57 results about "Multiplier accumulator" patented technology

An improved digital signal processing (DSP) processor architecture is presented in which word conditioning (e.g., rounding, saturation, etc.) and analysis operations (e.g., block floating point analysis) are implemented in the write datapath to memory. By moving word conditioning operations from the critical path to the write datapath, the throughput of common DSP functional blocks such as multiplier-accumulator (MAC) blocks may be improved. Delays may be further reduced by combining analysis operations with write or move operations.

The invention provides a convolution operation chip and communication equipment. The convolution operation chip comprises an M*N multiplier accumulator array, a data cache module and an output control module; the M*N multiplier accumulator array comprises a first multiplier accumulator window, and a processing unit PE<X, Y> of the first multiplier accumulator window is used for conducting multiplication operation on convolution data of the PE<X, Y> and convolution parameters of the PE<X, Y> and transmitting the convolution parameters of the PE<X, Y> to PE<X, Y+1> and transmitting the convolution data of the PE<X, Y> to PE<X-1, Y+1> to serve as multipliers of multiplication operation between the PE<X, Y+1> and the PE<X-1, Y+1>; the data cache module is used for transmitting the convolution data and the convolution parameters to the first multiplier accumulator window; the output control module is used for outputting a convolution result. According to the convolution operation chip and the communication equipment, the RAM access frequency can be decreased and the RAM access stress can be relieved while the array resource utilization rate is increased.

A programmable logic device is provided that includes a MAC block having mode splitting capabilities. Different modes of operation may be implemented simultaneously whereby the multipliers and other DSP circuitry of the MAC block may be allocated among the different modes of operation. For example, one multiplier may be used to implement a multiply mode while another two multipliers may be used to implement a sum of two multipliers mode.

The invention relates to a convolutional neural network acceleration method and apparatus. The method comprises the steps of reading an input feature graph of a convolutional layer; inputting the input feature graph to a processing array group of the convolutional layer, and according to a first reference pixel vector, a second reference pixel vector, a convolution kernel weight and finished inputchannel data, performing convolution multiplication and addition operation by utilizing a propagate partial multiplier-accumulator to obtain an output result of the processing array group; accordingto the output result of the processing array group, obtaining an output feature graph of the convolutional layer; writing the output feature graph of the final layer of the convolutional layer in an input cache of a full connection layer; executing the multiplication and addition operation by the full connection layer according to the output feature graph of the final layer of the convolutional layer to obtain an output eigenvector of the full connection layer; and outputting the output eigenvector of the final layer of the full connection layer to a fourth block of a first memory. The hardware resources and the power consumption are effectively reduced; and the processing speed of a convolutional neural network is increased.

An efficient implementation of DSP functions in a field programmable gate array (FPGA) using one or more computational blocks, each block having of a multiplier, an accumulator, and multiplexers. The structure implements most common DSP equations in a fast and a highly compact manner. A novel method for cascading these blocks with the help of dedicated DSP lines is provided, which leads to a very simple and proficient implementation of n-stage MAC operations.

A programmable logic integrated circuit device (“PLD”) includes programmable logic and a dedicated (i.e., at least partly hard-wired) processor object (or at least a high-functionality functional unit) for performing or at least helping to perform tasks that it is unduly inefficient to implement in the more general-purpose programmable logic and/or that, if implemented in the programmable logic, would operate unacceptably or at least undesirably slowly. The processor object includes an operating portion and a program sequencer that retrieves or at least helps to retrieve instructions for controlling or at least partly controlling the operating portion. The processor object may also include an address generator and/or a multi-ported register file for generating or at least helping to generate addresses of data on which the operating portion is to operate and/or destinations of data output by the operating portions. Examples of typical operating portions include multiplier-accumulators, arithmetic logic units, barrel shifters, and DSP circuitry of these or other kinds. The PLD may be provided with the capability to allow programs to be written for the device using local or “relative” addresses, and to automatically convert these addresses to actual or “absolute” addresses when the programs are actually performed by the device.

A virtual parallel multiplier-accumulator (VMAC) that can execute more than or less than one MAC operation in a single system clock cycle. The inventive VMAC advantageously employs a resource/time-sharing methodology with multiple sequential computational stages.

