Imagination Technologies Limited

Linling Zhang

Simon James Fenney

James Imber

Cagatay Dikici

Jeff Bond

Ilaria Martinelli

Greg Clark

Hierarchical methods for selecting fixed point number formats with reduced mantissa bit lengths for representing values input to, and/or output, from, the layers of a DNN. The methods begin with one or more initial fixed point number formats for each layer. The layers are divided into subsets of layers and the mantissa bit lengths of the fixed point number formats are iteratively reduced from the initial fixed point number formats on a per subset basis. If a reduction causes the output error of the DNN to exceed an error threshold, then the reduction is discarded, and no more reductions are made to the layers of the subset. Otherwise a further reduction is made to the fixed point number formats for the layers in that subset. Once no further reductions can be made to any of the subsets the method is repeated for continually increasing numbers of subsets until a predetermined number of layers per subset is achieved.

Hierarchical mantissa bit length selection for hardware implementation of deep neural network

A lossy method of compressing data, such as image data, which uses wrap-around wavelet compression is described. Each data value is divided into two parts and the first parts, which comprise the most significant bits from the data values, are compressed using wrap-around wavelet compression. Depending upon the target compression ratio and the compression ratio achieved by compressing just the first parts, none, one or more bits from the second parts, or from a data value derived from the second parts, may be appended to the compressed first parts. The method described may be lossy or may be lossless. A corresponding decompression method is also described.

Lossy data compression

A method of converting 10-bit pixel data (e.g. 10:10:10:2 data) into 8-bit pixel data involves converting the 10-bit values to 8-bits using a technique that is selected dependent upon the values of the MSBs of the 10-bit values and setting the value of an HDR flag dependent upon the values of the MSBs. The HDR flag is appended to the 3-bit channel.

Guaranteed data compression

Lossy methods and hardware for compressing data and the corresponding decompression methods and hardware are described. The lossy compression method comprises dividing a block of pixels into a number of sub-blocks and then analysing, for each sub-block, and selecting one of a candidate set of lossy compression modes. The analysis may, for example, be based on the alpha values for the pixels in the sub-block. In various examples, the candidate set of lossy compression modes comprises at least one mode that uses a fixed alpha channel value for all pixels in the sub-block and one or more modes that encode a variable alpha channel value.

Image synthesis methods and units are described which can be implemented in real-time on low-cost and mobile devices. A set of sample blocks from an input image are defined. For each combination of sample blocks, a matching block is determined, which provides a best match for that combination when positioned in conjunction with a predetermined arrangement of the sample blocks of the combination. A match table can then be created and stored which includes an indication of the determined matching block for each combination of sample blocks from the input image. The match table can be used to generate a synthesised image, which may be much larger than the input image. For example, the match table may be used to determine an index table which comprises indications of sample blocks to be positioned in a synthesis arrangement for generating the synthesised image.

Growing community of inventors

Watford, United Kingdom

Linling Zhang