(19)
(11)EP 3 451 160 A1

(12)EUROPEAN PATENT APPLICATION
published in accordance with Art. 153(4) EPC

(43)Date of publication:
06.03.2019 Bulletin 2019/10

(21)Application number: 16899916.7

(22)Date of filing:  10.05.2016
(51)International Patent Classification (IPC): 
G06F 9/30(2018.01)
(86)International application number:
PCT/CN2016/081526
(87)International publication number:
WO 2017/185405 (02.11.2017 Gazette  2017/44)
(84)Designated Contracting States:
AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR
Designated Extension States:
BA ME
Designated Validation States:
MA MD

(30)Priority: 26.04.2016 CN 201610266970

(71)Applicant: Cambricon Technologies Corporation Limited
Beijing 100190 (CN)

(72)Inventors:
  • LUO, Tao
    Beijing 100190 (CN)
  • ZHI, Tian
    Beijing 100190 (CN)
  • LIU, Shaoli
    Beijing 100190 (CN)
  • CHEN, Tianshi
    Beijing 100190 (CN)
  • CHEN, Yunji
    Beijing 100190 (CN)

(74)Representative: AWA Sweden AB 
P.O. Box 5117
200 71 Malmö
200 71 Malmö (SE)

  


(54)APPARATUS AND METHOD FOR PERFORMING VECTOR OUTER PRODUCT ARITHMETIC


(57) A device and a method for performing a vector outer product operation are provided. The device includes a storage unit, a register unit, and a vector outer product operation unit. The storage unit is configured to store vectors; the register unit is configured to store addresses of the vectors stored; the vector outer product operation unit is configured to acquire vector addresses from the register unit according to a corresponding instruction set, to acquire vectors from the storage unit according to the vector addresses acquired, and to perform the vector outer product operation on the vectors acquired to obtain a result of the vector outer product operation. According to the disclosure, vector data involved in calculation can be temporarily stored into a scratchpad memory, so that vector data of different widths can be supported more flexibly and effectively in a vector outer product operation process, thereby improving a performance of executing vector outer product operations.




Description

TECHNICAL FIELD



[0001] The disclosure relates to a vector outer product operation device and method for performing a vector outer product operation according to a vector outer product operation instruction.

BACKGROUND



[0002] In the vector space category, the homomorphism between objects is a linear map. In fact, however, we often encounter the concept of "bilinear map". For example, a dot product is a bilinear map V × VC. In order to attribute the property of "bilinear" to the vector space category, one way is to construct a vector space Z associated with V, W, so that each bilinear map defined on V × W can be replaced by a unique linear map defined on Z, where Z is called a tensor product of V and W(also known as an outer product). For example, a calculated value of a vector outer product a×b is equal to |a|·|b|·sin<a, b>.

[0003] In the current computer field, more and more algorithms are now involving vector operations. Taking machine learning algorithms (i.e., mainstream algorithms of artificial intelligence in popular application fields) as an example, almost all existing classical algorithms of the machine learning algorithms contain a large number of vector outer product operations.

[0004] In the related art, one solution to conduct a vector outer product operation is to use general-purpose processors by executing general instructions via general-purpose register files and general-purpose functional units to perform a vector outer product operation. However, one defect of the solution is the low operation performance during a vector outer product operation because a single general-purpose processor is primarily used for scalar calculations. On the other hand, adopting multiple general-purpose processors for concurrent execution may lead to a possible performance bottleneck resulting from the intercommunication among the multiple general-purpose processors. In another related art, a vector calculation is conducted by adopting a graphics processing unit (GPU), where general-purpose register files and general-purpose stream processing units are configured to execute general single instruction multiple data (SIMD) instructions, thereby performing a vector outer product operation. Nonetheless, in the above-mentioned solution, GPU's on-chip small cache in performing large-scale vector outer product operations requires a constant transportation of off-chip data, which makes off-chip bandwidth a main performance bottleneck. In another related art, a specialized vector outer product operation device is adopted to perform vector calculations, where customized register files and processing units are used to perform a vector outer product operation. Limited by the register files, however, the present specialized vector outer product operation device is unable to flexibly support outer product operations of vectors of different lengths.

