Method and chip for integrating negation and timeout retransmission strategies in chiplet interconnection interface
By integrating negation and timeout retransmission strategies in the core-grain interconnect interface and multiplexing hardware resources and state machines, the problems of high protocol compatibility and hardware overhead in the existing technology are solved, and the configurability of flexibility and compatibility is achieved .
Patent Information
- Application Number
- CN202411073144.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-08-06
AI Technical Summary
The retransmission mechanism of the existing core-particle interconnect interface is not effectively compatible with different protocols, resulting in the adapter layer logic redundancy, inability to reuse, and high hardware overhead.
A method of fusing negative and timeout retransmission strategies in the core-grain interconnect interface is adopted. By multiplexing the retransmission buffer bank and state machine, and using the timeout counter independently, the configurability of the two policies is achieved and hardware overhead is reduced.
It realizes the flexibility and compatibility of the chip-particle interconnection interface. Users can choose retransmission strategies according to their needs, reduce hardware overhead, and are compatible with the UCIe protocol.
Smart Images

Figure CN118656316B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of chiplet interconnection interface protocol and circuit technology, and in particular to a method and a chip for integrating negation and timeout retransmission strategies in a chiplet interconnection interface. Background Art
[0002] The chip interconnect interface is the basis for realizing die-to-die (D2D) interconnection. The hierarchical structure of the chip interconnect interface is different from the traditional chip-to-chip interconnection protocol. Its hierarchical structure has been simplified and is usually divided into three levels: protocol layer, adapter layer and physical layer. The protocol layer defines the transmission methods of multiple mainstream protocols on the interconnect interface for the chip-to-chip communication protocol under typical applications. The supported protocols include CXL2.0, PCIe6.0 and AXI. The adapter layer provides reliable link management for both communicating parties, including standard micropacket (Flit) format definition and link reliability transmission. The physical layer usually includes the logical physical layer and the electrical physical layer, providing bit stream transmission functions that conform to the channel characteristics of the package substrate.
[0003] At present, the reliability transmission of chiplets mainly draws on the retransmission mechanism between traditional chips. However, on the one hand, the chiplet interconnection link has new characteristics in transmission reliability due to its short distance and high line density; on the other hand, the interconnection of chiplets of different protocols puts forward higher compatibility requirements for the chiplet interconnection interface. For the strategies adopted by the retransmission mechanism of the chiplet interconnection interface, different chiplet interconnection standard initiators inherited them from their existing chiplet interconnection protocol ecosystems in order to be compatible with their original chiplet interconnection ecosystems. They did not consider compatibility with other chiplet interconnection standard protocols, resulting in the adapter layer retransmission mechanisms of different chiplets not being able to work together and the adapter layer logic being unable to be reused.
[0004] At present, the reliability transmission between chips is mainly based on the retransmission mechanism of the "sliding window", using the negative retransmission strategy (Nak+Ack, referred to as NA strategy) and the timeout retransmission strategy (TimeoutRetry+Ack, referred to as TA strategy). The adapter layer retransmission mechanism in the chip-to-chip interconnection standard is also mainly inherited from the two strategies. The negative retransmission strategy is currently adopted by Intel's PCIe6.0 and CXL2.0 protocols, and is inherited by Intel's newly released UCIe chip interconnection interface standard. The timeout retransmission strategy is used for long-distance reliability transmission of chips in the retransmission mechanism of the Transmission Control Protocol (TCP). It is used in the retransmission mechanism of the consistent interconnection interface between chips of the domestic Feiteng multi-way server CPU (FT-2500), supporting reliability interconnection between up to 16 CPUs. When formulating the domestic chip interconnection standard, different strategies need to be considered to build a prosperous ecosystem for chip interconnection.
[0005] The principle of the timeout retransmission strategy (TA strategy) is as follows: The TA strategy is implemented through a timeout counter and an ACK (Acknowledgment). The sender will add a sequence number to each data packet in the retransmission buffer and maintain a timeout counter. After the sender sends a data packet to the receiver, the timeout counter is started. When the receiver successfully receives the data packet, it will return an ACK message, indicating that the data has been successfully received and cleared from the retransmission buffer. When the timeout counter overflows, the sender does not receive the ACK message sent by the receiver, which triggers a timeout retransmission and resends the data packet for which the corresponding ACK was not received before.
[0006] The principle of negative retransmission strategy (NA strategy) is as follows: NA strategy is implemented through positive ACK and negative ACK (Negative Acknowledgment, NAK). The sender will add a sequence number to each data packet in the retransmission buffer and send the data packet to the sender. When the receiver successfully receives the data packet, it will return an ACK message, indicating that the data has been successfully received and cleared from the retransmission buffer; when the received data packet has a check error, a NAK message will be returned. When the sender receives the NAK message, it will trigger retransmission and resend the data packet in the retransmission buffer.
[0007] When the core adapter layer supports both TA and NA policies, if the two policies are simply implemented completely, they will be configured as one of them according to the negotiation between the interconnected cores in actual use, resulting in logic redundancy in the adapter layer. Comparing the retransmission mechanisms of the TA and NA policies, the retransmission buffer, which occupies the most logical resources, is necessary and the same, and the difference lies in the individual counters and management mechanisms. Therefore, it is technically feasible and meaningful to study the characteristics of the two mechanisms, find out the common content, and realize the integration support of the two mechanisms while reusing most logical resources and increasing the cost of small logical resources. Summary of the invention
[0008] Technical problem to be solved by the present invention: In view of the above-mentioned problems in the prior art, a method and chip for integrating negation and timeout retransmission strategies in a chip-grain interconnection interface are provided. The present invention aims to realize a reliable transmission mode with configurable negation retransmission and timeout retransmission at the chip-grain interconnection interface protocol adapter layer, so that users can choose the retransmission strategy to be used according to their needs, thereby reducing hardware overhead while maintaining flexibility and compatibility.
