36 results about "Mixed precision" patented technology

Embodiments detailed herein relate to arithmetic operations of float-point values. An exemplary processor includes decoding circuitry to decode an instruction, where the instruction specifies locations of a plurality of operands, values of which being in a floating-point format. The exemplary processor further includes execution circuitry to execute the decoded instruction, where the execution includes to: convert the values for each operand, each value being converted into a plurality of lower precision values, where an exponent is to be stored for each operand; perform arithmetic operations among lower precision values converted from values for the plurality of the operands; and generate a floating-point value by converting a resulting value from the arithmetic operations into the floating-point format and store the floating-point value.

Channel coupling for an AC-3 encoder, using mixed precision computations and 16-bit coupling coefficient calculations for channels with 32-bit frequency coefficients.

The resources needed—particularly in a programmable device—when carrying out a mixed-precision multiplication-based floating-point operation (i.e., multiplication or division) is reduced by maintaining the mantissas of the operands in their native precisions instead of promoting the lower-precision number to the higher precision. Exponents and other elements can be handled by the higher-precision logic as they do not consume significant resources.

The invention discloses a deep learning model training method, device and system based on mixed precision, and the method comprises the steps: carrying out the data processing of sample data in a sample data set based on a deep learning model obtained through the last training, and obtaining a set number of first weight gradient data with the data precision being the first data precision; determining the data precision as a scaling coefficient of the second data precision according to the set number of first weight gradient data and the second data precision, the first data precision being higher than the second data precision; based on the sample data set and the scaling coefficient, training the deep learning model to update the weight parameters of the deep learning model, obtaining thedeep learning model trained this time, and the scaling coefficient is used for amplifying the loss value with the data precision being the second data precision in the process of training the deep learning model, so that the training efficiency and the training precision are improved.

The invention discloses a deep neural network accelerator based on hybrid precision storage, and belongs to the technical field of calculation, reckoning and counting. The accelerator comprises an on-chip cache module, a control module, a bit-width-controllable multiply-add-batch calculation module, a nonlinear calculation module, a register array and a Huffman decoding module based on double lookup tables, effective bit and sign bit parameters of the weight are stored in the same memory, so that mixed-precision data storage and analysis are realized, and multiply-add operation of mixed-precision data and weight is realized. Through data storage analysis based on mixed precision and Huffman decoding based on a dual lookup table, compression and storage of data and weights under different precision are realized, data streams are reduced, and low-power-consumption data scheduling based on a deep neural network is realized.

The embodiment of the invention discloses a hybrid precision space-time multiplexing multiplier based on NAS search and a control method of the hybrid precision space-time multiplexing multiplier. Themethod comprises the steps of splitting a first multiplier and a first multiplicand to obtain one or more second multipliers and one or more second multiplicand, wherein the byte quantity of the second multipliers and the byte quantity of the second multiplicand are equal to the calculation byte quantity of each basic calculation unit; generating at least two groups of input parameters accordingto the second multiplier and the second multiplicand, wherein each group of input parameters comprises the second multiplier and the second multiplicand; inputting the at least two groups of input parameters into at least two basic calculation units in a one-to-one correspondence manner to obtain a first calculation result of each group of input parameters; and determining a target result of the first multiplier and the first multiplicand according to the first calculation result. According to the embodiment of the invention, the hybrid precision fixed-point multiply-accumulate calculation issupported through the hybrid serial-parallel design of time and space, so that the data stream is deeply optimized. The working speed of a traditional multiplier is improved, and the required area andenergy consumption are reduced.

A method and architecture with which to achieve efficient sub-word parallelism for multiplication resources is presented. In a preferred embodiment, a dual two's complement multiplier is presented, such that an n bit operand B can be split, and each portion of the operand B multiplied with another operand A in parallel. The intermediate products are combined in an adder with a compensation vector to correct any false negative sign on the two's complement sub-product from the multiplier handling the least significant, or lower, p bits of the split operand B, or B[p-1:0], where p=n/2. The compensation vector C is derived from the A and B operands using a simple circuit. The technique is easily extendible to 3 or more parallel multipliers, over which an n bit operand D can be split and multiplied with operand A in parallel. The compensation vector C' is similarly derived from the D and A operands in an analogous manner to the dual two's complement multiplier embodiment.

