Huawei Technologies Co., Limited

Kunlunxin Technology (beijing) Company Limited

Beijing Baidu Netcom Science and Technology Co., Ltd.

Beijing Baidu Netcom Science Technology Co., Ltd

Yuan Ruan

Yao Liu

Shaojie Chen

Wei Yang

Mingyu Chen

Zehan Cui

Yongbing Huang

Haoyang Li

Licheng Chen

Mingyang Chen

Yupeng Li

Tianyue Lu

The present disclosure relates to technical field of data access, and discloses a matrix storage method, a matrix access method, an apparatus and an electronic device in the technical field of data access. The matrix storage method includes: dividing a matrix into a plurality of data blocks with a preset segmentation granularity of N rows&#xd7;M columns; the plurality of data blocks includes at least one first data block of N rows&#xd7;M columns; if the column number of the matrix is not an integer multiple of M, the plurality of data blocks further includes at least one second data block of N rows&#xd7;P columns, the second data block is aligned with an adjacent row of first data block; and storing the data in each of the first data blocks and the second data blocks continuously in an off-chip storage.

Matrix storage method, matrix access method, apparatus and electronic device

A data output method, a data acquisition method, a device, and an electronic apparatus are provided, and a specific technical solution is: reading a first data sub-block, and splicing the first data sub-block into a continuous data stream, wherein the first data sub-block is a data sub-block in transferred data in a neural network; compressing the continuous data stream to acquire a second data sub-block; determining, according to a length of the first data sub-block and a length of the second data sub-block, whether there is a gain in compression of the continuous data stream; outputting the second data sub-block if there is the gain in the compression of the continuous data stream.

Data output method, data acquisition method, device, and electronic apparatus

A neural network data processing apparatus includes: an instruction parsing module, configured to split a DMA task into multiple subtasks and acquire configuration information of a data sub-block corresponding to each subtask, where the subtasks are in a one-to-one correspondence with data sub-blocks of transported neural network data; a data reading module, configured to read a first data sub-block according to the configuration information, where the first data sub-block is a data sub-block among data sub-blocks corresponding to multiple subtasks; a data processing module, configured to compress the first data sub-block; a data write-out module, configured to output compressed data resulting from the compression of the first data sub-block.

Neural network data processing apparatus, method and electronic device

In a method for accessing an extended memory, after receiving a first memory access request from a processor system in a computer, an extended memory controller sends a read request for obtaining to-be-accessed data to the extended memory and return, to the processor system, a first response message indicating the to-be-accessed data has not been obtained. The extended memory controller writes the to-be-accessed data into a data buffer after receiving the to-be-accessed data returned by the extended memory. After receiving, from the processor system, a second memory access request comprising a second access address, the extended memory controller returns, to the processor system, the to-be-accessed data in the data buffer in response to the second memory access request, wherein the second access address is different from the first access address and points to the physical address of the to-be-accessed data.

Method for accessing extended memory, device, and system

In the field of data reading and writing technologies, a data writing method is associated with a data writing apparatus and an electronic device. The data writing method includes: determining whether a start storage address of a first data block is aligned with a bus bit width of a storage; in response to that the start storage address of the first data block is not aligned with the bus bit width of the storage, determining whether a second data block which is a data block immediately before the first data block is compressed; in response to that the second data block is compressed, executing complete writing on a first beat of the first data block.

Data writing method and apparatus, and electronic device

In a non-volatile memory (NVM) persistence method, a memory controller maintains a plurality of memory write instruction sets that are locked separately for persistence of write operations. A credit value for monitoring a quantity of completed write instructions is configured for each of the plurality of memory write instruction sets, and a credit value is also configured for each of a plurality of medium write instruction sets maintained by a medium controller and corresponding respectively to the memory write instruction sets. After a memory write instruction set is locked in response to a persistence query, the credit value of a corresponding medium write instruction set is used as means for the memory controller to determine whether the write instructions in the locked memory write instruction set have been completed by the medium controller.

