Processing cores with data associative adaptive rounding and associated methods are disclosed herein. One disclosed processing core comprises an arithmetic logic unit cluster configured to generate a value for a unit of directed graph data using input directed graph data, a comparator coupled to a threshold register and a data register, a core controller configured to load a threshold value into the threshold register when the value for the unit of directed graph data is loaded into the data register, and a rounding circuit. The rounding circuit is configured to receive the value for the unit of directed graph data from the arithmetic logic unit cluster and conditionally round the value for the unit of directed graph data based on a comparator output from the comparator.

Disclosed herein is a method for measuring the weight of a discrete entity, performed in a neural network model configured with multiple layers, the method including receiving data configured with the indices of discrete entities, converting the data into embedding vectors corresponding to respective indices through an embedding layer, generating a masked vector through element-wise multiplication between a mask vector and the embedding vector, calculating a loss using output based on the masked vector, and training the model based on the loss.

A device for deep learning acceleration with mixed precision may include a precision mode port configured to receive an indication of an output precision mode, a data input port configured to receive an input value, and a truncation component configured to truncate the input value into a keep segment value and a truncate segment value. The device may be configured to add the keep segment value and a carry bit to generate a rounded keep segment value, and to generate a rounded output based on the rounded keep segment value and the output precision mode. The rounded output generation component may be configured to generate the rounded output to include a sign bit of the keep segment value and either a first quantity or a second quantity of lower bits of the keep segment value based on the output precision mode being either a first value or a second value.

Neural processing elements are configured with a hardware AND gate configured to perform a logical AND operation between a sign extend signal and a most significant bit (“MSB”) of an operand. The state of the sign extend signal can be based upon a type of a layer of a deep neural network (“DNN”) that generate the operand. If the sign extend signal is logical FALSE, no sign extension is performed. If the sign extend signal is logical TRUE, a concatenator concatenates the output of the hardware AND gate and the operand, thereby extending the operand from an N-bit unsigned binary value to an N+1 bit signed binary value. The neural processing element can also include another hardware AND gate and another concatenator for processing another operand similarly. The outputs of the concatenators for both operands are provided to a hardware binary multiplier.

A fixed binary adder adds an “N”-bit second operand to a first operand having an “N”-bit fixed value (N=2.sup.M, M is a natural number) to generate “N+1”-bit output data. The fixed binary adder includes a plurality of transfer logic stages each configured with at least one logic gate, and a summation addition logic configured to generate the “N+1”-bit output data by using the “N”-bit second operand and transfer data that is generated through the plurality of transfer logic stages. The logic gate is configured with one of an AND gate, an OR gate, and a buffer gate.

Disclosed herein are a floating-point computation apparatus and method using Computing-in-Memory (CIM). The floating-point computation apparatus performs a multiply-and-accumulation operation on pieces of input neuron data represented in a floating-point format, and includes a data preprocessing unit configured to separate and extract an exponent and a mantissa from each of the pieces of input neuron data, an exponent processing unit configured to perform CIM on input neuron exponents, which are exponents separated and extracted from the input neuron data, and a mantissa processing unit configured to perform a high-speed computation on input neuron mantissas, separated and extracted from the input neuron data, wherein the exponent processing unit determines a mantissa shift size for a mantissa computation and transfers the mantissa shift size to the mantissa processing unit, and the mantissa processing unit normalizes a result of the mantissa computation and transfers a normalization value to the exponent processing unit.

A method is provided that includes performing, by a processor in response to a vector finite impulse response (VFIR) filter instruction, generating of a plurality of filter outputs using a plurality of coefficients and a plurality of sequential data elements, the plurality of coefficients specified by a coefficient operand of the VFIR filter instruction and the plurality of sequential data elements specified by a data operand of the VFIR filter instruction, and storing the filter outputs in a storage location specified by the VFIR filter instruction.

A histogram-based method of selecting a fixed point number format for representing a set of values input to, or output from, a layer of a Deep Neural Network (DNN). The method comprises obtaining a histogram that represents an expected distribution of the set of values of the layer, each bin of the histogram is associated with a frequency value and a representative value in a floating point number format; quantising the representative values according to each of a plurality of potential fixed point number formats; estimating, for each of the plurality of potential fixed point number formats, the total quantisation error based on the frequency values of the histogram and a distance value for each bin that is based on the quantisation of the representative value for that bin; and selecting the fixed point number format associated with the smallest estimated total quantisation error as the optimum fixed point number format for representing the set of values of the layer.

Apparatus includes hardware logic arranged to normalise an n-bit input number. The hardware logic comprises at least a first hardware logic stage, an intermediate hardware logic stage and a final hardware logic stage. Each stage comprises a left shifting logic element, the first and intermediate stages each also comprise a plurality of OR-reduction logic elements and the intermediate and final stages each also comprise one or more multiplexers. The OR-reduction logic elements operate on different subsets of bits from the number input to the particular stage. In the intermediate and final hardware logic stages, a first of the multiplexers selects an OR-reduction result received from a previous hardware logic stage and the left shifting logic element is arranged to perform left shifting on the updated binary number received from an immediately previous hardware logic stage dependent upon the selected OR-reduction result.

An information processing device that executes calculation of a neural network, includes a memory; and a processor coupled to the memory and the processor configured to: set a division position for quantization of a variable to be used for the calculation so that a quantization error based on a difference between the variable before the quantization and the variable after the quantization is reduced; and quantize the variable based on the division position set.