor, causes the processor to perform at least one operation related to the neural network model. A phase change memory (PCM) that stores a first instruction that causes the processor to perform an operation, and a dynamic random memory (DRAM) that stores a second instruction that causes the processor to perform operations other than the at least one operation. Access Memory), wherein the at least one operation includes a first delta value corresponding to the difference between a current access address and a previous access address of the memory and a second delta value corresponding to a previous delta value of the first delta value. Checking the value, comparing the first delta value and the second delta value according to the first operation priority for prefetching the access address, and if the first delta value is equal to the second delta value, An operation of performing a first process of prefetching a first prediction address by applying the first delta value to the current access address. As a subordinate operation of the first operation priority, according to the second operation priority, ROB (Reorder Buffer ) A second process of searching whether the current access address exists in the table and, if the current access address exists in the ROB table, prefetching at least one predicted address recorded in the ROB table by a preset number of degrees. As an operation to be performed, and a subordinate operation of the second operation priority, according to the third operation priority, when the first delta value and the second delta value are different, the delta score corresponding to the first delta value is If it is greater than the offset threshold (OT), it may include performing a third process of prefetching a second prediction address by applying the first delta value to the current access address.

According to various embodiments, a method of operating an electronic device that performs a prefetcher algorithm suitable for a neural network model using a phase change memory includes a first access address corresponding to the difference between the current access address and the previous access address of the memory of the electronic device. Checking a delta value and a second delta value that is the previous delta value of the first delta value; Comparing the first delta value and the second delta value according to the first operation priority for prefetching the access address If the first delta value is the same as the second delta value, performing a first process of prefetching a first prediction address by applying the first delta value to the current access address, a lower priority of the first operation priority As an operation, according to the second operation priority, it is searched whether the current access address exists in the ROB (Reorder Buffer) table, and if the current access address exists in the ROB table, it is stored in the ROB table by a preset number of degrees. performing a second process of prefetching at least one predicted address that has been recorded, and a subordinate operation of the second operation priority, wherein according to a third operation priority, the first delta value and the second delta value In this different case, if the delta score corresponding to the first delta value is greater than or equal to OT (Offset Threshold), performing a third process of prefetching a second prediction address by applying the first delta value to the current access address. It can be included.

The present invention can be universally applied to CNN models, enabling power savings while maintaining high performance, providing an effect that can be used in various artificial intelligence industry fields.

1 shows a block diagram of an electronic device and a network according to various embodiments of the present invention.

FIG. 2 is a diagram illustrating an operation in which an electronic device updates records of tables used in a prefetch algorithm in a PCM system, according to various embodiments.

FIG. 3 is a diagram illustrating an example in which an electronic device updates a record of an OST, according to various embodiments.

FIG. 4 is a diagram illustrating a first embodiment in which an electronic device updates records of RHB and PST, according to various embodiments.

FIG. 5 is a diagram illustrating a second embodiment in which an electronic device updates records of RHB and PST, according to various embodiments.

FIG. 6 is a diagram illustrating an example in which an electronic device updates records of PST and ROB, according to various embodiments.

FIG. 7 is a flowchart illustrating a method by which an electronic device performs prefetching, according to various embodiments.

FIG. 8 is a diagram illustrating a method in which an electronic device performs prefetching using an OST, according to various embodiments.

FIG. 9 is a diagram illustrating a method in which an electronic device performs prefetching using ROB, according to various embodiments.

Hereinafter, various embodiments of this document are described with reference to the attached drawings. The examples and terms used herein are not intended to limit the technology described in this document to specific embodiments, and should be understood to include various modifications, equivalents, and/or substitutes for the examples. In connection with the description of the drawings, similar reference numbers may be used for similar components. Singular expressions may include plural expressions unless the context clearly indicates otherwise. In this document, expressions such as “A or B” or “at least one of A and/or B” may include all possible combinations of the items listed together. Expressions such as “first,” “second,” “first,” or “second,” can modify the corresponding components regardless of order or importance, and are used to distinguish one component from another. It is only used and does not limit the corresponding components. When a component (e.g., a first) component is said to be "connected (functionally or communicatively)" or "connected" to another (e.g., second) component, it means that the component is connected to the other component. It may be connected directly to a component or may be connected through another component (e.g., a third component).

In this document, “configured to” means “suitable for,” “having the ability to,” or “changed to,” depending on the situation, for example, in terms of hardware or software. ," can be used interchangeably with "made to," "capable of," or "designed to." In some contexts, the expression “a device configured to” may mean that the device is “capable of” working with other devices or components. For example, the phrase "processor configured (or set) to perform A, B, and C" refers to a processor dedicated to performing the operations (e.g., an embedded processor), or by executing one or more software programs stored on a memory device. , may refer to a general-purpose processor (e.g., CPU or application processor) capable of performing the corresponding operations.