Non-volatile memory persistence method and computing device

A memory management method and a device, where the method includes: receiving a memory access request, where the memory access request carries a virtual address; determining a page fault type of the virtual address if finding, in a translation lookaside buffer TLB and a memory, no page table entry corresponding to the virtual address; allocating a corresponding page to the virtual address if the page fault type of the virtual address is a blank-page-caused page fault, where the blank-page-caused page fault means that no corresponding page is allocated to the virtual address; and updating the page table entry corresponding to the virtual address to the memory and the TLB. The memory manager does not generate a page fault when a blank-page-caused page fault occurs, but allocates a corresponding page to the virtual address. Therefore, a quantity of occurrences of the page fault is reduced, thereby improving memory management efficiency.

Memory management and device

A method for accessing an extended memory, a device, and a system are disclosed. According to the method, after receiving a first memory access requests from a processor system in a computer, an extended memory controller sends a read request for obtaining to-be-accessed data to the extended memory and return, to the processor system, a first response message indicating the to-be-accessed data has not been obtained. The extended memory controller writes the to-be-accessed data into a data buffer after receiving the to-be-accessed data returned by the extended memory. After receiving, from the processor system, a second memory access request comprising a second access address, the extended memory controller returns, to the processor system, the to-be-accessed data in the data buffer in response to the second memory access request, wherein the second access address is different from the first access address and points to the physical address of the to-be-accessed data.

A method for reducing power consumption of a memory system and a memory controller are provided. The method for reducing power consumption of a memory system includes: determining whether a dynamic random access memory DRAM memory module with a low access frequency exists in a memory system; when a DRAM memory module with a low access frequency exists, transfer, according to a size of a working set in the memory system, page data that does not belong to the working set to a non-volatile memory NVM memory module, where the page data that does not belong to the working set is page data that does not need to be accessed when a process runs within preset time.

Method for reducing power consumption of memory system, and memory controller

A memory access processing method is based on memory chip interconnection, a memory chip, and a system, which relate to the field of electronic devices, and can shorten a time delay in processing a memory access request and improve a utilization rate of system bandwidth. The method of the present disclosure includes receiving, by a first memory chip, a memory access request; and if the first memory chip is not a target memory chip corresponding to the memory access request, sending, according to a preconfigured routing rule, the memory access request to a next memory chip connected with the first memory chip, until the target memory chip corresponding to the memory access request is determined. Embodiments of the present disclosure are mainly used in a process of processing a memory access request.

Memory access processing method, memory chip, and system based on memory chip interconnection

A message-based memory access apparatus and an access method thereof are disclosed, The message-based memory access apparatus includes: a message-based command bus, configured to transmit a message-based memory access instruction generated by the CPU to instruct a memory system to perform a corresponding operation; a message-based memory controller, configured to package a CPU request into a message packet and sent the packet to a storage module, and parse a message packet returned by the storage module and return data to the CPU; a message channel, configured to transmit a request message packet and a response message packet; and the storage module, including a buffer scheduler, and configured to receive the request packet from the message-based memory controller and process the corresponding request.

Message-based memory access apparatus and access method thereof

A first level buffer chip gates a target second level buffer chip according to a preset mapping relationship, a first chip select signal, and a first higher-order address signal, and forwards a memory access instruction and a lower-order address signal received from a memory controller to the target second level buffer chip. The target second level buffer chip determines a target memory module according to a second chip select signal and a delayed address signal obtained by delay processing on a second higher-order address signal, determines a target memory chip according to the lower-order address signal, acquires target data from the target memory chip according to the memory access instruction, and returns the target data to the memory controller. A cascading manner of a system memory is changed to a tree-like topological form, which avoids a protocol conversion problem and reduces the memory access time.

Reducing latency in an expanded memory system

An address compression method, an address decompression method, a compressor, and a decompressor are disclosed, wherein the address compression method includes after a compressor receives multiple operation request messages that are sent by a first processor, determining, according to an address feature formed by address information carried in all operation request messages that have a same stream number, a compression algorithm corresponding to the operation request messages that have a same stream number; and then compressing, according to the determined compression algorithm, addresses carried in the operation request messages that have a same stream number. The present invention is applicable to the computer field.

Growing community of inventors

Beijing, China

Yuan Ruan