SUMMARY


(1)Technical problems to be solved



[0005] The disclosure aims to provide a vector outer product operation device and method to solve the problems existing in the related art, such as limited inter-chip communication, insufficient on-chip cache, and inflexible support for different vector lengths.

(2)Technical solutions



[0006] The disclosure provides a vector outer product operation device, which is configured to perform a vector outer product operation according to a vector outer product operation instruction. The device includes a storage unit, a register unit, and a vector outer product operation unit.

[0007] The storage unit is configured to store vectors.

[0008] The register unit is configured to store vector addresses, where the vector addresses are addresses of the vectors stored in the storage unit.

[0009] The vector outer product operation unit is configured to acquire the vector outer product operation instruction, to acquire vector addresses from the register unit according to the vector outer product operation instruction, to acquire vectors from the storage unit according to the vector addresses acquired, and to perform the vector outer product operation on the vectors acquired to obtain a result of the vector outer product operation.

[0010] The disclosure further provides a vector outer product operation method, which is configured to perform a vector outer product operation according to a vector outer product operation instruction. The method includes the following operations.

[0011] At S1, vectors are stored.

[0012] At S2, vector addresses are stored.

[0013] At S3, the vector outer product operation instruction is acquired, vector addresses are acquired according to the vector outer product operation instruction, vectors are acquired according to the vector addresses acquired, and the vector outer product operation is performed on the vectors acquired, to obtain a result of the vector outer product operation.

(3)Advantageous effects



[0014] According to the vector outer product operation device and method of the disclosure, vector data involved in calculation can be temporarily stored into a scratchpad memory. In the case of only transmitting a same instruction, data of different widths can be supported more flexibly and effectively in the vector outer product operation unit, and correlation problems in data storing can be solved, therefore an execution performance of a task including a large number of vector calculations can be improved. In addition, an instruction set adopted by the disclosure is convenient to use and flexible in supporting vectors of different lengths.

[0015] The disclosure can be applied to the following (including but not limited to) scenarios: a data processing device, a robot, a computer, a printer, a scanner, a telephone, a tablet, a smart terminal, a mobile phone, a driving recorder, a navigator, a sensor, a webcam, a cloud server, a camera, a video camera, a projector, a watch, headphones, a mobile storage, a wearable device, and other electronic products; an airplane, a ship, a car, and other vehicles; a television (TV), an air conditioner, a microwave oven, a refrigerator, an electric cooker, a humidifier, a washing machine, an electric lamp, a gas stove, a hood, and other home appliances; a nuclear magnetic resonance spectrometer, a B-ultrasonic, an electrocardiograph, and other medical equipment.

BRIEF DESCRIPTION OF THE DRAWINGS



[0016] 

FIG. 1 is a schematic structural diagram illustrating a vector outer product operation device according to an embodiment of the present disclosure.

FIG. 2 is a schematic diagram of the format of an instruction set according to an embodiment of the present disclosure.

FIG. 3 is a schematic structural diagram illustrating a vector outer product operation device according to an embodiment of the present disclosure.

FIG. 4 is a schematic flow diagram illustrating a method for performing a vector outer product operation instruction through a vector outer product operation device according to an embodiment of the present disclosure.

FIG. 5 is a schematic structural diagram illustrating a vector outer product operation unit according to an embodiment of the present disclosure.


DETAILED DESCRIPTION



[0017] The disclosure provides a vector outer product operation device and a corresponding instruction set. The device includes a storage unit, a register unit, and a vector outer product operation unit. The storage unit is configured to store vectors. The register unit is configured to store addresses of the vectors stored. The vector outer product operation unit is configured to acquire vector addresses from the register unit according to a vector outer product operation instruction, to acquire vectors from the storage unit according to the vector addresses acquired, and to perform a vector outer product operation on the vectors acquired to obtain a result of the vector outer product operation. According to the disclosure, vector data involved in calculation can be temporarily stored into a scratchpad memory, so that data of different widths can be supported more flexibly and effectively in a vector outer product operation process, and it is possible to improve an execution performance of a task including a large number of vector calculations.

