CN111078291B

CN111078291B - Operation method, system and related product

Info

Publication number: CN111078291B
Application number: CN201811221778.XA
Authority: CN
Inventors: 不公告发明人
Original assignee: Cambricon Technologies Corp Ltd
Current assignee: Cambricon Technologies Corp Ltd
Priority date: 2018-10-19
Filing date: 2018-10-19
Publication date: 2021-02-09
Anticipated expiration: 2038-10-19
Also published as: CN111078291A

Abstract

The present disclosure relates to an operational method, system and related product. The system includes an instruction generating device and an executing device. The instruction generating device includes: the device determining module is used for calculating the macroinstruction according to the received matrix vector and determining the operation device for executing the matrix vector to calculate the macroinstruction; the instruction generating module is used for calculating the macro instruction and the operation equipment according to the matrix vector to generate an operation instruction. The operation device includes: the control module is used for acquiring required data, a neural network model and an operation instruction, analyzing the operation instruction and acquiring a plurality of analysis instructions; the execution module is used for executing the plurality of analysis instructions according to the data to obtain an execution result. The operation method, the operation system and the related products provided by the embodiment of the disclosure can be used in a cross-platform mode, and have the advantages of good applicability, high instruction conversion speed, high processing efficiency, low error probability and low development cost of manpower and material resources.

Description

Operation method, system and related product

Technical Field

The present disclosure relates to the field of information processing technologies, and in particular, to a method and a system for processing a matrix vector calculation instruction, and a related product.

Background

With the continuous development of science and technology, neural network algorithms are more and more widely used. The method is well applied to the fields of image recognition, voice recognition, natural language processing and the like. However, as the complexity of neural network algorithms is higher and higher, the scale of the neural network algorithms is continuously increased. A large-scale neural network model based on a Graphics Processing Unit (GPU) and a Central Processing Unit (CPU) takes a lot of computation time and consumes a lot of power. In the related art, the method for accelerating the processing speed of the neural network model has the problems of incapability of cross-platform processing, low processing efficiency, high development cost, easiness in making mistakes and the like.

Disclosure of Invention

In view of this, the present disclosure provides a matrix vector calculation instruction processing method, system and related product, which can be used across platforms, improve processing efficiency, and reduce error probability and development cost.

According to a first aspect of the present disclosure, there is provided a matrix vector computation instruction processing system, the system comprising an instruction generating device and an executing device,

the instruction generating apparatus includes:

the device determining module is used for determining running devices for executing the computing macro instructions according to the received computing macro instructions;

the instruction generating module is used for generating an operating instruction according to the computing macro instruction and the operating equipment; the operation device includes:

the control module is used for acquiring required data, a neural network model and the operation instruction, analyzing the operation instruction and acquiring a plurality of analysis instructions;

the execution module is used for executing the plurality of analysis instructions according to the data to obtain an execution result,

wherein the matrix-vector computation macro is a macro for computing a matrix and a vector,

the matrix vector calculation macro-instruction comprises an operation type, an input address and an output address, the operation instruction comprises the operation type, an operation input address and an operation output address, and the operation input address and the operation output address are determined according to the input address and the output address respectively.

According to a second aspect of the present disclosure, there is provided a machine learning arithmetic device, the device including:

one or more matrix vector computing instruction processing systems according to the first aspect, configured to obtain data to be computed and control information from another processing apparatus, execute a specified machine learning operation, and transmit an execution result to the other processing apparatus through an I/O interface;

when the machine learning arithmetic device comprises a plurality of matrix vector calculation instruction processing systems, the matrix vector calculation instruction processing systems can be connected through a specific structure and transmit data;

the matrix vector calculation instruction processing systems are interconnected through a PCIE bus of a fast peripheral equipment interconnection bus and transmit data so as to support operation of larger-scale machine learning; a plurality of matrix vector computing instruction processing systems share the same control system or own respective control systems; the matrix vector calculation instruction processing systems share a memory or have respective memories; the interconnection mode of the matrix vector calculation instruction processing system is any interconnection topology.

According to a third aspect of the present disclosure, there is provided a combined processing apparatus, the apparatus comprising:

the machine learning arithmetic device, the universal interconnect interface, and the other processing device according to the second aspect;

and the machine learning arithmetic device interacts with the other processing devices to jointly complete the calculation operation designated by the user.

According to a fourth aspect of the present disclosure, there is provided a machine learning chip including the machine learning network operation device of the second aspect or the combination processing device of the third aspect.

According to a fifth aspect of the present disclosure, there is provided a machine learning chip package structure, which includes the machine learning chip of the fourth aspect.

According to a sixth aspect of the present disclosure, a board card is provided, which includes the machine learning chip packaging structure of the fifth aspect.

According to a seventh aspect of the present disclosure, there is provided an electronic device, which includes the machine learning chip of the fourth aspect or the board of the sixth aspect.

According to an eighth aspect of the present disclosure, there is provided a matrix vector calculation instruction processing method applied to a matrix vector calculation instruction processing system including an instruction generating apparatus and an executing apparatus, the method including:

calculating a macro instruction according to the received matrix vector through the instruction generating equipment, determining operating equipment for executing the matrix vector calculation macro instruction, and generating an operating instruction according to the matrix vector calculation macro instruction and the operating equipment;

acquiring data, a neural network model and an operation instruction through the operation equipment, analyzing the operation instruction to obtain a plurality of analysis instructions, executing the plurality of analysis instructions according to the data to obtain an execution result,

In some embodiments, the electronic device comprises a data processing apparatus, a robot, a computer, a printer, a scanner, a tablet, a smart terminal, a cell phone, a tachograph, a navigator, a sensor, a camera, a server, a cloud server, a camera, a camcorder, a projector, a watch, a headset, a mobile storage, a wearable device, a vehicle, a household appliance, and/or a medical device.

In some embodiments, the vehicle comprises an aircraft, a ship, and/or a vehicle; the household appliances comprise a television, an air conditioner, a microwave oven, a refrigerator, an electric cooker, a humidifier, a washing machine, an electric lamp, a gas stove and a range hood; the medical equipment comprises a nuclear magnetic resonance apparatus, a B-ultrasonic apparatus and/or an electrocardiograph.

The matrix vector calculation instruction processing method, the matrix vector calculation instruction processing system and the related product provided by the embodiment of the disclosure comprise instruction generation equipment and operation equipment. The instruction generating device includes: the device determining module is used for calculating the macroinstruction according to the received matrix vector and determining the operation device for executing the matrix vector to calculate the macroinstruction; the instruction generating module is used for calculating the macro instruction and the operation equipment according to the matrix vector to generate an operation instruction. The operation device includes: the control module is used for acquiring required data, a neural network model and an operation instruction, analyzing the operation instruction and acquiring a plurality of analysis instructions; the execution module is used for executing the plurality of analysis instructions according to the data to obtain an execution result. The method, the system and the related products can be used in a cross-platform mode, the applicability is good, the instruction conversion speed is high, the processing efficiency is high, the error probability is low, and the cost of developing manpower and material resources is low.

Other features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments, which proceeds with reference to the accompanying drawings.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate exemplary embodiments, features, and aspects of the disclosure and, together with the description, serve to explain the principles of the disclosure.

Fig. 1 shows a block diagram of a matrix vector calculation instruction processing apparatus according to an embodiment of the present disclosure.

Fig. 2 illustrates a block diagram of a matrix vector computation instruction processing apparatus according to an embodiment of the present disclosure.

Fig. 3a and 3b are schematic diagrams illustrating application scenarios of a matrix vector calculation instruction processing apparatus according to an embodiment of the present disclosure.

Fig. 4a, 4b show block diagrams of a combined processing device according to an embodiment of the present disclosure.

Fig. 5 shows a schematic structural diagram of a board card according to an embodiment of the present disclosure.

Fig. 6 illustrates a flow diagram of a matrix vector compute instruction processing method according to an embodiment of the present disclosure.

Detailed Description

Various exemplary embodiments, features and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. In the drawings, like reference numbers can indicate functionally identical or similar elements. While the various aspects of the embodiments are presented in drawings, the drawings are not necessarily drawn to scale unless specifically indicated.

The word "exemplary" is used exclusively herein to mean "serving as an example, embodiment, or illustration. Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments.

Furthermore, in the following detailed description, numerous specific details are set forth in order to provide a better understanding of the present disclosure. It will be understood by those skilled in the art that the present disclosure may be practiced without some of these specific details. In some instances, methods, means, elements and circuits that are well known to those skilled in the art have not been described in detail so as not to obscure the present disclosure.

FIG. 1 shows a block diagram of a matrix vector computation instruction processing system according to an embodiment of the present disclosure. As shown in fig. 1, the system includes an instruction generating apparatus 100 and an executing apparatus 200.

The instruction generating device 100 includes a device determining module 11 and an instruction generating module 12. The device determining module 11 is configured to determine an operating device that executes the matrix vector calculation macro according to the received matrix vector calculation macro. The instruction generating module 12 is configured to calculate a macro instruction and an operation device according to the matrix vector, and generate an operation instruction.

