WO2024093112A1 - Computing engine communication method and apparatus, electronic device, and storage medium - Google Patents

Abstract

The present application discloses a computing engine communication method and apparatus, applied to the technical field of computing engine communications. The method comprises: acquiring device-side program data sent by a host-side computing engine, wherein the device-side program data is program data obtained by compiling a program code which contains a communication function and a computing task function, and the communication function is a function created on the basis of a communication demand; and executing the device-side program data by using a device-side computing engine, computing target data on the basis of the computing task function to obtain a computing result, and sending the computing result to a target receiving end on the basis of the communication function so as to implement communication between the device-side computing engine and a computing engine in the target receiving end. The method can prevent the host-side computing engine from allocating communication tasks, and release the computing power of the host-side computing engine, and is suitable for communication scenarios of different computing engines in the same node and computing engines among different nodes.

Description

Computing engine communication method and device, electronic device, and storage medium

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims priority to a Chinese patent application filed with the China Patent Office on October 31, 2022, with application number 202211347106.X and application name “Computing Engine Communication Method and Device”, the entire contents of which are incorporated by reference into this application.

Technical Field

The present application relates to the field of computing engine communication technology, and in particular to computing engine communication methods and devices, electronic devices, and storage media.

Background technique

With the development of traditional Moore's Law, the development of computing power has gone through the process of increasing the main frequency of a single processor, increasing the number of processors in a single chip, and then increasing the number of chips in a single computing network (such as a large-scale cloud computing network). At present, the dividends brought by the chip process have almost disappeared, and relying on the number of chips to stack up has also encountered a series of problems such as mismatched computing power and high energy consumption. In response to the current diversified and giant computing tasks, it has become an industry consensus and development trend to deploy computing engines with different architectures in the same system, such as CPU (Central Processing Unit), GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), etc. According to the characteristics of computing engines with different architectures, they can either serve different links of the same task or serve different tasks to maximize the use of computing power.

Due to the existence of historical legacy issues of network software protocol stacks and network hardware devices in traditional computer architecture systems, in a system with multiple heterogeneous computing engines, the engine chips, PCIE (Peripheral Component Interconnect Express, Transmission Control Protocol) or local interconnect buses, and network cards of various computing nodes are independent of each other, and the efficiency of mutual communication and computing performance have become very low, and the degree of freedom and flexibility are not high. As a result, the current large-scale computing power network not only has the problem of computing power performance of computing devices, but also faces the bottleneck caused by inefficient communication between computing engines of different architectures. The existing communication schemes between different computing engines are usually divided into communication schemes within nodes and communication schemes between different nodes. In addition, communication tasks usually require computing engines on the host side, and the host side still bears a lot of work.

Summary of the invention

In view of this, the purpose of this application is to provide a computing engine communication method and device, electronic device, and storage medium, which can avoid the host-side computing engine from allocating communication tasks, release the computing power of the host-side computing engine, and adapt to the communication scenarios of different computing engines in the same node and computing engines between different nodes. The specific scheme is as follows:

In a first aspect, the present application discloses a computing engine communication method, which is applied to a first device end, comprising:

Obtaining device-side program data sent by the host-side computing engine; wherein the device-side program data is program data obtained by compiling program codes including communication functions and computing task functions, and the communication function is a function created based on communication requirements;

The device-side computing engine is used to execute the device-side program data, the target data is calculated based on the computing task function to obtain the calculation result, and the calculation result is sent to the target receiving end based on the communication function to realize the communication between the device-side computing engine and the computing engine in the target receiving end.

In some embodiments, obtaining device-side program data sent by a host-side computing engine includes:

The RDMA module is used to obtain the device-side program data sent by the host-side computing engine; The standby computing engine is integrated into the same chip;

Correspondingly, sending the calculation result to the target receiving end based on the communication function includes: sending the calculation result to the target receiving end based on the communication function and through the RDMA module.

In some embodiments, obtaining device-side program data sent by a host-side computing engine through an RDMA module includes:

Obtaining target data, parameter information, and device-side program data sent by the host-side computing engine through the RDMA module; wherein the parameter information includes the first address and length information of the target data written into the memory of the first device;

Write target data into memory based on parameter information;

Accordingly, the device-side computing engine is used to execute the device-side program data, and the target data is calculated based on the computing task function to obtain the calculation result, including:

The device-side computing engine is used to execute the device-side program data, the target data is read from the memory based on the parameter information, and the target data is calculated based on the computing task function to obtain the calculation result.

In some embodiments, based on the communication function, the calculation result is sent to the target receiving end through the RDMA module, including:

Write the identification information of the target receiving end into the communication table based on the communication function, and notify the table parsing engine;

Obtain identification information of the target receiving end from the communication table through the table parsing engine, and obtain the IP address corresponding to the identification information;

The calculation results are sent to the target receiving end based on the IP address through the RDMA module.

In some embodiments, it also includes:

Write the calculation result into the data area in the memory, and write the write address and data length into the communication table;

The write address and data length are obtained from the communication table through the table parsing engine, and the write address, data length and IP address are written into the memory in the RDMA module.

In some embodiments, sending the calculation result to the target receiving end based on the IP address through the RDMA module includes:

The memory is detected through the RDMA module, and when the memory is not empty, the write address, data length and IP address are read from the memory, the calculation result is read from the memory according to the write address and data length, and the calculation result is sent to the target receiving end based on the IP address.

In some embodiments, obtaining the IP address corresponding to the identification information includes:

Obtain the IP address corresponding to the identification information from the IP identification comparison table.

In some embodiments, it also includes:

Through the RDMA module, a discovery message is sent to the host end, so that the host end allocates an IP address and identification information to each device end based on the discovery message of each device end;

Obtain the IP address and identification information of each device end and the IP address and identification information of the host end sent by the host end;

The IP address and identification information of each device end, the IP address and identification information of the host end are saved in the IP identification comparison table.

In some embodiments, after sending the discovery message to the host through the RDMA module, the method further includes:

Obtain the IP address and identification information assigned by the host end to the first device end through the RDMA module, and reply confirmation information to the host end, so that after receiving the confirmation information replied by each device end, the host end saves the IP address and identification information of each device end, and sends the IP address and identification information of each device end and the IP address and identification information of the host end to each device end;

The confirmation information carries the IP address and identification information allocated by the host to the corresponding device.

In some embodiments, the target receiving end is a host end or a second device end;

If the target receiving end is a plurality of second device ends, the identification information of the target receiving end is written into the communication table based on the communication function, including:

Based on the communication function, the identification information of the plurality of second device terminals is written into the communication table.