[0009] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0010] A method for integrating denial and timeout retransmission strategies in a chip interconnection interface comprises the following steps:
[0011] S101, configuring a selection retransmission strategy between a negative retransmission strategy and a timeout retransmission strategy for a sender and a receiver, wherein the negative retransmission strategy and the timeout retransmission strategy have a fused hardware resource design, wherein the fused hardware resource design includes the negative retransmission strategy and the timeout retransmission strategy multiplexing the same set of retransmission buffer storage bodies, and the timeout retransmission strategy independently uses a retransmission timeout counter;
[0012] S102, when the sender sends a retransmittable micro-packet to the receiver, if the micro-packet needs to be retransmitted, the sender and the receiver complete the micro-packet retransmission based on the integrated hardware resource design according to the pre-configured negative retransmission strategy or timeout retransmission strategy.
[0013] Optionally, the hardware resource design integrated in step S101 also includes variables reused by the negative retransmission strategy and the timeout retransmission strategy including WritePointer, ReadPointer, Ack_Info, RetryNum and LinkRetrainNum, the variables independently used by the negative retransmission strategy include AccAck and FreeRB, and the timeout retransmission strategy independently uses the variable AckSentTimeout, wherein WritePointer is the retransmission buffer write pointer, ReadPointer is the retransmission buffer read pointer, Ack_Info is the response information, RetryNum is the retransmission count counter, LinkRetrainNum is the physical layer reinitialization count counter, AccAck is the cumulative response count counter, FreeRB is the retransmission buffer idle count counter, and AckSentTimeout is the positive response timeout counter.
[0014] Optionally, the hardware resource design integrated in step S101 also includes a receiver state machine RSM that multiplexes for both a negative retransmission strategy and a timeout retransmission strategy. The receiver state machine RSM is a receiver state machine that includes at least a normal state NORMAL used in a negative retransmission strategy, and adds a receiver response state R_RSP required by the receiver state machine of a timeout retransmission strategy to send a retransmission management micropacket Retry.Rcvd to the adapter layer management micropacket control module; the normal state NORMAL and the receiver response state R_RSP are used under the timeout retransmission strategy. R_RSP performs receiver status management and enters the normal state NORMAL when requesting link recovery; in the normal state NORMAL, if a retransmittable micropacket is received and the affirmative response timeout counter overflows, the receiver enters the response state R_RSP, indicating that the receiver needs to send a retransmission management micropacket Retry.Rcvd at regular intervals for the received retransmittable micropacket; in the receiver response state R_RSP, if the adapter layer management micropacket control module receives the retransmission management micropacket Retry.Rcvd sent by the receiver, it returns to the normal state NORMAL.
[0015] Optionally, the hardware resource design integrated in step S101 also includes the receiver multiplexing the sender state machine SSM for both the negative retransmission strategy and the timeout retransmission strategy. The sender state machine SSM adds the receiver response state S_REQ, link retraining state LINK_RETRAIN and link termination state END required by the sender state machine of the timeout retransmission strategy on the basis of the sender state machine including at least the normal state NORMAL and the retransmission state RETRY used in the negative retransmission strategy, wherein the receiver response state S_REQ is used to send the retransmission management micropacket Retry.Timeout to the adapter layer management micropacket control module, the link retraining state LINK_RETRAIN is used for the link retraining state when the number of retransmissions exceeds the threshold, and the link termination state END is used for the link retraining state when the number of link retraining exceeds the threshold, resulting in link failure; under the timeout retransmission strategy, the normal state NORMAL, the retransmission state RETRY, the receiver response state S_REQ, the link retraining state LINK_RETRAIN and the link termination state END are used. There are five states in total for sender status management: when reset, it is in the link retraining state LINK_RETRAIN. In the link retraining state LINK_RETRAIN, if the link is successfully recovered and there is no link retraining request, it means that the link is recovered and enters the receiver response state S_REQ; in the receiver response state S_REQ, if the adapter layer management micropacket control module receives the retransmission management micropacket Retry.Timeout sent by the sender, it enters the retransmission state RETRY, if the link termination is requested, it enters the link termination state END; in the retransmission state RETRY, if the retransmission is completed, it enters the normal state NORMAL, if the timeout retransmission counter overflows, it returns to the receiver response state S_REQ; in the normal state NORMAL, if the timeout retransmission counter overflows, it is necessary to re-issue the retransmission management micropacket Retry.Timeout request micropacket, then enter the receiver response state S_REQ, if the link recovery request requires link retraining, then enter the link retraining state LINK_RETRAIN.
[0016] Optionally, the hardware resource design integrated in step S101 also includes the receiver multiplexing a retransmission buffer state machine RBSM for both a negative retransmission strategy and a timeout retransmission strategy. The retransmission buffer state machine RBSM uses a retransmission buffer state machine including an idle state IDLE and a read state READ in the negative retransmission strategy and is reused for the timeout retransmission strategy; under the timeout retransmission strategy, it is in the idle state IDLE when reset. In the idle state IDLE, if the retransmission timeout counter overflows and retransmission is required, the read state READ is entered; in the read state READ, if the retransmission is completed, the idle state IDLE is returned.
[0017] Optionally, the receiving state machine RSM, the sending state machine SSM and the retransmission buffer state machine RBSM all decide whether to use a negative retransmission strategy or a timeout retransmission strategy according to the enable signal TA_enable. If the enable signal TA_enable is 1, the timeout retransmission strategy is used, otherwise the negative retransmission strategy is used.
[0018] In addition, the present invention also provides a multi-chip chip, comprising a plurality of chips interconnected via a chip interconnection interface, wherein the chip is programmed or configured to execute a method for integrating negation and timeout retransmission strategies in the chip interconnection interface.
[0019] In addition, the present invention also provides a computer-readable storage medium, which stores a computer program or instruction, and the computer program or instruction is programmed or configured to execute the method of integrating the denial and timeout retransmission strategy in the chip interconnection interface through a processor.
