Cross-lane data processing method, target processor, system on chip and computing device
By parsing the control bits of channel instructions and integrating cross-channel data transmission and reception logic with the depth of instruction execution pipeline, the problem of insufficient scalability in the hardware design of cross-channel instruction execution is solved, achieving a higher physical implementation ceiling and lower hardware complexity, and supporting more complex data processing.
Patent Information
- Application Number
- PCT/CN2026/087836
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-12
- Filing Date
- 2026-03-31
- Publication Date
- 2026-11-19
AI Technical Summary
Existing processors require an additional cross-channel data dependency tracking mechanism when executing instructions across channels, resulting in increased instruction overhead and synchronization latency, insufficient hardware design scalability, and a low physical implementation ceiling.
By parsing the control bits of channel instructions and integrating cross-channel data transmission and reception logic with the depth of the instruction execution pipeline, automatic routing and targeted transmission of cross-channel data can be achieved, avoiding the additional overhead and latency introduced by cross-channel dependency tracking mechanisms and reducing hardware design complexity.
It improves the hardware design scalability of the target processor, raises the upper limit of physical implementation, simplifies hardware design, reduces power consumption, and supports more complex data processing tasks.
Smart Images

Figure CN2026087836_19112026_PF_FP_ABST
Abstract
Description
Cross-channel data processing methods, target processors, on-chip systems and computing devices
[0001] This disclosure claims priority to Chinese Patent Application No. 202510610567.9, filed with the China Patent Office on May 12, 2025, entitled “Cross-channel data processing method, target processor, system-on-a-chip and computing device”, the entire contents of which are incorporated herein by reference. Technical Field
[0002] This disclosure relates to the technical field of hardware development, and in particular to a cross-channel data processing method, a target processor, a system-on-a-chip, and a computing device. Background Technology
[0003] With the development of computing technology, the demands on processors for large-scale, high-speed computing capabilities are gradually increasing. For example, the widespread application of artificial intelligence and related technologies has led to the continuous development of vector computing technology, and the application's demand for processor computing power is increasing day by day.
[0004] Currently, the data processing bit width in processors is constantly increasing. For example, the data bit width of vector processing units has increased from 128, 256, and 512 bits to 1024, 2048, and even 4096 bits. Multi-channel designs offer better scalability compared to single-channel designs, supporting larger processing bit widths. Multi-channel architectures break down instruction execution into multiple sets of microinstruction sequences distributed to multiple channels. Each channel executes a set of microinstructions independently, maintaining independent data dependencies between microinstructions.
[0005] However, for instructions that require cross-channel execution, such as ablation and expansion instructions, where the source operands come from other channels or the destination operands need to be written back to other channels, an additional cross-channel data dependency tracking mechanism needs to be built and maintained. This leads to increased instruction overhead and synchronization latency. The dependency tracking hardware mechanism grows with the number of channels, resulting in insufficient scalability of the target processor's hardware design and consequently a lower physical implementation ceiling. Therefore, a highly scalable cross-channel data processing method with a high physical ceiling is urgently needed. Summary of the Invention
[0006] In view of this, embodiments of this disclosure provide a cross-channel data processing method. One or more embodiments of this disclosure simultaneously relate to a target processor, a system-on-a-chip, and a computing device, in order to address the technical deficiencies existing in the prior art.
[0007] According to a first aspect of the present disclosure, a cross-channel data processing method is provided, applied to a data execution engine of a target processor including multiple channels, comprising:
[0008] Parse the current command of the first channel and determine the control bits corresponding to the current command. The control bits include cross-channel receive valid bit, command valid bit, cross-channel transmit valid bit, source channel identifier bit, and destination channel identifier bit.
[0009] When the cross-channel receive valid bit is active, the source operand from the source channel corresponding to the source channel identifier bit is received and the register write-back is completed;
[0010] When the instruction validity bit is active, the current instruction is executed based on the source operand to obtain the destination operand;
[0011] When the cross-channel transmit valid bit is active, the destination operand is sent to the destination channel corresponding to the destination channel identifier bit, and the destination channel completes the register write-back.
[0012] According to a second aspect of the present disclosure, a target processor is provided, including a data execution engine with multiple channels, wherein a first channel is any one of the multiple channels;
[0013] The data execution engine is used to parse the current instruction of the first channel, determine the control bits corresponding to the current instruction, and when the cross-channel receive valid bit is active, receive the source operand from the source channel corresponding to the source channel identifier bit and complete the register write-back. When the instruction valid bit is active, execute the current instruction based on the source operand to obtain the destination operand. When the cross-channel transmit valid bit is active, send the destination operand to the destination channel corresponding to the destination channel identifier bit, and the destination channel completes the register write-back. The control bits include the cross-channel receive valid bit, instruction valid bit, cross-channel transmit valid bit, source channel identifier bit, and destination channel identifier bit.
[0014] According to a third aspect of the present disclosure, a system-on-a-chip is provided, including a plurality of on-chip components, wherein the plurality of on-chip components include a target processor;
[0015] The target processor is used to execute computer programs / instructions that, when executed by the target processor, implement the steps of the aforementioned cross-channel data processing method.
[0016] According to a fourth aspect of the present disclosure, a computing device is provided, comprising:
[0017] Memory and target processor;
[0018] The memory is used to store computer programs / instructions, and the target processor is used to execute the computer programs / instructions, which, when executed by the target processor, implement the steps of the above-described cross-channel data processing method.
[0019] According to a fifth aspect of the present disclosure, a computer-readable storage medium is provided that stores a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.
[0020] According to a sixth aspect of the present disclosure, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.
[0021] In one embodiment of this disclosure, a cross-channel data processing method is provided. By parsing the current instruction of the first channel, the control bits corresponding to the current instruction are determined. These control bits include a cross-channel receive valid bit, an instruction valid bit, a cross-channel transmit valid bit, a source channel identifier bit, and a destination channel identifier bit. The control bits embedded in the current instruction deeply integrate the cross-channel data transmission and reception logic with the instruction execution pipeline, avoiding the additional instruction overhead and increased synchronization latency introduced by the cross-channel data dependency tracking mechanism. When the cross-channel receive valid bit is active, the source operand from the source channel corresponding to the source channel identifier bit is received and register write-back is completed. Automatic routing of cross-channel data reception is achieved through the source channel identifier bit, directly locating the data source channel for targeted data transmission, reducing the hardware design complexity of the target processor. When the instruction valid bit is active, the current instruction is executed based on the source operand to obtain the destination operand. The instruction valid bit ensures that instruction execution is triggered after data readiness. When the cross-channel transmit validity bit is active, the destination operand is sent to the destination channel corresponding to the destination channel identifier bit. The destination channel then performs register write-back. Automatic routing of cross-channel data transmission is achieved through the destination channel identifier bit, directly locating the data destination channel for targeted data transmission. This reduces cross-channel transmission and lowers the hardware design complexity of the target processor. This cross-channel data processing method avoids direct register read / write operations across channels, and dependency maintenance only needs to be completed within the current channel. This enhances the scalability of the target processor hardware design, thereby increasing the physical implementation ceiling of the target processor. Attached Figure Description
[0022] Figure 1 is a flowchart of a cross-channel data processing method provided in an embodiment of this disclosure;
[0023] Figure 2 is a schematic diagram of the architecture of abbreviated instructions in a target processor for a cross-channel data processing method provided in an embodiment of this disclosure;
[0024] Figure 3 is one of the schematic diagrams of the architecture of the extended instruction in the target processor of a cross-channel data processing method provided in an embodiment of the present disclosure;
[0025] Figure 4 is a second schematic diagram of the architecture of the extended instruction in the target processor for a cross-channel data processing method provided in an embodiment of this disclosure;
[0026] Figure 5 is a schematic diagram of the structure of a target processor provided in an embodiment of this disclosure;
[0027] Figure 6 is a schematic diagram of the structure of a system-on-a-chip provided in an embodiment of this disclosure;
[0028] Figure 7 is a structural block diagram of a computing device provided in an embodiment of this disclosure. Detailed Implementation
[0029] Numerous specific details are set forth in the following description to provide a full understanding of this disclosure. However, this disclosure can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this disclosure. Therefore, this disclosure is not limited to the specific implementations disclosed below.
