Abstract: Systems, methods, and apparatuses relating to a matrix operations accelerator are described.

Abstract: A processing device is provided which comprises memory configured to store data and a processor configured to receive a portion of data of a first matrix comprising a first plurality of elements and receive a portion of data of a second matrix comprising a second plurality of elements. The processor is also configured to determine values for a third matrix by dropping a number of products from products of pairs of elements of the first and second matrices based on approximating the products of the pairs of elements as a sum of the exponents of the pairs of elements and performing matrix multiplication on remaining products of the pairs of elements of the first and second matrices.

Abstract: A digital signal processing (DSP) block includes a plurality of multipliers and a summation block separate from the plurality of multipliers. The DSP block is configurable to perform a first multiplication operation to determine a first product of a first floating-point value and a second floating-point value using only a first multiplier of the plurality of multipliers. Additionally, the DSP block is configurable to perform a second multiplication operation between a third floating-point value and a fourth floating-point value by receiving, at each of the plurality of multipliers, two integer values generated from the third floating-point value and the fourth floating-point value, generating, via the plurality of multipliers, a plurality of subproducts by multiplying, at each of the multipliers, the two integer values, and generating a second product of the second multiplication operation by adding, via the summation block, the plurality of subproducts.

Abstract: Disclosed embodiments relate to computing dot products of nibbles in tile operands. In one example, a processor includes decode circuitry to decode a tile dot product instruction having fields for an opcode, a destination identifier to identify a M by N destination matrix, a first source identifier to identify a M by K first source matrix, and a second source identifier to identify a K by N second source matrix, each of the matrices containing doubleword elements, and execution circuitry to execute the decoded instruction to perform a flow K times for each element (M,N) of the identified destination matrix to generate eight products by multiplying each nibble of a doubleword element (M,K) of the identified first source matrix by a corresponding nibble of a doubleword element (K,N) of the identified second source matrix, and to accumulate and saturate the eight products with previous contents of the doubleword element (M,N).

Abstract: A method for performing a convolution operation includes storing, a convolution kernel in a first storage device, the convolution kernel having dimensions x by y; storing, in a second storage device, a first subset of element values of an input feature map having dimensions n by m; performing a first simultaneous multiplication, of each value of the first subset of element values of the input feature map with a first element value from among the x*y elements of the convolution kernel; for each remaining value of the x*y elements of the convolution kernel, performing, a simultaneous multiplication of the remaining value with a corresponding subset of element values of the input feature map; for each simultaneous multiplication, storing, result of the simultaneous multiplication in an accumulator; and outputting, the values of the accumulator as a first row of an output feature map.

Abstract: Folded integer multiplier (FIM) circuitry includes a multiplier configurable to perform multiplication and a first addition/subtraction unit and a second addition/subtraction unit both configurable to perform addition and subtraction. The FIM circuitry is configurable to determine each product of a plurality of products for a plurality of pairs of input values having a first number of bits by performing, using the first and second addition/subtraction units, a plurality of operations involving addition or subtraction, and performing, using the multiplier, a plurality of multiplication operations involving values having fewer bits than the first number of bits. The plurality of multiplication operations includes a first number of multiplication operations, and the multiplier is configurable to begin performing all multiplication operations of the plurality of multiplication operations within a first number of clock cycles equal to the first number of multiplication operations.

Abstract: A unit for accumulating a plurality N of multiplied M bit values includes a receiving unit, a bit-wise multiplier and a bit-wise accumulator. The receiving unit receives a pipeline of multiplicands A and B such that, at each cycle, a new set of multiplicands is received. The bit-wise multiplier bit-wise multiplies bits of a current multiplicand A with bits of a current multiplicand B and to sum and carry between bit-wise multipliers. The bit-wise accumulator accumulates output of the bit-wise multiplier thereby to accumulate the multiplicands during the pipelining process.

