Huawei Technologies Co., Limited

Zte Corporation

Hu Liu

Heng Liao

Hou Fun Lam

Honghui Yuan

Jiajin Tu

Hao Wang

Hai Chen

Fan Zhu

Zhenjiang Dong

Baoqing Liu

Bin Huang

Yu Liu

Jun Liang

Bing Li

Hui Chen

Yincheng Zhang

Qinglong Chen

Chio In Ieong

Zhiqiang Zhang

Qing Su

Luping Cui

Xiancheng Guo

The present disclosure provides example operation accelerators and compression methods. One example operation accelerator includes a storage configured to store first input data, weight data, and a control instruction, and an operation circuit connected to the storage and configured to perform matrix multiplication on the first input data and the weight data, to obtain a computation result. The operation accelerator further includes a compression module configured to compress the computation result to obtain compressed data, as well as a controller connected to the storage and configured to obtain the control instruction from the storage, and when the control instruction includes instructions to compress the computation result, control the compression module to compress the computation result to obtain the compressed data. The operation accelerator further includes a direct memory access controller connected to the compression module and configured to store the compressed data in a memory outside the operation accelerator.

Operation accelerator and compression method

An operation accelerator, a processing method, and a related device, the operation accelerator including a first memory configured to store an input dataset, a matrix converter configured to perform reading M row vectors from the input dataset, generating a first instruction, and sending the M row vectors and the first instruction to a second memory configured to perform, according to the first instruction, preprocessing on the M row vectors to obtain n row vectors, and storing the n row vectors, where the n row vectors include the M row vectors and (n−M) padding row vectors, the n row vectors are N row vectors in a target matrix, and a storage sequence of the n row vectors in the second memory is consistent with a sequence of the N row vectors in the target matrix.

Operation accelerator, processing method, and related device

The present disclosure discloses example operation accelerators and compression methods. One example operation accelerator performs operations, including storing, in a first buffer, first input data. In a second buffer, weight data can be stored. A computation result is obtained by performing matrix multiplication on the first input data and the weight data by an operation circuit connected to the input buffer and the weight buffer. The computation result is compressed by a compression module to obtain compressed data. The compressed data can be stored into a memory outside the operation accelerator by a direct memory access controller (DMAC) connected to the compression module.

Embodiments of the present invention disclose a matrix multiplier, and relate to the field of data computing technologies, so as to divide two matrices into blocks for computation. The matrix multiplier includes: a first memory, a second memory, an operation circuit, and a controller, where the operation circuit, the first memory, and the second memory may perform data communication by using a bus; and the controller is configured to control, according to a preset program or instruction, a first matrix and a second matrix to be divided into blocks, and control the operation circuit to perform a multiplication operation on corresponding blocks in the first memory and the second memory based on block division results of the controller. The matrix multiplier may be configured to perform a multiplication operation on two matrices.

Matrix multiplier

A data processing method and apparatus are disclosed. In various embodiments, R groups of proposal region sequences are obtained. Each group of proposal region sequence includes a plurality of proposal regions. In those embodiments, a VRPAC instruction is invoked to calculate an area of each proposal region in each group of proposal region sequence. For a jgroup of proposal region sequence in the R groups of proposal region sequences, a VIOU instruction and a VAADD instruction are invoked to determine j suppression matrices of the jgroup of proposal region sequence and determine a suppression vector of the jgroup of proposal region sequence based on the j suppression matrices. In those embodiments, an unsuppressed proposal region is determined based on a suppression vector of each group of proposal region sequence.

Data processing method and apparatus

A matrix processing method performed by a graphics processing unit (GPU) includes: determining a plurality of non-zero elements in a to-be-processed matrix at a processor in the GPU; generating a distribution matrix of the to-be-processed matrix at the processor, where the distribution matrix comprises identities for indicating positions of the plurality of non-zero elements in the to-be-processed matrix; obtaining a target matrix from another matrix by using the distribution matrix at a logic circuit in the processor, where the target matrix comprises a plurality of target elements from the another matrix; and performing matrix processing on the plurality of non-zero elements and the target matrix to obtain an operation result at the processor.

Matrix processing method and apparatus, and logic circuit

The present invention in the field of data calculation technologies, discloses an operation accelerator, to reduce time for performing a multiplication operation on two N*N matrices. The operation accelerator includes: a first memory, a second memory, an operation circuit, and a controller. The operation circuit performs data communication with the first memory and the second memory by using a bus. The operation circuit is configured to: extract matrix data from the first memory and the second memory, and perform a multiplication operation. The controller is configured to control, according to a preset program or instruction, the operation circuit to complete the multiplication operation. The operation accelerator is configured to perform a multiplication operation on two matrices.

Operation accelerator

The present application is in the field of data calculation technologies, and discloses an operation accelerator, to reduce time for performing a multiplication operation on two N*N matrices. The operation accelerator includes: a first memory, a second memory, an operation circuit, and a controller. The operation circuit performs data communication with the first memory and the second memory by using a bus. The operation circuit is configured to: extract matrix data from the first memory and the second memory, and perform a multiplication operation. The controller is configured to control, according to a preset program or instruction, the operation circuit to complete the multiplication operation. The operation accelerator is configured to perform a multiplication operation on two matrices.

A matrix processing method includes: determining a quantity of non-zero elements in a to-be-processed matrix, where the to-be-processed matrix is a one-dimensional matrix; generating a distribution matrix of the to-be-processed matrix, where the distribution matrix is used to indicate a position of a non-zero element in the to-be-processed matrix; combining the quantity of non-zero elements, values of all non-zero elements in the to-be-processed matrix arranged sequentially, and the distribution matrix, to obtain a compressed matrix of the to-be-processed matrix.

Embodiments of the present disclosure provide a method and bus for accessing a dynamic random access memory (DRAM). The embodiments include receiving an access instruction, where the access instruction includes an access address, the access address includes a physical address, and a first field and a second field that are additionally set, the first field is used to indicate an interleaving mode, the interleaving mode indicates a manner of selecting an access channel, the second field is used to indicate an interleaving granularity, and the interleaving granularity indicates a capacity of an address space corresponding to the access channel; determining, according to the first field and the second field, the access channel and an address corresponding to the access channel; and accessing the DRAM according to the access channel and the address corresponding to the access channel.

Method and bus for accessing dynamic random access memory

Technical effects of a method, an apparatus, and a system for operating a shared resource in an asynchronous multiprocessing system that are provided in the present invention are as follows: A processor in an asynchronous multiprocessing system implements an operation on a shared resource by locking a hardware resource lock, and the hardware resource lock is implemented by a register; in this way, a bus in the asynchronous multiprocessing system does not need to support a synchronization operation, and the processor also does not need to have a feature of supporting a synchronization operation, and is capable of implementing the operation on the shared resource only in a manner of accessing the register, which simplifies the operation on the shared resource by the processor in the asynchronous multiprocessing system, enlarges a selection range of the processor in the asynchronous multiprocessing system, and further improves flexibility of the asynchronous multiprocessing system.

Growing community of inventors

Shenzhen, China

Hu Liu