[0030] The terminology used in one or more embodiments of this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this disclosure. The singular forms “a,” “the,” and “the” as used in one or more embodiments of this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this disclosure refers to and includes any or all possible combinations of one or more associated listed items.
[0031] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this disclosure, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this disclosure, and similarly, second may also be referred to as first. Depending on the context, the word “if” as used herein may be interpreted as “when”, “in response to a determination”, or “when…”.
[0032] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this disclosure are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.
[0033] First, the terms and concepts involved in one or more embodiments of this disclosure will be explained.
[0034] Reduced Instruction Set Computing (RISC): A processor architecture design philosophy that emphasizes using a smaller number of simple instructions to build a processor in order to improve the processor's execution efficiency and speed.
[0035] Reduced Instruction Set Computing (RISC) processors are processors designed based on the RISC concept. They are characterized by a shorter instruction pipeline, a fixed instruction format, and a simple addressing mode, enabling high-speed processing.
[0036] Complex Instruction Set Computing (CISC): A processor architecture design philosophy that emphasizes the use of a large number of complex instructions to build the processor in order to improve the processor's functionality and flexibility.
[0037] Complex Instruction Set Computing (CISC) processors: processors designed based on the CISC concept, characterized by having a large number of complex instructions, each of which can perform multiple operations.
[0038] A data lane is a collection of hardware logic units within a processor used for data transmission and manipulation, including but not limited to processing units, register files, and other key components. These units work together to execute the data processing flow defined by instructions. Each data lane functions as an independent working unit, supporting a specific type of data processing task while ensuring efficient data flow throughout the processor, thus enhancing overall computing performance.
[0039] Cross-lane instructions are microinstructions whose operands originate from other data channels or whose execution results need to be written to different channels. These instructions require an efficient internal mechanism within the processor to exchange data between channels, ensuring that information is accurately transferred from one LANE to another. This approach expands the processor's functionality and flexibility, allowing complex computational tasks to be decomposed and processed in parallel across multiple LANEs.
[0040] Microinstructions are low-level instructions used to control the specific operations of functional units within the processor. Microinstructions directly correspond to execution details at the hardware level, such as signal activation and data path configuration, and are the foundation for implementing higher-level instruction functions. They constitute the underlying support structure for the processor to perform any complex operation, ensuring the precise execution of instructions.
[0041] Instruction dependency: In an instruction sequence, the result of a previous instruction affects the correct execution of subsequent instructions. This includes read-after-write (RAW), write-after-write (WAW), and write-after-read (WAR) dependencies.
[0042] Read-write dependency: In an instruction stream (instruction sequence), a situation arises where a certain instruction needs to read data that was written by a previous instruction. This dependency requires the processor to wait for the previous instruction to complete its write operation before it can read the data; otherwise, data errors will occur. Identifying and properly handling this type of dependency is one of the key steps in ensuring program correctness.
[0043] Write-write dependency: In an instruction stream (instruction sequence), two or more instructions attempt to write data to the same location. To prevent data overwriting errors, the correct write order must be determined or measures must be taken to avoid conflicts. Properly scheduling write operations is crucial for maintaining data consistency and program stability.
[0044] Write-read dependency: In an instruction stream (instruction sequence), a value that one instruction expects to write is read prematurely by another subsequent instruction. This situation can lead to unexpected behavior or erroneous results. Therefore, special attention must be paid to the detection and handling of such dependencies during processor design to ensure the correctness of data flow and the reliability of program execution.
[0045] Pseudo-instructions are machine instructions that are not actually executed, but rather special markers used by the compiler to guide the code generation process. They are primarily used to simplify programming models, provide high-level abstractions, or assist in debugging. After compilation, pseudo-instructions are converted into a series of specific machine instructions or other forms of executable code to meet specific application requirements.
[0046] Vector register length (VLEN): The width of data a vector register can hold, reflecting the processor's ability to process vector data. A larger VLEN value means the processor can process a larger amount of data at the same time, making it suitable for large-scale data parallel processing tasks in high-performance computing scenarios.
[0047] The vector datapath length (DPLEN) of a vector processing unit is the amount of data processed by the unit each time, directly affecting the speed and efficiency of vector operations. It determines how the processing unit segments and processes the input data, as well as the format of the final output, and is one of the important indicators for measuring the performance of a vector processor.
[0048] Number of Data Channels in a Vector Processor: The vector processor employs a multi-channel design, where LANE_NUM represents the number of independent data processing channels the vector processor possesses. Each channel is equipped with an execution unit of a corresponding proportion of vector register bit width and data bit width, supporting parallel processing of different data segments. By increasing LANE_NUM, the overall throughput of the processor can be significantly improved while maintaining high efficiency, adapting to diverse computing needs.
[0049] Abbreviation instructions: Instructions with a large source operand and a small destination operand.
[0050] Extended instruction: An instruction with a small source operand width and a large destination operand width.
[0051] System-on-a-Chip (SoC): A design philosophy and implementation method for integrated circuits that integrates all components of a computer or other electronic system onto a single chip. In addition to a central processing unit (CPU), an SoC may also include a graphics processing unit (GPU), memory controller, communication interfaces (such as USB, PCIe, etc.), input / output interfaces, and various hardware accelerators and dedicated circuit modules, aiming to provide a comprehensive set of functions to meet the needs of specific applications. Through highly integrated design, SoCs can significantly reduce the physical size of the system, lower power consumption, and improve data processing speed and efficiency, making them suitable for a variety of applications from mobile devices to servers. Furthermore, by optimizing the interconnection between components, it reduces latency issues common in traditional multi-chip solutions, further enhancing the overall performance of the system. This design philosophy has facilitated the miniaturization, functional diversification, and performance improvement of consumer electronics products.
[0052] This disclosure provides a cross-channel data processing method, and also relates to a target processor, a system-on-a-chip, and a computing device, which will be described in detail in the following embodiments.
[0053] Referring to Figure 1, which shows a flowchart of a cross-channel data processing method provided in an embodiment of this disclosure, applied to a target processor of a multi-channel data execution engine, the method includes the following specific steps:
[0054] Step 102: Parse the current instruction of the first channel and determine the control bits corresponding to the current instruction. The control bits include the cross-channel receive valid bit, instruction valid bit, cross-channel transmit valid bit, source channel identifier bit, and destination channel identifier bit.
[0055] The target processor is a multi-channel architecture processor, which includes multiple independently operating channels with collaborative data processing capabilities, a unified instruction dispatch unit, a shared instruction cache, and a data execution engine with a distributed register file.
[0056] The target processor can be implemented using either the RISC instruction set or the CISC instruction set; no specific restriction is imposed here.
