Intel Corporation

Omid Azizi

Amin Firoozshahian

Mahesh Srinivasa Maddury

Mahesh Madhav

Chandan Egbert

Andreas Kleen

Ashok Raj

Stephen R Van Doren

Chen Koren

Alexandre Y Solomatnikov

Eric Gouldey

David Hansen

John Peter Stevenson

Rajesh M Sankaran

Eriko Nurvitadhi

Ishwar Agarwal

Vedaraman Geetha

Rupin H Vakharwala

Toby Opferman

Tony Lee Werner

Henk G Neefs

Nitin N Garegrat

Andrew Yang

Michael Rotzin

Vijay Bahirji

Guy Boudoukh

Mahesh Maddury

This disclosure relates matrix operation acceleration for different matrix sparsity patterns. A matrix operation accelerator may be designed to perform matrix operations more efficiently for a first matrix sparsity pattern rather than for a second matrix sparsity pattern. A matrix with the second sparsity pattern may be converted to a matrix with the first sparsity pattern and provided to the matrix operation accelerator. By rearranging the rows and/or columns of the matrix, the sparsity pattern of the matrix may be converted to a sparsity pattern that is suitable for computation with the matrix operation accelerator.

Machine learning architecture support for block sparsity

Various systems and methods for computer memory overcommitment management are described herein. A system for computer memory management includes a memory device to store data and a mapping table; and a memory overcommitment circuitry to: receive a signal to move data in a first block from a memory reduction area in the memory device to a non-memory reduction area in the memory device, the memory reduction area to store data using a memory reduction technique, and the non-memory reduction area to store data without any memory reduction techniques; allocate a second block in the non-memory reduction area; copy the data in the first block to the second block; and update the mapping table to revise a pointer to point to the second block, the mapping table used to store pointers to memory device in the memory reduction area and the non-memory reduction area.

Reservation architecture for overcommitted memory

Processing circuitry for computer memory management includes memory reduction circuitry to implement a memory reduction technique; and reference count information collection circuitry to: access a memory region, the memory region subject to the memory reduction technique; obtain an indication of memory reduction of the memory region; calculate metrics based on the indication of memory reduction of cache lines associated with the memory region; and provide the metrics to a system software component for use in memory management mechanisms.

Hardware-assisted paging mechanisms

A processing device includes a core to execute instructions, and memory management circuitry coupled to, memory, the core and an I/O device that supports page faults. The memory management circuitry includes an express invalidations circuitry, and a page translation permission circuitry. The memory management circuitry is to, while the core is executing the instructions, receive a command to pause communication between the I/O device and the memory. In response to receiving the command to pause the communication, modify permissions of page translations by the page translation permission circuitry and transmit an invalidation request, by the express invalidations circuitry to the I/O device, to cause cached page translations in the I/O device to be invalidated.

Pause communication from I/O devices supporting page faults

An apparatus and method for a masked multiply instruction to support neural network pruning operations. For example, one embodiment of a processor comprises: a decoder to decode a matrix multiplication with masking (GEMM) instruction identifying a destination matrix register to store a result, and source registers storing an A-matrix, a B-matrix, and a matrix mask; execution circuitry to execute the GEMM instruction, the execution circuitry to multiply a plurality of B-matrix elements with a plurality of A-matrix elements, each of the B-matrix elements associated with a mask value in the matrix mask, wherein if the mask value is set to a first value, then the execution circuitry is to multiply the B-matrix element with one or more of the A-matrix elements to generate a first partial result, and if the mask value is set to a second value, then the execution circuitry is to multiply an alternate B-matrix element with a one or more of the A-matrix elements to generate a second partial result.

Apparatus and method for a masked multiply instruction to support neural network pruning operations

Various systems and methods for computer memory management are described herein. A system for computer memory management includes a first memory device including a mapping table; a second memory device including a staging area; a third memory device including a dedup data region; and a controller operable to: receive a memory access request, the memory access request including an address and data; write the data to the staging area; and update the mapping table with the address.

Lazy memory deduplication

A system for computer memory management that implements a memory pool table, the memory pool table including entries that describe a plurality of memory pools, each memory pool representing a group of memory pages related by common attributes; a per-page tracking table, each entry in the per-page tracking table used to related a memory page with a memory pool of the memory pool table; and processing circuitry to: scan each entry in the per-page tracking table and, for each entry: determine an amount of memory released if the memory page related with the entry is swapped; aggregate the amount of memory for the respective memory pool related with the memory page related with the entry in the per-page tracking table, to produce a per-pool memory aggregate; and output the per-pool memory aggregate for the memory pools related with the memory pages in the per-page tracking table.

Physical page tracking for handling overcommitted memory

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

Various systems and methods for computer memory management are described herein. A system includes a memory controller to: monitor utilization of a memory device, the memory device used with a memory compression technique; determine that the utilization of the memory device violates a threshold; and initiate a system interrupt to provoke a response, responsive to the utilization of the memory device violating the threshold.

Memory pressure notifier

Growing community of inventors

Redwood City, CA, United States of America

Omid Azizi