39results about How to "Give full play to computing power" patented technology

The invention discloses a vector crossing multithread processing method and a vector crossing multithread microprocessor. The processing method comprises the following steps: using a multithread instruction-acquiring part to choose a vector thread from N vector threads for reading an instruction and storing the read instruction to a corresponding instruction buffer array of the vector thread; using a thread scheduling part to choose an instruction buffer array from N instruction buffer arrays and taking out an instruction from the instruction buffer array for the purpose of decoding; and sending a decoded instruction to a vector executing streamline or scalar executing streamline so as to execute. The method can be realized by using hardware structure by the vector crossing multithread microprocessor. The method and the microprocessor provided by the invention have the advantages that the vector processing technique and multithread technique are combined, the hardware structure is simple, the operation capability is strong, the compatibility and expansibility are excellent, and the like.

The invention discloses a data flow control method and apparatus. According to the technical scheme, the method mainly comprises the steps of obtaining task data; converting the task data into multiple pieces of sub-task data in a calculation format; caching the sub-task data in a circular buffer region; and according to a predetermined time sequence, obtaining the sub-task data from the circularbuffer region and forwarding the data to a calculation apparatus for performing task calculation. The technical scheme can ensure that the calculation apparatus always can obtain the sub-task data from a buffer to perform calculation, so that the calculation capability of the calculation apparatus can be brought into full play. The buffer adopts a circular buffer design, so that the problem of lock contention between threads can be avoided and the interruption of distributing the sub-task data to the calculation apparatus can be avoided. The technical scheme can be realized by adopting a bus structure, so that the mining hardware cost can be reduced and an excellent task data sending time sequence can be provided for a calculation capability board.

The invention discloses a service node backup method and a distributed system. Wherein, the service node backup method comprises following steps: a network management device processes to deploy for a mapping relationship between each physical processing unit in the distributed system and each logic service node, wherein, each physical processing unit exists the mapping relationship with at least one logic service node and each logic service node exists the mapping relationship with at least two logic service nodes; and each physical processing unit existing the mapping relationship with the same logic service node concurrent runs all service application program of the logic service node existing the mapping relationship with the physical processing unit. The invention can give full play to system hardware processing ability so as to improve system processing ability, at the same time cannot damage communication system safety backup mechanism.

The invention provides a new method for controlling flying shears for bars and wires based on T400, which is an advanced electrical control method for start-stop type flying shears for bars and wires. The method comprises the steps of accurate shearing-blade position control, shearing length control and automatic shearing optimization, and the like; in the process of production, when an upstream rolling mill has an accident, or the quality of a section of a rolled piece is poor, a fragmentation function can be started on line so as to carry out partial fragmentation or complete fragmentation on the rolled piece, and the fragmentation and shearing speed is in step with the running speed of the rolled piece automatically; and when the fragmentation operation is stopped, a flying shear is automatically positioned to a start position and enters into a standby state of the normal shear mode. In the invention, the powerful computation function of the T400 is fully used, core control parts of the flying shear are completely integrated into the T400, and high-speed counters, shaft positioning modules and touch screens are not required, therefore, the cost is reduced, and high accuracy is obtained; and when the flying shear carries out double-length shearing at a speed of 21 m / s, the double-length error is basically within 75 mm, and partial double-length error can be within 30 mm.

A three-dimensional CT image is obtained and output by sequentially storing image data of an image of an object to be processed according to a prescribed direction, and using a multicore processor, performing a back projection / projection process in parallel by a plurality of threads so that each thread sequentially processes in a direction that is orthogonal to the prescribed direction. Accordingly, by using a multicore processor such as a GPU, there is high flexibility and less restrictions and the technical problem of the process speed being limited by the band width is solved. At the same time, when processing the data using a multicore processor in storage and back projection / projection process, by taking into account the rules for multithread memory access of a multicore processor such as a GPU, cache hit ratio is improved, coalesced access is achieved and memory access speed is greatly improved.

This disclosure embodiment disclosed methods and devices for training deep learning models.A specific implementation method of this method includes: the model description information and configuration information of the deep learning model; the model description information is divided into at least two sections according to the cut points variable in the configuration information, and the model description information is loaded to the corresponding resources to the corresponding resourcesRun; enter a batch of training samples to enter the first section of the model to describe the resources corresponding to the information and start training and use the context information obtained as the input of resources corresponding to the resources corresponding to the information;Corresponding resource operation results; if you meet the training conditions of training, the deep learning model of the output training is completed; otherwise, continue to obtain the next batch of training samples to perform the above training steps until the training complete conditions are met.This implementation method realizes the free mix of heterogeneous equipment, gives full play to the computing capabilities of different computing devices, and improves training speed.

The invention relates to an SSL data packet rapid processing method and system, and an electronic device. The method comprises the steps of: selecting a preset byte number to group a plurality of bytes in the SSL data packet, in a first preset period, encrypting bytes of the first group through an encryption algorithm; and in the next preset period, judging whether a next group exists or not; andif so, encrypting bytes of the next group through the encryption algorithm, processing the encrypted bytes in the previous group through an HMAC algorithm; and if not, processing the encrypted bytes in the previous group through an HMAC algorithm, so that by that analogy, seamless connection and full-load work of the encryption algorithm and the HMAC algorithm can be guaranteed, the operational capability of an algorithm core is brought into full play, and through comparison experiment verification, when the SSL data packet rapid processing system is used, compared with a previous processing mode, the SSL data packet processing efficiency can be doubled.

The present invention provides a framework for performance optimization of grid application programs on many-core architectures, including an application field description layer, a programming model layer, a hardware architecture layer, and a converter; (1) the user uses the application field of the application field description layer Describe, in the case of only considering the domain programming, write the program to solve the domain problem; (2) The converter converts the user program into the code of the programming model layer based on the programming environment combination; (3) the programming model layer The code is compiled with the compiler of the corresponding programming environment, and linked with the corresponding dynamic library, and finally generates the code that can run at the hardware architecture layer; (4) through the mapping relationship between the performance model and the efficiency model, the converted programming The program at the model layer is manually optimized and adjusted, or even automatically iteratively optimized and adjusted through the converter.

The invention discloses a load balancing method and device using dual blade servers. The method includes: setting an input and output processing IOP module inside each blade server in an ATCA blade server system; PE bridges of each blade server and IOP The modules are connected, and the IOP module writes the input and output IO data into the memory through the PE bridge for processing by the central processing unit CPU of the blade server, and reads the data processed by the CPU from the memory of the blade server for transmission to the backplane through the IO device ;Two adjacent blade servers in the system are connected to each other's PE bridge through the backplane through the NT ports of their respective PE bridges. The two blade servers share resources with each other through the above connection, and the two blade server IOP modules Connected through the backplane, the IOP module provides a unified virtual IP address for the blade server. The invention improves the working efficiency of the blade server.