The operation device 200 includes a control module 21 and an execution module 22. The control module 21 is configured to obtain required data, a neural network model, and an operation instruction, and analyze the operation instruction to obtain a plurality of analysis instructions. The execution module 22 is configured to execute a plurality of parsing instructions according to the data to obtain an execution result.

The matrix-vector calculation macro is a macro for calculating a matrix and a vector. The matrix vector computing macro-instruction comprises an operation type, an input address and an output address, and the operation instruction comprises the operation type, the operation input address and the operation output address. The operation input address and the operation output address are determined according to the input address and the output address respectively.

In this implementation, a macro is a name for batch processing, and a macro may be a rule or pattern, or syntax replacement, which is automatically performed when the macro is encountered. The matrix vector calculation macro instruction can be formed by integrating commonly used matrix vector calculation instructions to be executed for performing calculation, control, transportation and other processing on data.

In one possible implementation, the matrix vector computation macro-instruction may include at least one of: the method comprises the following steps of calculating a macro instruction by a matrix multiplication vector, calculating a macro instruction by a vector multiplication matrix, calculating a macro instruction by a tensor, calculating a macro instruction by matrix addition and calculating a macro instruction by matrix subtraction.

In one possible implementation, the matrix vector computation macro-instruction may further include at least one of the following options: identification of a specified device for executing the matrix vector computation macro-instruction, input quantities, output quantities, operands, and instruction parameters. The instruction parameters may include at least one of an address, a length of the second operand.

The identifier of the specific device may be a physical address, an IP address, a name, a number, and the like of the specific device. The mark may comprise one or any combination of numbers, letters, symbols. When the position of the identifier of the specified device of the matrix vector calculation macro instruction is null, determining that the matrix vector calculation macro instruction has no specified device; or, when the field of "identification of the specified device" is not included in the matrix vector calculation macro, determining that the matrix vector calculation macro does not have the specified device. The operation type may refer to a type of the operation performed by the matrix vector calculation macro-instruction on the data, and represents a specific type of the matrix vector calculation macro-instruction, for example, when an operation type of a certain matrix vector calculation macro-instruction is "XXX", the specific type of the operation performed by the matrix vector calculation macro-instruction on the data may be determined according to "XXX". The instruction set required for executing the matrix vector calculation macro instruction may be determined according to the operation type, for example, when the operation type of a certain matrix vector calculation macro instruction is "XXX", the instruction set required for the certain matrix vector calculation macro instruction is all instruction sets required for performing the processing corresponding to "XXX". The input address may be an input address of data, an address for obtaining data such as a read address, and the output address may be an output address of processed data, an address for storing data such as a write address. The input amount may be information indicating the size of the data amount, such as the input size and the input length of the data. The output quantity may be information indicating the size of the data quantity, such as the output size and the output length of the data. The operands may include the length of the register, the address of the register, the identity of the register, the immediate, and the like. The immediate is the number given in the immediate addressing mode instruction. Instruction parameters may refer to parameters associated with the execution of the macro instruction corresponding to the matrix vector computation. For example, the instruction parameters may be the address and length of the second operand, etc. The instruction parameters may be the size of the convolution kernel, the step size of the convolution kernel, the filling of the convolution kernel, etc.

In this implementation, for a matrix vector computation macro-instruction, it must include an operation code, i.e. an operation type, and at least one operation field, which includes the identification of the specified device, input address, output address, input quantity, output quantity, operands, and instruction parameters. An opcode may be the portion of an instruction or field (usually denoted by a code) specified in a computer program that is to perform an operation, and is an instruction sequence number that tells the device executing the instruction which instruction specifically needs to be executed. The operation domain may be a source of all data required for executing the corresponding instruction, including parameter data, data to be operated on or processed, a corresponding operation method, or an address or the like storing the parameter data, the data to be operated on or processed, the corresponding operation method.

It should be understood that the instruction format and inclusion of the matrix vector computation macro-instruction may be configured as desired by those skilled in the art, and the present disclosure is not limited thereto.

In this embodiment, the device determining module 11 may determine one or more operating devices according to the matrix vector calculation macro. Instruction generation module 12 may generate one or more execution instructions. When a plurality of generated operating instructions are provided, the plurality of operating instructions may be executed in the same operating device or different operating devices, and the present disclosure is not limited thereto.

The matrix vector calculation instruction processing system provided by the embodiment of the disclosure comprises an instruction generation device and an operation device. The instruction generating device comprises a device determining module, a macro-instruction executing module and a macro-instruction executing module, wherein the device determining module is used for calculating the macro-instruction according to the received matrix vector and determining the running device for executing the matrix vector calculation macro-instruction; the instruction generating module is used for calculating the macro instruction and the operation equipment according to the matrix vector to generate an operation instruction. The operation equipment comprises a control module, a neural network model and an operation instruction, wherein the control module is used for acquiring required data, the neural network model and the operation instruction, analyzing the operation instruction and acquiring a plurality of analysis instructions; the execution module is used for executing the plurality of analysis instructions according to the data to obtain an execution result. The matrix vector calculation instruction processing system provided by the embodiment of the disclosure can be used in a cross-platform manner, and has the advantages of good applicability, high instruction conversion speed, high processing efficiency, low error probability and low development cost of manpower and material resources.

FIG. 2 illustrates a block diagram of a matrix vector computation instruction processing system, according to an embodiment of the present disclosure. In one possible implementation, as shown in fig. 2, the instruction generating device 100 may further include a macro instruction generating module 13. The matrix vector calculation macro-instruction generating module 13 is configured to receive a to-be-executed matrix vector calculation instruction, and generate a matrix vector calculation macro-instruction according to the determined identifier of the specific device and the to-be-executed matrix vector calculation instruction.

In this implementation, the specified device may be determined according to the operation type, the input amount, the output amount, and the like of the matrix vector calculation instruction to be executed. The received matrix vector calculation instruction to be executed can be one or more.

The matrix vector calculation instruction to be executed may include at least one of: the method comprises a to-be-executed matrix multiplication vector calculation instruction, a to-be-executed vector multiplication matrix calculation instruction, a to-be-executed tensor calculation instruction, a to-be-executed matrix addition calculation instruction and a to-be-executed matrix subtraction calculation instruction.

The matrix vector calculation instruction to be executed may include at least one of the following options: operation type, input address, output address, input quantity, output quantity, operands, and instruction parameters.

In this implementation, when there is one matrix vector calculation instruction to be executed, the determined identifier of the specific device may be added to the matrix vector calculation instruction to be executed, and a matrix vector calculation macro instruction is generated. For example, some matrix vector calculation instruction m to be executed is "XXX … … param". Where XXX is the operation type and param is the instruction parameter. Its designated device m-1 may be determined from the operation type "XXX" of the matrix vector calculation instruction m to be executed. Then, an identification (e.g., 09) specifying the device M-1 is added to the matrix vector calculation instruction M to be executed, and a matrix vector calculation macro instruction M "XXX 09, … … param" corresponding to the matrix vector calculation instruction M to be executed is generated. When the matrix vector calculation instruction to be executed is multiple, the identifier of the specified device corresponding to each determined matrix vector calculation instruction to be executed may be added to the matrix vector calculation instruction to be executed, and one matrix vector calculation macro instruction or multiple corresponding matrix vector calculation macro instructions may be generated according to the multiple matrix vector calculation instructions to be executed with the identifier of the specified device.

It should be understood that the instruction format and the content of the matrix vector calculation instruction to be executed may be set by those skilled in the art according to the needs, and the present disclosure does not limit this.

In one possible implementation, as shown in fig. 2, the device determining module 11 may include a first determining sub-module 111. The first determining submodule 111 is configured to determine the specified device as an operating device when it is determined that the matrix vector calculation macro includes an identifier of the specified device and a resource of the specified device meets an execution condition for executing the matrix vector calculation macro. Wherein, the execution condition may include: the specifying device contains an instruction set corresponding to a matrix vector computation macro instruction.

In this implementation, the matrix vector computation macro-instruction may include an identification of one or more specified devices that execute the matrix vector computation macro-instruction. When the matrix vector calculation macro instruction includes the identifier of the specified device and the resource of the specified device meets the execution condition, the first determining sub-module 111 may directly determine the specified device as the operating device, so as to save the generation time for generating the operating instruction based on the matrix vector calculation macro instruction and ensure that the generated operating instruction can be executed by the corresponding operating device.

In one possible implementation, as shown in fig. 2, the instruction generating apparatus 100 may further include a resource obtaining module 14. The device determination module 11 may also include a second determination submodule 112. The resource obtaining module 14 is configured to obtain resource information of the alternative device. The second determining submodule 112 is configured to, when it is determined that the matrix vector calculation macro does not include the identifier of the specified device, determine, from the candidate devices, an operating device for executing the matrix vector calculation macro according to the received matrix vector calculation macro and resource information of the candidate devices. Wherein the resource information may comprise a set of instructions contained by the alternative device. The instruction set included in the alternative device may be a set of instructions corresponding to one or more types of operation of the matrix vector computation macro-instructions. The more instruction sets the candidate device contains, the more types of matrix vector computation macro instructions the candidate device is capable of executing.