[0057] The target processor breaks down instruction execution into multiple parallel microinstruction sequences, which are then distributed to different channels for independent execution. For example, a 1024-bit RISC-V vector processor may employ an 8-channel parallel design, with each channel processing 128 bits of data. LANE_NUM is 8, VLEN is 1024, and DPLEN is 128.
[0058] Multiple channels are multiple independent data processing units physically divided within the processor. Each channel has an independent processing unit to complete instruction execution pipeline processing and an independent register file to complete data storage. Each channel receives instruction fragments through a unified instruction dispatch unit and performs cross-channel data transmission during execution via a dedicated data bus. The first channel can be any one of the multiple channels. For example, in eight parallel channels Lane_0, Lane_1, ..., Lane_7, Lane_0 can be considered the first channel.
[0059] The data execution engine is the core functional module in the target processor used for instruction scheduling, execution, and result write-back. It can be a scalar execution engine or a vector execution engine. The vector execution engine supports Single Instruction Multiple Data (SIMD) parallel processing and includes multiple parallel computing units and cross-channel data exchange interfaces.
[0060] The current instruction is the micro-instruction for cross-channel data transmission currently being executed on the first channel. This instruction includes a complete opcode field, operand field, and extended control field, where the extended control field contains the control bits required for cross-channel data transmission. The current instruction can be a scalar instruction or a vector instruction; there is no limitation on this. For example, the current instruction is a reduction instruction.
[0061] The control bits corresponding to the current instruction are pseudo-instruction control fields embedded in the microinstruction, including cross-channel receive valid bit, instruction valid bit, cross-channel transmit valid bit, source channel identifier bit, and destination channel identifier bit.
[0062] The cross-channel receive valid bit is a pseudo-instruction hardware signal bit that controls whether the current instruction activates the cross-channel data receive function. For example, receive_vld = 1 indicates activation, and receive_vld = 0 indicates deactivation.
[0063] The instruction validity bit is a microinstruction hardware signal bit that controls whether the current instruction activates the cross-channel data transfer function. For example, inst_vld = 1 indicates activation, and inst_vld = 0 indicates deactivation.
[0064] The cross-channel transmit validity bit is a pseudo-instruction hardware signal bit that controls whether the current instruction activates the cross-channel data transmit function. send_vld=1 indicates activation, and send_vld=0 indicates deactivation.
[0065] The source channel identifier is a pseudo-instruction hardware identifier that identifies the source operand sent from a specific channel among multiple channels. It can be represented using one-hot encoding, with a bit width equal to the total number of channels, or other binary encodings; no limitation is made here. For example, 8'b0000_0001 indicates Lane_0 as the source channel, 8'b0000_0010 indicates Lane_1 as the source channel, and 8'b0000_0111 indicates Lane_0, Lane_1, and Lane_2 as source channels.
[0066] The destination channel identifier bit identifies the pseudo-instruction hardware identifier for sending the destination operand to a specific channel among multiple channels. It can use one-hot encoding to represent the source channel, with a bit width equal to the total number of channels, or other binary encodings; no limitation is made here. For example, 8'b0001_0000 indicates Lane_4 as the destination channel, 8'b0100_0000 indicates Lane_6 as the destination channel, and 8'b1101_0000 indicates Lane_7, Lane_6, and Lane_4 as destination channels.
[0067] To parse the current instruction of the first channel and determine its corresponding control bits, one possible approach is to parse the pseudo-instruction corresponding to the current instruction of the first channel to determine its corresponding control bits. Here, the pseudo-instruction is a data transmission pseudo-instruction. For example, cross-channel instructions are split into micro-instructions in pairs. The sender's cross-channel transmit validity bit `send_vld` is valid, and the receiver's cross-channel receive validity bit `receive_vld` is valid. `send_vld` and `receive_vld` are sent to each channel sequentially according to the program, ensuring that register dependency maintenance is performed only on the local channel and used in pairs. Data transmission pseudo-instructions include the vector processor's data receive pseudo-instruction `receive` and data send pseudo-instruction `send`. The `receive` pseudo-instruction indicates cross-channel data transmission, with the current execution unit being the data receiver, and writes back the registers of the current channel. The `send` pseudo-instruction indicates cross-channel data transmission, with the current execution unit being the data sender, and reads the registers of the current channel.
[0068] For example, in a RISC-V processor with a VLEN of 1024 used in an artificial intelligence application scenario, the DPLEN is 128, divided into 8 parallel channels: Lane_0, Lane_1, ..., Lane_7. Each channel includes an independent processing unit and register file. Taking Lane_0 as the first channel as an example, the current instruction of the first channel is a vector addition ablation instruction. The pseudo-instructions corresponding to the ablation instruction include the data transmission pseudo-instruction `send`: the ablation calculation results of two adjacent channels are sent to one channel, and the data reception pseudo-instruction `receive`: after the receiving channel receives the ablation calculation results of all sending channels, it writes them back to the destination operand register of that channel. The vector execution engine parses the pseudo-instructions corresponding to the ablation instruction to determine the control bits corresponding to the current instruction, including the cross-channel receive valid bit (receive_vld), instruction valid bit (inst_vld), cross-channel transmit valid bit (send_vld), source channel identifier bit (rx_src_lane), and destination channel identifier bit (tx_dst_lane). Table 1 below shows the control bits of the current commands for the eight channels:
[0069] Table 1
[0070] In step 102, the cross-channel data transmission and reception logic is deeply integrated with the instruction execution pipeline through the control bits embedded in the current instruction, so as to avoid the additional instruction overhead and synchronization delay introduced by the cross-channel data dependency tracking mechanism.
[0071] Step 104: When the cross-channel receive valid bit is active, receive the source operand from the source channel corresponding to the source channel identifier bit and complete the register write-back.
[0072] The source channel is the channel through which the first channel receives the source operand. There can be at least one source channel. The source channel can be the first channel itself or other channels among multiple channels besides the first channel. No limitation is made here.
[0073] Source operands are data items provided as input to the first channel during the execution of the current instruction. Source operands can come from registers in the first channel or from other channels. Source operands form the basis of the computational task for executing the current instruction. There can be at least one source operand, corresponding to at least one source channel. For example, in the ablation instruction of vector addition C = A + B in Lane_0 of the first channel, the first source operand A comes from Lane_0 itself, and the second source operand B needs to be obtained from Lane_1 across channels.
[0074] When the cross-channel receive valid bit is active, the source operand from the source channel corresponding to the source channel identifier bit is received. One possible method is to receive the source operand from the source channel corresponding to the source channel identifier bit through a dedicated cross-channel data bus when the cross-channel receive valid bit is active. Another possible method is to receive the source operand from the source channel corresponding to the source channel identifier bit through a shared buffer when the cross-channel receive valid bit is active. Yet another possible method is to receive the source operand from the source channel corresponding to the source channel identifier bit through a message passing protocol when the cross-channel receive valid bit is active. No limitation is made here.
[0075] For example, in the first channel Lane_0, when the cross-channel receive valid bit is active (receive_vld=1), the source operand Src Op_0 of the source channel Lane_0 and the source operand Src Op_1 of Lane_1 corresponding to the source channel identifier bit (rx_src_lane=8'b0000_0011) are received through the dedicated cross-channel data bus.
[0076] In step 104, automatic routing for cross-channel data reception is achieved through the source channel identifier bit, directly locating the data source channel for directional data transmission, thereby reducing the hardware design complexity of the target processor.
[0077] Step 106: If the instruction validity bit is active, execute the current instruction based on the source operand to obtain the destination operand.