A convolution processing apparatus and method are disclosed. The convolution processing apparatus may include a controller configured to load a pixel of an input image and skip a process associated with the pixel in response to a value of the loaded pixel being 0, a filter bank including at least one filter and configured to extract at least one kernel element corresponding to the pixel from the at least one filter based on an index of the pixel and an input channel of the pixel, and a multiplier-accumulator (MAC) configured to perform a convolution operation based on the value of the pixel and a value of the at least one kernel element and accumulatively store an operation result of the convolution operation, the operation result corresponding to an output image.

A polyphase decimation FIR filter apparatus including a modulo integrator circuit configured to integrate input samples and to provide integrated input samples; and a polyphase FIR filter circuit configured to process the integrated input samples, the polyphase FIR filter circuit including a plurality of multiplier accumulator circuits, each configured to accumulate products of coefficients and respective integrated signal samples, wherein each of the multiplier accumulator circuits receives a subset of FIR filter coefficients, wherein the FIR filter coefficients are derived as the nth difference of original filter coefficients, where n is a number of integrators in the integrator circuit, and wherein the FIR filter circuit is configured to perform computation operations with modulo arithmetic.

This invention is a FFT butterfly circuit. This circuit includes four temporary data registers connected to three memories. The three memories include read/write X and Y memories and a read only twiddle coefficient memory. A multiplier-accumulator forms a product and accumulates the product with one of two accumulator registers. A register file with plural registers is loaded from one of the accumulator registers or the fourth temporary data register. An adder/subtracter forms a selected one of a sum of registers or a difference of registers. A write buffer with two buffers temporarily stores data from the adder/subtracter before storage in the first or second memory. The X and Y memories must be read/write but the twiddle memory may be read only.

The present invention relates to improved adaptive filtering techniques and architectures. Preferably, this filtering is performed as part of the digital processing that occurs with a digital signal processor. It is a feature of this invention that the adaptive filtering taught herein provides the advantages of both serial and parallel architectures, without the accompanying disadvantages thereof. In particular, an adaptive filter is taught that possesses low pin counts, fast processing times suitable for high-speed applications and reduced numbers of filter elements. In a preferred embodiment, the inputs and outputs of the adaptive filter are supplied to and from the adaptive filter in a serial manner while the processing is performed internally within the adaptive filter in a parallel manner. The parallel processing is preferably effected by a delayed least-means-squares algorithm implemented using a single adder, a single multiplier and a single multiplier-accumulator instead of by numerous such adders, multipliers and multiplier-accumulators.

The invention belongs to the technical field of low-power integrated circuits, and particularly relates to an FIR (finite impulse response) filter structure for a low-power fault-tolerant circuit. The FIR filter structure mainly includes transcoding logical units, multiplier-accumulators (MACs), and a residual bit transmission chain, wherein the transcoding logical units are used for conversion between the two's complement and sign-magnitude binary numbers; each MAC mainly includes a sign processing unit, a Wallace tree, a last-stage accumulator and a magnitude computing logical unit; a residual bit is produced by each MAC and transmitted individually, and all the residual bits are added up at the last stage, so as to shorten the critical path. The FIR filter structure for the low-power fault-tolerant circuit has the characteristics of simple structure, small area overhead, low critical path activation probability and the like, and is particularly suitable for a low-power fault-tolerant digital signal processing system based on the VOS (voltage over-scaling) technology.

A polyphase filter including M taps, each of the M taps including a filter coefficient. The filter also includes a multiplier-accumulator (MAC) shared by the M taps, a plurality of multiplexors for sequentially selecting a subset of the plurality of taps, and a scheduler for controlling the MAC to perform arithmetic operations on respective filter coefficients of the selected subset of the plurality of taps.

A matched filter able to reduce the size of a circuit is disclosed that includes a shift register, a spreading code storage unit, a multiplier, an accumulator, a spreading code generator, and an accumulation and storage unit. The shift register has a number of sixteen taps, which equals the square root of the spreading factor 256, and holds and bit-shifts received data. The spreading code generator generates spreading codes divided by the number of the taps at a calculation rate roughly sixteen times higher than a data rate of the received data of 3.84 MHz. The spreading code storage unit stores the divided spreading codes generated in the spreading code generator sequentially at the calculation rate. The multiplier 15 multiplies the received data stored in the shift register with the divided spreading codes stored in the spreading code storage unit at the calculation rate. The accumulator sums the products from the multiplier at the calculation rate. The accumulation and storage unit accumulates correlation values related to processing sequences of the same calculation phase, and stores the accumulation result.