Electronic devices according to various embodiments of this document may include, for example, at least one of a smartphone, tablet PC, desktop PC, laptop PC, netbook computer, workstation, and server.

1, an electronic device 101 within a network environment 100, in various embodiments, is described. The electronic device 101 may include a bus 110, a processor 120, a memory 130, an input/output interface 150, a display 160, and a communication interface 170. In some embodiments, the electronic device 101 may omit at least one of the components or may additionally include another component. The bus 110 connects the components 110 to 170 to each other and may include circuitry that transfers communication (eg, control messages or data) between the components. The processor 120 may include one or more of a central processing unit, an application processor, or a communication processor (CP). The processor 120 may, for example, perform operations or data processing related to control and/or communication of at least one other component of the electronic device 101.

Memory 130 may include volatile and/or non-volatile memory. For example, the memory 130 may store commands or data related to at least one other component of the electronic device 101. According to one embodiment, memory 130 may store software and/or program 140. The program 140 may include, for example, a kernel 141, middleware 143, an application programming interface (API) 145, and/or an application program (or “application”) 147, etc. . At least a portion of the kernel 141, middleware 143, or API 145 may be referred to as an operating system. Kernel 141, for example, provides system resources (e.g., middleware 143, API 145, or application program 147) used to execute operations or functions implemented in other programs (e.g., : Bus 110, processor 120, or memory 130, etc.) can be controlled or managed. In addition, the kernel 141 provides an interface for controlling or managing system resources by accessing individual components of the electronic device 101 in the middleware 143, API 145, or application program 147. You can.

The middleware 143 may, for example, perform an intermediary role so that the API 145 or the application program 147 can communicate with the kernel 141 to exchange data. Additionally, the middleware 143 may process one or more work requests received from the application program 147 according to priority. For example, the middleware 143 may use system resources (e.g., bus 110, processor 120, or memory 130, etc.) of the electronic device 101 for at least one of the application programs 147. Priority may be assigned and the one or more work requests may be processed. The API 145 is an interface for the application 147 to control functions provided by the kernel 141 or middleware 143, for example, at least for file control, window control, image processing, or character control. Can contain one interface or function (e.g. command). The input/output interface 150, for example, transmits commands or data input from a user or other external device to other component(s) of the electronic device 101, or to other component(s) of the electronic device 101 ( Commands or data received from (fields) can be output to the user or other external device.

Display 160 may include, for example, a liquid crystal display (LCD), a light-emitting diode (LED) display, an organic light-emitting diode (OLED) display, a microelectromechanical system (MEMS) display, or an electronic paper display. It can be included. For example, the display 160 may display various contents (e.g., text, images, videos, icons, and/or symbols, etc.) to the user. The display 160 may include a touch screen and may receive, for example, a touch, gesture, proximity, or hovering input using an electronic pen or a part of the user's body. The communication interface 170, for example, establishes communication between the electronic device 101 and an external device (e.g., the first external electronic device 102, the second external electronic device 104, or the server 106). You can. For example, the communication interface 170 may be connected to the network 162 through wireless or wired communication to communicate with an external device (eg, the second external electronic device 104 or the server 106).

Wireless communications include, for example, LTE, LTE Advance (LTE-A), code division multiple access (CDMA), wideband CDMA (WCDMA), universal mobile telecommunications system (UMTS), Wireless Broadband (WiBro), or Global GSM (GSM). It may include cellular communication using at least one of the System for Mobile Communications). According to one embodiment, wireless communication includes, for example, wireless fidelity (WiFi), Bluetooth, Bluetooth Low Energy (BLE), Zigbee, near field communication (NFC), Magnetic Secure Transmission, and radio. It may include at least one of frequency (RF) or body area network (BAN). According to one embodiment, wireless communications may include GNSS. GNSS may be, for example, Global Positioning System (GPS), Global Navigation Satellite System (Glonass), Beidou Navigation Satellite System (hereinafter “Beidou”), or Galileo, the European global satellite-based navigation system. Hereinafter, in this document, “GPS” may be used interchangeably with “GNSS.” Wired communication may include, for example, at least one of universal serial bus (USB), high definition multimedia interface (HDMI), recommended standard 232 (RS-232), power line communication, or plain old telephone service (POTS). there is. Network 162 may include at least one of a telecommunications network, for example, a computer network (e.g., a LAN or WAN), the Internet, or a telephone network.