In some embodiments, if the target receiving end is the second device end, the second device end receives the calculation result through its own RDMA module and stores the calculation result in its own memory.

In some embodiments, after sending the calculation result to the second device end based on the communication function, the method further includes:

The transmission end information is sent to the second device end, so that the RDMA module of the second device end returns the first address and length information of the calculation result in the memory to the calculation engine of the second device end after receiving the transmission end information.

In some embodiments, it also includes:

During the process of executing the program data on the device side, if the target program instruction is executed, the device waits to receive the data from the third device side, and after receiving the data from the third device side, continues to execute subsequent operations.

In a second aspect, the present application discloses a computing engine communication method, which is applied to a host side, comprising:

Sending device-side program data to the device side through the host-side computing engine, so that the device-side computing engine at the device side executes the device-side program data, calculates the target data based on the computing task function to obtain a calculation result, and sends the calculation result to the target receiving end based on the communication function;

The device-side program data is program data obtained by compiling program codes including communication functions and computing task functions, and the communication function is a function created based on communication requirements.

In some embodiments, sending device-side program data to the device-side through a host-side computing engine includes:

The target data, parameter information and device-side program data are sent to the device-side through the host-side computing engine; wherein the parameter information includes the first address and length information of the target data written into the device-side memory, so that the device-side writes the target data into the memory based on the parameter information.

In some embodiments, it also includes:

Get the discovery message sent by each device;

Based on the discovery messages sent by each device, configure an IP address and identification information for each device;

The IP address and identification information of each device end and the IP address and identification information of the host end are sent to each device end, so that each device end saves the IP address and identification information of each device end and the IP address and identification information of the host end into the IP identification comparison table.

In some embodiments, sending the IP address and identification information of each device end and the IP address and identification information of the host end to each device end further includes:

Send the IP address and identification information assigned by the host to each device;

After receiving the confirmation information replied by each device end, the IP address and identification information of each device end are saved, and the IP address and identification information of each device end and the IP address and identification information of the host end are sent to each device end.

In a third aspect, the present application discloses a computing engine communication device, applied to a first device end, comprising:

The communication engine is used to obtain the device-side program data sent by the host-side computing engine; wherein the device-side program data is program data compiled from program codes including communication functions and computing task functions, and the communication function is a function created based on communication requirements;

The device-side computing engine is used to execute device-side program data and calculate the target data based on the computing task function. The calculation result is obtained and sent to the target receiving end based on the communication function.

In a fourth aspect, the present application discloses a computing engine communication device, applied to a host side, comprising:

The host-side computing engine is used to send the device-side program data to the device-side, so that the device-side computing engine of the device-side executes the device-side program data, calculates the target data based on the computing task function to obtain a computing result, and sends the computing result to the target receiving end based on the communication function;

The device-side program data is program data obtained by compiling program codes including communication functions and computing task functions, and the communication function is a function created based on communication requirements.

In a fifth aspect, the present application discloses an electronic device, including:

Memory for storing computer programs;

The processor is used to execute a computer program to implement the computing engine communication method as described above.

In a sixth aspect, the present application discloses a non-volatile computer-readable storage medium, in which a computer program is stored. When the computer program is loaded and executed by a processor, the computing engine communication method as described above is implemented.

It can be seen that the present application first obtains the device-side program data sent by the host-side computing engine; wherein the device-side program data is the program data obtained by compiling the program code containing the communication function and the computing task function, and the communication function is a function created based on the communication demand, and then the device-side computing engine is used to execute the device-side program data, and the target data is calculated based on the computing task function to obtain the calculation result, and the calculation result is sent to the target receiving end based on the communication function, so as to realize the communication between the device-side computing engine and the computing engine in the target receiving end. That is, in the present application, a communication function is created based on the communication demand, and the program code containing the communication function and the computing task function is compiled to obtain the device-side program data, and after the device-side obtains the device-side program code sent by the host-side computing engine, the device-side computing engine completes the computing task and the communication task in the process of executing the device-side program code, and sends the calculation result to the target receiving end, so as to realize the communication between the computing engine in the host-side and the computing engine in the device-side, and the computing engine in the device-side and the computing engine in the target receiving end, without the need for the host-side computing engine to allocate the communication task, releasing the computing power of the host-side computing engine, and adapting to the communication scenarios of different computing engines in the same node and computing engines between different nodes.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

FIG1 is a flow chart of a computing engine communication method disclosed in the present application;

FIG2 is a schematic diagram of a specific distributed communication engine architecture disclosed in this application;

FIG3 is a flow chart of another computing engine communication method disclosed in the present application;

FIG4 is a schematic diagram of the structure of a computing engine communication device disclosed in the present application;

FIG5 is a schematic diagram of the structure of another computing engine communication device disclosed in the present application;

FIG6 is a structural block diagram of an electronic device disclosed in the present application;

FIG. 7 is a structural block diagram of a non-volatile computer-readable storage medium disclosed in the present application.

Detailed ways

The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

Due to the existence of historical legacy issues of network software protocol stacks and network hardware devices in traditional computer architecture systems, in a system with multiple heterogeneous and multi-computing engines, the engine chips, PCIE or local interconnection buses, and network cards of various computing nodes are independent of each other, and the communication efficiency and computing performance become very inefficient, and the degree of freedom and flexibility are not high, which makes the current large-scale computing power network not only have the problem of computing device computing performance, but also face the bottleneck caused by the inefficient communication between computing engines of different architectures. Take the common traditional communication between CPU and GPU as an example (other various xPUs are similar), inside a single device, when data needs to be transmitted between the application memory space and the GPU memory space, it needs to cross the system kernel space and the PCIE bus, that is, it needs to be copied once more, which is generally done by the CPU. When data needs to be transferred between the application memory space of device A and device B, the data first needs to be copied from the application memory space of device A to the system kernel space of device A, and then sent from the system kernel space of device A to the network device space of device A. After receiving the data, device B needs to send it from the network device space of device B to the system kernel space of device B, and then copy it from the system kernel space of device B to the application memory space of device B. When data needs to be transferred between the GPU memory space of device A and the GPU memory space of device B, the data must at least first be copied from the GPU memory space of device A to the system kernel space of device A, and then sent from the system kernel space of device A to the network device space of device A. After receiving the data, device B needs to send it from the network device space of device B to the system kernel space of device B, and then copy it from the system kernel space of device B to the GPU memory space of device B. Under normal circumstances, four data transmissions and copies are generally required. It can be seen that the communication architecture of traditional computing networks is an important factor causing the inefficiency and low speed of heterogeneous computing networks. It can be seen that in the non-traditional solution, the repeated and unnecessary data copying process between computing engines of different architectures leads to inefficiency problems. Moreover, although there are multiple computing engines in a single device, the data transmission control is completely dominated by the CPU, and the GPU only acts as a coprocessor of the system to share part of the computing work. There are two key points to solving the above problems. The first is to reduce the number of unnecessary cross-chip data copies as much as possible. The second is to give full play to the subjective initiative of each computing engine in transmitting data. Each computing engine is no longer a coprocessor. Each computing engine can freely communicate with any other type of computing engine, freeing up the CPU's computing power to serve the real core business.