Abstract: A method related to posit tensor processing can include receiving, by a plurality of multiply-accumulator (MAC) units coupled to one another, a plurality of universal number (unum) or posit bit strings organized in a matrix and to be used as operands in a plurality of respective recursive operations performed using the plurality of MAC units and performing, using the MAC units, the plurality of respective recursive operations. Iterations of the respective recursive operations are performed using at least one bit string that is a same bit string as was used in a preceding iteration of the respective recursive operations. The method can further include prior to receiving the plurality of unum or posit bit strings, performing an operation to organize the plurality of unum or posit bit strings to achieve a threshold bandwidth ratio, a threshold latency, or both during performance of the plurality of respective recursive operations.

Abstract: An apparatus and method for performing dual concurrent multiplications of packed data elements.

Abstract: An information processing apparatus comprises a control unit configured to set a shift amount based on a bit width of data, for each layer of a network including a plurality of layers, a plurality of MAC (multiply-accumulate) units configured to execute MAC operations on a plurality of data and a plurality of filter coefficients of the layer, a plurality of shift operation units configured to shift a plurality of MAC operation results obtained by the plurality of MAC units based on the shift amount, and an adding unit configured to calculate a total sum of the plurality of MAC operation results shifted by the plurality of shift operation units.

Abstract: A MAC operating device comprising a plurality of operation circuits respectively including an operation capacitor and a plurality of switches; and a division capacitor, wherein one end of the operation capacitor is respectively connected to a first operation switch connected to an input terminal and a first reset switch connected to a ground terminal, and the other end of the operation capacitor is connected to both a second operation switch connected to a division capacitor and a second reset switch connected to the ground terminal is provided.

Abstract: An integrated circuit with specialized processing blocks is provided. A specialized processing block may be optimized for machine learning algorithms and may include a multiplier data path that feeds an adder data path. The multiplier data path may be decomposed into multiple partial product generators, multiple compressors, and multiple carry-propagate adders of a first precision. Results from the carry-propagate adders may be added using a floating-point adder of the first precision. Results from the floating-point adder may be optionally cast to a second precision that is higher or more accurate than the first precision. The adder data path may include an adder of the second precision that combines the results from the floating-point adder with zero, with a general-purpose input, or with other dot product terms. Operated in this way, the specialized processing block provides a technical improvement of greatly increasing the functional density for implementing machine learning algorithms.

Abstract: A processing device is provided which includes memory configured to store data and a processor configured to determine, based on convolutional parameters associated with an image, a virtual general matrix-matrix multiplication (GEMM) space of a virtual GEMM space output matrix and generate, in the virtual GEMM space output matrix, a convolution result by matrix multiplying the data corresponding to a virtual GEMM space input matrix with the data corresponding to a virtual GEMM space filter matrix. The processing device also includes convolutional mapping hardware configured to map, based on the convolutional parameters, positions of the virtual GEMM space input matrix to positions of an image space of the image.

Abstract: Embodiments of systems, apparatuses, and methods for fused multiple add. In some embodiments, a decoder decodes a single instruction having an opcode, a destination field representing a destination operand, and fields for a first, second, and third packed data source operand, wherein packed data elements of the first and second packed data source operand are of a first, different size than a second size of packed data elements of the third packed data operand.

Abstract: Embodiments of systems, apparatuses, and methods for fused multiple add. In some embodiments, a decoder decodes a single instruction having an opcode, a destination field representing a destination operand, and fields for a first, second, and third packed data source operand, wherein packed data elements of the first and second packed data source operand are of a first, different size than a second size of packed data elements of the third packed data operand.

Abstract: Embodiments of systems, apparatuses, and methods for fused multiple add. In some embodiments, a decoder decodes a single instruction having an opcode, a destination field representing a destination operand, and fields for a first, second, and third packed data source operand, wherein packed data elements of the first and second packed data source operand are of a first, different size than a second size of packed data elements of the third packed data operand.

Abstract: Embodiments of systems, apparatuses, and methods for fused multiple add. In some embodiments, a decoder decodes a single instruction having an opcode, a destination field representing a destination operand, and fields for a first, second, and third packed data source operand, wherein packed data elements of the first and second packed data source operand are of a first, different size than a second size of packed data elements of the third packed data operand.

Abstract: An architecture is disclosed for an neural processing element having single instruction, multiple data (“SIMD”) compute lanes. The neural processing element includes compute lanes having multipliers configured to multiply a binary operand with another binary operand to generate a binary output. The neural processing element also includes a single adder tree for summing the binary outputs of the hardware binary multipliers. The neural processing element also includes a storage element for storing a binary output of the single hardware binary adder tree.