[0020] In addition, the present invention also provides a computer program product, including a computer program or instruction, which is programmed or configured to execute the method of integrating the denial and timeout retransmission strategies in the chip interconnection interface through a processor.
[0021] Compared with the prior art, the present invention mainly has the following advantages:
[0022] 1. Good flexibility. Compared with the single strategy of realizing reliable transmission in the traditional chip interconnection interface adapter layer, the present invention can realize the configurability of the traditional negative retransmission strategy and timeout retransmission strategy, and users can choose which strategy to use to realize reliable transmission according to their needs.
[0023] 2. Good compatibility. The configurable strategies of the negative retransmission and timeout retransmission strategies implemented by the present invention are compatible with the reliability transmission strategy of the UCIe protocol adapter layer and the adapter layer retransmission management micro-packet.
[0024] 3. Low hardware overhead. The negative retransmission strategy and the timeout retransmission strategy of the present invention have a fused hardware resource design, wherein the fused hardware resource design includes the negative retransmission strategy and the timeout retransmission strategy multiplexing the same set of retransmission buffer storage bodies, and the timeout retransmission strategy independently uses the retransmission timeout counter, thereby lowering the hardware overhead while maintaining flexibility and compatibility. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 Schematic diagram of the basic flow of the method of the embodiment of the present invention.
[0026] Figure 2 Schematic diagram of state changes of a multiplexed receiver state machine RSM in an embodiment of the present invention.
[0027] Figure 3 Schematic diagram of state transition of the multiplexed sender state machine SSM in an embodiment of the present invention.
[0028] Figure 4 It is a schematic diagram of the state transition of the retransmission buffer state machine RBSM multiplexed in an embodiment of the present invention. DETAILED DESCRIPTION
[0029] like Figure 1 As shown, the method for integrating the denial and timeout retransmission strategies in the chiplet interconnection interface of this embodiment includes the following steps:
[0030] S101, configures the sender and the receiver to select a retransmission strategy from the negative retransmission strategy (NA strategy) and the timeout retransmission strategy (TA strategy). The negative retransmission strategy and the timeout retransmission strategy have a fused hardware resource design, integrating the resources used by the two strategies of timeout retransmission and negative retransmission. These resources include: a retransmission buffer storage body reused by the two strategies; a timeout retransmission counter used only by the timeout retransmission strategy; variables that control the correct retransmission of micro-packets that need to be retransmitted, wherein the read pointer, write pointer, retransmission counter, link retraining counter, and response information are resources reused by the two strategies; the retransmission state machine reused by the two strategies is divided into a receiver state machine, a sender state machine, and a retransmission buffer state machine. By reusing the retransmission buffer storage body, the variables controlling retransmission, and the retransmission state machine, the hardware overhead can be reduced;
[0031] S102, when the sender sends a retransmittable micropacket to the receiver, if the micropacket retransmission is required, the sender and the receiver complete the micropacket retransmission based on the fused hardware resource design according to the pre-configured selected negative retransmission strategy or timeout retransmission strategy. When the micropacket retransmission is completed based on the fused hardware resource design according to the pre-configured selected negative retransmission strategy or timeout retransmission strategy, each time the sender sends a retransmittable micropacket, it will be placed in the retransmission buffer at the same time and a micropacket request sequence number (Seq) will be assigned to it. The receiver maintains a variable Rcvd_Seq for recording the expected sequence number of the correctly received retransmittable micropacket. The initial value of Rcvd_Seq is 0, and its maximum value depends on the depth of the retransmission buffer. Whenever the receiver correctly receives a retransmittable micropacket, Rcvd_Seq increases by 1; if the received retransmittable micropacket has an error, Rcvd_Seq remains unchanged. If the sequence number of the received retransmittable micropacket does not match Rcvd_Seq, it is directly discarded. The receiver needs to pass Rcvd_Seq to the sender through the retransmission management micro-packet. At the same time, the sender also maintains a variable Send_Seq to record the latest response sequence number in the received management micro-packet. The sender will update Send_Seq according to the Rcvd_Seq information contained in the received retransmission management micro-packet. The difference between the NA strategy and the TA strategy lies in the different triggering mechanisms of retransmission. The NA strategy is that when an erroneous micro-packet is received, the receiver generates a retransmission request micro-packet to the sender, and the retransmission request micro-packet must include Rcvd_Seq. When the sender receives the retransmission request micro-packet, all micro-packets stored in the retransmission buffer starting from the Rcvd_Seq sequence number are retransmitted. In the TA strategy, each micro-packet corresponds to a retransmission timeout counter. When the sender sends the micro-packet, the corresponding counter is started. When any retransmission timeout counter times out, retransmission will be triggered. The sender generates a retransmission request micro-packet, and the retransmission request micro-packet must include Send_Seq. When retransmitting, all micropackets stored in the retransmission buffer starting from the Send_Seq sequence number need to be retransmitted. Every time the receiver correctly receives a retransmittable micropacket, Rcvd_Seq will increase by 1, and some other information will be generated, collectively referred to as retransmission acknowledgment information. The retransmission acknowledgment information can be transmitted through a special retransmission acknowledgment management micropacket, or through a separate acknowledgment management micropacket. In the NA strategy, the retransmission acknowledgment information is the number of acknowledgments (AckNum) generated by correctly receiving the micropacket. The receiver counts the number of generated acknowledgments and returns a number of acknowledgments to the sender through a micropacket at an appropriate time. RtnAck is used to represent the number of acknowledgments carried in the micropacket. After the sender receives the management micropacket and obtains RtnAck, it sends a retransmission acknowledgment micropacket (Rerty.Ack) to the receiver and removes the corresponding number of micropackets from the retransmission buffer. In the TA strategy, the retransmission acknowledgment information is the acknowledgment sequence number (AckNo).Every time the receiver successfully receives a certain number of micropackets, it will send a retransmission response micropacket (Rerty.Rcvd) to the sender. After the sender receives the management micropacket and obtains the response sequence number, it removes the sequence number and all retransmittable micropackets before the sequence number from the retransmission buffer.