At present, in order to solve the above problems, some technical solutions have emerged:

For example, for the communication between GPU and CPU within a single node, NVIDIA launched GPU Direct Shared Memory (i.e., direct shared memory) technology in 2010, and gradually developed GPU Direct P2P (Peer-to-Peer), NVLink, and the latest NVSwitch technology. Among them, NVLink is a set of bus protocols developed by NVIDIA to solve the limitation of PCIe transmission rate when transmitting data between GPUs in a single node. Based on this bus protocol, the corresponding hardware bus implementation is also developed for physical connection; however, since this method is mainly used to solve the problem of one-to-one high-speed interconnection between GPUs within a single node, it is highly dependent on circuit customization and is powerless for communication between GPUs across nodes; and from the implementation method of NVLink, it can be seen that its data transmission depends on the direct circuit physical link binding between GPUs, and it is completely unable to communicate over long distances. NVSwitch is an independent communication chip designed by NVIDIA. It integrates multiple NVLinks to achieve many-to-many GPU communication within a single node at the higher speed of NVLink, thereby further improving the interconnection performance; for example, 12 NVSwitches connect 16 GPUs to achieve 0-hop arbitrary interconnection between these GPUs. NVSwitch is a dedicated data chip based on NVLink, so the limitations of NVLink still exist on NVSwitch. The node machines are completely customized, and the nodes are connected in close proximity, so they cannot be deployed in a distributed manner.

For the communication between GPUs in different nodes, the current solutions are basically implemented through RDMA network. Through RDMA network, the hardware network card can directly read the data to be sent in the user space without going through the system kernel space. The number of data copies can be reduced, thus greatly reducing network latency and CPU overhead on data copying and increasing transmission rates. Currently, there are solutions for multi-node GPU communication based on RDMA (Remote Direct Memory Access). In 2017, NVIDIA released the latest GPU Direct RDMA Async technology, which allows direct synchronization between GPU and third-party devices, while the CPU does not participate in the key communication path of GPU applications. Data is sent directly from GPU memory to RDMA NIC (Network Interface Controller), and the peer RDMA NIC receives it and directly transmits it to GPU memory, reducing CPU participation and the number of data copies. The working process is as follows: 1) The CPU dispatches computing and communication tasks to the GPU through NVIDIA's CUDA API (Application Programming Interface); 2) After the GPU completes the computing task, it automatically executes the communication task and directly triggers the communication operation to the RDMA NIC; 3) The RDMA NIC copies data directly from the GPU's memory or the host's memory to the RDMA NIC through the RDMA network; 4) The RDMA NIC sends data; From the process of this technology, it can be seen that except for the stage where the host starts to distribute computing and communication tasks, the rest of the process does not require the host to participate. In the process of sending data in step 3, the data is only copied once. However, the problem with this solution is that the GPU and RDMA NIC are two independent devices. The GPU needs to trigger the data movement of the RDMA NIC through PCIe. As a physical link, PCIe makes the GPU and RDMA NIC have certain physical distance restrictions and bindings. Although the data transmission reduces the number of temporary storage times, the physical PCIE cross-node signal transmission path does not change. And as a closed-source solution of a commercial company, this technical solution is difficult to modify again according to actual conditions.

In addition to GPU, another common heterogeneous computing engine is FPGA. For FPGA-based computing engines, Intel has developed IKL (Inter-Kernel Links) to achieve heterogeneous communication between FPGAs within a node or between nodes. User Kernel is a module written by the user in OpenCL language to implement specific computing tasks. When writing User Kernel, the following two main functions are mainly used: write_channel_intel(channel_id,data) and read_channel_intel(channel_id) to achieve communication between different FPGAs. These two functions will send the data to be sent from User Kernel through IKL I/O Channel (i.e. channel) to Inter-Kernel Logic RTL IP, which will package the data and finally forward it to Ethernet Switch, i.e. network card. From the way IKL realizes communication between FPGAs, it can be seen that it has the following main disadvantages: 1) The write_channel_intel and read_channel_intel functions used to send and receive data in OpenCL are based on channel_id, which is cumbersome to configure and each channel only supports fixed point-to-point communication; 2) In order to achieve reliable communication, Inter-Kernel Logic RTL IP needs to implement complex control functions such as timeout retransmission, fragmentation, and packet loss retransmission similar to the TCP/IP protocol stack, which occupies a large amount of FPGA resources, and the occupied resources increase with the number of channel_ids. At present, according to different FPGA board models, the maximum number of channel_ids supported is 48 to 256; This communication method only supports network communication between FPGAs, but not between FPGAs and the host. When communicating between FPGAs, it only supports exchanging data on the FPGA board, and cannot exchange data in the host memory;

Based on the above content, it can be found that the current main technical solutions have the following main shortcomings: 1) For the communication between computing engines within a single node, it is all achieved through traditional PCIe or NVLink-like methods with physical link distance limitations, and cannot be deployed in a large-scale distributed manner; 2) For the communication between computing engines between nodes, even GPU Direct RDMA Async technology requires the host to allocate communication tasks in the first step. In addition, it fails to completely decouple the physical link between the GPU and the RDMA NIC, and still needs to trigger data movement through the PCIe link. When deploying such GPUs, it takes up a large physical space and consumes higher energy; while the FPGA-based Intel IKL is more inclined to implement a point-to-point transmission between devices of the same type, and due to the number limit of channle_id, the number of objects that a device can exchange data is severely limited, the transmission is limited and the efficiency is extremely poor. In addition, this solution does not support the NIC in the node to directly read the local The memory of the host end within the node is sent to the other end; 3) Regardless of whether it is a solution based on a GPU or FPGA computing engine, most of them are closed-source, specialized solutions that are difficult to modify and lack a certain degree of versatility.