[0078] The destination operand is the output data obtained from executing the current instruction; it is the result of the instruction's execution. The destination operand is typically written to a specified destination operand register or memory address. The destination operand can serve as the source operand for subsequent instructions. In cases involving cross-channel data transfer, the destination operand is sent to another channel. For example, in an expansion instruction executed on the first channel, Lane_0, the destination operand's width is doubled, truncating the destination operand D into a high half (D_high) and a low half (D_low). The first destination operand, D_low, is sent to Lane_0 itself, while the second destination operand, D_high, needs to be sent across channels to Lane_1.
[0079] To obtain the destination operand by executing the current instruction based on the source operand, one possible method is to perform arithmetic or logical operations on the source operand to generate the destination operand. Another possible method is to perform data type conversion, truncation, or expansion operations on the source operand to generate the destination operand. Yet another possible method is to perform vector operations or bitwise operations on the source operand to generate the destination operand. No particular method is specified here.
[0080] For example, in the first channel Lane_0, when the instruction valid bit is active (inst_vld=1), a vector addition instruction is executed on the source operands Src Op_0 and Src Op_1 to obtain the destination operand Dst Op_0.
[0081] In step 106, the instruction execution is triggered after the data is ready by controlling the instruction validity bit, thus providing a data transmission foundation for subsequent cross-channel data transmission.
[0082] Step 108: When the cross-channel transmit valid bit is active, send the destination operand to the destination channel corresponding to the destination channel identifier bit, and the destination channel completes the register write-back.
[0083] The destination channel is the channel through which the first channel sends the destination operand. There can be at least one destination channel. The destination channel can be the first channel itself or other channels other than the first channel among multiple channels. No limitation is made here.
[0084] When the cross-channel transmit validity bit is active, the destination operand is sent to the destination channel corresponding to the destination channel identifier bit. One possible method is to send the destination operand to the destination channel corresponding to the destination channel identifier bit via a dedicated cross-channel data bus when the cross-channel transmit validity bit is active. Another possible method is to send the destination operand to the destination channel corresponding to the destination channel identifier bit via a shared buffer when the cross-channel transmit validity bit is active. Yet another possible method is to send the destination operand to the destination channel corresponding to the destination channel identifier bit via a message passing protocol when the cross-channel transmit validity bit is active. No further limitation is made here.
[0085] For example, in the first channel Lane_0, when the cross-channel transmit valid bit is active (send_vld=1), the destination operand Dst Op_0 is sent to the destination channel Lane_0 corresponding to the destination channel identifier bit (tx_dst_lane=8'b0000_0001) via the dedicated cross-channel data bus.
[0086] In this embodiment of the disclosure, the cross-channel data processing method avoids direct reading and writing of registers across channels and the correlation maintenance only needs to be completed in the current channel, which makes the target processor hardware design more scalable and thus improves the physical implementation limit of the target processor.
[0087] In one optional embodiment of this disclosure, after step 104, the following specific steps are further included:
[0088] Write the source operand to the register of the current channel, where the current channel is either the source channel or the destination channel;
[0089] Step 106 includes the following specific steps:
[0090] If the instruction validity bit is active, read the source operand from the register of the current channel;
[0091] The current instruction is executed based on the source operand to obtain the destination operand.
[0092] The registers for the current channel are hardware storage units used for temporary storage of operands within the current channel. Each channel has independent source operand registers and destination operand registers to ensure the parallelism and independence of data processing. The number and bit width of the registers can be adjusted according to the specific application scenario and processor design requirements.
[0093] To write the source operand to the register of the current channel, one option is to write the source operand to the register of the current channel within the receive clock cycle via a physical pin connection on the cross-channel data bus. Another option is to use a buffer queue to cache the source operand and write it to the register of the current channel when the buffer queue is not empty. Yet another option is to use Direct Memory Access (DMA) to write the source operand to the register of the current channel. No further restrictions are imposed here.
[0094] One possible way to read the source operand from the register of the current channel is to connect it via the physical pin of the data bus of the first channel and read the source operand from the register of the current channel within the processing clock cycle. Another possible way is to use a multiplexer to read the source operand from the register of the current channel. Yet another possible way is to use a bypass logic network to listen to the register of the current channel and read the source operand from the register of the current channel. No particular method is specified here.
[0095] For example, in the first channel Lane_0, after receiving the source operand Src Op_1 from Lane_1, it is directly written to the current channel register REG_0 of Lane_0 within the receive clock cycle via the physical pin connection of the cross-channel data bus. When executing a shortened addition instruction, if the instruction valid bit inst_vld is activated, the instruction decoding unit issues a read signal, and the source operand Src Op_1 is read from the current channel register REG_0 within the processing clock cycle via the physical pin connection of the data bus of the first channel Lane_0.
[0096] In this embodiment, by introducing a register for the current channel, the decoupling of cross-channel data storage and instruction execution pipeline is achieved. This allows the channel to avoid releasing the processing unit before executing the current instruction, while ensuring the timing controllability of cross-channel data transmission. This effectively improves data processing efficiency and system parallelism, reduces waiting time caused by cross-channel data dependencies, enhances the flexibility and performance of the multi-channel processor architecture, and enables the processor to support more complex data processing tasks without increasing complexity. This simplifies hardware design, reduces power consumption, and provides strong support for high-performance computing.
[0097] Although introducing the register of the current channel can improve the efficiency of temporary data storage, since microinstructions are executed independently within the channel, it is necessary to maintain the instruction dependencies between microinstructions, including read-write dependencies, write-write dependencies, and write-read dependencies. Therefore, before writing the source operand into the register of the current channel, it is necessary to determine whether the register of the current channel has completed the read-write operation to avoid data conflicts or overwriting caused by incomplete read-write operations, and to ensure the consistency of cross-channel data and the determinism of instruction execution timing.
[0098] In one optional embodiment of this disclosure, writing the source operand into the register of the current channel includes the following specific steps:
[0099] Determine whether the register of the current channel has completed a read or write operation, where a read or write operation is either a write operation or a read operation;
[0100] If so, write the source operand to the register of the current channel;
[0101] If not, wait for the current channel's register to complete the read / write operation, and then write the source operand into the current channel's register.
[0102] Read and write operations are data access behaviors performed on registers, including write operations and read operations.
[0103] A write operation is a data access behavior that stores data into a register. If a new write operation is attempted before the register has completed the previous write operation, it may cause data to be overwritten or lost, thus causing data inconsistency problems, i.e., write-write dependency.
[0104] A read operation is a data access behavior that reads data from a register. If a new write operation is attempted before the register has completed the previous read operation, it may lead to data races or the reading of incomplete or incorrect data, which in turn affects the accuracy of the calculation results. This is known as read-write dependency or write-read dependency.
[0105] To determine whether the registers of the current channel have completed a read / write operation, one possible method is to use a status flag to poll and check whether the registers of the current channel have completed a read / write operation. Another possible method is to use a hardware interlock signal to trigger the determination of whether the registers of the current channel have completed a read / write operation. Yet another possible method is to use a pipeline stage synchronization mechanism to determine whether the registers of the current channel have completed a read / write operation. No specific method is specified here.
[0106] For example, in the first channel Lane_0, after receiving the source operand Src Op_1 from Lane_1, the status flag (busy_flag) is used to poll and check whether the current channel's register REG_0 has completed a read / write operation during the receive clock cycle. If yes (busy_flag = 0), it is directly written to the current channel's register REG_0 in Lane_0 via a physical pin connection on the cross-channel data bus. If no (busy_flag = 1), it waits for the current channel's register to complete the read / write operation, and continues to poll and check whether the current channel's register REG_0 has completed the read / write operation. If it has, (busy_flag = 0), it is directly written to the current channel's register REG_0 in Lane_0 via a physical pin connection on the cross-channel data bus.