In this implementation, when determining that the matrix vector calculation macro does not include the identifier of the specific device, the second determining submodule 112 may determine, from the candidate devices, one or more operating devices capable of executing the matrix vector calculation macro. Wherein the determined instruction set of the running device comprises an instruction set corresponding to a matrix vector computation macro. For example, where the received matrix vector computation macro is a matrix multiplication vector computation macro, an alternate device containing an instruction set corresponding to the matrix multiplication vector computation macro may be determined to be the executing device to ensure that it can execute the generated executing instruction.

In one possible implementation, as shown in fig. 2, the device determining module 11 may further include a third determining sub-module 113. When it is determined that the matrix vector calculation macro instruction contains the identifier of the specified device and the resource of the specified device does not meet the execution condition for executing the matrix vector calculation macro instruction, the third determining submodule 113 determines the operating device according to the matrix vector calculation macro instruction and the resource information of the alternative device.

In this implementation, when it is determined that the matrix vector calculation macro includes the identifier of the specified device and the resource of the specified device does not satisfy the execution condition, the third determining sub-module 113 may determine that the specified device of the matrix vector calculation macro does not have the capability of executing the matrix vector calculation macro. The third determination submodule 113 may determine an operating device from among candidate devices, and may determine, as the operating device, a candidate device containing an instruction set corresponding to the matrix vector calculation macro instruction.

In a possible implementation manner, as shown in fig. 2, the matrix vector calculation macro may further include at least one of an input quantity and an output quantity, and the instruction generating module 12 is further configured to determine a data quantity of the matrix vector calculation macro, and generate the operation instruction according to the data quantity of the matrix vector calculation macro, and resource information of the operation device. The data amount of the matrix vector calculation macro instruction may be determined according to at least one of the input amount and the output amount, and the resource information of the operating device may further include at least one of a storage capacity and a remaining storage capacity.

The storage capacity of the operating device may refer to the amount of binary information that the memory of the operating device can accommodate. The remaining storage capacity of the operating device may refer to the storage capacity that the operating device is currently available for instruction execution after the occupied storage capacity is removed. The resource information of the running device can characterize the running capability of the running device. The larger the storage capacity and the larger the remaining storage capacity are, the stronger the operation capability of the operation device is.

In this implementation, the instruction generating module 12 may determine a specific manner of splitting the matrix vector calculation macro instruction according to the resource information of each operating device, the data amount of the matrix vector calculation macro instruction, and the like, so as to split the matrix vector calculation macro instruction and generate an operating instruction corresponding to the operating device.

In one possible implementation, as shown in fig. 2, the instruction generating module 12 may include a first instruction generating submodule 121. The first instruction generation submodule 121 is configured to split the matrix vector calculation macro instruction into a plurality of operation instructions according to the operation data amount and the data amount of the operation device when it is determined that the number of the operation devices is one and the operation data amount of the operation device is smaller than the data amount of the matrix vector calculation macro instruction, so that the operation device sequentially executes the plurality of operation instructions. The operation data amount of the operation device may be determined according to the resource information of the operation device, each operation instruction may further include at least one of an operation input amount and an operation output amount, and the operation input amount and the operation output amount may be determined according to the operation data amount.

In this implementation, the operation data amount of the operation device may be determined according to the storage capacity or the remaining storage capacity of the operation device. The operation input quantity and the operation output quantity are required to be less than or equal to the operation data quantity so as to ensure that the generated operation instruction can be executed by the operation equipment. The operation input amount (or operation output amount) of different operation instructions in the plurality of operation instructions may be the same or different, and the disclosure does not limit this.

In this implementation, when it is determined that there is one operation device and the operation data amount of the operation device is greater than or equal to the data amount of the matrix vector calculation macro, the first instruction generation sub-module 121 may directly convert the matrix vector calculation macro into one operation instruction, and may also split the matrix vector calculation macro into a plurality of operation instructions, which is not limited in this disclosure.

In one possible implementation, as shown in FIG. 2, the instruction generation module 12 may include a second instruction generation submodule 122. The second instruction generating submodule 122 is configured to split the matrix vector calculation macro instruction according to the operation data amount and the data amount of each operating device when it is determined that a plurality of operating devices are provided, and generate an operation instruction corresponding to each operating device. The operation data volume of each operating device may be determined according to the resource information of each operating device, the operation instruction may further include at least one of an operation input volume and an operation output volume, and the operation input volume and the operation output volume are determined according to the operation data volume of the operating device that executes the operation instruction.

In this implementation, the operation input quantity and the operation output quantity need to be less than or equal to the operation data quantity to ensure that the generated operation instruction can operate the device to execute. The second instruction generation sub-module 122 may generate one or more operation instructions for each operation device according to the operation data amount of each operation device, so as to be executed by the corresponding operation device.

In the above implementation, the operation instruction includes at least one of the operation input quantity and the operation output quantity, except that the data quantity of the operation instruction can be limited to be executed by the corresponding operation device. And special limiting requirements of different operation instructions on operation input quantity and/or operation output quantity can be met.

In a possible implementation manner, for some operation instructions that do not have a special limited requirement on the operation input amount and/or the operation output amount, where the operation input amount and/or the operation output amount may not be included, a default operation input amount and a default operation output amount may be preset, so that when it is determined that the operation input amount and the operation output amount do not exist in the received operation instructions, the operation device may use the default operation input amount and the default operation output amount as the operation input amount and the operation output amount of the operation instructions. By presetting the default operation input quantity and the default operation output quantity, the generation process of the operation instruction can be simplified, and the generation time of the operation instruction is saved.

In one possible implementation, the default input quantity and the default output quantity of the macro instruction may be preset for different types of matrix vectors. When the input quantity and the output quantity are not included in the matrix vector calculation macro instruction, the corresponding preset default input quantity and default output quantity can be used as the input quantity and the output quantity of the matrix vector calculation macro instruction. And then determining the data volume of the matrix vector calculation macro instruction according to the default input quantity and/or the default output quantity, calculating the resource information of the macro instruction and the operating equipment according to the data volume of the matrix vector calculation macro instruction and the matrix vector, and generating the operating instruction. When the matrix vector calculation macro instruction does not include the input quantity and the output quantity, the generated operation instruction may not include the operation input quantity and the operation output quantity, and may include at least one of the operation input quantity and the operation output quantity. When the operation instruction does not include the operation input amount and/or the operation output amount, the operation device may execute the operation instruction according to a preset default operation input amount and/or a default operation output amount.

In a possible implementation manner, the instruction generating module 12 may further split the matrix vector calculation macro instruction according to the matrix vector calculation macro instruction and a preset matrix vector calculation macro instruction splitting rule, so as to generate the operation instruction. The matrix vector calculation macro instruction splitting rule may be determined according to a conventional matrix vector calculation macro instruction splitting manner (for example, splitting according to a processing procedure of a matrix vector calculation macro instruction, and the like), and in combination with an operation data amount threshold of an instruction that can be executed by all the candidate devices. And splitting the matrix vector calculation macro instruction into operation instructions of which the operation input quantity and the operation output quantity are less than or equal to an operation data quantity threshold value so as to ensure that the generated operation instructions can be executed in corresponding operation equipment (the operation equipment is any one of the alternative equipment). The storage capacities (or remaining storage capacities) of all the candidate devices may be compared, and the determined minimum storage capacity (or remaining storage capacity) may be determined as an operation data amount threshold of the instruction that can be executed by all the candidate devices.

It should be understood that, the person skilled in the art can set the generation mode of the operation instruction according to the actual needs, and the present disclosure does not limit this.

In this embodiment, the operation instruction generated by the instruction generating module according to the matrix vector calculation macro instruction may be a to-be-executed matrix vector calculation instruction, or may be one or more analyzed instructions obtained by analyzing the to-be-executed matrix vector calculation instruction, which is not limited in this disclosure.

In one possible implementation, as shown in fig. 2, the instruction generating apparatus 100 may further include a queue building module 15. The queue building module 15 is configured to sort the operation instructions according to a queue sorting rule, and build an instruction queue corresponding to the operation device according to the sorted operation instructions.

In this implementation, an instruction queue uniquely corresponding to each execution device may be constructed for each execution device. The operating instructions can be sequentially sent to the operating equipment uniquely corresponding to the instruction queue according to the sequence of the operating instructions in the instruction queue; or the instruction queue may be sent to the execution device, so that the execution device sequentially executes the execution instructions in the instruction queue according to the order of the execution instructions in the instruction queue. By the mode, the operation equipment can execute the operation instruction according to the instruction queue, the operation instruction is prevented from being executed mistakenly and delayed, and the operation instruction is prevented from being omitted.

In this implementation, the queue sorting rule may be determined according to information such as a predicted execution time for executing the operation instruction, a generation time of the operation instruction, an operation input amount, an operation output amount, and an operation type related to the operation instruction itself, which is not limited by this disclosure.

In one possible implementation, as shown in FIG. 2, instruction generation device 100 may also include an instruction dispatch module 16. The instruction dispatch module 16 is configured to send the execution instruction to the execution device, so that the execution device executes the execution instruction.

