Arm Limited

Erik Persson

Ola Hugosson

Andreas Björklund

John Wakefield Brothers

Dominic Hugo Symes

Graeme Leslie Ingram

Sven Ola Johannes Hugosson

Tomas Fredrik Edsö

Rune Holm

Pontus Borg

Stefan Johannes Frid

Daren Croxford

Roberto Lopez Mendez

Mats Petter Wallander

Sean Tristram LeGuay Ellis

Michal Bogusz

Fredrik Peter Stolt

Elliot Maurice Simon Rosemarine

Davide Marani

Philip Gregory Hall

Per Edgren

Viet-Hoa Do

Robert Norberg

Dan Staffan Henrik Johansson

Derek David Whiteman

A method of processing image data of a neural network is performed by a data processing apparatus and comprises writing a first tensor to first storage of the data processing apparatus using a row stride, wherein the first tensor comprises at least one data group, the at least one data group comprising a plurality of data samples and having height, width, and depth dimensions [h, w, c]. The method further comprises transforming the first tensor into a second tensor using a first stride such that the second tensor is a column tensor comprising a plurality of rows, and writing the second tensor to second storage using a second stride that is related to a multiple of the first stride, δ, such that the second stride covers a first set of memory elements in the second storage into which data samples of a first row of the second tensor are stored and a second set of memory elements into which no data samples from the second tensor are stored.

Processing data of a neural network

The present disclosure relates generally to multi-processor arrangements and, more particularly, to a point of serialization for broadcast communications within multi-processor arrangements.

Multi-processor arrangement with point of serialization for broadcast communications

The present disclosure relates generally to multi-processor arrangements and, more particularly, to broadcast regions for multi-processor arrangements.

Broadcast regions for multi-processor arrangement

A computer implemented method of storing and retrieving feature map data of a neural network the method comprising receiving a first portion of feature map data from local storage, selecting a first set of subportions of the first portion of feature map data, compressing the subportions to produce a first plurality of sections of compressed feature map data and instructing the storage of the sections into external storage. The method also comprises receiving a second plurality of sections of compressed feature map data from the external storage, decompressing the sections to produce a second set of subportions of the second portion of feature map data and storing the second portion of feature map data in local storage. The first and second sets of subportions each correspond to a predetermined format of subdivision and the method comprises selecting the predetermined format of subdivision from a plurality of predetermined formats of subdivision.

Data transfers in neural processing

The present disclosure relates generally to multi-processor arrangements and, more particularly, to broadcast hubs for multi-processor arrangements. A processing tile may comprise a broadcast hub to obtain a plurality of parameters applicable in a particular operation from at least one of a plurality of processing tiles and initiate distribution of the plurality of parameters to the plurality of processing tiles, wherein the plurality of processing tiles may execute the particular operation based at least in part on the plurality of distributed parameters.

Broadcast hub for multi-processor arrangement

A method for adapting a trained neural network is provided. Input data is input to the trained neural network and a plurality of filters are applied to generate a plurality of channels of activation data. Differences between corresponding activation values in the plurality of channels of activation data are calculated and an order of the plurality of channels is determined based on the calculated differences. The neural network is adapted so that it will output channels of activation data in the determined order. The ordering of the channels of activation data is subsequently used to compress activation data values by taking advantage of a correlation between activation data values in adjacent channels.

System and method for compressing activation data

A head-mounted display (HMD) comprising a first side for facing a user of the HMD, a second side opposite to the first side, and a reflective layer for at least partially reflecting incident light incident on the second side. At least one processor of the HMD is configured to obtain luminance data indicative of a luminance of the incident light and control a display device, based on the luminance data, to control a luminance of a portion of emitted light directed towards the user of the HMD during the display of the image. Further examples relate to an HMD with a display device configured to emit light of at least one predetermined wavelength range during display of an image by the display device, and a layer arranged to at least partially prevent transmission of the light of the at least one predetermined wavelength range outward from the HMD.