[0107] In this embodiment of the disclosure, during the writing phase of the source operand, a precise read / write operation status judgment mechanism is introduced, and instruction dependency is used to avoid problems such as data inconsistency and inconsistent processing results.
[0108] Although introducing the register of the current channel can improve the efficiency of temporary data storage, since microinstructions are executed independently within the channel, it is necessary to maintain the instruction dependencies between microinstructions, including read-write dependencies and write-read dependencies. Therefore, before reading the register of the current channel from the source operand, it is necessary to determine whether the register of the current channel has completed the read-write operation to avoid data conflicts or overwriting due to incomplete read-write operations, and to ensure the consistency of cross-channel data and the determinism of instruction execution timing.
[0109] In one optional embodiment of this disclosure, reading the source operand from the register of the current channel includes the following specific steps:
[0110] Determine whether the register of the current channel has completed a read or write operation, where a read or write operation is either a write operation or a read operation;
[0111] If so, read the source operand from the register of the current channel;
[0112] If not, wait for the current channel's register to complete the read / write operation, and then read the source operand from the current channel's register.
[0113] To determine whether the registers of the current channel have completed a read / write operation, one possible method is to use a status flag to poll and check whether the registers of the current channel have completed a read / write operation. Another possible method is to use a hardware interlock signal to trigger the determination of whether the registers of the current channel have completed a read / write operation. Yet another possible method is to use a pipeline stage synchronization mechanism to determine whether the registers of the current channel have completed a read / write operation. No specific method is specified here.
[0114] For example, in the first channel Lane_0, when a shortened addition instruction is executed, if the instruction validity bit inst_vld is activated, the instruction decoding unit sends a read signal. Through the physical pin connection of the data bus of the first channel Lane_0, the status flag bit (busy_flag) is used to determine whether the current channel's register REG_0 has completed a read / write operation within the processing clock cycle. If yes (busy_flag = 0), the source operand is directly read from the current channel's register REG_0 and the shortened addition instruction is executed through the physical pin connection of the cross-channel data bus. If no (busy_flag = 1), the system waits for the current channel's register to complete the read / write operation, and continues to use the status flag bit to poll and check whether the current channel's register REG_0 has completed the read / write operation. If it has, (busy_flag = 0), the source operand is directly read from the current channel's register REG_0 and the shortened addition instruction is executed through the physical pin connection of the cross-channel data bus.
[0115] In this embodiment of the disclosure, during the source operand reading stage, a precise read / write operation status judgment mechanism is introduced, and instruction dependency is used to avoid problems such as data inconsistency and inconsistent processing results.
[0116] In this embodiment of the disclosure, by introducing a precise read / write operation status judgment mechanism and through instruction dependency, problems of data inconsistency and inconsistent processing results are avoided.
[0117] In one optional embodiment of this disclosure, after step 106, the following specific steps are further included:
[0118] Write the destination operand to the register of the current channel;
[0119] Step 108 includes the following specific steps:
[0120] If the cross-channel transmit valid bit is active, read the destination operand from the register of the current channel;
[0121] The destination operand is sent to the destination channel corresponding to the destination channel identifier bit, and the destination channel completes the register write-back.
[0122] The registers for the current channel are hardware storage units used for temporary storage of operands within the current channel. Each channel has independent source operand registers and destination operand registers to ensure the parallelism and independence of data processing. The number and bit width of the registers can be adjusted according to the specific application scenario and processor design requirements.
[0123] To write the destination operand to the register of the current channel, one option is to write the destination operand to the register of the current channel within the receive clock cycle via a physical pin connection on the cross-channel data bus. Another option is to use a buffer queue to cache the destination operand and write it to the register of the current channel when the buffer queue is not empty. Yet another option is to use direct memory access to write the destination operand to the register of the current channel. No further restrictions are imposed here.
[0124] One possible way to read the destination operand from the register of the current channel is to connect it via the physical pin of the data bus of the first channel and read the destination operand from the register of the current channel within the processing clock cycle. Another possible way is to use a multiplexer to read the destination operand from the register of the current channel. Yet another possible way is to use a bypass logic network to listen to the register of the current channel and read the destination operand from the register of the current channel. No particular method is specified here.
[0125] For example, in the first channel Lane_0, after the vector addition ablation instruction generates the destination operand Dst Op_0, it is written to the current channel's register REG_0 within the write-back clock cycle via a physical pin connection on the cross-channel data bus. When the cross-channel send valid bit (send_vld = 1) is triggered, the destination operand Dst Op_0 is read from the current channel's register REG_0 via a multiplexer, and the data is transmitted to the destination channel Lane_0 via the physical line of the dedicated cross-channel data bus according to the configuration of the destination channel identifier bit (tx_dst_lane = 8'b0000_0001).
[0126] In this embodiment, by introducing a register for the current channel, the decoupling of cross-channel data storage and instruction execution pipeline is achieved. This allows the channel to release the processing unit immediately after completing the calculation of the current instruction, while ensuring the timing controllability of cross-channel data transmission. This effectively improves data processing efficiency and system parallelism, reduces the waiting time caused by cross-channel data dependencies, enhances the flexibility and performance of the multi-channel processor architecture, and enables the processor to support more complex data processing tasks without increasing complexity. This simplifies hardware design, reduces power consumption, and provides strong support for high-performance computing.
[0127] Although introducing the register of the current channel can improve the efficiency of temporary data storage, since microinstructions are executed independently within the channel, it is necessary to maintain the instruction dependencies between microinstructions, including read-write dependencies, write-write dependencies, and write-read dependencies. Therefore, before writing the destination operand into the register of the current channel, it is necessary to determine whether the register of the current channel has completed the read-write operation to avoid data conflicts or overwriting caused by incomplete read-write operations, and to ensure the consistency of cross-channel data and the determinism of instruction execution timing.
[0128] In one optional embodiment of this disclosure, writing the destination operand into the register of the current channel includes the following specific steps:
[0129] Determine whether the register of the current channel has completed a read or write operation, where a read or write operation is either a write operation or a read operation;
[0130] If so, write the destination operand to the register of the current channel;
[0131] If not, wait for the current channel's register to complete the read / write operation, and then write the destination operand into the current channel's register.
[0132] A write operation is a data access behavior that stores data into a register. If a new write operation is attempted before the register has completed the previous write operation, it may cause data to be overwritten or lost, thus causing data inconsistency problems, i.e., write-write dependency.
[0133] A read operation is a data access behavior that reads data from a register. If a new write operation is attempted before the register has completed the previous read operation, it may lead to data races or the reading of incomplete or incorrect data, which in turn affects the accuracy of the calculation results. This is known as read-write dependency or write-read dependency.
[0134] To determine whether the registers of the current channel have completed a read / write operation, one possible method is to use a status flag to poll and check whether the registers of the current channel have completed a read / write operation. Another possible method is to use a hardware interlock signal to trigger the determination of whether the registers of the current channel have completed a read / write operation. Yet another possible method is to use a pipeline stage synchronization mechanism to determine whether the registers of the current channel have completed a read / write operation. No specific method is specified here.