[0032] In this embodiment, based on the comparative analysis of the two strategies, under the premise of reusing the same set of retransmission buffer storage bodies, the common counters, pointers, management micro-packets and other logics under the two strategies are integrated and designed, that is, only one set of common logic is designed, and the logical resources owned by each are designed separately; in particular, for the retransmission management state machine, it is expanded on the basis of the implementation of the NA strategy, by adding some TA strategy states, and configurable design of the transition conditions between the original states, or adding new transition conditions to the newly added states, to achieve configurable support for NA and TA strategies with relatively small hardware overhead.
[0033] The hardware resource design integrated in step S101 of this embodiment also includes the variables reused by the negative retransmission strategy and the timeout retransmission strategy, including WritePointer, ReadPointer, Ack_Info, RetryNum and LinkRetrainNum. The variables independently used by the negative retransmission strategy include AccAck and FreeRB. The timeout retransmission strategy independently uses the variable AckSentTimeout, where WritePointer is the retransmission buffer write pointer, ReadPointer is the retransmission buffer read pointer, Ack_Info is the response information, RetryNum is the retransmission count counter, LinkRetrainNum is the physical layer reinitialization count counter, AccAck is the cumulative response count counter, FreeRB is the retransmission buffer idle count counter, and AckSentTimeout is the positive response timeout counter. It can be seen that under the strategy of integrating NA and TA, the hardware resources in the retransmission mechanism structure are divided into three categories.
[0034] The first category is fully reused and uniformly controlled resources: retransmission buffer memory. It stores micropackets that have been sent but not yet confirmed. The width of the memory is the micropacket width, and the depth is the number of micropackets that can be stored.
[0035] The second category is resources that are reused but configured with different functions according to different policies. Each functional component in this part of resources uses the same hardware, but stores different content or implements different functions under different policies. These resources include: Retransmission buffer write pointer (WritePointer): points to the position where the retransmittable micropacket to be sent is written into the retransmission buffer. The write pointer uses a wrap-around type counter, with the maximum value being the depth of the retransmission buffer, and is reset to 0 after reset. When a retransmission occurs, the retransmission buffer stops receiving new micropackets, the write pointer remains unchanged, and all micropackets in the retransmission buffer are retransmitted. After the retransmission is completed, new micropackets can be written into the retransmission buffer, and the write pointer can continue to increase. Retransmission buffer read pointer (ReadPointer): points to the position where the micropacket needs to be read from the retransmission buffer. The read pointer uses a wrap-around type counter, with the maximum value being the depth of the retransmission buffer, and the reset value is 0. During the retransmission process, the sender increases the read pointer by 1 each time it reads a micropacket from the retransmission buffer. The retransmission ends when all micropackets in the retransmission buffer are sent (when the next position of the read pointer is the write pointer). Response information (Ack_Info): This information comes from the receiver. The receiver verifies the data of the receiver's micropacket, verifies the latest sequence number (AckNo) of the correct micropacket or the number of correctly received micropackets to be returned (AckNum), and transmits it to the sender. This information needs to be implemented as different functions according to the configuration of the policy. When configured as NA policy, the response information is AckNum, and when configured as TA policy, the response information is AckNo. Retransmission counter (RetryNum): records the number of retransmission operations performed by the retransmission buffer (each retransmission retransmits all valid micropackets in the retransmission buffer). When the number of retransmissions exceeds the set maximum number of consecutive retransmissions and no response is received, it is considered that the physical link is faulty and the physical layer needs to be initialized. Physical layer reinitialization counter (LinkRetrainNum): records the number of physical layer reinitializations. When the number of physical layer initializations exceeds the set maximum number of consecutive initializations and the physical link failure is still not resolved, it is considered that the link cannot be restored. Retransmission management state machine: When implementing two configurable retransmission strategies, the state machine is designed to be a sender state machine (SSM), a receiver state machine (RSM), and a retry buffer state machine (RBSM). These three state machines extend the TA strategy while satisfying the NA strategy. This part of the design content will be introduced in detail later.
[0036] The third category is the independent and non-reusable resources used under different strategies. These resources can only be used or not used according to the configuration after the corresponding NA or TA strategy is configured. The NA resources will be idle after the TA strategy is configured, and vice versa. The resources of TA will be idle after the NA strategy is configured. These resources include: Retransmission buffer free number counter (FreeRB): This value indicates how much free space is left in the retransmission buffer, and is set to the depth of the retransmission buffer when reset. When a new micropacket is stored in the buffer, the counter is reduced by 1, and when an old micropacket is correctly received by the receiver, the counter increases the number of correctly received micropackets (RtnAck). When the value of FreeRB is 1, it means that there is only one free position left in the retransmission buffer, and new data micropackets are not allowed to be written at this time. Accumulated acknowledgment number counter (AccAck): Records the number of micropackets correctly received by the receiver and accumulated waiting to be sent. When a retransmittable micropacket is correctly received, AccAck is incremented by 1. When a certain number of responses are returned to the other party through the retransmission management micropacket, the AccAck counter is deducted by the returned value (RtnAck). Retransmission timeout counter (RetryTimeout): Each micropacket stored in the sender's retransmission buffer corresponds to a timeout counter. Initially, all timeout counters stop timing; when the micropacket enters the retransmission buffer, the timeout counter corresponding to the sequence number of the micropacket is started; when the sender receives the response sequence number AckNo, the timeout counters corresponding to AckNo and all previous sequence numbers are cleared to 0; if any timeout counter times out, all timeout counters stop timing and prepare for retransmission; after the retransmission is started, the timeout counter corresponding to the sequence number of the retransmitted micropacket is started, and the count of whether the micropacket times out during the retransmission process is continued. AckSentTimeout counter: The receiver does not return a retransmission management message for each micropacket received, but only returns a retransmission management message when a certain number of correctly received micropackets are accumulated. Each time the receiver sends a retransmission management message, it resets the positive response timeout counter and restarts counting. After the timeout, it immediately sends a retransmission management message (Retry.Rcvd). If the received micropacket is erroneous or out of order, there is no need to wait for the timeout counter to overflow. It immediately sends a retransmission management message (Retry.Req) and resets the positive response timeout counter.