As shown in FIG. 1 , an embodiment of the present application discloses a computing engine communication method, which is applied to a first device end and includes:

Step S11: Obtain device-side program data sent by the host-side computing engine; wherein the device-side program data is program data obtained by compiling program codes including communication functions and computing task functions, and the communication function is a function created based on communication requirements.

Among them, the host-side computing engine is the CPU, and, in the embodiment of the present application, a specific computing task is described in the OpenCL language to obtain a computing task function. In addition, the difference between the device-side program data in the embodiment of the present application and the ordinary device-side OpenCL program is that a custom communication function library is integrated. When writing code, the functions in the custom communication function library add corresponding communication functions to the OpenCL code of the cl file according to the communication requirements of the task, and are compiled together with the cl file to form the final device-side binary code, that is, the device-side program data.

Furthermore, the embodiment of the present application can obtain the device-side program data sent by the host-side computing engine through the RDMA module; wherein the RDMA module and the device-side computing engine are integrated in the same chip. In this way, by integrating the module responsible for RDMA network communication and the computing engine into one chip, the connection between the computing engine and the communication engine is further accelerated, eliminating the problem of physical link distance limitation in the prior art, and facilitating large-scale distributed deployment. The device-side computing engine is an xPU (i.e., a general computing engine), a general term for various computing engines, including GPU, NPU (i.e., Neural-network Processing Unit), DPU (i.e., Data Processing Unit), and IPU (i.e., Infrastructure Processing Unit). The embodiment of the present application can integrate the RDMA module and the computing engine in the same FPGA chip.

Furthermore, in one implementation, the target data, parameter information, and device-side program data sent by the host-side computing engine can be obtained through the RDMA module; wherein the parameter information includes the first address and length information of the target data written into the memory of the first device; and the target data is written into the memory based on the parameter information. The memory can be DDR (i.e., DDR SDRAM, Double Data Rate Synchronous Dynamic Random Access Memory).

Step S12: Utilize the device-side computing engine to execute the device-side program data, calculate the target data based on the computing task function to obtain the calculation result, and send the calculation result to the target receiving end based on the communication function to realize the communication between the device-side computing engine and the computing engine in the target receiving end.

In one implementation, the device-side computing engine can be used to execute the device-side program data, read the target data from the memory based on the parameter information, and calculate the target data based on the computing task function to obtain the calculation result. Based on the communication function, the calculation result is sent to the target receiving end through the RDMA module.

Among them, based on the communication function, the calculation result is sent to the target receiving end through the RDMA module, which specifically includes the following steps:

Step 00: Write the identification information of the target receiving end into the communication table based on the communication function, and notify the table parsing engine.

Among them, the table parsing engine and the RDMA module are both underlying hardware IP (i.e., intellectual property). In one implementation, a custom IP can be written in HDL (i.e., Hardware Description Language) to obtain the table parsing engine and the RDMA module. It can be seen that in this application, the custom communication library of the upper-layer software can be implemented based on the common C/C++ language, while the underlying hardware IP is implemented through HDL code. The solution provided in the embodiment of this application has good versatility and portability.

Furthermore, in the embodiment of the present application, when creating a communication function, a corresponding target receiving end can be specified, and specifically, the target receiving end can be specified based on the identification information of the device.

Step 01: Obtain the identification information of the target receiving end from the communication table through the table parsing engine, and obtain the IP address corresponding to the identification information.

In addition, the embodiment of the present application also writes the calculation result into the data area in the memory, and writes the write address and data length into the communication table; obtains the write address and data length from the communication table through the table parsing engine, and writes the write address, data length and IP address into the memory in the RDMA module. Among them, the memory can be a FIFO (First Input First Output) memory.

Among them, the embodiment of the present application can obtain the IP address corresponding to the identification information from the IP identification comparison table.

Further, in a specific implementation, during the driver initialization phase, a discovery message can be sent to the host side through the RDMA module, so that the host side can assign IP addresses and identification information to each device side based on the discovery message of each device side; obtain the IP address and identification information of each device side sent by the host side, and the IP address and identification information of the host side; save the IP address and identification information of each device side, and the IP address and identification information of the host side to the IP identification comparison table. Among them, after sending a discovery message to the host side through the RDMA module, the IP address and identification information assigned by the host side to the first device side can be obtained through the RDMA module, and confirmation information can be replied to the host side, so that after receiving the confirmation information replied by each device side, the host side saves the IP address and identification information of each device side, and sends the IP address and identification information of each device side, and the IP address and identification information of the host side to each device side. Among them, the confirmation information carries the IP address and identification information assigned by the host side to the corresponding device side.

It should be pointed out that when powered on, each device is actually a DHCP (Dynamic Host Configuration Protocol) client, that is, the RDMA module will send DHCP discovery information to find a DHCP server, that is, send specific broadcast information to the broadcast IP address 255.255.255.255. The host side, as a DHCP server, will receive DHCP discovery messages from each device side and assign the device side's IP address, device side identification and other information to each device side. After receiving the information assigned to it by the host side, the RDMA module of each device side will reply with a confirmation message, which will also contain all the information assigned to it by the host side, in order to declare to the host side that the device side will use this information for communication. After receiving the confirmation information from each device side, the host side will send the information assigned to all device sides and the IP and identification information of the host side in a custom format to each device side, and also keep a copy in the host side. After receiving the communication information of all devices sent by the host side, the RDMA module of each device side saves it in the IP identification comparison table.

Furthermore, in an embodiment of the present application, a node may include a host end and at least one device end. If the same network includes multiple nodes, the host end in a node can be determined to allocate IP addresses and identification information, that is, allocate IP addresses and identification information of other host ends and all device ends in the network. Ultimately, each host end and each device end has an IP identification comparison table that includes the IP addresses and identification information of all host ends and all device ends in the entire network.

In this way, any device in the same network can automatically discover and obtain the communication information of all devices in the network without human intervention. It is scalable and flexible, and provides a physical basis for realizing the expansion of arbitrary distributed computing power networks with high degrees of freedom.

Furthermore, in the embodiment of the present application, the addresses of the IP identification comparison table and the information table are determined, and the device-side program data and the host-side program data are written, wherein the host-side program data is the program data executed on the host side. In one embodiment, the program mainly written in C/C++ uses the C/C++ standard library and the API function of OpenCL on the host side to obtain the final host-side program data.