In this implementation, when there is one execution instruction executed by the execution device, the execution instruction may be directly sent to the execution device. When the number of the operation instructions executed by the operation device is multiple, all of the multiple operation instructions may be sent to the operation device, so that the operation device sequentially executes the multiple operation instructions. The plurality of operation instructions can also be sequentially sent to the corresponding operation equipment, wherein after the operation equipment completes the current operation instruction, the next operation instruction corresponding to the current operation instruction is sent to the operation equipment each time. The manner in which the person skilled in the art can send the operation instruction to the operation device is set, and the present disclosure does not limit this.

In one possible implementation, as shown in FIG. 2, the instruction dispatch module 16 may include an instruction assembly submodule 161, an assembly translation submodule 162, and an instruction issue submodule 163. The instruction assembling sub-module 161 is used for generating an assembling file according to the operation instruction. The assembly translation sub-module 162 is used to translate the assembly file into a binary file. The instruction sending submodule 163 is configured to send the binary file to the operating device, so that the operating device executes the operating instruction according to the binary file.

By the mode, the data volume of the operation instruction can be reduced, the time for sending the operation instruction to the operation equipment is saved, and the conversion and execution speed of the matrix vector calculation macro instruction is increased.

In this implementation manner, after the binary file is sent to the running device, the running device may decode the received binary file to obtain a corresponding running instruction, and execute the obtained running instruction to obtain an execution result.

In a possible implementation manner, the running device may be one or any combination of a CPU, a GPU, and an embedded Neural-Network Processing Unit (NPU). Thus, the speed of the instruction generating device for generating the operation instruction according to the matrix vector calculation macro instruction is improved.

In one possible implementation, the instruction generating apparatus 100 may be provided in the CPU and/or the NPU. The process of generating the operation instruction according to the matrix vector calculation macro instruction is realized by the CPU and/or the NPU, and more possible modes are provided for the realization of the instruction generating equipment.

In a possible implementation, the execution device 200 further comprises a storage module 23. The memory module 23 may include at least one of a register and a cache, and the cache may include a scratch pad cache. The cache may be used to store data. The registers may be used to store scalar data within the data.

In one possible implementation, the control module 21 may include an instruction storage submodule 211 and an instruction processing submodule 212. The instruction storage submodule 211 is used for storing the operation instruction. The instruction processing submodule 212 is configured to parse the operation instruction to obtain a plurality of parsing instructions.

In one possible implementation, the control module 21 may further include a store queue submodule 213. The storage queue submodule 213 is configured to store an operation instruction queue, where the operation instruction queue includes an operation instruction that needs to be executed by the operation device and a plurality of parsing instructions. And all the instructions in the operation instruction queue are sequentially arranged according to the execution sequence.

In a possible implementation, the execution module 22 may further include a dependency processing sub-module 221. The dependency relationship processing submodule 221 is configured to cache the first parsing instruction in the instruction storage submodule when it is determined that the first parsing instruction has an association relationship with a zeroth parsing instruction before the first parsing instruction, and extract the first parsing instruction from the instruction storage submodule after the zeroth parsing instruction is executed and send the first parsing instruction to the execution module.

The association relationship between the first parsing instruction and the zeroth parsing instruction before the first parsing instruction may include: the first storage address interval for storing the data required by the first analysis instruction and the zeroth storage address interval for storing the data required by the zeroth analysis instruction have an overlapping area. Conversely, the no association relationship between the first parse instruction and the zeroth parse instruction may be that the first memory address interval and the zeroth memory address interval have no overlapping area.

In one possible implementation, the matrix-vector computation macro-instruction may refer to a macro-instruction used to compute matrices and vectors. For example, the macro instruction performs calculation such as matrix multiplication vector calculation, vector multiplication matrix calculation, tensor calculation, matrix addition calculation, and matrix subtraction calculation on the matrix and the vector. The different types of matrix vector calculation macro-instructions correspond to different operation types, for example, the operation type corresponding to the matrix addition calculation macro-instruction is MADD, and the operation type corresponding to the matrix multiplication vector calculation macro-instruction is MMV.

It should be noted that, although the matrix vector calculation instruction processing system is described above by taking the above-mentioned embodiment as an example, those skilled in the art can understand that the present disclosure should not be limited thereto. In fact, the user can flexibly set each module according to personal preference and/or actual application scene, as long as the technical scheme of the disclosure is met.

Application example

An application example according to the embodiment of the present disclosure is given below in conjunction with "working process of a matrix vector calculation instruction processing system" as an exemplary application scenario to facilitate understanding of the flow of the matrix vector calculation instruction processing system. It is to be understood by those skilled in the art that the following application examples are for the purpose of facilitating understanding of the embodiments of the present disclosure only and are not to be construed as limiting the embodiments of the present disclosure.

First, an instruction format of the matrix vector calculation macro instruction, an instruction format of the matrix vector calculation instruction to be executed, and a process of the execution device executing the execution instruction are described, and a specific example is as follows.

The instruction format of the matrix vector compute macro may be:

Type device_id,input_addr,output_addr,input_size,output_size,[param1,param2,…]

the Type is an operation Type, idevic _ id is an identifier of a designated device, input _ addr is an input address, output _ addr is an output address, input _ size is the size of an input vector (i.e., an input quantity), output _ size is the size of an output vector (i.e., an output quantity), and param1 and param2 are instruction parameters. The instruction parameter may be an address and length of the second operand.

For the matrix vector computing macro-instruction, the macro-instruction must include an operation type, an input address and an output address, and an operation instruction generated by computing the macro-instruction according to the matrix vector must also include an operation type, an operation input address and an operation output address, wherein the operation input address and the operation output address are determined according to the input address and the output address respectively.

Taking "@ MADD #502, #8, #34, # 5" as an example of the operation instruction generated by calculating a macro instruction from a certain matrix vector. After the operation device receives the operation instruction, the execution process is as follows: an input matrix vector of size 34 is obtained from the input address 502. And performing matrix addition calculation on the input matrix vector to obtain an output matrix vector with the size of 5. The output matrix vector of size 5 is stored as the execution result at memory address 8.

The instruction format of the matrix vector calculation instruction to be executed may be:

Type input_addr,output_addr,input_size,output_size,[param1,param2,…]

the Type is an operation Type, input _ addr is an input address, output _ addr is an output address, input _ size is the size of an input vector (i.e., an input quantity), output _ size is the size of an output vector (i.e., an output quantity), and param1 and param2 are instruction parameters. The instruction parameter may be an address and length of the second operand.

Fig. 3a and 3b are schematic diagrams illustrating application scenarios of a matrix vector calculation instruction processing system according to an embodiment of the present disclosure. The matrix vector calculation instruction processing system may include the above-described alternative device and the instruction generating device. As shown in fig. 3a and fig. 3b, there may be a plurality of candidate devices for executing the matrix vector calculation macro instruction, the candidate devices may be CPU-1, CPU-2, …, CPU-n, NPU-1, NPU-2, …, NPU-n, GPU-1, GPU-2, …, and GPU-n, and the candidate device is selected to execute the corresponding operation instruction, i.e. the operation device. The working process and principle of the matrix vector calculation instruction processing system for generating and executing the operation instruction according to a certain matrix vector calculation macro instruction are as follows.

Resource acquisition module 14

And acquiring resource information of the alternative device, wherein the resource information comprises the residual storage capacity and the storage capacity of the alternative device and an instruction set contained in the alternative device. The resource obtaining module 14 sends the obtained resource information of the candidate device to the device determining module 11 and the instruction generating module 12.

The device determination module 11 (including a first determination sub-module 111, a second determination sub-module 112, and a third determination sub-module 113)

And when receiving the matrix vector calculation macro instruction, determining the operation equipment for executing the matrix vector calculation macro instruction according to the received matrix vector calculation macro instruction. For example, the following matrix vector computation macro is received. Where the matrix vector computation macro-instructions may be from different platforms.

Matrix vector calculation macroinstruction 1: @ XXX #01 … …

Matrix vector computation macroinstruction 2: @ SSS #02 … …

Matrix vector computation macroinstruction 3: @ DDD #04 … …

Matrix vector computation macroinstruction 4: @ NNN … …

When the first determining sub-module 111 determines that the matrix vector calculation macro instruction contains the identifier of the specified device and determines that the specified device contains the instruction set corresponding to the matrix vector calculation macro instruction, the first determining sub-module 111 may determine the specified device as an operating device for executing the matrix vector calculation macro instruction and send the identifier of the determined operating device to the instruction generating module 12. For example, the first determination sub-module 111 may determine a specific device corresponding to the identifier 01, such as CPU-2 (the CPU-2 includes an instruction set corresponding to the matrix vector calculation macro 1), as an execution device for executing the matrix vector calculation macro 1. A specific device such as CPU-1(CPU-1 contains an instruction set corresponding to the matrix vector calculation macro instruction 2) to which the identifier 02 corresponds may be determined as an execution device for executing the matrix vector calculation macro instruction 2.