Abstract: An adaptive matrix multiplier. In some embodiments, the matrix multiplier includes a first multiplying unit a second multiplying unit, a memory load circuit, and an outer buffer circuit. The first multiplying unit includes a first inner buffer circuit and a second inner buffer circuit, and the second multiplying unit includes a first inner buffer circuit and a second inner buffer circuit. The memory load circuit is configured to load data from memory, in a single burst of a burst memory access mode, into the first inner buffer circuit of the first multiplying unit; and into the first inner buffer circuit of the second multiplying unit.

Abstract: An information processing apparatus comprises a control unit configured to set a shift amount based on a bit width of data, for each layer of a network including a plurality of layers, a plurality of MAC (multiply-accumulate) units configured to execute MAC operations on a plurality of data and a plurality of filter coefficients of the layer, a plurality of shift operation units configured to shift a plurality of MAC operation results obtained by the plurality of MAC units based on the shift amount, and an adding unit configured to calculate a total sum of the plurality of MAC operation results shifted by the plurality of shift operation units.

Abstract: A method and apparatus for performing a multi-precision computation in a plurality of arithmetic logic units (ALUs) includes pairing a first Single Instruction/Multiple Data (SIMD) block channel device with a second SIMD block channel device to create a first block pair having one-level staggering between the first and second channel devices. A third SIMD block channel device is paired with a fourth SIMD block channel device to create a second block pair having one-level staggering between the third and fourth channel devices. A plurality of source inputs are received at the first block pair and the second block pair. The first block pair computes a first result, and the second block pair computes a second result.

Abstract: An apparatus and method for multiplying packed unsigned words.

Abstract: Embodiments described herein provide a graphics processor that can perform a variety of mixed and multiple precision instructions and operations. One embodiment provides a streaming multiprocessor that can concurrently execute multiple thread groups, wherein the streaming multiprocessor includes a single instruction, multiple thread (SIMT) architecture and the streaming multiprocessor is to execute multiple threads for each of multiple instructions. The streaming multiprocessor can perform concurrent integer and floating-point operations and includes a mixed precision core to perform operations at multiple precisions.

Abstract: A data processing apparatus is provided to convert a plurality of signed digits to an output value, the data processing apparatus comprising: receiver circuitry to receive, at each of a plurality of iterations, a signed digit from the plurality of signed digits, and previous intermediate data. Conversion circuitry performs a negative-output conversion from the signed digit to an unsigned digit, such that the output value comprising the unsigned digit is negative. Concatenation circuitry concatenate bits of the unsigned digit and bits of the previous intermediate data to produce updated intermediate data and output circuitry provides the updated intermediate data as the previous intermediate data of a next iteration. After the plurality of iterations, the output circuitry outputs at least part of the updated intermediate data as the output value.

Abstract: An apparatus to facilitate compute optimization is disclosed. The apparatus includes a mixed precision core to perform a mixed precision multi-dimensional matrix multiply and accumulate operation on 8-bit and/or 32 bit signed or unsigned integer elements.

Abstract: A tile of an FPGA includes a multiple mode arithmetic circuit. The multiple mode arithmetic circuit is configured by control signals to operate in an integer mode, a floating-point mode, or both. In some example embodiments, multiple integer modes (e.g., unsigned, two's complement, and sign-magnitude) are selectable, multiple floating-point modes (e.g., 16-bit mantissa and 8-bit sign, 8-bit mantissa and 6-bit sign, and 6-bit mantissa and 6-bit sign) are supported, or any suitable combination thereof. The tile may also fuse a memory circuit with the arithmetic circuits. Connections directly between multiple instances of the tile are also available, allowing multiple tiles to be treated as larger memories or arithmetic circuits. By using these connections, referred to as cascade inputs and outputs, the input and output bandwidth of the arithmetic circuit is further increased.

