Intel Corporation

Cisco Technology, Inc.

Tony Lee Werner

Aravind Kalaiah

Amir Khosrowshahi

Andrew Yang

Michael Rotzin

Horace Lau

Nitin N Garegrat

Carey Kevin Kloss

Vijay Anand R Korthikanti

Naveen Gandham Rao

Anne Q Ye

Ujwal Basavaraj Sajjanar

Jeff DelChiaro

Vijay Korthikanti

Robert T Rhoades

Earl T Cohen

Eriko Nurvitadhi

Omid Azizi

Chen Koren

Guy Boudoukh

Krishnakumar Nair

Tom Schebye

Prashant Arora

Olivia K Wu

Horace H Lau

In one embodiment, a matrix processor comprises a memory to store a matrix operand and a strided read sequence, wherein: the matrix operand is stored out of order in the memory; and the strided read sequence comprises a sequence of read operations to read the matrix operand in a correct order from the memory. The matrix processor further comprises circuitry to: receive a first instruction to be executed by the matrix processor, wherein the first instruction is to instruct the matrix processor to perform a first operation on the matrix operand; read the matrix operand from the memory based on the strided read sequence; and execute the first instruction by performing the first operation on the matrix operand.

Multi-variate strided read operations for accessing matrix operands

In one embodiment, a matrix operation may be performed using a plurality of input matrices, wherein the matrix operation is associated with one or more convolution operations. The plurality of input matrices may be partitioned into a plurality of input partitions, wherein the plurality of input matrices is partitioned based on a number of available processing elements. The plurality of input partitions may be distributed among a plurality of processing elements, wherein each input partition is distributed to a particular processing element of the plurality of processing elements. A plurality of partial matrix operations may be performed using the plurality of processing elements, and partial matrix data may be transmitted between the plurality of processing elements while performing the plurality of partial matrix operations. A result of the matrix operation may be determined based on the plurality of partial matrix operations.

Distributed convolution for neural networks

In one embodiment, a matrix operation may be performed to reorder a plurality of dimensions of an input matrix stored in two-dimensional memory. Data associated with the input matrix may be accessed using one or more strided memory operations, wherein the one or more strided memory operations are configured to access the two-dimensional memory at a plurality of locations that are separated by a particular interval. The data accessed using the one or more strided memory operations may be stored in a result matrix, wherein the data accessed using each strided memory operation is stored in the result matrix in non-transpose form or transpose form.

Dimension shuffling using matrix processors

In one embodiment, a matrix operation may be performed on one or more matrix operands. For example, matrix data may be received from a multi-dimensional memory, wherein the matrix data is associated with the one or more matrix operands. The one or more matrix operands may be extracted from the matrix data. A matrix routine associated with the matrix operation may be identified. The matrix routine may be executed on a matrix processor using the one or more matrix operands. A result of the matrix operation may be obtained based on the matrix routine executed by the matrix processor.

Programmable matrix processing engine

An apparatus and method for a converting tensor data. For example, one embodiment of a method comprises: fetching source tensor blocks of a source tensor data structure, each source tensor block comprising a plurality of source tensor data elements having a first numeric representation, wherein the source tensor data structure comprises a predefined structural arrangement of source tensor blocks; converting the one or more source tensor blocks into one or more destination tensor blocks comprising a plurality of destination tensor data elements having a second numeric representation different from the first numeric representation, wherein the sets of one or more source tensor blocks are converted to one or more corresponding destination tensor blocks in a specified order based on the first and second numeric representations; and storing each individual destination tensor block in a designated memory region to maintain coherency with the predefined structural arrangement of the source tensor blocks.

Apparatus and method for coherent, accelerated conversion between data representations

Disclosed embodiments relate to sparse matrix multiplication (SMM) acceleration using column folding and squeezing. In one example, a processor, in response to a SMM instruction having fields to specify locations of first, second, and output matrices, the second matrix being a sparse matrix, uses execution circuitry to pack the second matrix by replacing one or more zero-valued elements with non-zero elements yet to be processed, each of the replaced elements further including a field to identify its logical position within the second matrix, and, the execution circuitry further to, for each non-zero element at row M and column K of the specified first matrix, generate a product of the element and each corresponding non-zero element at row K, column N of the packed second matrix, and accumulate each generated product with a previous value of a corresponding element at row M and column N of the specified output matrix.