When the third determining submodule 113 determines that the matrix vector calculation macro instruction contains the identifier of the specified device and determines that the specified device does not contain the instruction set corresponding to the matrix vector calculation macro instruction, the third determining submodule 113 may determine the candidate device containing the instruction set corresponding to the matrix vector calculation macro instruction as the operating device, and send the determined identifier of the operating device to the instruction generating module 12. For example, when it is determined that the instruction set corresponding to the matrix vector calculation macro 3 is not included in the specified device corresponding to the identifier 04, the third determination sub-module 113 may determine, as an execution device for executing the matrix vector calculation macro 3, an alternative device, such as NPU-n or NPU-2, that includes an instruction set corresponding to the operation type DDD of the matrix vector calculation macro 3.

When the second determining submodule 112 determines that the identifier of the specified device does not exist in the matrix vector calculation macro instruction (the position corresponding to the identifier of the specified device is empty, or the field of "identifier of the specified device" is not included in the matrix vector calculation macro instruction), the second determining submodule 112 may determine the operating device from the candidate devices according to the matrix vector calculation macro instruction and the resource information of the candidate devices (the specific determination process is detailed in the related description of the second determining submodule 112 above), and send the determined identifier of the operating device to the instruction generating module 12. For example, since the identifier of the specified device does not exist in the matrix vector calculation macro 4, the second determination sub-module 112 may determine, from the candidate devices, the operating device, for example, GPU-n (the GPU-n includes an instruction set corresponding to the operating type NNN) for executing the matrix vector calculation macro 4, according to the operation type NNN of the matrix vector calculation macro 4 and the resource information (the included instruction set) of the candidate devices.

Instruction generation module 12 (including first instruction generation module 121 and second instruction generation module 122)

When the number of the operating devices is one and the amount of the operating data of the operating device is smaller than the amount of the data of the matrix vector calculation macro, the first instruction generating module 121 splits the matrix vector calculation macro into a plurality of operating instructions according to the amount of the operating data and the amount of the data of the operating device, and sends the plurality of operating instructions to the queue constructing module 15. For example, a plurality of execution instructions 2-1, 2-2, …, 2-n are generated based on the amount of data of the matrix vector calculation macro instruction 2 and the amount of execution data of the execution device CPU-1. And calculating the data volume of the macro instruction 4 and the operation data volume of the operation equipment GPU-n according to the matrix vector to generate a plurality of operation instructions 4-1, 4-2, … and 4-n.

When it is determined that there is one operating device and the data amount of the operating device is greater than or equal to the data amount of the matrix vector calculation macro, the first instruction generating module 121 may generate one operating instruction according to the matrix vector calculation macro, and send the operating instruction to the queue building module 15. For example, one operation instruction 1-1 is generated by calculating the data amount of the macro instruction 1 and the operation data amount of the operation device CPU-2 according to the matrix vector.

When determining that a plurality of operating devices are provided, the second instruction generating module 122 splits the matrix vector calculation macro instruction according to the operating data amount of each operating device and the data amount of the matrix vector calculation macro instruction, generates an operating instruction corresponding to each operating device, and sends the operating instruction to the queue constructing module 15. For example, the data size of the macroinstruction 3, the operation data size of the operation device NPU-n, and the operation data size of the operation device NPU-2 are calculated from the matrix vector, a plurality of operation instructions 3-1, 3-2, …, 3-n are generated for the operation device NPU-n, and a plurality of operation instructions 3 ' -1, 3 ' -2, …, 3 ' -n are generated for the operation device NPU-2.

Queue building Block 15

When receiving the operation instruction, all the operation instructions to be executed by each operation device are sorted according to the queue sorting rule, a unique corresponding instruction queue is constructed for each operation device according to the sorted operation instructions, and the instruction queue is sent to the instruction dispatching module 16. In particular, the amount of the solvent to be used,

for an operation instruction 1-1 executed by the operation device CPU-2. The instruction queue CPU-2 "constructed corresponding to the execution device CPU-2 includes only the execution instructions 1-1.

For a plurality of execution instructions 2-1, 2-2, …, 2-n executed by the execution device CPU-1. And sequencing the plurality of operating instructions 2-1, 2-2, … and 2-n according to a queue sequencing rule, and constructing an instruction queue CPU-1' corresponding to the operating equipment CPU-1 according to the sequenced plurality of operating instructions 2-1, 2-2, … and 2-n.

For a plurality of execution instructions 3-1, 3-2, …, 3-n executed by the execution device NPU-n. The multiple operating instructions 3-1, 3-2, …, 3-n are sorted according to a queue sorting rule, and an instruction queue NPU-n' corresponding to the operating equipment NPU-n is constructed according to the sorted multiple operating instructions 3-n, …, 3-2, 3-1.

For the plurality of execution instructions 3 ' -1, 3 ' -2, …, 3 ' -n executed by the execution device NPU-2. The plurality of operating instructions 3 '-1, 3' -2, …, 3 '-n are ordered according to a queue ordering rule, and an instruction queue NPU-2 "corresponding to the operating device NPU-2 is constructed according to the ordered plurality of operating instructions 3' -n, …, 3 '-2, 3' -1.

For the plurality of execution instructions 4-1, 4-2, …, 4-n executed by the execution device GPU-n. And sequencing the plurality of operating instructions 4-1, 4-2, … and 4-n according to a queue sequencing rule, and constructing an instruction queue GPU-n' corresponding to the operating equipment GPU-n according to the sequenced plurality of operating instructions 4-1, 4-2, … and 4-n.

Instruction dispatch module 16

After the instruction queues are received, the operation instructions in each instruction queue are sequentially sent to corresponding operation equipment, so that the operation equipment executes the operation instructions. For example, the execution instruction 1-1 included in the instruction queue CPU-2 ″ is sent to its corresponding execution device CPU-2. And sequentially sending a plurality of running instructions 2-1, 2-2, … and 2-n in the instruction queue CPU-1' to the corresponding running equipment CPU-1. And sequentially sending the plurality of operating instructions 3-n, …, 3-2 and 3-1 in the instruction queue NPU-n' to the corresponding operating equipment NPU-n. And sequentially sending the plurality of running instructions 3 '-n, …, 3' -2 and 3 '-1 in the instruction queue NPU-2' to the corresponding running equipment NPU-2. And sequentially sending the multiple operating instructions 4-1, 4-2, … and 4-n in the queue GPU-n' to the corresponding operating equipment GPU-n.

After receiving the instruction queue, the operation device CPU-2, the operation device CPU-1, the operation device NPU-n and the operation device NPU-2 execute the operation instructions in sequence according to the arrangement sequence of the operation instructions in the instruction queue. Taking the operating device CPU-2 as an example, a specific process of executing the received operating instruction will be described. The operating device CPU-2 includes a control module 21, an execution module 22, and a storage module 23. The control module 21 includes an instruction storage sub-module 211, an instruction processing sub-module 212, and a storage queue sub-module 213, and the execution module 22 includes a dependency processing sub-module 221, which is described in detail above with respect to the operating device.

Assume that the run instruction 1-1 generated from the matrix vector computation macroinstruction 1 is "@ XXX … …". After receiving the operation instruction 1-1, the operation device CPU-2 executes the operation instruction 1-1 as follows:

the control module 21 of the operating device CPU-2 obtains data, a neural network model, and an operating instruction 1-1. The instruction storage submodule 211 is configured to store an operation instruction 1-1. The instruction processing sub-module 212 is configured to parse the operation instruction 1-1 to obtain a plurality of parsing instructions, such as parsing instruction 0, parsing instruction 1, and parsing instruction 2, and send the plurality of parsing instructions to the storage queue sub-module 213 and the execution module 22. The storage queue submodule 213 is configured to store an operation instruction queue, where the operation instruction queue includes an analysis instruction 0, an analysis instruction 1, an analysis instruction 2, and other operation instructions that need to be executed by the operating device CPU-2, and all the instructions are sequentially arranged in the operation instruction queue according to the execution sequence. For example, the obtained sequence of the execution of the multiple analysis instructions is analysis instruction 0, analysis instruction 1, and analysis instruction 2, and there is an association relationship between analysis instruction 1 and analysis instruction 0.

After the execution module 22 of the operating device CPU-2 receives the multiple parsing instructions, the dependency processing sub-module 221 therein determines whether there is an association relationship between the multiple parsing instructions. The dependency relationship processing sub-module 221 determines that the parsing instruction 1 and the parsing instruction 0 have an association relationship, caches the parsing instruction 1 in the instruction storage sub-module 211, and after it is determined that the parsing instruction 0 is executed, extracts the parsing instruction 1 from the cache and sends the parsing instruction 1 to the execution module 22, so that the execution module 22 can execute the parsing instruction.

Execution module 22 receives and executes parse instruction 0, parse instruction 1, and parse instruction 2 to complete execution of execute instructions 1-1.

The working process of the above modules can refer to the above related description.

Therefore, the system can be used in a cross-platform mode, the applicability is good, the instruction conversion speed is high, the processing efficiency is high, the error probability is low, and the cost of developing manpower and material resources is low.