Step 02: Send the calculation results to the target receiving end based on the IP address through the RDMA module.

In a specific implementation, the memory can be detected through the RDMA module, and when the memory is not empty, the write address, data length and IP address are read from the memory, the calculation result is read from the memory according to the write address and data length, and the calculation result is sent to the target receiving end based on the IP address.

Among them, the target receiving end is the host end or the second device end; if the target receiving end is multiple second device ends, the identification information of the multiple second device ends is written into the communication table based on the communication function, and the calculation results are sent to the multiple second device ends through the table parsing engine.

Furthermore, if the target receiving end is the second device end, the second device end receives the calculation result through its own RDMA module and saves the calculation result in its own memory. Moreover, after the first device end sends the calculation result to the second device end based on the communication function, a transmission end message can also be sent to the second device end, so that after receiving the transmission end message, the RDMA module of the second device end returns the first address and length information of the calculation result in the memory to the calculation engine of the second device end, so that the calculation engine of the second device end can perform subsequent operations according to the first address and length information of the calculation result in the memory.

Furthermore, when the first device executes the device program data, if the target program instruction is executed, the first device waits to receive the data from the third device, and after receiving the data from the third device, continues to perform subsequent operations. The target program instruction is a program instruction compiled based on the communication function. It is understandable that the multiple second device also executes the corresponding program instruction and waits to receive the calculation result of the first device.

It can be understood that the embodiments of the present application can provide a communication solution for active interconnection between computing engines of any different architectures, eliminate the distinction between inter-node and intra-node communications, and do not require the CPU to allocate communication tasks, and have one-to-many communication capabilities. Whether it is between XPUs or between XPUs and the host, data only needs to be moved once during the data transmission process, which has higher efficiency and effectiveness.

For example, as shown in FIG. 2, FIG. 2 is a schematic diagram of a specific distributed communication engine architecture provided by an embodiment of the present application. Taking a fully software programmable hardware FPGA device as an example, there are two computing engines in the architecture-the ordinary CPU on the host side, and the GPU based on the RISC-V processor extension function in the acceleration card (i.e., the device side). Among them, the acceleration card is an FPGA board, which includes an FPGA chip, the RDMA module and the GPU are integrated into the chip, and in addition to the FPGA chip, there are a series of peripherals, such as network ports, DDR, etc., and there is no master-slave relationship between the host side and the device side. The host-side program is a program mainly written in C/C++, in which the C/C++ standard library and the API functions of OpenCL on the host side are used; the device-side program is mainly a specific computing task described in the OpenCL language, and a custom communication function library is integrated. The functions in the custom communication function library, when writing the code, add the corresponding communication function to the OpenCL code of the cl file according to the communication requirements of the task, and compile it together with the cl file to form the final binary code of the device side. The upper CL compiler and the xPU chip jointly determine the addresses of the communication table and the IP identification comparison table of both parties, and write the above binary code. That is, the addresses of the two-way communication table and IP identification comparison table required for communication between the host and the device are determined in advance, and the host program data and the device program data are written. In addition, custom IPs are written in HDL language, including RDMA, table parsing engine, etc., and customized communication function libraries implemented in the upper layer C/C++ language and some standard protocol libraries are used to realize any interconnection between any computing engines based on the public Ethernet protocol, without the need for the host to control the communication process.

Furthermore, the main working process of the above distributed communication engine architecture includes:

(1) Driver initialization phase: The RDMA module implements the functions of the RDMA NIC and automatically completes the following tasks: a) When powered on, each accelerator card is actually a DHCP client, that is, the RDMA module will send a DHCP discovery message to find a DHCP server, that is, send a specific broadcast message to the broadcast IP address 255.255.255.255; b) The host side acts as The DHCP server will receive the DHCP discovery message from each accelerator card and assign the IP address, device ID and other information of the accelerator card to each accelerator card; c) After receiving the information assigned to it by the host, the RDMA module of each accelerator card will reply with a confirmation message, which will also contain all the information assigned to it in b) to declare to the host that the accelerator card will use this information for communication; d) After receiving the confirmation information from each accelerator card, the host will send the information assigned to all accelerator cards and the IP and ID information of the host in a custom format to each accelerator card, and also keep a copy in the host; e) After receiving the communication information of all devices sent by the host in the previous step, the RDMA module of each accelerator card saves it in the IP identification comparison table.

(2) The host downloads relevant data to the accelerator card: The compiled host binary program is executed on the host. A series of OpenCL API functions in the program will perform a series of tasks such as searching for devices and initializing device parameters. In addition, the program will also send the following three types of data to the accelerator card: initial data required for calculation, some parameters (such as the first address and length of the initial data in the accelerator card DDR, and the addresses of these parameters in the accelerator card DDR are fixed), and the compiled device binary program data on the disk (the first address of this data in the accelerator card DDR is also fixed and has nothing to do with the specific task). The host sends this data to the accelerator card through RDMA Ethernet.

(3) There are two scenarios for communication between the accelerator card and the host:

a) The accelerator card reads data from the host: Usually, in the above (2), the host has downloaded the relevant data to the accelerator card. The data required for the GPU calculation in the accelerator card has been downloaded to the DDR of the accelerator card, and the GPU does not need to read data from the host.

b) The accelerator card writes data to the host: After the GPU in the accelerator card completes the calculation, it writes the calculation result into the data area in the accelerator card DDR, and saves the written address, length, and ID of the other end (i.e., the target receiving end) in the communication table; then the GPU notifies the table parsing engine, which takes out information from the communication table and marks the information that has been taken out from the communication table to prevent repeated retrieval; the table parsing engine queries the IP identification comparison table for the IP address corresponding to the ID of the other end in the obtained information; after the table parsing engine obtains the IP address corresponding to the ID, it writes this information into a FIFO in the RDMA module; when the RDMA module detects that the FIFO is not empty, it will continuously take out the information in it, read the corresponding data from the accelerator card DDR according to the accelerator card DDR address and length in the information, and then send the data to the other end according to the IP address; after the data is sent, the accelerator card also sends a transmission completion message to the host.