Abstract: A method related to posit tensor processing can include receiving, by a plurality of multiply-accumulator (MAC) units coupled to one another, a plurality of universal number (unum) or posit bit strings organized in a matrix and to be used as operands in a plurality of respective recursive operations performed using the plurality of MAC units and performing, using the MAC units, the plurality of respective recursive operations. Iterations of the respective recursive operations are performed using at least one bit string that is a same bit string as was used in a preceding iteration of the respective recursive operations. The method can further include prior to receiving the plurality of unum or posit bit strings, performing an operation to organize the plurality of unum or posit bit strings to achieve a threshold bandwidth ratio, a threshold latency, or both during performance of the plurality of respective recursive operations.

Abstract: The present invention discloses an instruction processing apparatus, comprising a first register adapted to store first source data, a second register adapted to store second source data, a third register adapted to store accumulated data, a decoder adapted to receive and decode a multiply-accumulate instruction, and an execution unit. The multiply-accumulate instruction indicates that the first register serves as a first operand, the second register serves as a second operand, the third register serves as a third operand, and a shift flag.

Abstract: A texture filtering unit includes a datapath block and a control block. The datapath block includes one or more parallel computation pipelines, each containing at least one hardware logic component configured to receive a plurality of inputs and generate an output value as part of a texture filtering operation. The control block includes a plurality of sequencers and an arbiter. Each sequencer executes a micro-program that defines a sequence of operations to be performed by the one or more pipelines in the datapath block as part of a texture filtering operation and the arbiter controls access, by the sequencers, to the one or more pipelines in the datapath based on predefined prioritization rules.

Abstract: This application concerns methods, apparatus, and systems for performing quantum circuit synthesis and/or for implementing the synthesis results in a quantum computer system. In certain example embodiments: a universal gate set, a target unitary described by a target angle, and target precision is received (input); a corresponding quaternion approximation of the target unitary is determined; and a quantum circuit corresponding to the quaternion approximation is synthesized, the quantum circuit being over a single qubit gate set, the single qubit gate set being realizable by the given universal gate set for the target quantum computer architecture.

Abstract: A vector reduction circuit configured to reduce an input vector of elements comprises a plurality of cells, wherein each of the plurality of cells other than a designated first cell that receives a designated first element of the input vector is configured to receive a particular element of the input vector, receive, from another of the one or more cells, a temporary reduction element, perform a reduction operation using the particular element and the temporary reduction element, and provide, as a new temporary reduction element, a result of performing the reduction operation using the particular element and the temporary reduction element. The vector reduction circuit also comprises an output circuit configured to provide, for output as a reduction of the input vector, a new temporary reduction element corresponding to a result of performing the reduction operation using a last element of the input vector.

Abstract: An accelerating device includes a signal detector that converts a first input signal and a second input signal into a first converted input signal and a second converted input signal, respectively, and that generates a final zero-value flag signal, a first one-value flag signal, and a second one-value flag signal. The accelerating device further includes a processing element (PE) that processes the first converted input signal and the second converted input signal based on the final zero-value flag signal, the first one-value flag signal, and the second one-value flag signal and that skips a first arithmetic operation and a second arithmetic operation when the final zero-value flag signal has a first value. The first value of the final zero-value flag signal indicates that the first input signal, or the second input signal, or both have a value of 0.

Abstract: Embodiments of the present disclosure pertain to digital circuits with compressed carries. In one embodiment, an adder circuit generates a sum and carry. The carry is compressed to reduce the number of bits required to represent the carry. In one embodiment, a multiplier circuit generates output product values. The output product values may be summed to produce a sum and carry. The carry may be compressed. In other embodiments, a multiplier circuit receives an input sum and compressed carry. The compressed input carry is decompressed and added to output product values and the input sum, and a resulting carry is compressed. The output of such a multiplier is another sum and compressed carry.

Abstract: When executing a program on a data processor comprising an execution unit for executing instructions in a program to be executed by the data processor, the execution unit being associated with one or more hardware units operable to execute instructions, at least one instruction in a program is associated with an indication of whether the instruction should be issued directly for execution by a hardware unit or should be intercepted during its execution by the execution unit. The execution unit then, when decoding the instruction for execution by a hardware unit in the program, determines from the indication associated with the instruction whether the instruction should be issued directly for execution by a hardware unit or intercepted during its execution by the execution unit, and issues the instruction for execution by a hardware unit directly, or pauses execution of the instruction and performs another operation, accordingly.