[0018] FIG. 1 is a schematic structural diagram illustrating a vector outer product operation device according to an embodiment of the present disclosure. As illustrated in FIG. 1, the device includes a storage unit, a register unit, and a vector outer product operation unit.

[0019] The storage unit is configured to store vectors. In one implementation, the storage unit can be a scratchpad memory capable of supporting vector data of different sizes. The disclosure is able to temporarily store required calculation data into a scratchpad memory, to enable the operation device to more flexibly and effectively support data of different widths during vector outer product operations. It may be understood that, the storage unit includes, but is not limited to, a variety of different memory devices, such as a static random access memory (SRAM), an enhanced dynamic random access memory (eDRAM), a dynamic random access memory (DRAM), a memristor, a 3D-DRAM, a non-transitory memory, or the like.

[0020] The register unit is configured to store vector addresses, where the vector addresses are addresses of the vectors stored in the storage unit. In one implementation, the register unit can be a scalar register file, to provide scalar registers required for an operation. It should be noted that, the scalar register is configured to store not only vector addresses but also scalar data. When an operation of a vector and a scalar is performed, the vector outer product operation unit acquires not only a vector address but also a corresponding scalar from the register unit.

[0021] The vector outer product operation unit is configured to acquire a vector outer product operation instruction, to acquire vector addresses from the register unit according to the vector outer product operation instruction, to acquire vectors from the storage unit according to the vector addresses acquired, to perform a vector outer product operation on the vectors acquired to obtain a result of the vector outer product operation, and to store the result of the vector outer product operation into the storage unit.

[0022] According to one implementation of the disclosure, the vector outer product operation unit includes an addition section, a nonlinear operation section, and a multiplication section. Meanwhile, the vector outer product operation unit has a structure of multiple-stage pipeline, where the addition section is arranged in a first stage pipeline, the nonlinear operation section is arranged in a second stage pipeline, and the multiplication section is arranged in a third stage pipeline. For example, when calculating a× b=|a|•|b|•sin<a, b>, the addition section of the first stage pipeline is configured to calculate |a| and |b|, the nonlinear operation section of the second stage pipeline is configured to calculate sin<a, b>, and the multiplication section of the third stage pipeline is configured to calculate |a|•|b|•sin<a, b>. Due to the fact that the sections mentioned-above are arranged in different pipeline stages, when sequential order of successive serial multiple vector outer product operation instructions is consistent with sequential order of pipeline stages of corresponding sections, operations as demanded by the serial multiple vector outer product operation instructions can be implemented more efficiently.

[0023] According to one implementation of the disclosure, the vector outer product operation device further includes an instruction cache unit. The instruction cache unit is configured to store vector outer product operation instructions to be executed. It should be understood that, one instruction (for example, a vector outer product operation instruction) will be cached in the instruction cache unit while being executed. When execution of an instruction is completed, if the instruction is also the earliest one among instructions that have not been submitted and are in the instruction cache unit, the instruction will be submitted. Once the instruction above is submitted, status change of the device (i.e., the vector outer product operation device) caused by operations performed by the instruction cannot be cancelled. In an implementation, the instruction cache unit can be a reordering cache.

[0024] According to one implementation of the disclosure, the vector outer product operation device further includes an instruction processing unit. The instruction processing unit is configured to acquire a vector outer product operation instruction from the instruction cache unit, to process the vector outer product operation instruction acquired, and to provide the vector outer product operation instruction processed for the vector outer product operation unit. In an implementation, the instruction processing unit includes an instruction fetching module, a decoding module, and an instruction queue.