(4) The reading and writing between accelerator cards is different from the reading and writing between the accelerator card and the host. The former occurs in pairs, that is, when one end performs a write operation, the other end or multiple ends must perform a corresponding read operation: suppose that accelerator card 2 and accelerator card 3 need to read the data of accelerator card 1, that is, after the GPUs in accelerator card 2 and accelerator card 3 perform tasks to a certain stage, they may need the data of accelerator card 1. The fundamental reason for this situation is that the GPUs in these two accelerator cards have executed the program instructions compiled by the read function in the custom communication function library. The program instructions indicate that the GPU needs to wait to receive data from Device ID = 1 (that is, the device ID of the accelerator card is 1) before continuing to execute subsequent tasks.

a) In accelerator card 1, after GPU completes calculation, it writes the calculation result into the data area in accelerator card 1 DDR, and saves the written address, length, peer ID, etc. in the communication table with fixed address in DDR. Since there are two device ends that need to be sent, two pieces of information with different peer IDs are written into the communication table here, and then GPU notifies the table parsing engine, and the table parsing engine takes out the two pieces of information from the communication table; the table parsing engine queries the IP identification comparison table for the IP address corresponding to the two peer IDs (i.e., device ID = 2 and 3) in the obtained information, and after the table parsing engine obtains the IP addresses corresponding to the two IDs, it writes the two pieces of information and its IP into a FIFO in the RDMA module; when the RDMA module detects that the FIFO is not empty, it will continuously take out the information therein, read the corresponding data from the accelerator card DDR according to the accelerator card DDR address and length in the information, and then send the data to the peer device (i.e., accelerator card 2 and accelerator card 3) according to the IP address; in data sending After completion, a data transmission completion message is sent to accelerator card 2 and accelerator card 3 respectively;

b) In accelerator card 2 and accelerator card 3, after their respective GPUs execute the instructions generated by the compiled read function, they will wait for the device ID given in the instructions to send data. When the RDMA modules of accelerator card 2 and accelerator card 3 receive the data sent by accelerator card 1, they will save it in their respective DDRs. After receiving the transmission completion information sent by accelerator card 1, they will return the first address and length of the received data in DDR to their respective GPUs. The GPUs can perform subsequent operations based on the first address and length of the data.

It should be pointed out that in other communication engine architectures, the following situation may occur: since accelerator card 1 and accelerator card 2 cannot communicate directly, accelerator card 1 will send the intermediate result calculated by itself back to a certain address in the host side, and then the host side will notify accelerator card 2 to let accelerator card 2 read the data at the address in the host side; since accelerator cards can communicate with each other in this communication architecture, this situation will not exist.

In this way, a custom communication function library of the upper-level software written in a high-level language and the underlying xPU hardware IP written in HDL language are used, based on the physical FPGA chip as an example, to realize an efficient and universal communication engine between computing engines of different architectures. According to the specific type of the underlying xPU (GPU based on RISC-V extension in this case), a custom communication function library is written in a standard high-level language, combined with the RDMA module, table parsing engine, etc. written in HDL language, and the communication engine integrates the module responsible for RDMA network communication with the computing engine into one chip, further accelerating the connection between the computing engine and the communication engine, and realizing the active interconnection between computing engines of any different architectures in the same RDMA network, completely eliminating the need for CPU control to participate in communication control, and having one-to-many communication capabilities; whether it is between accelerator cards and accelerator cards or between accelerator cards and host terminals, only data needs to be moved once during data transmission, effectively solving the problem of inefficient and low-speed communication between computing engines. The solution provided in this application is applicable to various xPU computing engines, such as GPU, NPU, IPU, DPU, VPU, etc.

Further, referring to FIG. 3 , an embodiment of the present application discloses a computing engine communication method, which is applied to a host side and includes:

Step S31, sending device-side program data to the device side through the host-side computing engine, so that the device-side computing engine of the device side executes the device-side program data, calculates the target data based on the computing task function to obtain a calculation result, and sends the calculation result to the target receiving end based on the communication function;

The device-side program data is program data obtained by compiling program codes including communication functions and computing task functions, and the communication function is a function created based on communication requirements.

In some embodiments, sending device-side program data to the device-side through a host-side computing engine includes:

The target data, parameter information and device-side program data are sent to the device-side through the host-side computing engine; wherein the parameter information includes the first address and length information of the target data written into the device-side memory, so that the device-side writes the target data into the memory based on the parameter information.

In some embodiments, it also includes:

Get the discovery message sent by each device;

Based on the discovery messages sent by each device, configure an IP address and identification information for each device;

The IP address and identification information of each device end and the IP address and identification information of the host end are sent to each device end, so that each device end saves the IP address and identification information of each device end and the IP address and identification information of the host end into the IP identification comparison table.

In some embodiments, sending the IP address and identification information of each device end and the IP address and identification information of the host end to each device end further includes:

Send the IP address and identification information assigned by the host to each device;

After receiving the confirmation information replied by each device end, the IP address and identification information of each device end are saved, and the IP address and identification information of each device end and the IP address and identification information of the host end are sent to each device end.

It can be seen that in the present application, a communication function is created based on the communication needs, and the program code containing the communication function and the computing task function is compiled to obtain the device-side program data. After the device side obtains the device-side program code sent by the host-side computing engine, the device-side computing engine completes the computing task and the communication task in the process of executing the device-side program code, and sends the computing result to the target receiving end. In this way, the communication between the computing engine in the host side and the computing engine in the device side, and between the computing engine in the device side and the computing engine in the target receiving end is realized. There is no need for the host computing engine to allocate communication tasks, thus releasing the computing power of the host-side computing engine, and it is suitable for communication scenarios between different computing engines in the same node and between computing engines in different nodes.

As shown in FIG. 4 , an embodiment of the present application discloses a computing engine communication device, which is applied to a first device end and includes:

The communication engine 11 is used to obtain the device-side program data sent by the host-side computing engine; wherein the device-side program data is program data compiled from program codes including communication functions and computing task functions, and the communication function is a function created based on communication requirements;

The device-side computing engine 12 is used to execute device-side program data, calculate the target data based on the computing task function to obtain the calculation result, and send the calculation result to the target receiving end based on the communication function.

It can be seen that the present application first obtains the device-side program data sent by the host-side computing engine; wherein the device-side program data is the program data obtained by compiling the program code containing the communication function and the computing task function, and the communication function is a function created based on the communication demand, and then the device-side computing engine is used to execute the device-side program data, and the target data is calculated based on the computing task function to obtain the calculation result, and the calculation result is sent to the target receiving end based on the communication function, so as to realize the communication between the device-side computing engine and the computing engine in the target receiving end. That is, in the present application, a communication function is created based on the communication demand, and the program code containing the communication function and the computing task function is compiled to obtain the device-side program data, and after the device-side obtains the device-side program code sent by the host-side computing engine, the device-side computing engine completes the computing task and the communication task in the process of executing the device-side program code, and sends the calculation result to the target receiving end, so as to realize the communication between the computing engine in the host-side and the computing engine in the device-side, and the computing engine in the device-side and the computing engine in the target receiving end, without the need for the host-side computing engine to allocate the communication task, releasing the computing power of the host-side computing engine, and adapting to the communication scenarios of different computing engines in the same node and computing engines between different nodes.