Each of the first and second external

electronic devices

102 and 104 may be of the same or different type as the electronic device 101. According to various embodiments, all or part of the operations executed in the electronic device 101 may be executed in one or more electronic devices (e.g., the

electronic devices

102 and 104, or the server 106). In one embodiment, According to this, when the electronic device 101 is to perform a certain function or service automatically or upon request, the electronic device 101 performs at least some functions associated therewith instead of or in addition to executing the function or service on its own. The other electronic device (e.g.,

electronic device

102, 104, or server 106) may request the requested function or Additional functions may be executed and the results may be transmitted to the electronic device 101. For this purpose, the electronic device 101 may process the received results as is or additionally to provide the requested function or service. For example, cloud computing, distributed computing, or client-server computing technologies may be used.

FIG. 2 is a diagram illustrating an operation of an electronic device (e.g., the electronic device 101 of FIG. 1) updating records of tables used in a prefetch algorithm in a PCM system, according to various embodiments.

FIG. 3 is a diagram illustrating an example in which the electronic device 101 updates a record of an OST, according to various embodiments.

FIG. 4 is a diagram for explaining a first embodiment in which the electronic device 101 updates records of RHB and PST, according to various embodiments.

FIG. 5 is a diagram illustrating a second embodiment in which the electronic device 101 updates records of RHB and PST, according to various embodiments.

FIG. 6 is a diagram illustrating an example in which the electronic device 101 updates PST and ROB records, according to various embodiments.

According to various embodiments, the electronic device 101 may include a processor (e.g., processor 120 in FIG. 1) and a memory (e.g., memory 130 in FIG. 1), and the memory 130 is a processor. A phase change memory (PCM) storing a first instruction that, when executed by 120, causes the processor 120 to perform at least one operation related to a neural network model (e.g., a CNN model) and the processor It may include a dynamic random access memory (DRAM) that stores a second instruction that causes 120 to perform operations other than the at least one operation.

According to various embodiments, referring to FIG. 2, the electronic device 101 checks the current access address (e.g., trigger access) and updates the OST (Offset Score Table) 201 by using the current access address. You can check the first delta value corresponding to the difference between and the previous access address and the second delta value, which is the previous delta value of the first delta value. For example, referring to Figure 2, the OST 201 consists of a first delta value (Delta) item, a second delta value (Last Delta) item, and a delta score (Score) item corresponding to the first delta value. It can be. For example, referring to FIG. 3, the electronic device 101 has a current access address 301 (e.g., 315) and a previous access address (e.g., 315) for a memory (e.g., memory 130 in FIG. 1) at a specific point in time. 302) (e.g. 310) can be checked, and the first delta value (e.g. 5) corresponding to the difference between the current access address 301 and the previous access address 302 and the previous delta value of the first delta value A second delta value (e.g., 5) corresponding to the difference between the access address 302 and the previous access address 303 can be confirmed.

According to one embodiment, the electronic device 101 may record the confirmed first delta value and the confirmed second delta value in the OST 201. For example, referring to FIG. 3 , the electronic device 101 may update the OST 201 record of the confirmed first delta value 311 and second delta value 312. According to one embodiment, the electronic device 101 may store the second delta values 312 one by one, and may delete the previous record when recording a new value.

According to one embodiment, when the first delta value 311 and the second delta value 312 are the same, the electronic device 101 sets the delta score corresponding to the first delta value 311 to be the same as the previous delta score. This can be recorded in OST (201). For example, referring to FIG. 3, when the first delta value 311 (e.g., 5) and the second delta value 312 (e.g., 5) are the same, the electronic device 101 displays the first delta value (e.g., 5). The delta score 313 corresponding to 311) may be recorded in the OST 201 as the same value as the previous delta score (e.g., 1).

According to one embodiment, when the first delta value 311 and the second delta value 312 are different, the electronic device 101 determines the previous delta score 313 corresponding to the first delta value 311. (e.g. 1) can be recorded in the OST (201) as a value (e.g. 2) increased by a predetermined score (e.g. 1).

According to various embodiments, referring to FIG. 2, as the electronic device 101 checks the current access address (e.g., trigger access), the Region-History Buffer (RHB) table 202 (hereinafter referred to as 'RHB') In order to update the address, it can be checked whether the block number corresponding to the current access address exists in the block number item of the RHB 202. The RHB 202 can track delta histories for block addresses, use the block address as an index when a trigger address occurs, and sequentially record offsets based on the trigger address. The RHB 202 stores a stream of cache misses history to enable prediction of temporal patterns. For example, referring to FIG. 2, the RHB 202 may be composed of at least one delta value as a block number (Block Num) item, a block address (Block Addr) item, and a delta pattern (Delta Seq) item. . For example, referring to FIG. 4, the electronic device 101 may check the current access address 301 (e.g., 315), and the block number (e.g., 3) corresponding to the current access address 301 is RHB. You can check whether it exists in the block number item (411) of (202). At this time, the block number may represent the identification number of the corresponding block, and in the example of FIG. 4, the block unit is set to 100, but this is only an example and is not limited to this.