[0025] The instruction fetching module is configured to acquire the vector outer product operation instruction from the instruction cache unit.

[0026] The decoding module is configured to decode the vector outer product operation instruction acquired from the instruction cache unit.

[0027] The instruction queue is configured to sequentially store the vector outer product operation instruction decoded, and, considering that different instructions may have dependency relationships on registers included, to cache the vector outer product operation instruction decoded and then to send the vector outer product operation instruction cached when a dependency relationship is eliminated.

[0028] According to one implementation of the disclosure, the vector outer product operation device further includes a dependency relationship processing unit. The dependency relationship processing unit is configured to: determine whether the vector outer product operation instruction accesses the same vector as a previous vector outer product operation instruction (in other words, determine whether the vector outer product operation instruction and a previous vector outer product operation instruction access a same vector), before the vector outer product operation unit acquires the vector outer product operation instruction; and store the vector outer product operation instruction into a storage queue based on a determination that the vector outer product operation instruction accesses the same vector as the previous vector outer product operation instruction, and provide the vector outer product operation instruction in the storage queue for the vector outer product operation unit after execution of the previous vector outer product operation instruction is completed; or provide the vector outer product operation instruction for the vector outer product operation unit directly, based on a determination that the vector outer product operation instruction does not access the same vector as the previous vector outer product operation instruction. As an example, when the vector outer product operation instruction (i.e., a current vector outer product operation instruction) accesses a scratchpad memory, the previous vector outer product operation instruction and the vector outer product operation instruction may access a same memory space. In order to ensure the correctness of an execution result of the vector outer product operation instruction, if data of the current vector outer product operation instruction is detected to have a dependency relationship with data of the previous vector outer product operation instruction, the vector outer product operation instruction must wait within the storage queue until the dependency relationship is eliminated.

[0029] According to one implementation of the disclosure, the vector outer product operation device further includes an input/output (I/O) unit. The I/O unit is configured to store the vectors into the storage unit, or to acquire the result of the vector outer product operation from the storage unit. In an implementation, the I/O unit can be a direct storage unit and responsible for reading vector data from a memory or writing vector data into the memory.

[0030] The disclosure further provides a vector outer product operation method, which is configured to perform a vector outer product operation according to a vector outer product operation instruction. The method includes the following operations.

[0031] At S1, vectors are stored.

[0032] At S2, vector addresses (i.e., addresses of the vectors stored) are stored.

[0033] At S3, the vector outer product operation instruction is acquired, vector addresses are acquired according to the vector outer product operation instruction, vectors are acquired according to the vector addresses acquired, and the vector outer product operation is performed on the vectors acquired to obtain a result of the vector outer product operation.

[0034] According to one implementation of the disclosure, the method further includes the follows prior to the operation S3.

[0035] Vector outer product operation instructions are stored.

[0036] A vector outer product operation instruction is acquired from the vector outer product operation instructions stored.

[0037] The vector outer product operation instruction acquired is decoded.

[0038] The vector outer product operation instruction decoded is sequentially stored.

[0039] According to one implementation of the disclosure, the method further includes the follows prior to the operation S3.

[0040] Whether the vector outer product operation instruction accesses the same vector as a previous vector outer product operation instruction is determined (in other words, whether the vector outer product operation instruction and a previous vector outer product operation instruction access a same vector is determined). When the vector outer product operation instruction accesses the same vector as the previous vector outer product operation instruction, store the vector outer product operation instruction into a storage queue, and advance to the operation S3 after execution of the previous vector outer product operation instruction is completed; otherwise, advance to the operation S3 directly.

[0041] According to one implementation of the disclosure, the method further includes an operation of storing the result of the vector outer product operation.

[0042] According to one implementation of the disclosure, the operation S1 includes an operation of storing the vectors into a scratchpad memory.

[0043] According to one implementation of the disclosure, the vector outer product operation instruction includes an operation code and at least one operation field, where the operation code is configured to indicate a function of the vector outer product operation instruction, and the operation field is configured to indicate data information of the vector outer product operation instruction.