The invention provides a method for quantitatively improving model precision through low-bit mixing precision, and the method comprises the steps: carrying out the qualitative analysis of a model network structure, and guaranteeing that reasoning does not cross a boundary through the calculation and analysis of a model channel; and the network model precision is improved in a mixed precision configuration form. According to the method, by configuring the mixing precision mode, it is ensured that the precision of the model is the same as that of 8 bits and full precision in the low-bit process. And the mixed precision quantization is carried out, and the feature and the weight of the network are qualitatively analyzed, so that the low-bit quantization precision of the model is higher.

The invention relates to the technical field of artificial intelligence, and provides an object recognition method and device based on edge calculation, equipment and a medium. The method comprises the steps of obtaining a to-be-processed original model, performing pruning operation on the original model based on weights of all convolution kernels of each layer in the original model to obtain a pruned model, training the pruned model based on a first training sample set to obtain a pruned model, obtaining a second training sample set; and performing mixed precision training on the pruned model based on the second training sample set, and converting the trained model into a binary model file of a semi-precision floating-point number to obtain a target compression model, so that a client loads the target compression model to identify a to-be-identified object. According to the invention, the timeliness of model identification can be improved, and the risk of server downtime is reduced. The invention further relates to the technical field of block chains. The target compression model can be stored in a node of a block chain.

Technology related to training a neural network accelerator using mixed precision data formats is disclosed. In one example of the disclosed technology, a neural network accelerator is configured to accelerate a given layer of a multi-layer neural network. An input tensor for the given layer can be converted from a normal-precision floating-point format to a quantized-precision floating-point format. A tensor operation can be performed using the converted input tensor. A result of the tensor operation can be converted from the block floating-point format to the normal-precision floating-point format. The converted result can be used to generate an output tensor of the layer of the neural network, where the output tensor is in normal-precision floating-point format.

A neural network operation device includes an input feature map buffer to store an input feature map, a weight buffer to store a weight, an operator including an adder tree unit to perform an operation between the input feature map and the weight by a unit of a reference bit length, and a controller to map the input feature map and the weight to the operator to provide one or both of a mixed precision operation and data parallelism.

The use of mixed precision values when training an artificial neural network (ANN) can increase performance while reducing cost. Certain portions and/or steps of an ANN may be selected to use higher or lower precision values when training. Additionally, or alternatively, early phases of training are accurate enough with lower levels of precision to quickly refine an ANN model, while higher levels of precision may be used to increase accuracy for later steps and epochs. Similarly, different gates of a long short-term memory (LSTM) may be supplied with values having different precisions.

The invention discloses a dynamic mixing precision model construction method and system, and relates to a deep neural network technology. Aiming at the convertible state number and precision problems in the prior art, the scheme is provided, and the method is characterized by constructing a mixed precision state conversion table according to an average trace of a Hessian matrix of parameters of different blocks in a full-precision model and an optional parameter precision table S; randomly sampling a plurality of mixed precision sub-models by adopting each training iteration in the training process, and performing quantization operation by using an improved quantization function to obtain a mixed precision model; and forming a mixed bit deployment state table according to the actual deployment requirements to carry out actual deployment. The method has the advantages that the search space and the calculation amount required by training are reduced by utilizing the Hessian matrix average trace and the random sampling, and meanwhile, the improved quantization function can be directly migrated among different quantization bits, and the quantization error is small, so that the mixed precision model can be adaptively deployed in a lower bit scene; and meanwhile, the precision in a higher bit scene is also improved.

The invention discloses a fixed-point multiply-add operation unit and method suitable for a mixed precision neural network, and the method comprises the steps: inputting different input data precisions into a multiplier from different positions, controlling the multiplier to shield a partial product of a designated region according to a mode signal, and then outputting a partial product generation part, and carrying out summation operation on the output partial product generation part according to methods corresponding to different precisions, so as to achieve point multiplication operation of mixed precisions. According to the invention, the point multiplication operation of the mixed precision neural network can be realized by adopting one multiplier, and the problems of overlarge hardware overhead, redundant idle resources and the like caused by the need of adopting various processing units with different precisions to process the mixed precision operation in the prior art are solved.

