Intel Corporation

Christopher J Hughes

Robert Valentine

Mark Jay Charney

Elmoustapha Ould-Ahmed-Vall

Alexander F Heinecke

Bret L Toll

Raanan Sade

Milind Baburao Girkar

Amit Gradstein

Zeev Sperber

Jesus Corbal

Simon Rubanovich

Sanjeev Kumar

Yen-Cheng Liu

Kshitij Arun Doshi

Prasoonkumar Surti

Jason W Brandt

Vinodh Gopal

Wajdi K Feghali

James D Guilford

Rajesh M Sankaran

Guei-Yuan Lueh

Edward Thomas Grochowski

Shih-Lien L Lu

Gilbert A Neiger

Subramaniam M Maiyuran

Martin Guy Dixon

Abhishek R Appu

Altug Koker

Jill MacDonald Boyce

Rajesh Poornachandran

Yen-Kuang Chen

Daehyun Kim

Dan Baum

Jong Soo Park

Mikhail Plotnikov

Jonathan Cannon Hall

Andrey Naraikin

Andrew Thomas Forsyth

Michael J Espig

Brian J Hickmann

Victor W Lee

Partha Kundu

Menachem Adelman

Ronen Zohar

David A Koufaty

Naveen K Mellempudi

Nadathur Rajagopalan Satish

Stephen R Van Doren

Wei Wu

Jonathan D Pearce

Sara S Baghsorkhi

Rajat Agarwal

Srinivas Chennupaty

Robert S Chappell

Jorge E Parra

Mikhail Smelyanskiy

Zeshan A Chishti

Rajesh S Parthasarathy

Zhe Wang

Chris J Newburn

Mark J Buxton

Vasanth Ranganathan

Tsung-Yuan Charles Tai

Ren Wang

Nikos Kaburlasos

Christian Maciocco

Stephen H Gunther

Andrew J Herdrich

Adam T Lake

Lidong Xu

Alexander W Min

Jr-Shian Tsai

James M Holland

Joseph Nuzman

David B Papworth

Herbert H Hum

Deborah T Marr

Lihu Rappoport

Jatin Chhugani

Ronak Singhal

Ehud Cohen

Sheng C Li

Niall D McDonnell

Mesut A Ergin

Cristina S Anderson

Aamer Jaleel

Varghese George

Hong Wang

Cheng C Wang

Youfeng Wu

Bryan K Casper

Sreenivas Subramoney

Wenlong Li

Yu Chen

John Thomas Shen

Alaa R Alameldeen

Ralph Michael Kling

Anant V Nori

Jamison D Collins

Disclosed embodiments relate to systems for performing instructions to quickly convert and use matrices (tiles) as one-dimensional vectors. In one example, a processor includes fetch circuitry to fetch an instruction having fields to specify an opcode, locations of a two-dimensional (2D) matrix and a one-dimensional (1D) vector, and a group of elements comprising one of a row, part of a row, multiple rows, a column, part of a column, multiple columns, and a rectangular sub-tile of the specified 2D matrix, and wherein the opcode is to indicate a move of the specified group between the 2D matrix and the 1D vector, decode circuitry to decode the fetched instruction; and execution circuitry, responsive to the decoded instruction, when the opcode specifies a move from 1D, to move contents of the specified 1D vector to the specified group of elements.

Systems for performing instructions to quickly convert and use tiles as 1D vectors

Techniques for converting FP16 data elements to BF8 data elements using a single instruction are described. An exemplary apparatus includes decoder circuitry to decode a single instruction, the single instruction to include a one or more fields to identify a source operand, one or more fields to identify a destination operand, and one or more fields for an opcode, the opcode to indicate that execution circuitry is to convert packed half-precision floating-point data from the identified source to packed bfloat8 data and store the packed bfloat8 data into corresponding data element positions of the identified destination operand; and execution circuitry to execute the decoded instruction according to the opcode to convert packed half-precision floating-point data from the identified source to packed bfloat8 data and store the packed bfloat8 data into corresponding data element positions.

Instructions to convert from FP16 to BF8

Systems, methods, and apparatuses relating sparsity based FMA. In some examples, an instance of a single FMA instruction has one or more fields for an opcode, one or more fields to identify a source/destination matrix operand, one or more fields to identify a first plurality of source matrix operands, one or more fields to identify a second plurality of matrix operands, wherein the opcode is to indicate that execution circuitry is to select a proper subset of FP8 data elements from the first plurality of source matrix operands based on sparsity controls from a first matrix operand of the second plurality of matrix operands and perform a FMA.

Instructions for structured-sparse tile matrix FMA

Embodiments for a matrix transpose and multiply operation are disclosed. In an embodiment, a processor includes a decoder and execution circuitry. The decoder is to decode an instruction having a format including an opcode field to specify an opcode, a first destination operand field to specify a destination matrix location, a first source operand field to specify a first source matrix location, and a second source operand field to specify a second source matrix location. The execution circuitry is to, in response to the decoded instruction, transpose the first source matrix to generate a transposed first source matrix, perform a matrix multiplication using the transposed first source matrix and the second source matrix to generate a result, and store the result in a destination matrix location.