The present disclosure provides a machine learning arithmetic device, which may include one or more of the above matrix vector calculation instruction processing systems, and is configured to acquire data to be operated and control information from other processing devices, and execute a specified machine learning operation. The machine learning arithmetic device can obtain a matrix vector calculation macro instruction or a matrix vector calculation instruction to be executed from other machine learning arithmetic devices or non-machine learning arithmetic devices, and transmit an execution result to peripheral equipment (also called other processing devices) through an I/O interface. Peripheral devices such as cameras, displays, mice, keyboards, network cards, wifi interfaces, servers. When more than one matrix vector calculation instruction processing system is included, the matrix vector calculation instruction processing systems can be linked and transmit data through a specific structure, for example, the matrix vector calculation instruction processing systems are interconnected and transmit data through a PCIE bus, so as to support larger-scale operation of the neural network. At this time, the same control system may be shared, or there may be separate control systems; the memory may be shared or there may be separate memories for each accelerator. In addition, the interconnection mode can be any interconnection topology.

The machine learning arithmetic device has high compatibility and can be connected with various types of servers through PCIE interfaces.

Fig. 4a shows a block diagram of a combined processing device according to an embodiment of the present disclosure. As shown in fig. 4a, the combined processing device includes the machine learning arithmetic device, the universal interconnection interface, and other processing devices. The machine learning arithmetic device interacts with other processing devices to jointly complete the operation designated by the user.

Other processing devices include one or more of general purpose/special purpose processors such as Central Processing Units (CPUs), Graphics Processing Units (GPUs), neural network processors, and the like. The number of processors included in the other processing devices is not limited. The other processing devices are used as interfaces of the machine learning arithmetic device and external data and control, and comprise data transportation to finish basic control of starting, stopping and the like of the machine learning arithmetic device; other processing devices may cooperate with the machine learning computing device to perform computing tasks.

And the universal interconnection interface is used for transmitting data and control instructions between the machine learning arithmetic device and other processing devices. The machine learning arithmetic device acquires required input data from other processing devices and writes the input data into a storage device on the machine learning arithmetic device; control instructions can be obtained from other processing devices and written into a control cache on a machine learning arithmetic device chip; the data in the storage module of the machine learning arithmetic device can also be read and transmitted to other processing devices.

Fig. 4b shows a block diagram of a combined processing device according to an embodiment of the present disclosure. In a possible implementation manner, as shown in fig. 4b, the combined processing device may further include a storage device, and the storage device is connected to the machine learning operation device and the other processing device respectively. The storage device is used for storing data stored in the machine learning arithmetic device and the other processing device, and is particularly suitable for data which is required to be calculated and cannot be stored in the internal storage of the machine learning arithmetic device or the other processing device.

The combined processing device can be used as an SOC (system on chip) system of equipment such as a mobile phone, a robot, an unmanned aerial vehicle and video monitoring equipment, the core area of a control part is effectively reduced, the processing speed is increased, and the overall power consumption is reduced. In this case, the generic interconnect interface of the combined processing device is connected to some component of the apparatus. Some parts are such as camera, display, mouse, keyboard, network card, wifi interface.

The present disclosure provides a machine learning chip, which includes the above machine learning arithmetic device or combined processing device.

The present disclosure provides a machine learning chip package structure, which includes the above machine learning chip.

Fig. 5 shows a schematic structural diagram of a board card according to an embodiment of the present disclosure. As shown in fig. 5, the board includes the above-mentioned machine learning chip package structure or the above-mentioned machine learning chip. The board may include, in addition to the machine learning chip 389, other kits including, but not limited to: memory device 390, interface device 391 and control device 392.

The memory device 390 is coupled to a machine learning chip 389 (or a machine learning chip within a machine learning chip package structure) via a bus for storing data. Memory device 390 may include multiple sets of memory cells 393. Each group of memory cells 393 is coupled to a machine learning chip 389 via a bus. It is understood that each group 393 may be a DDR SDRAM (Double Data Rate SDRAM).

DDR can double the speed of SDRAM without increasing the clock frequency. DDR allows data to be read out on the rising and falling edges of the clock pulse. DDR is twice as fast as standard SDRAM.

In one embodiment, memory device 390 may include 4 groups of memory cells 393. Each group of memory cells 393 may include a plurality of DDR4 particles (chips). In one embodiment, the machine learning chip 389 may include 4 72-bit DDR4 controllers therein, where 64bit is used for data transmission and 8bit is used for ECC check in the 72-bit DDR4 controller. It is appreciated that when DDR4-3200 particles are used in each group of memory cells 393, the theoretical bandwidth of data transfer may reach 25600 MB/s.

In one embodiment, each group 393 of memory cells includes a plurality of double rate synchronous dynamic random access memories arranged in parallel. DDR can transfer data twice in one clock cycle. A controller for controlling DDR is provided in the machine learning chip 389 for controlling data transfer and data storage of each memory unit 393.

Interface device 391 is electrically coupled to machine learning chip 389 (or a machine learning chip within a machine learning chip package). The interface device 391 is used to implement data transmission between the machine learning chip 389 and an external device (e.g., a server or a computer). For example, in one embodiment, the interface device 391 may be a standard PCIE interface. For example, the data to be processed is transmitted to the machine learning chip 289 by the server through the standard PCIE interface, so as to implement data transfer. Preferably, when PCIE 3.0X 16 interface transmission is adopted, the theoretical bandwidth can reach 16000 MB/s. In another embodiment, the interface device 391 may also be another interface, and the disclosure does not limit the specific representation of the other interface, and the interface device can implement the switching function. In addition, the calculation result of the machine learning chip is still transmitted back to the external device (e.g., server) by the interface device.

The control device 392 is electrically connected to a machine learning chip 389. The control device 392 is used to monitor the state of the machine learning chip 389. Specifically, the machine learning chip 389 and the control device 392 may be electrically connected through an SPI interface. The control device 392 may include a single chip Microcomputer (MCU). For example, machine learning chip 389 may include multiple processing chips, multiple processing cores, or multiple processing circuits, which may carry multiple loads. Therefore, the machine learning chip 389 can be in different operation states such as a multi-load and a light load. The control device can regulate and control the working states of a plurality of processing chips, a plurality of processing circuits and/or a plurality of processing circuits in the machine learning chip.

The present disclosure provides an electronic device, which includes the above machine learning chip or board card.

The electronic device may include a data processing apparatus, a robot, a computer, a printer, a scanner, a tablet, a smart terminal, a cell phone, a tachograph, a navigator, a sensor, a camera, a server, a cloud server, a camera, a video camera, a projector, a watch, an earphone, a mobile storage, a wearable device, a vehicle, a household appliance, and/or a medical device.

The vehicle may include an aircraft, a ship, and/or a vehicle. The household appliances may include televisions, air conditioners, microwave ovens, refrigerators, electric rice cookers, humidifiers, washing machines, electric lamps, gas cookers, and range hoods. The medical device may include a nuclear magnetic resonance apparatus, a B-mode ultrasound apparatus and/or an electrocardiograph.

Fig. 6 illustrates a flow diagram of a matrix vector compute instruction processing method according to an embodiment of the present disclosure. As shown in fig. 6, the method is applied to the above-described matrix vector calculation instruction processing system including the instruction generating apparatus and the executing apparatus, and includes step S41 and step S42.

In step S41, the execution device that executes the matrix vector calculation macro is determined by the instruction generation device calculating the macro from the received matrix vector, and the execution device and the macro are generated from the matrix vector calculation macro.

In step S42, the operation device acquires the data, the neural network model, and the operation instruction, analyzes the operation instruction to obtain a plurality of analysis instructions, and executes the plurality of analysis instructions according to the data to obtain an execution result.

In one possible implementation, step S41 may include: and when the matrix vector calculation macro instruction is determined to contain the identification of the specified equipment, and the resource of the specified equipment meets the execution condition for executing the matrix vector calculation macro instruction, determining the specified equipment as the operating equipment. Wherein, the execution condition may include: the specifying device contains an instruction set corresponding to a matrix vector computation macro instruction.

In one possible implementation, the method may further include: and acquiring resource information of the alternative equipment.

Wherein, step S41 may further include: and when determining that the matrix vector calculation macro-instruction does not contain the identifier of the specified device, determining running equipment for executing the matrix vector calculation macro-instruction from the alternative devices according to the received matrix vector calculation macro-instruction and the resource information of the alternative devices. Wherein the resource information may comprise a set of instructions contained by the alternative device.

In one possible implementation, step S41 may further include: and when the matrix vector calculation macro instruction is determined to contain the identification of the specified equipment and the resource of the specified equipment does not meet the execution condition for executing the matrix vector calculation macro instruction, determining the running equipment according to the matrix vector calculation macro instruction and the resource information of the alternative equipment.

In a possible implementation manner, the matrix vector calculation macro may further include at least one of an input quantity and an output quantity, and the step S41 of generating the operation instruction according to the matrix vector calculation macro and the operation device may include: and determining the data volume of the matrix vector calculation macro instruction, calculating the data volume of the macro instruction according to the matrix vector, calculating the macro instruction according to the matrix vector and resource information of the running equipment, and generating the running instruction. Wherein the data amount is determined according to at least one of the input amount and the output amount, and the resource information of the operating device further includes at least one of a storage capacity and a remaining storage capacity.