The invention discloses a hybrid precision arithmetic unit applied to a reconfigurable array driven by a data flow, and relates to the field of arithmetic unit design. The invention provides the arithmetic unit which supports hybrid precision and multiple working modes. Compared with an arithmetic unit provided at a present stage, the invention designs the high-energy-efficiency fixed-point arithmetic unit supporting multiple specifications and hybrid precision for general computation intensive application, and working modes are selected according to requirements; and based on the low-power-consumption low-overhead hybrid precision arithmetic unit and a reasonable data flow scheduling mode designed by the invention, the problems of low resource utilization rate and precision loss of the hybrid unit in a neural network application-oriented low-precision operation mode of a coarse-grained reconfigurable array fixed-specification calculation unit are solved, and the neural network application-oriented performance of the reconfigurable array is greatly improved.

A circuit for calculating the fused sum of an addend and product of two multiplication operands, the addend and multiplication operands being binary floating-point numbers represented in a standardized format as a mantissa and an exponent is provided. The multiplication operands are in a lower precision format than the addend, with q>2p, where p and q are the mantissa size of the multiplication operand and addend precision formats. The circuit includes a p-bit multiplier receiving the mantissas of the multiplication operands; a shift circuit aligning the mantissa of the addend with the product output by the multiplier based on the exponent values of the addend and multiplication operands; and an adder processing q-bit mantissas, receiving the aligned mantissa of the addend and the product, the input lines of the adder corresponding to the product being completed to the right by lines at 0 to form a q-bit mantissa.

A mixed-precision quantization method for a neural network is provided. The neural network has a first precision and includes several layers and an original final output. For a particular layer, quantization of second precision on the particular layer and an input is performed. An output of the particular layer is obtained according to the particular layer of second precision and the input. De-quantization on the output of the particular layer is performed, and the de-quantized output is inputted to a next layer to obtain a final output. A value of an objective function is obtained according to the final output and the original final output. Above steps are repeated until the value of the objective function of each layer is obtained. A precision of quantization for each layer is decided according to the value of the objective function. The precision of quantization is one of first to fourth precision.

Computationally efficient mixed precision floating point waveform generation takes advantage of the high-speed generation of waveforms with single-precision floating point numbers while reducing the generally unacceptable loss of precision of pure single-precision floating point to generate any waveform that repeats in 2π. This approaches computes a reference phase in double precision as the modulus of the phase with 2π and then computes offsets to that value in single precision. The double precision reference phase is recomputed as needed depending on how quickly the phase grows and how large a machine epsilon is desired.

The invention discloses an error-controllable hybrid precision operator automatic optimization method, which comprises the following steps of: first optimization: under given input, optimizing a general matrix multiplication function through an ID (Identity) conversion method, and recording an operation sequence and a corresponding dimension of the general matrix multiplication function; second optimization: according to an optimized general matrix multiplication GEMM function ID list and a general matrix multiplication function sequence list, transmitting the input of each non-optimized general matrix multiplication calling function to a preset fast matrix multiplication function of different parameter combinations, and carrying out performance timing and error calculation; and third optimization: analyzing the obtained data, converting a single-precision matrix multiplication algorithm into a mixed-precision matrix multiplication operator, and outputting an optimized code. By using the method, the hybrid precision GEMM operator can help to improve the performance in a more complex and wider high-performance computing program. The method can be widely applied to the field of compilers.

A mixed-precision artificial intelligence (AI) processor and an operating method thereof are provided. The AI processor includes a first calculation module, a second calculation module and a control module. The first calculation module is configured to perform calculation based on the data with a first format. The second calculation module is configured to perform calculation based on the data with a second format different from the first format. The control module is coupled to the first calculation module and the second calculation module to select one of the first calculation module or the second calculation module to perform calculation based on an input data according to a calculation strategy.

The invention discloses an automatic studying and judging method for emotional tendency of Internet information, relates to the technical field of language emotion analysis. According to the method, a method for pre-training by using a RoBERTa model on general corpora and finely adjusting downstream tasks is adopted; in a deep learning training process, a mixed precision and multi-machine multi-GPU training mode is used; and after super parameter searching training is completed, a model is deployed and an interface is provided to complete the work of automatic studying and judging. The method solves the problems that in traditional public opinion emotion studying and judging work, the accuracy is not high, the studying and judging model generalization effect is not good enough, and the performance is not good enough when coping with complex Chinese contexts such as obscure and ambiguity.