Abstract: The present embodiments relate to circuitry that efficiently performs floating-point arithmetic operations and fixed-point arithmetic operations. Such circuitry may be implemented in specialized processing blocks. If desired, the specialized processing blocks may include configurable interconnect circuitry to support a variety of different use modes. For example, the specialized processing block may efficiently perform a fixed-point or floating-point addition operation or a portion thereof, a fixed-point or floating-point multiplication operation or a portion thereof, a fixed-point or floating-point multiply-add operation or a portion thereof, just to name a few. In some embodiments, two or more specialized processing blocks may be arranged in a cascade chain and perform together more complex operations such as a recursive mode dot product of two vectors of floating-point numbers or a Radix-2 Butterfly circuit, just to name a few.

Abstract: A reconfigurable matrix multiplier (RMM) system/method allowing tight or loose coupling to supervisory control processor application control logic (ACL) in a system-on-a-chip (SOC) environment is disclosed. The RMM provides for C=A*B matrix multiplication operations having A-multiplier-matrix (AMM), B-multiplicand-matrix (BMM), and C-product-matrix (CPM), as well as C=A*B+D operations in which D-summation-matrix (DSM) represents the result of a previous multiplication operation or another previously defined matrix. The RMM provides for additional CPM LOAD/STORE paths allowing overlapping of compute/data transfer operations and provides for CPM data feedback to the AMM or BMM operand inputs from a previously calculated CPM result.

Abstract: An apparatus and method for multiplying packed real and imaginary components of complex numbers.

Abstract: A microprocessor with dynamically adjustable bit width is provided, which has a bit width register, a datapath, a statistical register, and a bit width adjuster. The bit width register stores at least one bit width. The datapath operates according to the bit width stored in the bit width register to acquire input operands from received data and process input operands. The statistical register collects calculation results of the datapath. The bit width adjuster adjusts the bit width stored in the bit width register based on the calculation results collected in the statistical register.

Abstract: A data queuing and format apparatus is disclosed. A first selection circuit may be configured to selectively couple a first subset of data to a first plurality of data lines dependent upon control information, and a second selection circuit may be configured to selectively couple a second subset of data to a second plurality of data lines dependent upon the control information. A storage array may include multiple storage units, and each storage unit may be configured to receive data from one or more data lines of either the first or second plurality of data lines dependent upon the control information.

Abstract: An apparatus and method for performing signed multiplication of packed signed/unsigned doublewords and accumulation with a quadword.

Abstract: Techniques and circuits are provided for stochastic rounding. In an embodiment, a circuit includes carry-save adder (CSA) logic having three or more CSA inputs, a CSA sum output, and a CSA carry output. One of the three or more CSA inputs is presented with a random number value, while other CSA inputs are presented with input values to be summed. The circuit further includes adder logic having adder inputs and a sum output. The CSA carry output of the CSA logic is coupled with one of the adder inputs of the adder logic, and the CSA sum output of the CSA logic is coupled with another input of the adder inputs of the adder logic. A particular number of most significant bits of the sum output of the adder logic represent a stochastically rounded sum of the input values.

Abstract: Systems and methods are provided for performing a dot product. Each of a first series of numbers is divided into a first value, comprising the N most significant bits of the number, and a second value to form first and second sets of values. Each of a second series of numbers is divided into a third value, comprising the N most significant bits of the number, and a fourth value to form third and fourth sets of values. A dot product of the first and fourth sets of values is computed to provide a first partial sum. A dot product of the first and third sets of values is computed to provide a second partial sum. A dot product of the second and third sets of values is computed to provide a third partial sum. The partial sums are summed to provide a result for the dot product.