According to one embodiment, referring to FIG. 4, when a block number (e.g., 3) is present in the block number item 411 of the RHB 202, the electronic device 101 selects the RHB 202 corresponding to the block number. The RHB delta value (e.g., 15) corresponding to the current access address (301) (e.g., 315) can be recorded in the delta pattern item 413 of ). At this time, the block address item 412 may be an item recording the first access address (e.g., starting address) in the block, and the RHB delta value is the address of the block address item 412 corresponding to the block number ( Example: It may be the difference between 300) and the current access address (301).

According to one embodiment, referring to FIG. 4, after recording the RHB delta value in the delta pattern item 413, the electronic device 101 determines in advance the length of the delta pattern item 413 corresponding to the block number. If it is longer than a set length (e.g. 5), the RHB record corresponding to the block number 411 is recorded in the PST (Pattern Sequence Table) 203, and the corresponding RHB record can be deleted. The length of the delta pattern item 413 means the number of delta values in the item, and the PST 203 will be explained in detail below.

According to one embodiment, referring to FIG. 5, the electronic device 101 determines that the block number (e.g., 3) corresponding to the current access address (501) (e.g., 350) is the block number item (411) of the RHB (202). ), it is possible to check whether the block number corresponding to the current access address (501) exists in the block number item (421) of the PST (203). The PST 203 may be a table in which delta sequences of a certain sequence length of the RHB 202 are loaded identically. For example, referring to Figure 2, the PST 203 has at least one delta value as a block number (Block Num) item, a block address (Block Addr) item, a patch number item, and a delta pattern (Delta Seq) item. It can be configured. The concepts of the block number item, block address item, and delta pattern item of the PST 203 may be the same as the concepts of the block number item, block address item, and delta pattern item of the RHB 202.

According to one embodiment, referring to FIG. 5, the electronic device 101 includes the block number (e.g., 3) of the current access address 501 (e.g., 350) in the block number item 421 of the PST 203. If it exists, it can be checked whether the PST delta value (e.g., 50) corresponding to the current access address exists in the delta pattern item 424 of the PST 203. At this time, the PST delta value may be the difference value (e.g., 50) between the address (e.g., 300) of the block address item 422 corresponding to the block number and the current access address (501) (e.g., 350).

According to one embodiment, referring to FIG. 5, when the PST delta value (e.g., 50) does not exist in the delta pattern item 424 of the PST 203, the electronic device 101 selects a patch corresponding to the block number. If the value of the number item (e.g., 3) is greater than or equal to a specific value (e.g., 3), the PST delta value may be additionally recorded in the delta pattern item 424 of the PST 203. In this case, the specific value may be a predetermined ratio (eg, 60%) to the length (eg, 5) of the delta pattern item 424 of the PST 203 corresponding to the block number. According to one embodiment, referring to FIG. 5, the electronic device 101 stores the current access address 502 (e.g., 430) PST delta value (e.g., 30) in the delta pattern item 424 of the PST 203. If it does not exist, and the value (e.g. 1) of the patch number entry 423 of the PST 203 is less than a certain value (e.g. 3), a new block number entry 411 corresponding to the current access address 502 is created. value (e.g. 4), a new value (e.g. 430) of the block address entry 412 corresponding to the current access address 502, and a new value of the delta pattern entry 413 corresponding to the current access address 502. A value (e.g. 0) may be written to RHB 202.

According to one embodiment, referring to FIG. 6, the electronic device 101 determines that the PST delta value (e.g., 25) corresponding to the current access address 601 (e.g., 325) is a delta pattern item ( 424), record the current access address (601) in the ROB (Reorder Buffer) table (hereinafter referred to as 'ROB'), and correspond to the block number (e.g. 3) for the current access address (601). A value (e.g., 2) obtained by adding a predetermined value (e.g., 1) to the value (e.g., 1) of the patch number item 423 of the PST 203 may be recorded. ROB (204) is a buffer that sorts the delta sequence in PST (203) once more. For example, referring to FIG. 2, the ROB 204 may include an entry for recording an access address to perform prefetching.

According to one embodiment, the electronic device 101 is configured when the block number corresponding to the current access address does not exist in both the block number item 411 of the RHB 202 and the block number item 421 of the PST 203. , a new value of the block number entry 411 corresponding to the current access address, a new value of the block address entry 412 corresponding to the current access address, a new value of the delta pattern entry 413 corresponding to the current access address. can be recorded in RHB (202).