The communication engine 11 is an RDMA module, and the RDMA module and the device-side computing engine are integrated in the same chip;

Correspondingly, the device-side computing engine 12 is specifically used to send the computing result to the target receiving end based on the communication function and through the RDMA module.

Further, the RDMA module is specifically used to obtain target data, parameter information and device-side program data sent by the host-side computing engine; wherein the parameter information includes the first address and length information of the target data written into the memory of the first device; and write the target data into the memory based on the parameter information;

Correspondingly, the device-side computing engine 12 is specifically used to execute device-side program data using the device-side computing engine, read target data from the memory based on parameter information, and calculate the target data based on the computing task function to obtain a calculation result.

Further, the device-side computing engine 12 writes the identification information of the target receiving end into the communication table based on the communication function, and notifies the table parsing engine; obtains the identification information of the target receiving end from the communication table through the table parsing engine, and obtains the identification information The corresponding IP address; the RDMA module is used to send the calculation results to the target receiving end based on the IP address.

Furthermore, the RDMA module is also used to write the calculation result into the data area in the memory, and write the write address and data length into the communication table; correspondingly, the device-side calculation engine 12 is used to obtain the write address and data length from the communication table through the table parsing engine, and write the write address, data length and IP address into the memory in the RDMA module. The RDMA module is used to detect the memory, and when the memory is not empty, read the write address, data length and IP address from the memory, read the calculation result from the memory according to the write address and data length, and send the calculation result to the target receiving end based on the IP address.

Further, obtaining the IP address corresponding to the identification information includes: obtaining the IP address corresponding to the identification information from an IP identification comparison table.

Correspondingly, the RDMA module is also used to send a discovery message to the host side, so that the host side can assign an IP address and identification information to each device side based on the discovery message of each device side; obtain the IP address and identification information of each device side and the IP address and identification information of the host side sent by the host side; save the IP address and identification information of each device side and the IP address and identification information of the host side to the IP identification comparison table.

Among them, the RDMA module is also used to obtain the IP address and identification information assigned by the host side to the first device side after sending a discovery message to the host side, and reply confirmation information to the host side, so that after the host side receives the confirmation information replied by each device side, it saves the IP address and identification information of each device side, and sends the IP address and identification information of each device side and the IP address and identification information of the host side to each device side.

Further, the target receiving end is a host end or a second device end;

If the target receiving end is multiple second device ends, the device-end computing engine writes the identification information of the multiple second device ends into the communication table based on the communication function.

Furthermore, if the target receiving end is the second device end, the second device end receives the calculation result through its own RDMA module and stores the calculation result in its own memory.

Furthermore, the RDMA module is also used to send transmission end information to the second device end after sending the calculation result to the second device end based on the communication function, so that after receiving the transmission end information, the RDMA module on the second device end returns the starting address and length information of the calculation result in the memory to the computing engine on the second device end.

Furthermore, when executing the device-side program data, if the device-side computing engine executes the target program instruction, it waits to receive the data from the third device side, and continues to execute subsequent operations after receiving the data from the third device side.

Further, referring to FIG. 5 , an embodiment of the present application discloses a computing engine communication device, which is applied to a host side and includes:

The host-side computing engine 51 is used to send the device-side program data to the device-side, so that the device-side computing engine of the device-side executes the device-side program data, calculates the target data based on the computing task function to obtain the calculation result, and sends the calculation result to the target receiving end based on the communication function;

The device-side program data is program data obtained by compiling program codes including communication functions and computing task functions, and the communication function is a function created based on communication requirements.

In some embodiments, the host-side computing engine is specifically used to send target data, parameter information and device-side program data to the device-side; wherein the parameter information includes the starting address and length information of the target data written into the device-side memory, so that the device-side writes the target data into the memory based on the parameter information.

In some embodiments, the apparatus further comprises:

A discovery message acquisition module is used to acquire discovery messages sent by each device end;

A configuration module, used to configure an IP address and identification information for each device end based on the discovery message sent by each device end;

The configuration information sending module is used to send the IP address and identification information of each device end and the IP address and identification information of the host end to each device end, so that each device end saves the IP address and identification information of each device end and the IP address and identification information of the host end to the IP identification comparison table.

In some embodiments, the configuration information sending module further includes:

The first sending module sends the IP address and identification information assigned by the host end to each device end in English;

The second sending module is used to save the IP address and identification information of each device end after receiving the confirmation information replied by each device end, and send the IP address and identification information of each device end and the IP address and identification information of the host end to each device end.

Referring to FIG. 6 , an embodiment of the present application further provides an electronic device, including: a memory 601 for storing a computer program; and a processor 602 for executing the computer program to implement the steps of the computing engine communication method of any of the above embodiments.

Referring to FIG. 7 , an embodiment of the present application further provides a non-volatile computer-readable storage medium 7 on which a computer program 701 is stored. When the computer program 701 is executed by a processor, the steps of the computing engine communication method of any of the above embodiments are implemented.

Since the embodiments of the electronic device and the non-volatile computer-readable storage medium part correspond to the embodiments of the computing engine communication method part, please refer to the description of the embodiments of the computing engine communication method part for the embodiments of the non-volatile computer-readable storage medium part, which will not be repeated here.