Abstract: Disclosed embodiments relate to instructions for fused multiply-add (FMA) operations with variable-precision inputs. In one example, a processor to execute an asymmetric FMA instruction includes fetch circuitry to fetch an FMA instruction having fields to specify an opcode, a destination, and first and second source vectors having first and second widths, respectively, decode circuitry to decode the fetched FMA instruction, and a single instruction multiple data (SIMD) execution circuit to process as many elements of the second source vector as fit into an SIMD lane width by multiplying each element by a corresponding element of the first source vector, and accumulating a resulting product with previous contents of the destination, wherein the SIMD lane width is one of 16 bits, 32 bits, and 64 bits, the first width is one of 4 bits and 8 bits, and the second width is one of 1 bit, 2 bits, and 4 bits.

Abstract: Disclosed herein is a computer implemented method for performing multiply-add operations of binary numbers P, Q, R, S, B in an arithmetic unit of a processor, the operation calculating a result as an accumulated sum, which equals to B+n×P×Q+m×R×S, where n and m are natural numbers. Further disclosed herein is an arithmetic unit configured to implement multiply-add operations of binary numbers P, Q, R, S, B comprising at least a first binary arithmetic unit for calculating an aligned high part result and a second binary arithmetic unit for calculating an aligned low part result of the multiply-add operations.

Abstract: A circuit that includes a plurality of array cores, each array core of the plurality of array cores comprising: a plurality of distinct data processing circuits; and a data queue register file; a plurality of border cores, each border core of the plurality of border cores comprising: at least a register file, wherein: [i] at least a subset of the plurality of border cores encompasses a periphery of a first subset of the plurality of array cores; and [ii] a combination of the plurality of array cores and the plurality of border cores define an integrated circuit array.

Abstract: Tensor register files in a hardware accelerator are disclosed. An apparatus may comprise tensor operation calculators each configured to perform a type of tensor operation. The apparatus may also comprises tensor register files, each of which is associated with one of the tensor operation calculators. The apparatus may also comprises logic configured to store respective ones of the tensors in the plurality of tensor register files in accordance with the type of tensor operation to be performed on the respective tensors. The apparatus may also control read access to tensor register files based on a type of tensor operation that a machine instruction is to perform.

Abstract: A processor and a method for executing an instruction on a processor are provided. In the method, a to-be-executed instruction is fetched, the instruction including a source address field, a destination address field, an operation type field, and an operation parameter field; in at least one execution unit, an execution unit controlled by a to-be-generated control signal according to the operation type field is determined, a source address and a destination address of data operated by the execution unit are determined according to the source address field and the destination address field, and a data amount of the data operated by the execution unit controlled by the to-be-generated control signal is determined according to the operation parameter field; the control signal is generated; and the execution unit in the at least one execution unit is controlled by using the control signal.

Abstract: Apparatus and methods are disclosed for nullifying one or more registers identified in a target field of a nullification instruction. In some examples of the disclosed technology, an apparatus can include memory and one or more block-based processor cores configured to fetch and execute a plurality of instruction blocks. One of the cores can include a control unit configured, based at least in part on receiving a nullification instruction, to obtain a register identification of at least one of a plurality of registers, based on a target field of the nullification instruction. A write to the at least one register associated with the register identification is nullified. The nullification instruction is in a first instruction block of the plurality of instruction blocks. Based on the nullified write to the at least one register, a subsequent instruction is executed from a second, different instruction block.

Abstract: In one example in accordance with the present disclosure a resistive memory array is described. The array includes a number of resistive memory elements to receive a common-valued read signal. The array also includes a number of multiplication engines to perform a multiply operation by receiving a memory element output from a corresponding resistive memory element, receiving an input signal, and generating a multiplication output based on a received memory element output and a received input signal. The array also includes an accumulation engine to sum multiplication outputs from the number of multiplication engines.

Abstract: A correlation operation circuit includes a first SRAM storing a plurality of pieces of detection pattern data, product-sum operators, a second SRAM storing intermediate data, and a comparator. When time series data is sequentially input, the intermediate data of all correlation functions referring to one time series data in a period during which the one time series data is input. When one time series data is input, the product-sum operator multiplies the detection pattern data sequentially read from the first SRAM by the one input time series data. The corresponding intermediate data is read from the second SRAM in synchronization with the multiplication, and the sequentially-calculated products are cumulatively added to the read intermediate data to be written back into the second SRAM as the intermediate data. As a result, the calculated correlation function data is supplied to the comparator to be compared with a predetermined specified value.