Head-mounted display

A producer processing unit of a data processing system that is producing a stream of data for use by one or more consumer processing units of the data processing system maintains a record that is accessible to the consumer processing units of a position in the data stream for which it has written data to memory. The consumer processing units then control their reading of the data stream from the memory in accordance with the write position record maintained by the producer processing unit.

Data processing systems

A method for compressing activation data of a neural network to be written to a storage is provided. The activation data is formed into a plurality of groups and a state indicator indicates whether there are any data elements within each group that have a non-zero value. A second state indicator indicates, for groups having a non-zero value, whether sub-groups within the group contain a data element having a non-zero value. A sub-group state indicator indicates, for each sub-group having a non-zero value, which data elements within that sub-group have a non-zero value. Non-zero values of data elements in the activation data are encoded and a compressed data set is formed comprising the first state indicators, any second state indicators, any sub-group state indicators and the encoded non-zero values.

Compression and/or decompression of activation data

A data processing system is provided including a processing element and a storage. The storage stores code portions that, when executed by the processing element, cause the data processing system to obtain a plurality of variable length codes. Each variable length code has a first portion of a variable length and a second portion. The variable length codes are separated into first portions of the variable length codes and second portions of the variable length codes. The data processing system forms a processed stream in chunks using a set of control rules so that, for each chunk of the processed stream containing data from the first portions of the variable length codes, the data from the first portions forms a sub-stream within the chunk of the processed stream that has a size that is determined in accordance with the control rules.

Data processing system

A scaling apparatus for scaling data arrays, such as images, comprises first horizontal scaling stage circuitry operable to scale a data array input to the scaling apparatus in the horizontal direction, one or more line memories for storing horizontal lines for a data array, wherein the or each line memory is for storing a horizontal line of data for the data array, vertical scaling stage circuitry operable to read data stored in the one or more line memories and to scale the read data in the vertical direction, and second horizontal scaling stage circuitry operable to scale a data array in the horizontal direction.

Method of and apparatus for scaling data arrays

A data processing system comprises a video processor (). The data processing system is configured to, when a new frame () is to be encoded by the video processor (), determine for a sub-region of a set of plural sub-regions that the new frame () is divided into, whether the sub-region has changed from a previous frame (), and to control the encoding operation for the new frame () on the basis of the determination, e.g. to avoid performing motion estimation.

To perform motion estimation for a video frame block to be encoded, a difference measure is determined for each of a plurality of reference frame block positions at a first, coarser resolution. The determined difference measures are then used estimate difference measures for reference frame blocks at positions at a second resolution that is finer than the first resolution. The estimated second, finer position resolution difference measures are then used to select a set of reference frame block positions for which to determine 'full' difference measures. The determined 'full' difference measures for each of the selected reference frame block positions are then used to select the reference frame block position to use when encoding the frame block and a motion vector corresponding to that reference frame block position is associated with and encoded for the frame block being encoded.

Methods of and apparatus for encoding data arrays

A video encoder and method of video encoding are provided. At an encoding stage a selected degree of quantization is applied to the encoding of macroblocks of the input video sequence and quantized part-encoded macroblocks are generated. Quantization circuitry in the encoding stage is configured to select the selected degree of quantization for each macroblock in a current slice in dependence on a complexity estimate indicative of the expected entropy encoding complexity of a predetermined set of the quantized part-encoded macroblocks defined for that macroblock.

Rate control in video encoding

A video encoding apparatus for encoding a video stream comprises a reference frame cache for reference frame video data retrieved from a reference frame storage unit in external memory, which is derived from an individual frame of the video stream. First and second source frame storage units store first and second blocks of unencoded video data taken from first and second source frames of the video stream, respectively. First and second video encoders perform first and second encoding operations to encode the first and second blocks of unencoded video data with reference to the reference frame video data cached in the reference frame cache, respectively. The first video encoder and the second video encoder perform the first encoding operation and the second encoding operation in parallel with one another.

Growing community of inventors

Lund, Sweden

Erik Persson