In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

The computing engine communication method and device, electronic device, and storage medium provided by the present application are introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the idea of the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims (21)

A computing engine communication method, characterized in that it is applied to a first device end, comprising: Acquire device-side program data sent by a host-side computing engine; wherein the device-side program data is program data obtained by compiling a program code including a communication function and a computing task function, and the communication function is a function created based on communication requirements; The device-side computing engine is used to execute the device-side program data, the target data is calculated based on the computing task function to obtain a calculation result, and the calculation result is sent to the target receiving end based on the communication function to realize communication between the device-side computing engine and the computing engine in the target receiving end. The computing engine communication method according to claim 1, wherein obtaining the device-side program data sent by the host-side computing engine comprises: Acquire device-side program data sent by a host-side computing engine through an RDMA module; wherein the RDMA module and the device-side computing engine are integrated in the same chip; Correspondingly, sending the calculation result to the target receiving end based on the communication function includes: sending the calculation result to the target receiving end based on the communication function and through the RDMA module. The computing engine communication method according to claim 2, characterized in that the obtaining of the device-side program data sent by the host-side computing engine through the RDMA module comprises: Acquire target data, parameter information, and device-side program data sent by the host-side computing engine through the RDMA module; wherein the parameter information includes the first address and length information of the target data written into the memory of the first device side; Writing the target data into the memory based on the parameter information; Accordingly, the using of the device-side computing engine to execute the device-side program data and computing the target data based on the computing task function to obtain a computing result includes: The device-side program data is executed using a device-side computing engine, the target data is read from the memory based on the parameter information, and the target data is calculated based on the computing task function to obtain a calculation result. The computing engine communication method according to claim 2, characterized in that the step of sending the computing result to the target receiving end based on the communication function and through the RDMA module comprises: Writing the identification information of the target receiving end into a communication table based on the communication function, and notifying a table parsing engine; Acquire identification information of the target receiving end from the communication table through the table parsing engine, and acquire an IP address corresponding to the identification information; The calculation result is sent to a target receiving end through the RDMA module based on the IP address. The computing engine communication method according to claim 4, further comprising: Writing the calculation result into the data area in the memory, and writing the write address and data length into the communication table; The write address and the data length are acquired from the communication table by the table parsing engine, and the write address, the data length and the IP address are written into the memory in the RDMA module. The computing engine communication method according to claim 5, characterized in that the sending of the computing result to the target receiving end based on the IP address through the RDMA module comprises: The memory is detected by the RDMA module, and when the memory is not empty, the write address, the data length and the IP address are read from the memory, and the write address and the data length are read according to the write address and the data length. The calculation result is read from the memory at a certain time, and the calculation result is sent to a target receiving end based on the IP address. The computing engine communication method according to claim 4, characterized in that the step of obtaining the IP address corresponding to the identification information comprises: Obtain the IP address corresponding to the identification information from the IP identification comparison table. The computing engine communication method according to claim 7, further comprising: Sending a discovery message to the host end through the RDMA module, so that the host end allocates an IP address and identification information to each device end based on the discovery message of each device end; Obtaining the IP addresses and identification information of the devices and the IP address and identification information of the host sent by the host; The IP addresses and identification information of each device end and the IP address and identification information of the host end are saved in the IP identification comparison table. The computing engine communication method according to claim 8, characterized in that after sending the discovery message to the host side through the RDMA module, it also includes: Through the RDMA module, the IP address and identification information assigned by the host end to the first device end are obtained, and confirmation information is replied to the host end, so that after receiving the confirmation information replied by each device end, the host end saves the IP address and identification information of each device end, and sends the IP address and identification information of each device end and the IP address and identification information of the host end to each device end. The computing engine communication method according to claim 4, characterized in that the target receiving end is a host end or a second device end; If the target receiving end is a plurality of second device ends, writing the identification information of the target receiving end into a communication table based on the communication function includes: Based on the communication function, the identification information of the plurality of second device terminals are all written into the communication table. The computing engine communication method according to claim 5 is characterized in that if the target receiving end is a second device end, the second device end receives the calculation result through its own RDMA module and saves the calculation result in its own memory. The computing engine communication method according to claim 11, characterized in that after sending the computing result to the second device end based on the communication function, it also includes: Sending transmission end information to the second device end, so that the RDMA module of the second device end returns the first address and length information of the calculation result in the memory to the calculation engine of the second device end after receiving the transmission end information. The computing engine communication method according to claims 1 to 12, characterized in that it also includes: During the process of executing the device-side program data, if the target program instruction is executed, the device waits for receiving data from the third device side, and after receiving the data from the third device side, continues to execute subsequent operations. A computing engine communication method, characterized in that it is applied to a host side and comprises: Sending device-side program data to the device side through the host-side computing engine, so that the device-side computing engine of the device side executes the device-side program data, calculates the target data based on the computing task function to obtain a calculation result, and sends the calculation result to the target receiving end based on the communication function; The device-side program data is program data obtained by compiling program codes including communication functions and computing task functions, and the communication function is a function created based on communication requirements. The computing engine communication method according to claim 14, wherein the sending of device-side program data to the device-side through the host-side computing engine comprises: The target data, parameter information and device-side program data are sent to the device-side through the host-side computing engine; wherein the parameter information includes the starting address and length information of the target data written into the device-side memory, so that the device-side writes the target data into the memory based on the parameter information. The computing engine communication method according to claim 14, further comprising: Get the discovery message sent by each device; Based on the discovery messages sent by each device, configure an IP address and identification information for each device; The IP address and identification information of each device end and the IP address and identification information of the host end are sent to each device end, so that each device end saves the IP address and identification information of each device end and the IP address and identification information of the host end to the IP identification comparison table. The computing engine communication method according to claim 16, characterized in that the sending of the IP address and identification information of each device end and the IP address and identification information of the host end to each device end further comprises: Sending to each device the IP address and identification information assigned by the host end to it; After receiving the confirmation information replied by each device end, the IP address and identification information of each device end are saved, and the IP address and identification information of each device end and the IP address and identification information of the host end are sent to each device end. A computing engine communication device, characterized in that it is applied to a first device end, comprising: A communication engine, used to obtain device-side program data sent by a host-side computing engine; wherein the device-side program data is program data obtained by compiling a program code including a communication function and a computing task function, and the communication function is a function created based on communication requirements; The device-side computing engine is used to execute the device-side program data, calculate the target data based on the computing task function to obtain a calculation result, and send the calculation result to the target receiving end based on the communication function. A computing engine communication device, characterized in that it is applied to a host side and comprises: A host-side computing engine, configured to send device-side program data to a device-side, so that the device-side computing engine of the device-side executes the device-side program data, calculates the target data based on the computing task function to obtain a calculation result, and sends the calculation result to a target receiving end based on the communication function; The device-side program data is program data obtained by compiling program codes including communication functions and computing task functions, and the communication function is a function created based on communication requirements. An electronic device, comprising: Memory for storing computer programs; A processor, configured to execute the computer program to implement the computing engine communication method according to any one of claims 1 to 17. A non-volatile computer-readable storage medium, characterized in that a computer program is stored in the non-volatile computer-readable storage medium, and when the computer program is loaded and executed by a processor, the computing engine communication method according to any one of claims 1 to 17 is implemented.