[0044] According to one implementation of the disclosure, the vector outer product operation includes an addition operation, a nonlinear operation, and a multiplication operation.

[0045] According to one implementation of the disclosure, a vector outer product operation unit has a structure of multiple-stage pipeline, and the structure of multiple-stage pipeline includes a first stage pipeline, a second stage pipeline, and a third stage pipeline. In an implementation, the addition operation is performed at the first stage pipeline, the nonlinear operation is performed at the second stage pipeline, and the multiplication operation is performed at the third stage pipeline.

[0046] According to one implementation of the disclosure, the instruction set adopts fixed length instructions.

[0047] In the process of performing a vector outer product operation, the vector outer product operation device fetches a vector outer product operation instruction for decoding, and then sends the vector outer product operation instruction decoded to the instruction queue for storing. The vector outer product operation device acquires parameters of the vector outer product operation instruction stored according to a decoding result. In an implementation, the parameters can be parameters which are directly written in an operation field of the vector outer product operation instruction, or can be parameters which are read from a specified register according to a register number of the operation field of the vector outer product operation instruction. One advantage of using a register to store the parameters is that there is no need to change the vector outer product operation instruction itself. As long as values of the register are changed by the vector outer product operation instruction, most of loops can be realized, which greatly saves the number of instructions required to solve some practical problems. After all the parameters for operating are acquired, whether the vector outer product operation instruction is immediately sent to the vector outer product operation unit can be determined according to a result determined by the dependency relationship processing unit (i.e., "data required by the vector outer product operation instruction has a dependency relationship with data required by the previous vector outer product operation instruction" or "data required by the vector outer product operation instruction does not have a dependency relationship with data required by the previous vector outer product operation instruction"). Once a dependency relationship is found between data of the vector outer product operation instruction and data of the previous vector outer product operation instruction, the vector outer product operation instruction must wait until execution of the depended previous vector outer product operation instruction is completed, and then the vector outer product operation instruction can be sent to the vector outer product operation unit for execution. In a specialized operation unit, the vector outer product operation instruction will be executed quickly, and a result (i.e., a result of a vector outer product operation) is written back to an address provided by the vector outer product operation instruction. In this situation, the execution of the vector outer product operation instruction is completed.

[0048] FIG. 2 is a schematic diagram of the format of an instruction set according to an embodiment of the present disclosure. As illustrated in FIG. 2, a vector outer product operation instruction includes an operation code and a plurality of operation fields. The operation code is configured to indicate a function of the vector outer product operation instruction, and a vector outer product operation unit can perform a vector outer product operation by identifying the operation code. The operation field is configured to indicate data information of the vector outer product operation instruction, where the data information can be immediate data or a register number (in other words, a register unit number). For example, to acquire a vector, a start address and a length of the vector can be obtained from a register according to a register number, and then the vector stored in a corresponding address can be obtained from the storage unit according to the start address and the length of the vector.

[0049] The instruction set contains vector outer product operation instructions different in functions. The vector outer product operation instruction includes, but is not limited to, a tensor (TENS) product instruction (i.e. vector outer product operation instruction), a vector retrieval (VR) instruction, a vector load (VLOAD) instruction, a vector storage (VS) instruction, and a vector movement (VMOVE) instruction.

[0050] For the tensor (TENS) product instruction, a device (for example, a vector outer product operation device) is configured to respectively extract vector data of a specified size from specified addresses of a scratchpad memory. Then, in a vector calculation unit, an outer product operation is performed on two vectors. Finally, a result is written back into a specified address of a scalar register file.

[0051] For the vector retrieval (VR) instruction, a device is configured to retrieve vector data of a specified size from a specified address of a scratchpad memory. Then, in a vector calculation unit, an element of a vector is extracted according to a specified position to serve as an output. Finally, a result is written back into a specified address of a scalar register file.