[0135] For example, in the first channel Lane_0, after the vector multiplication extension instruction generates the destination operand Dst Op_0, the read / write status of the current channel's register REG_0 is polled using the status flag bit (busy_flag) during the write-back clock cycle. If it is (busy_flag = 0), it is directly written to the current channel's register REG_0 through the output port of the execution unit; if it is not (busy_flag = 1), the write-back clock cycle continues until busy_flag is reset to 0, and then the data write operation is completed through the physical pin connection of the cross-channel data bus. When the cross-channel send valid bit (send_vld = 1) is triggered, a read signal is generated according to the destination channel identifier bit (tx_dst_lane = 8'b0000_0001), and the destination operand Dst Op_0 is read from the current channel's register REG_0 through a multiplexer and transmitted to the corresponding register storage location in Lane_0 via the dedicated data bus.
[0136] In this embodiment of the disclosure, during the writing phase of the destination operand, a precise read / write operation status judgment mechanism is introduced, and instruction dependency is used to avoid problems such as data inconsistency and inconsistent processing results.
[0137] Although introducing the register of the current channel can improve the efficiency of temporary data storage, since microinstructions are executed independently within the channel, it is necessary to maintain the instruction dependencies between microinstructions, including read-write dependencies and write-read dependencies. Therefore, before reading the register of the current channel from the destination operand, it is necessary to determine whether the register of the current channel has completed the read-write operation to avoid data conflicts or overwriting caused by incomplete read-write operations, and to ensure the consistency of cross-channel data and the determinism of instruction execution timing.
[0138] In one optional embodiment of this disclosure, reading the destination operand from the register of the current channel includes the following specific steps:
[0139] Determine whether the register of the current channel has completed a read or write operation, where a read or write operation is either a write operation or a read operation;
[0140] If so, read the destination operand from the register of the current channel;
[0141] If not, wait for the current channel's register to complete the read / write operation, and then read the destination operand from the current channel's register.
[0142] To determine whether the registers of the current channel have completed a read / write operation, one possible method is to use a status flag to poll and check whether the registers of the current channel have completed a read / write operation. Another possible method is to use a hardware interlock signal to trigger the determination of whether the registers of the current channel have completed a read / write operation. Yet another possible method is to use a pipeline stage synchronization mechanism to determine whether the registers of the current channel have completed a read / write operation. No specific method is specified here.
[0143] For example, in the first channel Lane_0, after the vector multiplication extension instruction generates the destination operand Dst Op_0, the read / write status of the current channel's register REG_0 is polled using the status flag (busy_flag) during the write-back clock cycle. If it is (busy_flag = 0), the data is directly written to the current channel's register REG_0 through the output port of the execution unit; if not (busy_flag = 1), the data is written after the busy_flag is reset to 0, through the physical pin connection of the cross-channel data bus. When the cross-channel send valid bit (send_vld = 1) is triggered, a read signal is generated based on the destination channel identifier bit (tx_dst_lane = 8'b0000_0001), and it is determined whether DST_REG_0 has completed the read / write operation during the processing clock cycle. If busy_flag=0, the destination operand Dst Op_0 is read from the register REG_0 of the current channel through the multiplexer and transmitted to the register storage location corresponding to the destination channel Lane_0 via the dedicated data bus; if busy_flag=1, the read and send operations are performed only after the register of the current channel has completed the current operation.
[0144] In this embodiment of the disclosure, during the reading phase of the destination operand, a precise read / write operation status judgment mechanism is introduced, and by using instruction dependency, problems of data inconsistency and inconsistent processing results are avoided.
[0145] In one optional embodiment of this disclosure, the first channel includes a first processing unit;
[0146] When the instruction validity bit is active, the current instruction is executed based on the source operand to obtain the destination operand, including:
[0147] When the instruction validity bit is active, the first processing unit is scheduled to execute the current instruction based on the source operand to obtain the destination operand.
[0148] The first processing unit is an independent computing module integrated within the first channel, which includes an arithmetic logic unit (ALU), a vector operation unit, and a register read / write controller, used to execute micro-operations after instruction decoding.
[0149] The first processing unit is scheduled to execute the current instruction based on the source operand to obtain the destination operand. One possible approach is to schedule the arithmetic logic unit of the first channel to perform arithmetic or logical operations on the source operand to generate the destination operand. Another possible approach is to schedule the register file of the first channel to perform data type conversion, truncation, or expansion operations on the source operand to generate the destination operand. Yet another possible approach is to schedule the vector processing unit of the first channel to perform vector operations or bitwise operations on the source operand to generate the destination operand. No particular limitation is made here.
[0150] For example, when the vector multiply-accumulate instruction is executed in the first channel Lane_0, if the instruction valid bit is active (inst_vld=1), the vector multiply-accumulate unit in the first processing unit is scheduled to read the source operands A and B from the register of the current channel, perform the multiplication operation, and then perform the multiplication operation with the initial result in the accumulation register. The final result is then written as the destination operand into the register REG_0 of the current channel.
[0151] In this embodiment of the disclosure, the processing unit is directly associated with the instruction control signal through a hardware-level scheduling mechanism, which ensures that the calculation operation is triggered immediately after the data is ready. This eliminates the communication delay between instruction dispatch and execution units in the traditional multi-channel architecture, and effectively improves instruction-level parallelism and resource utilization.
[0152] Referring to the method embodiment in Figure 1 above, Figure 2 shows a schematic diagram of the architecture of ablation instructions in a target processor for a cross-channel data processing method provided by an embodiment of this disclosure, as shown in Figure 2:
[0153] In a vector processor, this architecture employs a design with eight parallel processing channels (Lane_0 to Lane_7). During the execution of vector ablation instructions, after each channel completes its local computation, a hardware pseudo-instruction-controlled cross-channel data routing mechanism transmits intermediate results to the target channel. Specifically:
[0154] For example, for ablation instructions, after reading the source operands from the source operand vector registers of the 8 channels and performing ablation processing, four destination operands are generated and written back to the register of the current channel. Except for Lane_0, all channels need to synchronously write back the results to the corresponding destination channels (the starting position of the arrow in the diagram is the source channel, and the end of the arrow points to the destination channel where the destination operands are written back). The source operands come from the 8 channels. Each channel calculates the result of DPLEN / 2 using the data of DPLEN. Then, the results need to be written back compactly. The results of L0 and L1 need to be concatenated and written back to L0, the results of L2 and L3 need to be concatenated and written back to L1, the results of L4 and L5 need to be concatenated and written back to L2, and the results of L6 and L7 need to be concatenated and written back to L3.
[0155] The scheme shown in Figure 2 enables each channel to independently maintain the instruction dependency (data dependency) of the vector register in a multi-channel vector processor in a software-invisible manner without changing the programming model, thereby improving the overall architecture's scalability and physical implementation ceiling.
[0156] Referring to the method embodiment in Figure 1 above, Figure 3 shows one of the schematic diagrams of the architecture of the extended instruction in the target processor for a cross-channel data processing method provided by this disclosure, as shown in Figure 3:
[0157] In a vector processor, this architecture employs a design with eight parallel processing channels (Lane_0 to Lane_7). During the execution of vector extension instructions, after each channel completes its local computation, a hardware pseudo-instruction-controlled cross-channel data routing mechanism transmits intermediate results to the target channel. Specifically:
[0158] For example, for an expansion instruction, after reading the source operands from the source operand vector registers of the four channels and performing expansion processing, eight destination operands are generated and written back to the destination operand registers. Except for Lane_0, all channels need to synchronously write the results back to the corresponding destination channels (the starting position of the arrow in the diagram is the source channel, and the tail of the arrow points to the destination channel for writing back the destination operands). The source operands come from the four channels. Each channel calculates a 2DPLEN result using the data from DPLEN. Then, the results need to be written back compactly. The result of L0 needs to be split and written back to L0 and L1, the result of L1 needs to be split and written back to L2 and L3, the result of L2 needs to be split and written back to L4 and L5, and the result of L3 needs to be split and written back to L6 and L7.
