-
公开(公告)号:US20210117231A1
公开(公告)日:2021-04-22
申请号:US17136020
申请日:2020-12-29
Applicant: HUAWEI TECHNOLOGIES CO., LTD.
Inventor: Xiaogang Pan , Chao Zhou
Abstract: Various embodiments provide a task processing method and apparatus. In those embodiments, when a first device determines that an available computing resource is less than a computing resource required by a to-be-executed AI task group, at least one second device from a valid device in a first area is selected. The first area is an area that uses the first device as a center and a valid communication distance of the first device as a radius, and the valid device in the first area is a collaborative computing device that is in the first area and that is connected to the first device. Still in those embodiments, a task in the to-be-executed AI task group is sent by the first device to each of the at least one second device; and an execution result of the sent task from the second device is received by the first device.
-
2.
公开(公告)号:US12166685B2
公开(公告)日:2024-12-10
申请号:US18324742
申请日:2023-05-26
Applicant: HUAWEI TECHNOLOGIES CO., LTD.
Inventor: Qiang Chen , Sicong Li , Xiaogang Pan , Yidu Chen
Abstract: A method for implementing collective communication, a computer device, and a communication system are provided, to resolve problems of high communication latency and resource consumption in the conventional technology. The method includes: identifying a work request having no communication dependency, and when forwarding work requests, directly forwarding the work request identified as having no communication dependency, and forwarding, after queue management, a work request that is not identified as having no communication dependency. According to the application, communication latency caused when the work request having no communication dependency is managed by using a queue can be avoided, and resource consumption caused by performing related management can be reduced, to improve overall communication performance of collective communication.
-
公开(公告)号:US20240232281A1
公开(公告)日:2024-07-11
申请号:US18616470
申请日:2024-03-26
Applicant: Huawei Technologies Co., Ltd.
Inventor: Xiaogang Pan , Tengyi Lin , Zhaohui Ding
IPC: G06F17/14
CPC classification number: G06F17/142
Abstract: A method includes: receiving a fast Fourier transformation FFT execution request sent by an application program; obtaining first input data on which an FFT calculation needs to be performed; decomposing the FFT calculation into a plurality of calculation stages based on a cardinality of the FFT calculation, where each calculation stage includes a DFT calculation; determining a DFT calculation matrix corresponding to the DFT calculation in each calculation stage; determining, based on input data of a corresponding calculation stage when a calculation of each calculation stage is performed, a first data matrix on which a matrix operation is performed with the DFT calculation matrix; and performing a matrix multiplication calculation on the DFT calculation matrix and the first data matrix to implement the DFT calculation, and returning a calculation result to the application program after the plurality of calculation stages are completed.
-
4.
公开(公告)号:US20230300080A1
公开(公告)日:2023-09-21
申请号:US18324742
申请日:2023-05-26
Applicant: HUAWEI TECHNOLOGIES CO., LTD.
Inventor: Qiang Chen , Sicong Li , Xiaogang Pan , Yidu Chen
Abstract: This application provides a method for implementing collective communication, a computer device, and a communication system, to resolve problems of high communication latency and resource consumption in the conventional technology. The method provided in this application includes: identifying a work request having no communication dependency, and when forwarding work requests, directly forwarding the work request identified as having no communication dependency, and forwarding, after queue management, a work request that is not identified as having no communication dependency. In this way, communication latency caused when the work request having no communication dependency is managed by using a queue can be avoided, and resource consumption caused by performing related management can be reduced, to improve overall communication performance of collective communication.
-
-
-