[0037] To ensure reliable transmission, the sender and receiver will exchange retransmission-related information and use retransmission management packets to store this information. In addition to retransmission management packets, the adapter also has link management packets and power management packets. In order to unify the management of these packets with different functions, the adapter layer management packet control module is introduced. All retransmission management packets sent between the sender and the receiver are sent to the adapter layer management packet control module. The module acts as an intermediary and returns the corresponding signal to the sender or receiver to realize the interaction between the sender, the receiver and the retransmission buffer. The integrated hardware resource design in step S101 of this embodiment further includes defining two retransmission management micro-packets Retry.Err and Retry.Ack for the negative retransmission strategy, and defining two retransmission management micro-packets Retry.Timeout and Retry.Rcvd for the timeout retransmission strategy. The fields of the retransmission management micro-packets Retry.Err, Retry.Ack, Retry.Timeout, and Retry.Rcvd include the corresponding retransmission strategy, the adapter layer micro-packet type, the retransmission management micro-packet type, and the filled fields. The filled fields in the retransmission management micro-packet Retry.Err include Rcvd_Seq, Ack_Info, RetryNum, and LinkRetrainNum; the filled fields in the retransmission management micro-packet Retry.Ack include WritePointer, Rcvd_Seq, Ack_Info, RetryNum, and LinkRetrainNum. eq and FreeRB; the fields filled in the retransmission management micropacket Retry.Err include WritePointer, Send_Seq, RetryNum, LinkRetrainNum and FreeRB; the fields filled in the retransmission management micropacket Retry.Err include Send_Seq and Ack_Info; among them, Rcvd_Seq is the expected sequence number maintained by the receiver, Ack_Info is the response information, RetryNum is the retransmission counter, LinkRetrainNum is the physical layer reinitialization counter, WritePointer is the retransmission buffer write pointer, FreeRB is the retransmission buffer idle number counter, and Send_Seq is the sequence number maintained by the sender to record the latest response sequence number AckNo in the received management micropacket. When configured as NA policy, Retry.Err and Retry.Ack are used. Retry.Err is a retransmission management micropacket sent by the receiver when a CRC check error occurs. Its function is to issue a retransmission request. Retry.Ack is a response micropacket sent by the sender when a retransmission request occurs. It also controls the retransmission. When configured as a TA strategy, Retry.Timeout and Retry.Rcvd are used.Retry.Timeout is a micro-packet for requesting retransmission when the timeout counter of a data micro-packet overflows, and controls the retransmission process; Retry.Rcvd is a micro-packet for responding to a certain number of ACKs and containing the sequence number of the received micro-packet. In this embodiment, the retransmission management micro-packets Retry.Err, Retry.Ack, Retry.Timeout, and Retry.Rcvd are specifically shown in Table 1.
[0038] Table 1 Field table of four types of retransmission management micropackets at the adapter layer
[0039]
[0040] The integrated hardware resource design in step S101 of this embodiment also includes a receiver state machine RSM that multiplexes the receiver state machine for both the negative retransmission strategy and the timeout retransmission strategy, a sender state machine SSM that multiplexes the receiver state machine for both the negative retransmission strategy and the timeout retransmission strategy, and a retransmission buffer state machine RBSM that multiplexes the receiver state machine for both the negative retransmission strategy (NA strategy) and the timeout retransmission strategy (TA strategy). The functions of the three state machines are as follows: the function of the receiver state machine RSM is to send a retransmission type management micropacket Retry.Err (NA strategy) or Retry.Rcvd (TA strategy) to the adapter layer control micropacket management module in time when receiving a retransmittable micropacket; maintain a receiver timeout counter to control how long it takes to send a Retry.Rcvd; under the NA strategy, maintain the RetryNum counter, and retrain the link when it overflows, and maintain the LinkRetrainNum counter, and the link fails when it overflows. The function of the sender state machine SSM is to send a retransmission type management micropacket Retry.Ack (NA policy) or Retry.Timeout (TA policy) to the adapter layer control micropacket management module in time when the retransmission condition is triggered; retransmit all micropackets in RB from the updated read pointer; maintain the RetryNum counter under the TA policy, retrain the link when it overflows, and maintain the LinkRetrainNum counter, which fails the link when it overflows. The function of the retransmission buffer state machine (Retry Buffer State Machine, RBSM) is to write RB through the write pointer when there is a retransmittable micropacket; when there is a retransmission request, control the read pointer to point to the micropacket that needs to be retransmitted, and the read pointer points to the micropacket corresponding to Rcvd_Seq (NA policy) or Send_Seq (TA policy); manage the Rcvd_Seq and Send_Seq registers.