[0159] The scheme shown in Figure 3 enables each channel to independently maintain the instruction dependency (data dependency) of vector registers in a multi-channel vector processor in a software-invisible manner without changing the programming model, thereby improving the overall architecture's scalability and physical implementation ceiling.
[0160] Referring to the method embodiment in Figure 1 above, Figure 4 shows a second schematic diagram of the architecture of the extended instruction in the target processor for a cross-channel data processing method provided by this disclosure, as shown in Figure 4:
[0161] In a vector processor, this architecture employs a design with eight parallel processing channels (Lane_0 to Lane_7). During the execution of vector extension instructions, after each channel completes its local computation, a hardware pseudo-instruction-controlled cross-channel data routing mechanism transmits intermediate results to the target channel. Specifically:
[0162] For example, for an expansion instruction, after reading the source operands from the source operand vector registers of the four channels and performing expansion processing, eight destination operands are generated and written back to the destination operand registers. Except for Lane_0, all channels need to synchronously write the results back to the corresponding destination channels (the starting position of the arrow in the diagram is the source channel, and the tail of the arrow points to the destination channel for writing back the destination operands). The source operands come from the four channels. Each channel calculates a 2DPLEN result using the data from DPLEN. Then, the results need to be written back compactly. The result of L7 needs to be split and written back to L6 and L7, the result of L6 needs to be split and written back to L4 and L5, the result of L5 needs to be split and written back to L2 and L3, and the result of L4 needs to be split and written back to L0 and L1.
[0163] The scheme shown in Figure 4 enables each channel to independently maintain the instruction dependency (data dependency) of the vector register in a multi-channel vector processor in a software-invisible manner without changing the programming model, thereby improving the overall architecture's scalability and physical implementation ceiling.
[0164] It should be noted that although the solutions in the above embodiments use a RISC-V architecture processor for illustrative purposes, the processor architecture is not limited. The execution logic of the above cross-channel data processing method can also be implemented in processor architectures such as x86 and ARM, and is not limited here.
[0165] Corresponding to the above method embodiments, this disclosure also provides a target processor embodiment. FIG5 shows a schematic diagram of the structure of a target processor provided in one embodiment of this disclosure. As shown in FIG5, the target processor 500 includes a data execution engine 520 with multiple channels 510, wherein the first channel 5110 is any one of the multiple channels 510;
[0166] The data execution engine 520 is used to parse the current instruction of the first channel 5110, determine the control bits corresponding to the current instruction, and when the cross-channel receive valid bit is active, receive the source operand from the source channel corresponding to the source channel identifier bit and complete the register write-back. When the instruction valid bit is active, execute the current instruction based on the source operand to obtain the destination operand. When the cross-channel send valid bit is active, send the destination operand to the destination channel corresponding to the destination channel identifier bit, and the destination channel completes the register write-back. The control bits include the cross-channel receive valid bit, the instruction valid bit, the cross-channel send valid bit, the source channel identifier bit, and the destination channel identifier bit.
[0167] In one optional embodiment of this disclosure, the first channel 5110 includes a first processing unit and a register;
[0168] The data execution engine 520 is specifically used to parse the current instruction of the first channel 5110, determine the control bit corresponding to the current instruction, and when the cross-channel receive valid bit is active, receive the source operand from the source channel corresponding to the source channel identifier bit and complete the register write-back, write the source operand into the register of the current channel, and when the instruction valid bit is active, read the source operand from the register of the current channel, schedule the first processing unit, execute the current instruction based on the source operand, obtain the destination operand, write the destination operand into the destination operand register, and when the cross-channel send valid bit is active, read the destination operand from the destination operand register, send the destination operand to the destination channel corresponding to the destination channel identifier bit, and have the destination channel complete the register write-back.
[0169] In this embodiment of the disclosure, the target processor directly triggers data transmission at the hardware level, realizing cross-channel data transmission invisible to software without changing the programming model. This avoids direct reading and writing of registers across channels, and correlation maintenance only needs to be completed in the current channel, making the target processor hardware design more scalable and thus improving the physical implementation limit of the target processor.
[0170] The above is an illustrative scheme of a target processor according to this embodiment. It should be noted that the technical solution of this target processor and the technical solution of the cross-channel data processing method described above belong to the same concept. For details not described in detail in the technical solution of the target processor, please refer to the description of the technical solution of the cross-channel data processing method described above.
[0171] Corresponding to the above method embodiments, this disclosure also provides an on-chip system embodiment. FIG6 shows a schematic diagram of the structure of an on-chip system provided in one embodiment of this disclosure. As shown in FIG6, the on-chip system 600 includes a plurality of on-chip components 610, and the plurality of on-chip components 610 include a target processor 500;
[0172] The target processor 500 is used to execute computer programs / instructions that, when executed by the target processor 500, implement the steps of the above-described cross-channel data processing method.
[0173] In this embodiment of the disclosure, the target processor in the system-on-a-chip directly triggers data transmission at the hardware level. Without changing the programming model, it achieves cross-channel data transmission that is invisible to software, avoids direct reading and writing of registers across channels, and ensures that dependency maintenance is completed only in the current channel. This makes the hardware design of the target processor more scalable, thereby increasing the physical implementation limit of the target processor. Through hardware-level autonomous control and end-to-end direct transmission mechanism, the system-on-a-chip solves the problems of complex cross-channel dependency tracking, poor scalability, and high latency in traditional multi-channel architectures, while maintaining software transparency. It provides a hardware solution with high scalability and high physical upper limit for high-performance, large-scale parallel computing.
[0174] The above is an illustrative scheme of a system-on-a-chip (SoC) according to this embodiment. It should be noted that the technical solution of this SoC and the technical solution of the aforementioned cross-channel data processing method belong to the same concept. Details not described in detail in the SoC technical solution can be found in the description of the aforementioned cross-channel data processing method.
[0175] Figure 7 shows a structural block diagram of a computing device according to an embodiment of the present disclosure. The components of the computing device 700 include, but are not limited to, a memory 710 and a target processor 500. The target processor 500 is connected to the memory 710 via a bus.
[0176] Optionally, the computing device 700 also includes an access device that enables the computing device 700 to communicate via one or more networks. Examples of such networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device may include one or more of any type of wired or wireless network interface (e.g., a Network Interface Controller (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.
[0177] In one embodiment of this disclosure, the aforementioned components of the computing device 700, as well as other components not shown in FIG. 7, may be interconnected, for example, via a bus. It should be understood that the computing device block diagram shown in FIG. 7 is merely for illustrative purposes and is not intended to limit the scope of this disclosure. Those skilled in the art can add or replace other components as needed.
[0178] The computing device 700 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 700 can also be a mobile or stationary server.
[0179] The target processor 500 is used to execute the following computer program / instructions, which, when executed by the target processor, implement the steps of the above-described cross-channel data processing method.
[0180] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the cross-channel data processing method described above belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the cross-channel data processing method described above.
[0181] An embodiment of this disclosure also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.
[0182] The various embodiments in this disclosure are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the computer-readable storage medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.
[0183] An embodiment of this disclosure also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.