[0052] For the vector load (VLOAD) instruction, a device is configured to load vector data of a specified size from a specified external source address to a specified address of a scratchpad memory.

[0053] For the vector storage (VS) instruction, a device is configured to store vector data of a specified size from a specified address of a scratchpad memory into an external destination address.

[0054] For the vector movement (VMOVE) instruction, a device is configured to store (in other words, move) vector data of a specified size from a specified address of a scratchpad memory into another specified address of the scratchpad memory.

[0055] To illustrate objects, technical solutions, and advantageous effects of the disclosure more clearly, the disclosure will be further described in detail below with reference to specific embodiments of the disclosure and accompanying drawings.

[0056] FIG. 3 is a schematic structural diagram illustrating a vector outer product operation device according to an embodiment of the present disclosure. As illustrated in FIG. 3, the vector outer product operation device includes an instruction fetching module, a decoding module, an instruction queue, a scalar register file, a dependency relationship processing unit, a storage queue, a reordering cache, a vector outer product operation unit, a scratchpad memory, and an input/output (I/O) direct memory access module.

[0057] The instruction fetching module is configured to fetch a next instruction (i.e., a vector outer product operation instruction) to be executed from an instruction sequence, and to transmit the instruction fetched to the decoding module.

[0058] The decoding module is configured to decode the instruction transmitted, and to transmit the instruction decoded to the instruction queue.

[0059] Considering that different instructions may have dependency relationships on scalar registers included, the instruction queue is configured to cache the instruction decoded, and to transmit the instruction cached after a dependency relationship is eliminated.

[0060] The scalar register file is configured to provide scalar registers required by the vector outer product operation device during an operation.

[0061] For the dependency relationship processing unit, a vector outer product operation instruction (also known as a current instruction) processed by the dependency relationship processing unit may have a dependency relationship with a previous vector outer product operation instruction (also known as a previous instruction). When the vector outer product operation instruction accesses a scratchpad memory, the previous vector outer product operation instruction and the vector outer product operation instruction may access a same memory space. In order to ensure the correctness of an execution result of the vector outer product operation instruction, if data of the vector outer product operation instruction is detected to have a dependency relationship with data of the previous vector outer product operation instruction, the vector outer product operation instruction must wait within the storage queue until the dependency relationship is eliminated.

[0062] The storage queue is a sequential queue, when the data of the vector outer product operation instruction has a dependency relationship with the data of the previous vector outer product operation instruction, the vector outer product operation instruction will be stored in the sequential queue until the dependency relationship is eliminated.

[0063] For the reordering cache, the vector outer product operation instruction is cached in the reordering cache while being executed. When execution of a vector outer product operation instruction is completed, if the vector outer product operation instruction is also the earliest one among instructions that have not been submitted and are in the reordering cache, the vector outer product operation instruction will be submitted. Once the vector outer product operation instruction is submitted, status change of the vector outer product operation device caused by operations performed by the vector outer product operation instruction cannot be cancelled.

[0064] The vector outer product operation unit is responsible for all outer product operations of the vector outer product operation device, and the vector outer product operation instruction is sent to the vector outer product operation unit for execution.

[0065] The scratchpad memory is a temporary storage device dedicated to vector data, and can support vector data of different sizes.

[0066] The I/O direct memory access module is configured to directly access the scratchpad memory and responsible for reading data from or writing data into the scratchpad memory.

[0067] FIG. 4 is a schematic flow diagram illustrating a method for performing a vector outer product operation instruction through a vector outer product operation device according to an embodiment of the present disclosure. As illustrated in FIG. 4, a manner of performing a tensor (TENS) product instruction (i.e. vector outer product operation instruction) is as follows.

[0068] At S0, an instruction fetching module fetches the tensor (called as TENS for short) product instruction, and sends the TENS product instruction to a decoding module.

[0069] At S1, the decoding module decodes the TENS product instruction, and sends the TENS product instruction decoded to an instruction queue.