FIG. 7 is a flowchart illustrating a method in which an electronic device (e.g., the electronic device 101 of FIG. 1 ) performs prefetching, according to various embodiments.

FIG. 8 is a diagram illustrating a method in which the electronic device 101 performs prefetching using an OST (e.g., the OST 201 of FIG. 2) according to various embodiments.

FIG. 9 is a diagram illustrating a method by which the electronic device 101 performs prefetching using a ROB (eg, ROB 204 in FIG. 2) according to various embodiments.

In operation 701, according to various embodiments, the electronic device 101 generates a first delta value corresponding to the difference between the current access address and the previous access address of the memory (e.g., the memory 130 of FIG. 1). You can check the second delta value, which is the previous delta value of the first delta value. According to one embodiment, the electronic device 101 includes a first delta value (e.g., 5) of the delta entry 311 recorded in the OST 201 corresponding to the current access address 301 and the first current access address 301. ), the second delta value (e.g. 5) recorded in the previous delta item 312 of the OST 201 can be confirmed.

In operation 703, according to various embodiments, the electronic device 101 compares the first delta value and the second delta value according to the first operation priority for prefetching the access address, and the first delta value is the first delta value. If the delta value is equal to 2, a first process of prefetching the first prediction address by applying the first delta value to the currently accessed address may be performed. For example, referring to FIG. 8 , the electronic device 101 displays the first delta value (e.g., 5) of the OST (201) corresponding to the current access address (301) (e.g., 315) and the current access address (301). ), if the second delta value (e.g., 5) of the OST (201) corresponding to ) is the same, the first prediction address obtained by applying (e.g., addition) the first delta value to the currently accessed address (301) is stored in the memory (130). A first process of prefetching (e.g., prefetcher cache memory) may be performed. The electronic device 101 can predict addresses by recording and scoring differences between consecutive addresses using the OST 201 to cope with sequential locality of data.

In operation 705, according to various embodiments, the electronic device 101 currently accesses the ROB (e.g., ROB 204 in FIG. 2) according to the second operation priority, as a subordinate operation of the first operation priority. A second process may be performed to check whether an address exists, and if a current access address exists in the ROB 204, prefetching at least one predicted address recorded in the ROB 204 by a preset number of degrees. For example, referring to FIG. 9, when the first delta value and the second delta value of the OST 201 are not the same, the electronic device 101 performs the second operation priority as a subordinate operation of the first operation priority. According to the ranking, it is searched whether the current access address 901 (e.g. 310) exists in the ROB 204, and if the current access address 901 exists in the ROB 204, a preset number of degrees (e.g. 4 ) A second process of prefetching the predicted addresses recorded in the ROB 204 (e.g., 325, 415, 777, 999) may be performed. At this time, the preset number of degrees can be freely set by a request from a user of the electronic device 101 or an application running on the electronic device 101. The electronic device 101 records repetitive access addresses based on block addresses through the RHB 202 and PST 203 to cope with the temporal locality of data, and makes predictions using the ROB 204. The order can be rearranged.

In operation 707, according to various embodiments, the electronic device 101 performs a lower-priority operation of the second operation priority and, according to the third operation priority, the first delta value and the second delta value of the OST 201. In a different case, if the delta score corresponding to the first delta value is more than OT (Offset Threshold), a third process of prefetching the second prediction address by applying the first delta value to the currently accessed address may be performed. For example, referring to FIG. 8, when there is no current access address in the ROB 204, the electronic device 101 performs OST (OST) according to the third operation priority as a subordinate operation of the second operation priority. Based on the difference between the first delta value (e.g., 30) and the second delta value (e.g., 5) of 201), the delta score (e.g., 11) corresponding to the first delta value (e.g., 30) is OT ( Example: If it is 10) or more, a third process of prefetching the second prediction address by applying (e.g., addition) the first delta value to the currently accessed address can be performed.

According to one embodiment, after performing the second process according to the third operation priority, the electronic device 101 initializes the entries recorded in the OST 201 or corresponds to the first delta value. The delta score can be reduced from the OT by a predetermined value and recorded in the OST 201.

According to various embodiments, the at least one operation may further include recording the first delta value and the second delta value in an offset score table (OST).

According to various embodiments, the at least one operation may include, when the first delta value and the second delta value are the same, making the delta score corresponding to the first delta value the same as the previous delta score and recording it in the OST. an operation of increasing the delta score corresponding to the first delta value by a predetermined score to the previous delta score and recording it in the OST when the first delta value and the second delta value are different. It can be included.