In one possible implementation manner, generating the operation instruction according to the data amount of the matrix vector calculation macro instruction, and the resource information of the operation device may include: when the number of the operation devices is determined to be one and the operation data volume of the operation devices is smaller than the data volume of the matrix vector calculation macro instruction, the matrix vector calculation macro instruction is split into a plurality of operation instructions according to the operation data volume and the data volume of the operation devices, so that the operation devices sequentially execute the plurality of operation instructions. The operation data volume of the operation device may be determined according to the resource information of the operation device, each operation instruction may further include at least one of an operation input volume and an operation output volume, and the operation input volume and the operation output volume are determined according to the operation data volume.

In one possible implementation manner, generating the operation instruction according to the data amount of the matrix vector calculation macro instruction, and the resource information of the operation device may include: when a plurality of operating devices are determined, the matrix vector calculation macro-instruction is split according to the operating data volume and the data volume of each operating device, and an operating instruction corresponding to each operating device is generated. The operation data volume of each operating device may be determined according to the resource information of each operating device, the operation instruction may further include at least one of an operation input volume and an operation output volume, and the operation input volume and the operation output volume are determined according to the operation data volume of the operating device that executes the operation instruction.

In one possible implementation, the method may further include: and sequencing the operating instructions according to a queue sequencing rule through the instruction generating equipment, and constructing an instruction queue corresponding to the operating equipment according to the sequenced operating instructions.

In one possible implementation, the method may further include: and receiving a matrix vector calculation instruction to be executed through the instruction generating equipment, and generating a matrix vector calculation macro instruction according to the determined identifier of the specified equipment and the matrix vector calculation instruction to be executed.

In one possible implementation, the method may further include: and sending the operation instruction to the operation equipment through the instruction generation equipment so as to enable the operation equipment to execute the operation instruction.

In a possible implementation manner, sending, by the instruction generating device, the execution instruction to the executing device to cause the executing device to execute the execution instruction may include:

generating an assembly file according to the operation instruction;

translating the assembly file into a binary file;

and sending the binary file to the running equipment so that the running equipment executes the running instruction according to the binary file.

In one possible implementation, the method may further include: the data and scalar data in the data are stored by the running device. The running equipment comprises a storage module, wherein the storage module comprises any combination of a register and a cache, and the cache comprises a temporary cache for high-speed storage and is used for storing data; and the register is used for storing scalar data in the data.

In one possible implementation, the method may further include:

storing the operation instruction through the operation equipment;

analyzing the operation instruction through the operation equipment to obtain a plurality of analysis instructions;

and storing an operation instruction queue through the operation equipment, wherein the operation instruction queue comprises an operation instruction and a plurality of analysis instructions, and the operation instruction queue operation instruction and the plurality of analysis instructions are sequentially arranged according to the executed sequence.

In one possible implementation, the method may further include:

the method comprises the steps that when the running equipment determines that the first analysis instruction and a zero analysis instruction before the first analysis instruction have an incidence relation, the first analysis instruction is cached, and after the execution of the zero analysis instruction is finished, the cached first analysis instruction is executed.

The method for analyzing the data comprises the following steps that an incidence relation exists between a first analysis instruction and a zeroth analysis instruction before the first analysis instruction: the first storage address interval for storing the data required by the first resolving instruction and the zeroth storage address interval for storing the data required by the zeroth resolving instruction have an overlapped area.

In one possible implementation, the running device may be one or any combination of a CPU, a GPU and an NPU.

In one possible implementation, the instruction generating device may be provided in the CPU and/or the NPU.

In one possible implementation, the matrix vector computation macro-instruction may include at least one of the following instructions: the method comprises the following steps of calculating a macro instruction by multiplying a matrix by a vector, calculating a macro instruction by multiplying a vector by a matrix, calculating a macro instruction by tensor, calculating a macro instruction by adding a matrix, and calculating a macro instruction by subtracting a matrix.

In one possible implementation, the matrix vector computation macro-instruction may further include at least one of the following options: identification of a specified device for executing the matrix vector computation macro-instruction, input quantities, output quantities, operands, and instruction parameters. The instruction parameters may include at least one of an address, a length of a second one of the operands.

According to the matrix vector calculation instruction processing method provided by the embodiment of the disclosure, the macro instruction is calculated according to the received matrix vector through the instruction generating equipment, the operation equipment for executing the matrix vector calculation macro instruction is determined, and the macro instruction and the operation equipment are calculated according to the matrix vector to generate the operation instruction; the data, the neural network model and the operation instruction are obtained through the operation equipment, the operation instruction is analyzed to obtain a plurality of analysis instructions, and the plurality of analysis instructions are executed according to the data to obtain an execution result. The method can be used in a cross-platform mode, and is good in applicability, high in instruction conversion speed, high in processing efficiency, low in error probability, and low in development labor and material cost.

It should be noted that, for simplicity of description, the above-mentioned method embodiments are described as a series of acts or combination of acts, but those skilled in the art will recognize that the present application is not limited by the order of acts described, as some steps may occur in other orders or concurrently depending on the application. Further, those skilled in the art should also appreciate that the embodiments described in the specification are exemplary embodiments and that acts and modules referred to are not necessarily required by the disclosure.

In the foregoing embodiments, the descriptions of the respective embodiments have respective emphasis, and for parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.

In the embodiments provided in the present disclosure, it should be understood that the disclosed apparatus may be implemented in other manners. For example, the above-described embodiments of the apparatus are merely illustrative, and for example, a division of modules is merely a division of logical functions, and an actual implementation may have another division, for example, a plurality of modules may be combined or may be integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection of devices or modules through some interfaces, and may be in an electrical or other form.

Modules described as separate parts may or may not be physically separate, and parts displayed as modules may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment.

In addition, functional modules in the embodiments of the present disclosure may be integrated into one processing unit, or each module may exist alone physically, or two or more modules are integrated into one module. The integrated module can be realized in a form of hardware or a form of a software program module.

The integrated modules, if implemented in the form of software program modules and sold or used as a stand-alone product, may be stored in a computer readable memory. Based on such understanding, the technical solution of the present disclosure may be embodied in the form of a software product, which is stored in a memory and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present disclosure. And the aforementioned memory comprises: a U-disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a removable hard disk, a magnetic or optical disk, and other various media capable of storing program codes.

Those skilled in the art will appreciate that all or part of the steps in the methods of the above embodiments may be implemented by associated hardware instructed by a program, which may be stored in a computer-readable memory, which may include: flash Memory disks, Read-Only memories (ROMs), Random Access Memories (RAMs), magnetic or optical disks, and the like.

The foregoing detailed description of the embodiments of the present application has been presented to illustrate the principles and implementations of the present application, and the above description of the embodiments is only provided to help understand the method and the core concept of the present application; meanwhile, for a person skilled in the art, according to the idea of the present application, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present application.

Claims

1. A matrix vector calculation instruction processing system, characterized in that the system comprises an instruction generating device and an executing device,

the instruction generating apparatus includes:

the device determining module is used for calculating the macroinstruction according to the received matrix vector and determining the running device for executing the matrix vector calculation macroinstruction;

the instruction generating module is used for calculating a macro instruction and the operation equipment according to the matrix vector to generate an operation instruction;

the operation device includes:

2. The system of claim 1, wherein the instruction generating device further comprises:

a resource obtaining module for obtaining resource information of the alternative device,

the device determination module comprises at least one of the following sub-modules:

the first determining submodule is used for determining the specified equipment as the running equipment when the matrix vector computing macro instruction is determined to contain the identification of the specified equipment and the resource of the specified equipment meets the execution condition of executing the matrix vector computing macro instruction;

a second determining submodule, configured to determine, when it is determined that the matrix vector calculation macro does not include the identifier of the designated device, an operating device for executing the matrix vector calculation macro from the candidate device according to the received matrix vector calculation macro and the resource information of the candidate device;

a third determining submodule, configured to determine an operating device according to the matrix vector calculation macro and the resource information of the candidate device when it is determined that the matrix vector calculation macro includes the identifier of the specified device and the resource of the specified device does not satisfy the execution condition for executing the matrix vector calculation macro,

wherein the execution condition includes: the specified device contains an instruction set corresponding to the matrix vector computing macro instruction, and the resource information includes the instruction set contained in the alternative device.

3. The system of claim 2 wherein the matrix vector computation macro instruction further includes at least one of an input quantity and an output quantity,

the instruction generating module is further configured to determine a data size of the matrix vector calculation macro instruction, generate an operation instruction according to the data size of the matrix vector calculation macro instruction, and the resource information of the operation device,

wherein the data amount is determined according to at least one of the input amount and the output amount, and the resource information of the operating device further includes at least one of a storage capacity and a remaining storage capacity.