[0070] At S2, in the instruction queue, data corresponding to four operation fields of the TENS product instruction is acquired from a scalar register file according to the TENS product instruction decoded. In an implementation, the data acquired from the scalar register file includes a start address of a vector vin0, a length of the vector vin0, a start address of a vector vin1, and a length of the vector vin1.

[0071] At S3, after scalar data needed is acquired, the TENS product instruction is sent to a dependency relationship processing unit. The dependency relationship processing unit analyzes whether there is a dependency relationship between data of the TENS product instruction and data of an uncompleted previous tensor (TENS) product instruction. The TENS product instruction needs to wait (in other words, the TENS product instruction is cached) in a storage queue until there is no dependency relationship between the data of the TENS product instruction and the data of the uncompleted previous TENS product instruction.

[0072] At S4, after the dependency relationship between the data of the TENS product instruction and the data of the uncompleted previous TENS product instruction does not exist any longer, the TENS product instruction is transmitted to a vector outer product operation unit. The vector outer product operation unit retrieves needed vectors from a scratchpad memory according to addresses and lengths of needed data, and then a vector outer product operation is performed in the vector outer product operation unit.

[0073] At S5, after the above operations are completed, a result is written back to a specified address of the scratchpad memory, and the TENS product instruction of a reordering cache is submitted at the same time.

[0074] FIG. 5 is a schematic structural diagram illustrating a vector outer product operation unit according to an embodiment of the present disclosure. As illustrated in FIG. 5, the vector outer product operation unit includes an addition section, a nonlinear operation section, and a multiplication section. Meanwhile, the vector outer product operation unit has a structure of multiple-stage pipeline, where the addition section is arranged at a pipeline stage 1, the nonlinear operation section is arranged at a pipeline stage 2, and the multiplication section is arranged at a pipeline stage 3. Since these sections are arranged in different pipeline stages, when sequential order of successive serial multiple vector outer product operations is consistent with sequential order of pipeline stages of corresponding sections, operations as demanded by the series of vector outer product operation instructions can be achieved more effectively.

[0075] Although the objects, technical solutions and advantageous effects of the present disclosure has been described in detail according to certain embodiments, it should be understood that the disclosure is not to be limited to the disclosed embodiments but, on the contrary, is intended to cover various modifications and equivalent arrangements included within the scope of the appended claims, which scope is to be accorded the broadest interpretation so as to encompass all such modifications and equivalent structures as is permitted under the law.


Claims

1. A vector outer product operation device, configured to perform a vector outer product operation according to a vector outer product operation instruction, the device comprising:

a storage unit, configured to store vectors;

a register unit, configured to store vector addresses, wherein the vector addresses are addresses of the vectors stored in the storage unit; and

a vector outer product operation unit, configured to acquire the vector outer product operation instruction, to acquire vector addresses from the register unit according to the vector outer product operation instruction, to acquire vectors from the storage unit according to the vector addresses acquired, and to perform the vector outer product operation on the vectors acquired to obtain a result of the vector outer product operation.


 
2. The device of claim 1, further comprising:

an instruction cache unit, configured to store vector outer product operation instructions to be executed.


 
3. The device of claim 2, further comprising:

an instruction processing unit, configured to acquire a vector outer product operation instruction from the instruction cache unit, to process the vector outer product operation instruction acquired from the instruction cache unit, and to provide the vector outer product operation instruction processed for the vector outer product operation unit.


 
4. The device of claim 3, wherein the instruction processing unit comprises:

an instruction fetching module, configured to acquire the vector outer product operation instruction from the instruction cache unit;

a decoding module, configured to decode the vector outer product operation instruction acquired from the instruction cache unit; and

an instruction queue, configured to sequentially store the vector outer product operation instruction decoded.