According to various embodiments, the at least one operation may be, according to the third operation priority, after performing the third process, initializing entries recorded in the OST or obtaining the delta score from the OT. An operation of decreasing the value by a predetermined value and recording it in the OST may be further included.

According to various embodiments, the at least one operation may include checking whether a block number corresponding to the current access address exists in a block number item of a Region-History Buffer (RHB) table, and the block number is the block number. If present in the item, recording the RHB delta value corresponding to the current access address in the delta pattern item of the RHB corresponding to the block number - the block address item records the first access address in the block , and the RHB delta value is a difference value between the address of the block address item corresponding to the block number and the current access address, and after recording the RHB delta value in the delta pattern item, the delta corresponding to the block number If the length of the pattern item is longer than a predetermined length, the operation of recording the RHB record corresponding to the block number in a PST (Pattern Sequence Table) and deleting the RHB record may be further included.

According to various embodiments, the at least one operation is such that, when the block number does not exist in the block number item of the RHB table, the block number corresponding to the current access address is added to the block number item of the PST. An operation to check whether the block number exists in the block number item of the PST, an operation to check whether the PST delta value corresponding to the current access address exists in the delta pattern item of the PST, the PST delta value is If it does not exist in the delta pattern item of the PST, if the value of the patch number item of the PST is greater than a certain value, the PST delta value is additionally recorded in the delta pattern item of the PST, and the value of the patch number item of the PST is additionally recorded. If this value is less than the specific value, recording a new block number corresponding to the current access address, a new block address corresponding to the current access address, and a new delta value corresponding to the current access address in the RHB table - the specific value is a predetermined ratio value to the length of the delta pattern item of the PST -, and if the PST delta value exists in the delta pattern item of the PST, record the current access address in the ROB table and correspond to the block number The operation of adding a predetermined value to the value of the patch number item of the PST may be further included.

According to various embodiments, the at least one operation is to generate a new block number corresponding to the current access address if the block number is not present in the block number entry of the RHB table and the block number entry of the PST. , may further include recording a new block address corresponding to the current access address and a new delta value corresponding to the current access address in the RHB table.

According to various embodiments, the method of operating the electronic device may further include recording the first delta value and the second delta value in an offset score table (OST).

According to various embodiments, a method of operating the electronic device may include, when the first delta value and the second delta value are the same, setting the delta score corresponding to the first delta value to be the same as the previous delta score and adding it to the OST. An operation of recording, and when the first delta value and the second delta value are different, an operation of increasing the delta score corresponding to the first delta value by a predetermined score to the previous delta score and recording it in the OST. More may be included.

According to various embodiments, a method of operating the electronic device may include, after performing the third process according to the third operation priority, initializing the entries recorded in the OST or using the OT as the delta score. It may further include an operation of decreasing by a predetermined value and recording it in the OST.

According to various embodiments, a method of operating the electronic device includes checking whether a block number corresponding to the current access address exists in a block number item of a Region-History Buffer (RHB) table, and the block number is the block number. If present in the number item, recording the RHB delta value corresponding to the current access address in the delta pattern item of the RHB corresponding to the block number - the block address item records the first access address in the block item, and the RHB delta value is a difference value between the address of the block address item corresponding to the block number and the current access address, and after recording the RHB delta value in the delta pattern item, the RHB delta value corresponding to the block number If the length of the delta pattern item is longer than a predetermined length, the operation of recording the RHB record corresponding to the block number in a PST (Pattern Sequence Table) and deleting the RHB record may be further included.

According to various embodiments, a method of operating the electronic device may include, when the block number does not exist in the block number item of the RHB table, the block number corresponding to the current access address is the block number item of the PST. Checking whether the block number exists in the block number item of the PST, checking whether the PST delta value corresponding to the current access address exists in the delta pattern item of the PST, the PST delta value If it does not exist in the delta pattern item of the PST, if the value of the patch number item of the PST is greater than a certain value, the PST delta value is additionally recorded in the delta pattern item of the PST, and the patch number item of the PST If the value is less than the specific value, recording a new block number corresponding to the current access address, a new block address corresponding to the current access address, and a new delta value corresponding to the current access address in the RHB table - the specific The value is a predetermined ratio value to the length of the delta pattern item of the PST, and if the PST delta value exists in the delta pattern item of the PST, the current access address is recorded in the ROB table and the block number is The operation of adding a predetermined value to the value of the corresponding patch number item of the PST may be further included.

According to various embodiments, a method of operating the electronic device may include, if the block number does not exist in the block number item of the RHB table and the block number item of the PST, a new block corresponding to the current access address. The method may further include recording a number, a new block address corresponding to the current access address, and a new delta value corresponding to the current access address in the RHB table.