[0041] like Figure 2As shown, the receiver state machine RSM is a receiver state machine including the three states of the normal state NORMAL, the receiver retransmission request state R_REQ, and the link retraining state LINK_RETRAIN used in the negative retransmission strategy, and adds the receiver response state R_RSP required by the receiver state machine of the timeout retransmission strategy to send the retransmission management micro-packet Retry.Rcvd to the adapter layer management micro-packet control module; under the timeout retransmission strategy, the normal state NORMAL and the receiver response state R_RSP are used to manage the receiver state, and the normal state NORMAL state is entered when the link is requested to be restored; in the normal state NORMAL state, if a retransmittable micro-packet is received and the positive response timeout counter overflows, the receiver response state R_RSP is entered, indicating that the receiver needs to send a retransmission management micro-packet Retry.Rcvd to the received retransmittable micro-packet at regular intervals; in the receiver response state R_RSP, if the adapter layer management micro-packet control module receives the retransmission management micro-packet Retry.Rcvd sent by the receiver, the normal state NORMAL is returned. Figure 2 In the figure, the state machine used by the negative retransmission strategy (NA strategy) is in the box, which should include the normal state NORMAL, the receiver retransmission request state R_REQ, the link retraining state LINK_RETRAIN, etc. Its functions are as follows: in the normal state NORMAL, the micropacket is received normally. When a micropacket with a check error appears, the receiver enters the retransmission request state R_REQ state to request retransmission. If the number of retransmissions exceeds the threshold, the link retraining state LINK_RETRAIN is entered; when the TA strategy is integrated, NORMAL is the normal working state machine of the receiver, which is used in both the NA strategy and the TA strategy. The receiver response state R_RSP is added to send the Retry.Rcvd management micropacket to the adapter layer management micropacket control module. The relevant signals and state transitions of the TA strategy state machine are as follows: ①: rcvd_timeout&&retryable_flit, indicating that a retransmittable micropacket is received and the positive response timeout counter overflows. ②: retry_rcvd_sent, indicating that the adapter layer management micropacket control module receives the Retry.Rcvd sent by the receiver. ③: req_recovery, indicating requesting link recovery. It is in the NORMAL state when reset; if the req_recovery signal is set to 1, the state machine enters the NORMAL state; when rcvd_timeout is set to 1 and retryable_flit is set to 1, it enters the R_RSP state, indicating that the receiver of the received retransmittable micro-packet needs to send Retry.Rcvd at regular intervals; when retry_rcvd_sent is set to 1, it enters the NORMAL state, indicating that the adapter layer management micro-packet control module has received the Retry.Rcvd response micro-packet. It should be noted that, refer to Figure 2 , the receiver state machine used by the negative retransmission strategy (NA strategy) in the box only needs the normal state NORMAL to realize the multiplexing of the receiver state machine used by the TA strategy, so Figure 2 The receiving state machine used by the negative retransmission strategy (NA strategy) in the box shown is only an example in this embodiment. In addition, other receiving state machines used by negative retransmission strategies (NA strategies) including the normal state NORMAL can also be used, and can also be multiplexed with the receiving state machine used by the timeout retransmission strategy (TA strategy).
[0042] like Figure 3 As shown, the sender state machine SSM adds the receiver response state S_REQ, link retraining state LINK_RETRAIN and link termination state END required by the sender state machine of the timeout retransmission strategy on the basis of the sender state machine including the normal state NORMAL, the sender retransmission response state S_RSP and the retransmission state RETRY used in the negative retransmission strategy, wherein the receiver response state S_REQ is used to send the retransmission management micropacket Retry.Timeout to the adapter layer management micropacket control module, the link retraining state LINK_RETRAIN is used for the link retraining state when the number of retransmissions exceeds the threshold, and the link termination state END is used for the link retraining state when the number of link retraining exceeds the threshold and causes the link to fail; under the timeout retransmission strategy, the normal state NORMAL, the retransmission state RETRY, the receiver response state S_REQ, the link retraining state LINK_RETRAIN and the link termination state END are used to perform the sender state Management: When reset, it is in the link retraining state LINK_RETRAIN. In the link retraining state LINK_RETRAIN, if the link is successfully recovered and there is no link retraining request, it means that the link is recovered and enters the receiver response state S_REQ; in the receiver response state S_REQ, if the adapter layer management micropacket control module receives the retransmission management micropacket Retry.Timeout sent by the sender, it enters the retransmission state RETRY, if the link termination is requested, it enters the link termination state END; in the retransmission state RETRY, if the retransmission is completed, it enters the normal state NORMAL, if the timeout retransmission counter overflows, it returns to the receiver response state S_REQ; in the normal state NORMAL, if the timeout retransmission counter overflows, it is necessary to re-issue the retransmission management micropacket Retry.Timeout request micropacket, then enter the receiver response state S_REQ, if the link recovery request requires link retraining, then enter the link retraining state LINK_RETRAIN. Figure 3The box is the state machine used by the negative retransmission strategy (NA strategy), which should include the normal state NORMAL, the sender retransmission response state S_RSP, the retransmission state RETRY, etc. Its functions are as follows: in the NORMAL state, there is no need to retransmit the micropacket. When receiving the retransmission request micropacket, it enters the retransmission response state, and when sending the retransmission response micropacket, it enters the retransmission state. When the TA strategy is integrated, the NORMAL and RETRY states are used in both the NA strategy and the TA strategy. The receiver response state S_REQ is added to send the Retry.Timeout management micropacket to the adapter layer management micropacket control module; the link retraining state LINK_RETRAIN is added to be used for link retraining when the number of retransmissions exceeds the threshold; the link termination state END is added to be used for link failure when the number of link retraining exceeds the threshold. The relevant signals and state transitions of the TA strategy state machine are as follows: ①: req_recovery, indicating a request for link recovery. ②: rsp_active&&!req_retrain, indicating that the link recovery is successful and there is no link retraining request. ③: req_retrain, indicating link retraining request. ④: ta_timeout. Indicates that the timeout retransmission counter of the TA policy has overflowed. ⑤: retry_timeout_sent, indicating that the adapter layer management micro-packet control module has received the Retry.Timeout sent by the sender. ⑥: retry_done, indicating that the retransmission is completed. ⑦: ta_timeout, indicating that the timeout retransmission counter of the TA policy has overflowed. ⑧: req_end, indicating the request for link termination. When reset, it is in the LINK_RETRAIN state; if rsp_active is set to 1 and req_retrain is 0, it indicates that the link is restored and enters the S_REQ state; if retry_timeout_sent is set to 1, it indicates that the adapter layer management packet control module has received the Retry.Timeout request packet and enters the RETRY state; if retry_done is set to 1, it indicates that the retransmission is completed and enters the NORMAL state; if ta_timeout is set to 1 in the RETRY state, it indicates that the timeout retransmission counter overflows again in this retransmission, and the Retry.Timeout request packet needs to be reissued to enter the S_REQ state; if ta_timeout is set to 1 in the NORMAL state, it indicates that the timeout retransmission counter overflows, and the Retry.Timeout request packet needs to be issued to enter the S_REQ state; if req_retrain is set to 1, it indicates that link retraining is required and enters the LINK_RETRAIN state; if req_end is set to 1, it indicates that the link fails and enters the termination state. It should be noted that, refer to Figure 3, the sender state machine used by the negative retransmission strategy (NA strategy) in the box only needs the normal state NORMAL and the retransmission state RETRY to realize the reuse of the sender state machine used by the timeout retransmission strategy (TA strategy). Figure 3 The sender state machine used by the negative retransmission strategy (NA strategy) in the box shown is only an example in this embodiment. In addition, other sender state machines used by NA strategies including the normal state NORMAL and the retransmission state RETRY can also be used, and can also be reused with the sender state machine used by the timeout retransmission strategy (TA strategy).