 
5. The device of claim 1, further comprising a dependency relationship processing unit configured to:

determine whether the vector outer product operation instruction accesses the same vector as a previous vector outer product operation instruction, before the vector outer product operation unit acquires the vector outer product operation instruction; and

store the vector outer product operation instruction into a storage queue based on a determination that the vector outer product operation instruction accesses the same vector as the previous vector outer product operation instruction, and provide the vector outer product operation instruction in the storage queue for the vector outer product operation unit after execution of the previous vector outer product operation instruction is completed; or

provide the vector outer product operation instruction for the vector outer product operation unit directly, based on a determination that the vector outer product operation instruction does not access the same vector as the previous vector outer product operation instruction.


 
6. The device of claim 1, wherein the storage unit is further configured to store the result of the vector outer product operation.
 
7. The device of claim 6, further comprising:

an input/output (I/O) unit, configured to store the vectors into the storage unit, or to acquire the result of the vector outer product operation from the storage unit.


 
8. The device of claim 1, wherein the storage unit is a scratchpad memory.
 
9. The device of claim 1, wherein the vector outer product operation instruction comprises:

an operation code; and

at least one operation field;

wherein the operation code is configured to indicate a function of the vector outer product operation instruction; and wherein the operation field is configured to indicate data information of the vector outer product operation instruction.
 
10. The device of claim 9, wherein the data information comprises a register unit number, and the vector outer product operation unit is configured to access a register unit and to acquire a vector address according to the register unit number.
 
11. The device of claim 1, wherein the vector outer product operation unit comprises an addition section, a nonlinear operation section, and a multiplication section.
 
12. The device of claim 11, wherein the vector outer product operation unit has a structure of multiple-stage pipeline, and the structure of multiple-stage pipeline comprises a first stage pipeline, a second stage pipeline, and a third stage pipeline; wherein the addition section is arranged in the first stage pipeline, the nonlinear operation section is arranged in the second stage pipeline, and the multiplication section is arranged in the second stage pipeline.
 
13. A vector outer product operation method, configured to perform a vector outer product operation according to a vector outer product operation instruction, the method comprising:

S1, storing vectors;

S2, storing vector addresses; and

S3, acquiring the vector outer product operation instruction, acquiring vector addresses according to the vector outer product operation instruction, acquiring vectors according to the vector addresses acquired, and performing the vector outer product operation on the vectors acquired to obtain a result of the vector outer product operation.


 
14. The method of claim 13, further comprising the follows prior to the operation S3:

storing vector outer product operation instructions;

acquiring a vector outer product operation instruction from the vector outer product operation instructions stored;

decoding the vector outer product operation instruction acquired; and

sequentially storing the vector outer product operation instruction decoded.


 
15. The method of claim 13, further comprising the follows prior to the operation S3:

determining whether the vector outer product operation instruction accesses the same vector as a previous vector outer product operation instruction; and

storing the vector outer product operation instruction into a storage queue based on a determination that the vector outer product operation instruction accesses the same vector as the previous vector outer product operation instruction, and advancing to the operation S3 after execution of the previous vector outer product operation instruction is completed; or

advancing to the operation S3 directly, based on a determination that the vector outer product operation instruction does not access the same vector as the previous vector outer product operation instruction.


 
16. The method of claim 13, further comprising:

storing the result of the vector outer product operation.


 
17. The method of claim 13, wherein the operation S1 comprises:

storing the vectors into a scratchpad memory.


 
18. The method of claim 13, wherein the vector outer product operation instruction comprises an operation code and at least one operation field; wherein the operation code is configured to indicate a function of the vector outer product operation instruction and the operation field is configured to indicate data information of the vector outer product operation instruction.
 
19. The method of claim 13, wherein the vector outer product operation comprises an addition operation, a nonlinear operation, and a multiplication operation.
 
20. The method of claim 19, wherein a vector outer product operation unit has a structure of multiple-stage pipeline, and the structure of multiple-stage pipeline comprises a first stage pipeline, a second stage pipeline, and a third stage pipeline; wherein the addition operation is performed at the first stage pipeline, the nonlinear operation is performed at the second stage pipeline, and the multiplication operation is performed at the third stage pipeline.
 




Drawing



















Search report