As used in this document, the term “module” or “unit” includes a unit comprised of hardware, software, or firmware, and can be used interchangeably with terms such as logic, logic block, component, or circuit, etc. there is. “Module” or “unit” may be an integrated part, a minimum unit that performs one or more functions, or a part thereof. A “module” or “unit” may be implemented mechanically or electronically, e.g., known or to be developed, application-specific integrated circuit (ASIC) chips, field-programmable circuits (FPGAs), that perform certain operations. gate arrays), or programmable logic devices, and may be executed by the processor 120. At least a portion of the device (e.g., modules or functions thereof) or method (e.g., operations) according to various embodiments includes instructions stored in a computer-readable storage medium (e.g., memory 130) in the form of a program module. It can be implemented as: When the instruction is executed by a processor (eg, processor 120), the processor may perform the function corresponding to the instruction. Computer-readable recording media include hard disks, floppy disks, magnetic media (e.g. magnetic tape), optical recording media (e.g. CD-ROM, DVD, magneto-optical media (e.g. floptical disks), built-in memory, etc. The instruction may include code generated by a compiler or code that can be executed by an interpreter. A module or program module according to various embodiments may include at least one of the above-described components. , some parts may be omitted, or other components may be further included. Operations performed by modules, program modules, or other components may be executed sequentially, parallel, iteratively, or heuristically. Some operations may be executed in a different order, omitted, or other operations may be added.

Additionally, the embodiments disclosed in this document are presented for explanation and understanding of the disclosed technical content, and do not limit the scope of the present disclosure. Accordingly, the scope of the present disclosure should be interpreted as including all changes or various other embodiments based on the technical idea of the present disclosure.

Claims

In an electronic device that performs a prefetcher algorithm suitable for a neural network model using a phase change memory,

processor, and

A phase change memory (PCM) storing a first instruction that, when executed by the processor, causes the processor to perform at least one operation related to a neural network model and causes the processor to perform the at least one operation. a memory including dynamic random access memory (DRAM) that stores second instructions that cause other operations to be performed,

The at least one operation is,

An operation of checking a first delta value corresponding to the difference between the current access address and the previous access address of the memory and a second delta value that is the previous delta value of the first delta value,

According to the first operation priority for prefetching of the access address, the first delta value and the second delta value are compared, and if the first delta value is equal to the second delta value, the current access address is connected to the second delta value. An operation of performing a first process of prefetching a first prediction address applying a delta value of 1;

As a subordinate operation of the first operation priority, according to the second operation priority, it is searched whether the current access address exists in the ROB (Reorder Buffer) table, and if the current access address exists in the ROB table, in advance An operation of performing a second process of prefetching at least one prediction address recorded in the ROB table by a set number of degrees, and

As a subordinate operation of the second operation priority, according to the third operation priority, when the first delta value and the second delta value are different, the delta score corresponding to the first delta value is OT (Offset Threshold) If it is greater than or equal to the current access address, it includes performing a third process of prefetching a second prediction address by applying the first delta value to the current access address.

Electronic devices.
According to claim 1,

The at least one operation is,

Further comprising recording the first delta value and the second delta value in an OST (Offset Score Table),

Electronic devices.
According to clause 2,

The at least one operation is,

When the first delta value and the second delta value are the same, recording the delta score corresponding to the first delta value as the previous delta score in the OST;

When the first delta value and the second delta value are different, increasing the delta score corresponding to the first delta value by a predetermined score to the previous delta score and recording it in the OST,

Electronic devices.
According to clause 3,

The at least one operation is,

According to the third operation priority, after performing the third process, an operation of initializing the entries recorded in the OST or decreasing the delta score by a predetermined value from the OT and recording the entries in the OST is further performed. containing,

Electronic devices.
According to claim 1,

The at least one operation is,

An operation to check whether the block number corresponding to the current access address exists in the block number item of the RHB (Region-History Buffer) table,

If the block number is present in the block number item, recording the RHB delta value corresponding to the current access address in the delta pattern item of the RHB corresponding to the block number - the block address item is It is an item that records the first access address, and the RHB delta value is the difference between the address of the block address item corresponding to the block number and the current access address -, and

After recording the RHB delta value in the delta pattern item, if the length of the delta pattern item corresponding to the block number is more than a predetermined length, the RHB record corresponding to the block number is recorded in a PST (Pattern Sequence Table) , including the operation of deleting the RHB record,

Electronic devices.
According to clause 5,

The at least one operation is,

If the block number does not exist in the block number item of the RHB table, checking whether the block number corresponding to the current access address exists in the block number item of the PST,

If the block number is present in the block number item of the PST, checking whether the PST delta value corresponding to the current access address is present in the delta pattern item of the PST;