[0043] like Figure 4 As shown in the figure, the retransmission buffer state machine RBSM, which includes the idle state IDLE and the read state READ used in the negative retransmission strategy, is reused in the timeout retransmission strategy; in the timeout retransmission strategy, it is in the idle state IDLE when reset. In the idle state IDLE, if the retransmission timeout counter overflows and retransmission is required, it enters the read state READ; in the read state READ, if the retransmission is completed, it returns to the idle state IDLE. In the retransmission buffer state machine RBSM, the NA strategy and the TA strategy reuse all state machines, but the signals of state machine transition are different. Both the negative retransmission strategy (NA strategy) and the timeout retransmission strategy (TA strategy) use these two state machines. In the idle state IDLE, there is no retransmission request, only the write pointer increases; in the read state READ, a retransmission request appears, the retransmission buffer is stopped from being written, and retransmission starts from the position of the read pointer. The relevant signals and state transitions of the TA strategy state machine are as follows: ①: retry_done, indicating that the retransmission is completed. ②: na_err || ta_timeout, indicating that retransmission is triggered under NA policy or request is triggered under TA policy. It is in IDLE state when reset; in NA policy, if na_err is set to 1, it indicates that a check error occurs and retransmission is required, and the state enters READ; in TA policy, if ta_timeout is set to 1, it indicates that retransmission timeout counter overflows and retransmission is required, and the state enters READ; if retry_done is set to 1, it indicates that retransmission is completed and the state enters IDLE.
[0044] In this embodiment, the receiving state machine RSM, the sending state machine SSM and the retransmission buffer state machine RBSM all decide whether to use a negative retransmission strategy or a timeout retransmission strategy according to the enable signal TA_enable. If the enable signal TA_enable is 1, the timeout retransmission strategy is used, otherwise the negative retransmission strategy is used.
[0045] In summary, the method for integrating the negation and timeout retransmission strategies in the core-grain interconnection interface of this embodiment is that the sender and the receiver configure and select the retransmission strategy from the negation retransmission strategy (NA strategy) and the timeout retransmission strategy (TA strategy). The negation retransmission strategy and the timeout retransmission strategy have a fused hardware resource design. When the sender sends a retransmittable micro-packet to the receiver, if the micro-packet retransmission is required, the sender and the receiver complete the micro-packet retransmission based on the fused hardware resource design according to the pre-configured selected negation retransmission strategy or the timeout retransmission strategy. The fused hardware resource design has the following advantages: (1) Compared with the single strategy of the traditional core-grain interconnection interface adapter layer to realize reliable transmission, this embodiment can realize the configurability of the traditional negation retransmission strategy and the timeout retransmission strategy, and the user can choose which strategy to use to realize reliable transmission according to their own needs. (2) The configurable strategy of the negation retransmission and timeout retransmission strategies realized in this embodiment is compatible with the reliability transmission strategy of the UCIe protocol adapter layer and the adapter layer retransmission management micro-packet. (3) In this embodiment, the negation retransmission strategy and the timeout retransmission strategy have a fused hardware resource design. The fused hardware resource design includes the negation retransmission strategy and the timeout retransmission strategy reusing the same set of retransmission buffer storage bodies, and the timeout retransmission strategy independently using the retransmission timeout counter, thereby reducing the hardware overhead while maintaining flexibility and compatibility.
[0046] In addition, this embodiment also provides a multi-chiplet chip, comprising a plurality of chiplets interconnected via a chiplet interconnection interface, wherein the chiplets are programmed or configured to execute a method for integrating a negation and a timeout retransmission strategy in the chiplet interconnection interface.
[0047] In addition, this embodiment also provides a computer-readable storage medium, which stores a computer program or instruction, and the computer program or instruction is programmed or configured to execute the method of integrating the denial and timeout retransmission strategy in the chip interconnection interface through a processor.
[0048] In addition, this embodiment also provides a computer program product, including a computer program or instructions, which are programmed or configured to execute the method of integrating the denial and timeout retransmission strategies in the chip interconnection interface through a processor.
[0049] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.