4. The system of claim 3, wherein the instruction generation module comprises at least one of the following sub-modules:

a first instruction generation submodule, configured to split the matrix vector calculation macro instruction into multiple operation instructions according to the operation data amount of the operation device and the data amount when it is determined that the number of the operation devices is one and the operation data amount of the operation device is smaller than the data amount of the matrix vector calculation macro instruction, so that the operation device sequentially executes the multiple operation instructions,

a second instruction generation submodule, configured to split the matrix vector calculation macro instruction according to the operation data amount and the data amount of each operating device when it is determined that a plurality of operating devices are provided, and generate an operation instruction corresponding to each operating device,

the operation data volume of the operation equipment is determined according to the resource information of the operation equipment, the operation instruction further comprises at least one of operation input volume and operation output volume, and the operation input volume and the operation output volume are determined according to the operation data volume of the operation equipment executing the operation instruction.

5. The system of claim 1, wherein the instruction generating device further comprises:

and the queue construction module is used for sequencing the operating instructions according to a queue sequencing rule and constructing an instruction queue corresponding to the operating equipment according to the sequenced operating instructions.

6. The system of claim 2, wherein the instruction generating device further comprises:

and the macro instruction generating module is used for receiving a matrix vector calculation instruction to be executed and generating the matrix vector calculation macro instruction according to the determined identifier of the specified equipment and the matrix vector calculation instruction to be executed.

7. The system of claim 1, wherein the instruction generating device further comprises:

an instruction dispatching module, configured to send the execution instruction to the execution device, so that the execution device executes the execution instruction,

wherein the instruction dispatch module comprises:

the instruction assembly submodule is used for generating an assembly file according to the operation instruction;

the assembly translation submodule is used for translating the assembly file into a binary file;

and the instruction sending submodule is used for sending the binary file to the operating equipment so as to enable the operating equipment to execute the operating instruction according to the binary file.

8. The system of claim 1, wherein the operational equipment further comprises:

a storage module, the storage module comprises any combination of a register and a cache, the cache comprises a temporary cache,

the cache is used for storing the data;

and the register is used for storing scalar data in the data.

9. The system of claim 1, wherein the control module comprises:

the instruction storage submodule is used for storing the operation instruction;

the instruction processing submodule is used for analyzing the operation instruction to obtain a plurality of analysis instructions;

and the storage queue submodule is used for storing an operation instruction queue, the operation instruction queue comprises the operation instructions and the plurality of analysis instructions, and the operation instructions and the plurality of analysis instructions in the operation instruction queue are sequentially arranged according to the executed sequence.

10. The system of claim 9, wherein the execution module comprises:

the dependency relationship processing submodule is used for caching the first analysis instruction in the instruction storage submodule when determining that the first analysis instruction and a zero analysis instruction before the first analysis instruction have an incidence relationship, extracting the first analysis instruction from the instruction storage submodule after the zero analysis instruction is executed, and sending the first analysis instruction to the execution module,

wherein the association relationship between the first parsing instruction and a zeroth parsing instruction before the first parsing instruction comprises:

and a first storage address interval for storing the data required by the first resolving instruction and a zeroth storage address interval for storing the data required by the zeroth resolving instruction have an overlapped area.

11. The system of claim 1,

the running equipment is one or any combination of a CPU, a GPU and an NPU;

the instruction generation device is arranged in the CPU and/or the NPU;

the matrix vector computation macro-instruction comprises at least one of the following instructions: calculating a macroinstruction by multiplying a vector by a matrix, calculating a macroinstruction by multiplying the vector by the matrix, calculating a macroinstruction by tensor, calculating a macroinstruction by adding the matrix, and calculating a macroinstruction by subtracting the matrix;

the matrix vector computation macro further comprises at least one of the following options: and the identifier, the input quantity, the output quantity, the operand and the instruction parameter of a specified device for executing the matrix vector calculation macro instruction, wherein the instruction parameter comprises at least one of the address and the length of the second operand.

12. A machine learning arithmetic device, the device comprising:

one or more matrix vector computing instruction processing systems according to any one of claims 1 to 11, configured to obtain data to be computed and control information from other processing apparatuses, perform a specified machine learning operation, and transmit an execution result to the other processing apparatuses through an I/O interface;

13. A combined processing apparatus, characterized in that the combined processing apparatus comprises:

the machine learning computing device, universal interconnect interface, and other processing device of claim 12;

the machine learning arithmetic device interacts with the other processing devices to jointly complete the calculation operation designated by the user,

wherein the combination processing apparatus further comprises: and a storage device connected to the machine learning arithmetic device and the other processing device, respectively, for storing data of the machine learning arithmetic device and the other processing device.

14. The utility model provides a board card, its characterized in that, the board card includes: memory device, interface device and control device and machine learning chip comprising a machine learning arithmetic device according to claim 12 or a combined processing device according to claim 13;

wherein the machine learning chip is connected with the storage device, the control device and the interface device respectively;

the storage device is used for storing data;

the interface device is used for realizing data transmission between the machine learning chip and external equipment;

and the control device is used for monitoring the state of the machine learning chip.

15. A matrix vector calculation instruction processing method applied to a matrix vector calculation instruction processing system including an instruction generation device and an execution device, the method comprising:

16. The method of claim 15, further comprising:

acquiring resource information of the alternative device through the instruction generating device,

the method comprises the following steps that through the instruction generation equipment, a macro instruction is calculated according to a received matrix vector, and the operation equipment for executing the matrix vector calculation macro instruction is determined, wherein the operation equipment comprises at least one of the following items:

when determining that the matrix vector calculation macro instruction comprises an identifier of a specified device and the resource of the specified device meets the execution condition for executing the matrix vector calculation macro instruction, determining the specified device as the running device;

when determining that the matrix vector calculation macro instruction does not contain the identifier of the specified device, determining running equipment for executing the matrix vector calculation macro instruction from the candidate device according to the received matrix vector calculation macro instruction and the resource information of the candidate device;

when the matrix vector computing macro instruction is determined to contain the identification of the specified equipment and the resource of the specified equipment does not meet the execution condition for executing the matrix vector computing macro instruction, determining operating equipment according to the matrix vector computing macro instruction and the resource information of the alternative equipment,

17. The method of claim 16 wherein the matrix vector computation macro instruction further includes at least one of an input quantity and an output quantity,

according to the matrix vector calculation macro and the operation device, generating an operation instruction, comprising:

determining the data quantity of the matrix vector calculation macro instruction, generating an operation instruction according to the data quantity of the matrix vector calculation macro instruction, the matrix vector calculation macro instruction and the resource information of the operation equipment,

18. The method of claim 17, wherein generating the run instruction according to the data amount of the matrix vector computation macro instruction, and the resource information of the run device comprises at least one of:

when the number of the operating devices is determined to be one and the operating data volume of the operating devices is smaller than the data volume of the matrix vector calculation macro-instruction, splitting the matrix vector calculation macro-instruction into a plurality of operating instructions according to the operating data volume and the data volume of the operating devices, so that the operating devices sequentially execute the plurality of operating instructions;

when a plurality of operating devices are determined, splitting the matrix vector calculation macro instruction according to the operating data volume and the data volume of each operating device to generate an operating instruction corresponding to each operating device,

19. The method of claim 15, further comprising:

and sequencing the operating instructions according to a queue sequencing rule through the instruction generating equipment, and constructing an instruction queue corresponding to the operating equipment according to the sequenced operating instructions.

20. The method of claim 16, further comprising:

and receiving a matrix vector calculation instruction to be executed through the instruction generating equipment, and generating the matrix vector calculation macro instruction according to the determined identifier of the specified equipment and the matrix vector calculation instruction to be executed.

21. The method of claim 15, further comprising:

sending, by the instruction generating apparatus, the execution instruction to the execution apparatus to cause the execution apparatus to execute the execution instruction,

the sending of the operation instruction to the operation device through the instruction generation device so as to enable the operation device to execute the operation instruction includes:

generating an assembly file according to the operation instruction;

translating the assembly file into a binary file;

and sending the binary file to the operating equipment so that the operating equipment executes the operating instruction according to the binary file.

22. The method of claim 15, further comprising:

storing, by the runtime device, the data and scalar data in the data,

wherein the running device comprises a storage module, the storage module comprises any combination of a register and a cache, the cache comprises a temporary cache,

the cache is used for storing the data;

and the register is used for storing scalar data in the data.

23. The method of claim 15, further comprising:

storing, by the operational device, the operational instructions;

and storing an operation instruction queue through the operation equipment, wherein the operation instruction queue comprises the operation instructions and the plurality of analysis instructions, and the operation instructions and the plurality of analysis instructions in the operation instruction queue are sequentially arranged according to the executed sequence.

24. The method of claim 23, further comprising:

caching the first analysis instruction when the running equipment determines that the first analysis instruction and a zeroth analysis instruction before the first analysis instruction have an incidence relation, executing the cached first analysis instruction after the zeroth analysis instruction is executed,

25. The method of claim 15,

the running equipment is one or any combination of a CPU, a GPU and an NPU;

the instruction generation device is arranged in the CPU and/or the NPU;

the matrix vector computation macro instruction further comprises at least one of the following options: and the identifier, the input quantity, the output quantity, the operand and the instruction parameter of a specified device for executing the matrix vector calculation macro instruction, wherein the instruction parameter comprises at least one of the address and the length of the second operand.