[0184] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the above method belong to the same concept, and all details not described in detail in the technical solution of the computer program product can be referred to the description of the technical solution of the above method.
[0185] The foregoing has described specific embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0186] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.
[0187] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of this disclosure are not limited to the described order of actions, because according to the embodiments of this disclosure, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments of this disclosure.
[0188] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0189] The preferred embodiments disclosed above are merely illustrative of this disclosure. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments of this disclosure. These embodiments are selected and specifically described in this disclosure to better explain the principles and practical applications of the embodiments of this disclosure, thereby enabling those skilled in the art to better understand and utilize this disclosure. This disclosure is limited only by the claims and their full scope and equivalents.
Claims
1. A cross-channel data processing method, applied to a target processor of multiple channel data execution engines, comprising: Parse the current instruction of the first channel and determine the control bits corresponding to the current instruction, wherein the control bits include a cross-channel receive valid bit, an instruction valid bit, a cross-channel transmit valid bit, a source channel identifier bit, and a destination channel identifier bit; When the cross-channel receive valid bit is active, the source operand from the source channel corresponding to the source channel identifier bit is received and register write-back is completed; When the instruction validity bit is active, the current instruction is executed based on the source operand to obtain the destination operand; When the cross-channel transmission valid bit is active, the destination operand is sent to the destination channel corresponding to the destination channel identifier bit, and the destination channel completes the register write-back.
2. The method according to claim 1, further comprising, after receiving the source operand from the source channel corresponding to the source channel identifier bit when the cross-channel receive valid bit is active: Write the source operand into the register of the current channel, where the current channel is either the source channel or the destination channel; When the instruction validity bit is active, the current instruction is executed based on the source operand to obtain the destination operand, including: When the instruction validity bit is active, the source operand is read from the register of the current channel; The current instruction is executed based on the source operand to obtain the destination operand.
3. The method according to claim 2, wherein writing the source operand into the register of the current channel comprises: Determine whether the register of the current channel has completed a read / write operation, wherein the read / write operation is a write operation or a read operation; If so, write the source operand into the register of the current channel; If not, wait for the register of the current channel to complete the read / write operation, and then write the source operand into the register of the current channel.
4. The method of claim 3, wherein determining whether the register of the current channel is complete with a read or write operation comprises: The status flag is used to poll and detect whether the register of the current channel has completed a read / write operation.
5. The method of claim 3, wherein determining whether the register of the current channel is complete with a read or write operation comprises: The hardware interlock signal triggers a determination of whether the register of the current channel has completed a read / write operation.
6. The method of claim 3, wherein determining whether the register of the current channel is complete with a read or write operation comprises: The pipeline stage synchronization mechanism is used to determine whether the registers of the current channel have completed read / write operations.
7. The method according to any one of claims 2-6, wherein reading the source operand from the register of the current channel comprises: Determine whether the register of the current channel has completed a read / write operation, wherein the read / write operation is a write operation or a read operation; If so, read the source operand from the register of the current channel; If not, wait for the register of the current channel to complete the read / write operation, and then read the source operand from the register of the current channel.
8. The method according to any one of claims 1-7, after executing the current instruction based on the source operand to obtain the destination operand when the instruction validity bit is active, further comprising: Write the destination operand into the register of the current channel, where the current channel is either the source channel or the destination channel; When the cross-channel transmission valid bit is active, the destination operand is sent to the destination channel corresponding to the destination channel identifier bit, and the destination channel completes the register write-back, including: If the cross-channel transmission valid bit is active, the destination operand is read from the register of the current channel; The destination operand is sent to the destination channel corresponding to the destination channel identifier, and the destination channel completes the register write-back.
9. The method according to claim 8, wherein writing the destination operand into the register of the current channel comprises: Determine whether the register of the current channel has completed a read / write operation, wherein the read / write operation is a write operation or a read operation; If so, write the destination operand into the register of the current channel; If not, wait for the current channel's register to complete the read / write operation, and then write the destination operand into the current channel's register.
10. The method according to claim 8, wherein reading the destination operand from the register of the current channel comprises: Determine whether the register of the current channel has completed a read / write operation, wherein the read / write operation is a write operation or a read operation; If so, read the destination operand from the register of the current channel; If not, wait for the register of the current channel to complete the read / write operation, and then read the destination operand from the register of the current channel.
11. The method according to any one of claims 1-10, wherein the first channel comprises a first processing unit; When the instruction validity bit is active, executing the current instruction based on the source operand to obtain the destination operand includes: When the instruction validity bit is active, the first processing unit is scheduled to execute the current instruction based on the source operand to obtain the destination operand.
12. The method of claim 11, wherein, The scheduling of the first processing unit, based on the source operand, to execute the current instruction and obtain the destination operand includes: The arithmetic logic unit of the first channel is scheduled to perform arithmetic or logical operations on the source operand to generate the destination operand.
13. The method of claim 11, wherein, The scheduling of the first processing unit, based on the source operand, to execute the current instruction and obtain the destination operand includes: The register file of the first channel is scheduled to perform data type conversion, truncation, or expansion operations on the source operand to generate the destination operand.
14. The method of claim 11, wherein, The scheduling of the first processing unit, based on the source operand, to execute the current instruction and obtain the destination operand includes: The vector processing unit of the first channel is scheduled to perform vector operations or bit operations on the source operand to generate the destination operand.
15. A target processor, comprising a data execution engine with multiple channels, wherein a first channel is any one of the multiple channels; The data execution engine is configured to parse a current instruction of the first lane, determine a control bit corresponding to the current instruction, receive a source operand from a source lane corresponding to a source lane identification bit and complete register write back in a case that a cross-lane receiving valid bit is activated, execute the current instruction based on the source operand and obtain a destination operand in a case that an instruction valid bit is activated, send the destination operand to a destination lane corresponding to a destination lane identification bit and complete register write back by the destination lane in a case that a cross-lane sending valid bit is activated, and The control bits include the cross-channel receive valid bit, the instruction valid bit, the cross-channel transmit valid bit, the source channel identifier bit, and the destination channel identifier bit.
16. The target processor according to claim 15, wherein the first channel comprises a first processing unit and a register; The data execution engine is specifically configured to parse a current instruction of the first lane, determine a control bit corresponding to the current instruction, in a case where a cross-lane receiving valid bit is activated, receive a source operand from a source lane corresponding to a source lane identification bit and complete register write back, write the source operand into a register of the current lane, in a case where an instruction valid bit is activated, read the source operand from the register of the current lane, schedule the first processing unit, execute the current instruction based on the source operand, obtain a destination operand, write the destination operand into the register of the current lane, in a case where a cross-lane sending valid bit is activated, read the destination operand from the register of the current lane, and send the destination operand to a destination lane corresponding to a destination lane identification bit, and complete register write back by the destination lane. 17.A system on chip, comprising a plurality of on-chip components, the plurality of on-chip components comprising a target processor; The target processor is configured to execute a computer program / instruction, which, when executed by the target processor, implements the steps of the method of any one of claims 1 to 14. 18.A computing device, comprising: a memory and a target processor; The memory is configured to store a computer program / instruction, and the target processor is configured to execute the computer program / instruction, which, when executed by the target processor, implements the steps of the method of any one of claims 1 to 14. 19.A computer-readable storage medium storing a computer program / instruction, which, when executed by a processor, implements the steps of the method of any one of claims 1 to 14. 20.A computer program product comprising a computer program / instruction, which, when executed by a processor, implements the steps of the method of any one of claims 1 to 14.