Claims
1. A method for integrating negation and timeout retransmission strategies in a chiplet interconnection interface, characterized in that: The steps include: S101, configuring a selection retransmission strategy between a negative retransmission strategy and a timeout retransmission strategy for a sender and a receiver, wherein the negative retransmission strategy and the timeout retransmission strategy have a fused hardware resource design, wherein the fused hardware resource design includes the negative retransmission strategy and the timeout retransmission strategy multiplexing the same set of retransmission buffer storage bodies, and the timeout retransmission strategy independently uses a retransmission timeout counter; S102, when the sender sends a retransmittable micro-packet to the receiver, if the micro-packet needs to be retransmitted, the sender and the receiver complete the micro-packet retransmission based on the integrated hardware resource design according to the pre-configured negative retransmission strategy or timeout retransmission strategy; The integrated hardware resource design in step S101 also includes a receiver state machine RSM multiplexing for both a negative retransmission strategy and a timeout retransmission strategy, wherein the receiver state machine RSM is a receiver state machine including at least a normal state NORMAL used by the negative retransmission strategy and adds a receiver response state R_RSP required by the receiver state machine of the timeout retransmission strategy for sending a retransmission management micro-packet Retry.Rcvd to the adapter layer management micro-packet control module; Under the timeout retransmission strategy, the normal state NORMAL and the receiver response state R_RSP are used to manage the receiver state. When the link recovery is requested, the normal state NORMAL is entered; in the normal state NORMAL, if a retransmittable micropacket is received and the affirmative response timeout counter overflows, the receiver response state R_RSP is entered, indicating that the receiver needs to send a retransmission management micropacket Retry.Rcvd at regular intervals for the received retransmittable micropacket; in the receiver response state R_RSP, if the adapter layer management micropacket control module receives the retransmission management micropacket Retry.Rcvd sent by the receiver, it returns to the normal state NORMAL.
2. The method for integrating negation and timeout retransmission strategies in a chiplet interconnection interface according to claim 1, characterized in that: The hardware resource design integrated in step S101 also includes the variables reused by the negative retransmission strategy and the timeout retransmission strategy, including WritePointer, ReadPointer, Ack_Info, RetryNum and LinkRetrainNum. The variables independently used by the negative retransmission strategy include AccAck and FreeRB. The timeout retransmission strategy independently uses the variable AckSentTimeout, wherein WritePointer is the retransmission buffer write pointer, ReadPointer is the retransmission buffer read pointer, Ack_Info is the response information, RetryNum is the retransmission count counter, LinkRetrainNum is the physical layer reinitialization count counter, AccAck is the cumulative response count counter, FreeRB is the retransmission buffer idle count counter, and AckSentTimeout is the positive response timeout counter.
3. The method for integrating negation and timeout retransmission strategies in a chiplet interconnection interface according to claim 2, characterized in that: The hardware resource design integrated in step S101 also includes that the receiver multiplexes the sender state machine SSM for both the negative retransmission strategy and the timeout retransmission strategy. The sender state machine SSM adds the receiver response state S_REQ, link retraining state LINK_RETRAIN and link termination state END required by the sender state machine of the timeout retransmission strategy on the basis of the sender state machine including at least the normal state NORMAL and the retransmission state RETRY used in the negative retransmission strategy, wherein the receiver response state S_REQ is used to send the retransmission management micropacket Retry.Timeout to the adapter layer management micropacket control module, the link retraining state LINK_RETRAIN is used for the link retraining state when the number of retransmissions exceeds the threshold, and the link termination state END is used for the link retraining state when the number of link retraining exceeds the threshold and causes the link to fail; under the timeout retransmission strategy, the normal state NORMAL, the retransmission state RETRY, the receiver response state S_REQ, the link retraining state LINK_RETRAIN and the link termination state END are used. There are five states in total for sender status management: when reset, it is in the link retraining state LINK_RETRAIN. In the link retraining state LINK_RETRAIN, if the link is recovered successfully and there is no link retraining request, it means that the link is recovered and enters the receiver response state S_REQ; in the receiver response state S_REQ, if the adapter layer management micropacket control module receives the retransmission management micropacket Retry.Timeout sent by the sender, it enters the retransmission state RETRY, if the link termination is requested, it enters the link termination state END; in the retransmission state RETRY, if the retransmission is completed, it enters the normal state NORMAL, if the timeout retransmission counter overflows, it returns to the receiver response state S_REQ; in the normal state NORMAL, if the timeout retransmission counter overflows, it is necessary to re-issue the retransmission management micropacket Retry.Timeout request micropacket, then enter the receiver response state S_REQ, if the link recovery request requires link retraining, then enter the link retraining state LINK_RETRAIN.
4. The method for integrating negation and timeout retransmission strategies in a chiplet interconnection interface according to claim 3, characterized in that: The integrated hardware resource design in step S101 also includes the receiver multiplexing a retransmission buffer state machine RBSM for both a negative retransmission strategy and a timeout retransmission strategy, wherein the retransmission buffer state machine RBSM includes a retransmission buffer state machine in an idle state IDLE and a read state READ used in the negative retransmission strategy and is reused in the timeout retransmission strategy; Under the timeout retransmission strategy, it is in the idle state IDLE when reset. In the idle state IDLE, if the retransmission timeout counter overflows and needs to be retransmitted, it enters the read state READ; In the read state READ, if the retransmission is completed, the idle state IDLE is returned.
5. The method for integrating negation and timeout retransmission strategies in a chiplet interconnection interface according to claim 4, characterized in that: The receiver state machine RSM, the sender state machine SSM and the retransmission buffer state machine RBSM all decide to use a negative retransmission strategy or a timeout retransmission strategy according to the enable signal TA_enable. If the enable signal TA_enable is 1, the timeout retransmission strategy is used, otherwise the negative retransmission strategy is used.
6. A multi-chip chip, comprising a plurality of chips interconnected by a chip interconnection interface, characterized in that: The chiplet is programmed or configured to execute the method for integrating the negation and timeout retransmission strategies in the chiplet interconnection interface according to any one of claims 1 to 5.
7. A computer-readable storage medium having a computer program or instruction stored therein, characterized in that: The computer program or instruction is programmed or configured to execute the method for integrating the denial and timeout retransmission strategies in the chip interconnection interface described in any one of claims 1 to 5 through a processor.
8. A computer program product comprising a computer program or instructions, characterized in that The computer program or instruction is programmed or configured to execute the method for integrating the denial and timeout retransmission strategies in the chip interconnection interface described in any one of claims 1 to 5 through a processor.