When the PST delta value does not exist in the delta pattern item of the PST, if the value of the patch number item of the PST is greater than a certain value, the PST delta value is additionally recorded in the delta pattern item of the PST, and the PST delta value is additionally recorded in the delta pattern item of the PST. If the value of the patch number item is less than the specific value, a new block number corresponding to the current access address, a new block address corresponding to the current access address, and a new delta value corresponding to the current access address are recorded in the RHB table. Operation - the specific value is a predetermined ratio value to the length of the delta pattern item of the PST -

If the PST delta value exists in the delta pattern item of the PST, recording the current access address in the ROB table and adding a predetermined value to the value of the patch number item of the PST corresponding to the block number. Including more,

Electronic devices.
According to clause 6,

The at least one operation is,

If the block number does not exist in the block number item of the RHB table and the block number item of the PST, a new block number corresponding to the current access address, a new block address corresponding to the current access address, and the current Further comprising recording a new delta value corresponding to the access address in the RHB table,

Electronic devices.
In a method of operating an electronic device that performs a prefetcher algorithm suitable for a neural network model using a phase change memory,

An operation of checking a first delta value corresponding to the difference between the current access address and the previous access address of the memory of the electronic device and a second delta value that is the previous delta value of the first delta value,

According to the first operation priority for prefetching of the access address, the first delta value and the second delta value are compared, and if the first delta value is equal to the second delta value, the current access address is connected to the second delta value. An operation of performing a first process of prefetching a first prediction address applying a delta value of 1;

As a subordinate operation of the first operation priority, according to the second operation priority, it is searched whether the current access address exists in the ROB (Reorder Buffer) table, and if the current access address exists in the ROB table, in advance An operation of performing a second process of prefetching at least one prediction address recorded in the ROB table by a set number of degrees, and

As a subordinate operation of the second operation priority, according to the third operation priority, when the first delta value and the second delta value are different, the delta score corresponding to the first delta value is OT (Offset Threshold) If it is greater than or equal to the current access address, it includes performing a third process of prefetching a second prediction address by applying the first delta value to the current access address.

How electronic devices work.
According to clause 8,

Further comprising recording the first delta value and the second delta value in an OST (Offset Score Table),

How electronic devices work.
According to clause 9,

When the first delta value and the second delta value are the same, recording the delta score corresponding to the first delta value as the previous delta score in the OST, and

When the first delta value and the second delta value are different, increasing the delta score corresponding to the first delta value by a predetermined score to the previous delta score and recording it in the OST,

How electronic devices work.
According to claim 10,

According to the third operation priority, after performing the third process, an operation of initializing the entries recorded in the OST or decreasing the delta score by a predetermined value from the OT and recording the entries in the OST is further performed. containing,

How electronic devices work.
According to clause 8,

An operation to check whether the block number corresponding to the current access address exists in the block number item of the RHB (Region-History Buffer) table,

If the block number is present in the block number item, recording the RHB delta value corresponding to the current access address in the delta pattern item of the RHB corresponding to the block number - the block address item is It is an item that records the first access address, and the RHB delta value is the difference between the address of the block address item corresponding to the block number and the current access address -, and

After recording the RHB delta value in the delta pattern item, if the length of the delta pattern item corresponding to the block number is more than a predetermined length, the RHB record corresponding to the block number is recorded in a PST (Pattern Sequence Table) , including the operation of deleting the RHB record,

How electronic devices work.
According to claim 12,

If the block number does not exist in the block number item of the RHB table, checking whether the block number corresponding to the current access address exists in the block number item of the PST,

If the block number is present in the block number item of the PST, checking whether the PST delta value corresponding to the current access address is present in the delta pattern item of the PST;

When the PST delta value does not exist in the delta pattern item of the PST, if the value of the patch number item of the PST is greater than a certain value, the PST delta value is additionally recorded in the delta pattern item of the PST, and the PST delta value is additionally recorded in the delta pattern item of the PST. If the value of the patch number item is less than the specific value, a new block number corresponding to the current access address, a new block address corresponding to the current access address, and a new delta value corresponding to the current access address are recorded in the RHB table. operation - the specific value is a predetermined ratio value to the length of the delta pattern item of the PST, and

If the PST delta value exists in the delta pattern item of the PST, recording the current access address in the ROB table and adding a predetermined value to the value of the patch number item of the PST corresponding to the block number. Including more,

How electronic devices work.
According to claim 13,

If the block number does not exist in the block number item of the RHB table and the block number item of the PST, a new block number corresponding to the current access address, a new block address corresponding to the current access address, and the current Further comprising recording a new delta value corresponding to the access address in the RHB table,

How electronic devices work.