Matrix multiplication acceleration of sparse matrices using column folding and squeezing

In one embodiment, a matrix operation may be performed, wherein the matrix operation comprises a matrix multiplication operation on a plurality of matrix operands. Matrix data may be received from a multi-dimensional memory, wherein the matrix data is associated with the plurality of matrix operands. The plurality of matrix operands may be extracted from the matrix data, wherein the plurality of matrix operands comprises a first matrix operand and a second matrix operand. A first transform may be performed on the first matrix operand to obtain a transformed matrix operand, wherein performing matrix multiplication using the transformed matrix operand is faster than performing matrix multiplication using the first matrix operand. Matrix multiplication may be performed on the transformed matrix operand to obtain a partial result. A second transform may be performed on the partial result to obtain a result of the matrix multiplication operation.

Winograd algorithm on a matrix processing architecture

An apparatus may comprise a multi-dimensional memory, a plurality of matrix processors, and a matrix routine memory. The matrix routine memory may store a plurality of programmable matrix routines, wherein each programmable matrix routine comprises a plurality of instructions associated with a particular matrix operation, wherein the plurality of instructions is to be executed by the plurality of matrix processors. Further, the plurality of matrix processors may be configured to: receive a command to perform a matrix operation; receive matrix data from the multi-dimensional memory; extract one or more matrix operands from the matrix data; identify a programmable matrix routine associated with the matrix operation; receive the programmable matrix routine from the matrix routine memory; execute the programmable matrix routine using the one or more matrix operands; and obtain a result of the matrix operation based on execution of the programmable matrix routine.

In one embodiment, an apparatus comprises a multi-dimensional memory and a plurality of processing elements to perform a matrix operation, wherein the matrix operation comprises a max pooling operation on one or more matrix operands. The plurality of processing elements comprises one or more matrix processors, and the plurality of processing elements is configured to: receive matrix data from the multi-dimensional memory, wherein the matrix data is associated with the one or more matrix operands; extract the one or more matrix operands from the matrix data; perform the max pooling operation using the one or more matrix operands; and obtain a result of the max pooling operation.

Max pooling in a matrix processing architecture

Described herein are methods, systems, and apparatuses to utilize a matrix operation by accessing each of the operation's matrix operands via a respective single memory handle. This use of a single memory handle for each matrix operand eliminates significant overhead in memory allocation, data tracking, and subroutine complexity present in prior art solutions. The result of the matrix operation can also be accessible via a single memory handle identifying the matrix elements of the result.

Matrix operands for linear algebra operations

Described herein are one or more integrated circuits (ICs) comprising controller circuitry to receive a command to execute an operation for data inputs stored in an external memory or a local memory, and convert the operation into a set of matrix operations to operate on sub-portions of the data inputs. The IC(s) further comprise at least one processing circuitry to execute the set of matrix operations, the processing circuitry to include ALUs, a local memory external to the ALUs and accessible by the ALUs, and processing control circuitry to create at least one matrix operand in the local memory (from the data inputs of the operation) comprising at least one of a scalar, a vector, or a 2D matrix, and provide memory handles corresponding to each of the matrix operands to one of the ALUs to access the respective matrix operands when executing a matrix operation.

Pipelined convolutional operations for processing clusters

Methods and apparatus are disclosed for identifying a variable number of items first in sequence from a variable starting position which may be particularly useful by packet or other scheduling mechanisms, such as, but not limited to the SLIP/I SLIP scheduling algorithms or variants thereof. Each of the groups of items is typically identified with a number of items the group desires to be selected. Based on an identified starting position, a progressive sum value is initialized, with progressive sum values corresponding to successive groups of items in the sequence being adjusted typically based on the corresponding number of items each successive group desires to be selected. The number of items a particular group is authorized to select can then be determined, such as, but not limited to, by being based on its corresponding progressive sum value, the progressive sum value of the immediately prior group in the sequence, and its desired number of items to be selected.

Methods and apparatus for identifying a variable number of items first in sequence from a variable starting position which may be particularly useful by packet or other scheduling mechanisms

Growing community of inventors

Los Altos, CA, United States of America

Tony Lee Werner