Matrix transpose and multiply

Techniques for scale and reduction of BF16 data elements are described. An exemplary instruction includes fields for an having fields for an opcode, an identification of a location of a first packed data source operand, an identification of a location of a second packed data source operand, and an identification of a packed data destination operand, wherein the opcode is to indicate that execution circuitry is to perform, for each data element position of the packed data source operands, a floating point scale operation of a BF16 data element of the first packed data source by multiplying the data element by a power of 2 value, wherein a value of the exponent of the power of 2 value is a floor value of a BF16 data element of the second packed data source, and store a result of the floating point scale operation into a corresponding data element position of the packed data destination operand.

BFLOAT16 scale and/or reduce instructions

According to one embodiment, a processor includes an instruction decoder to decode a first instruction to gather data elements from memory, the first instruction having a first operand specifying a first storage location and a second operand specifying a first memory address storing a plurality of data elements. The processor further includes an execution unit coupled to the instruction decoder, in response to the first instruction, to read contiguous a first and a second of the data elements from a memory location based on the first memory address indicated by the second operand, and to store the first data element in a first entry of the first storage location and a second data element in a second entry of a second storage location corresponding to the first entry of the first storage location.

Coalescing adjacent gather/scatter operations

Techniques for matrix multiplication are described. In some examples, decode circuitry is to decode a single instruction having fields for an opcode, an indication of a location of a first source operand, an indication of a location of a second source operand, and an indication of a location of a destination operand, wherein the opcode is to indicate that execution circuitry is to at least convert data elements of the first and second source operands from a first floating point representation to a second floating point representation, perform matrix multiplication with the converted data elements, and accumulate results of the matrix multiplication in the destination operand in the first floating point representation; and the execution circuitry is to execute to the decoded instruction as specified by the opcode.

Apparatuses, methods, and systems for instructions for matrix multiplication instructions

In one embodiment, a processor includes interconnect circuitry, processing circuitry, a first cache, and cache controller circuitry. The interconnect circuitry communicates over a processor interconnect with a second processor that includes a second cache. The processing circuitry generates a memory read request for a corresponding memory address of a memory. Based on the memory read request, the cache controller circuitry detects a cache miss in the first cache, which indicates that the first cache does not contain a valid copy of data for the corresponding memory address. Based on the cache miss, the cache controller circuitry requests the data from the second cache or the memory based on a current bandwidth utilization of the processor interconnect.

Dynamic cache coherence protocol based on runtime interconnect utilization

Disclosed embodiments relate to accelerating multiplication of sparse matrices. In one example, a processor is to fetch and decode an instruction having fields to specify locations of first, second, and third matrices, and an opcode indicating the processor is to multiply and accumulate matching non-zero (NZ) elements of the first and second matrices with corresponding elements of the third matrix, and executing the decoded instruction as per the opcode to generate NZ bitmasks for the first and second matrices, broadcast up to two NZ elements at a time from each row of the first matrix and each column of the second matrix to a processing engine (PE) grid, each PE to multiply and accumulate matching NZ elements of the first and second matrices with corresponding elements of the third matrix. Each PE further to store an NZ element for use in a subsequent multiplications.

Systems and methods of instructions to accelerate multiplication of sparse matrices using bitmasks that identify non-zero elements

Systems, methods, and apparatuses relating to instructions to convert 16-bit floating-point formats are described. In one embodiment, a processor includes fetch circuitry to fetch a single instruction having fields to specify an opcode and locations of a source vector comprising N plurality of 16-bit half-precision floating-point elements, and a destination vector to store N plurality of 16-bit bfloat floating-point elements, the opcode to indicate execution circuitry is to convert each of the elements of the source vector from 16-bit half-precision floating-point format to 16-bit bfloat floating-point format and store each converted element into a corresponding location of the destination vector, decode circuitry to decode the fetched single instruction into a decoded single instruction, and the execution circuitry to respond to the decoded single instruction as specified by the opcode.

Apparatuses, methods, and systems for instructions to convert 16-bit floating-point formats

Techniques for performing BF16 FMA in response to an instruction are described. In some examples, an instruction has fields for an opcode, an identification of location of a packed data source/destination operand (a first source), an identification of a location of a second packed data source operand, an identification of a location of a third packed data source operand, and an identification of location of a packed data source/destination operand, wherein the opcode is to indicate operand ordering and that execution circuitry is to, per data element position, perform a BF16 value fused multiply-accumulate operation using the first, second, and third source operands and store a result in a corresponding data element position of the source/destination operand.

BFLOAT16 fused multiply instructions

Disclosed embodiments relate to atomic memory operations. In one example, an apparatus includes multiple processor cores, a cache hierarchy, a local execution unit, and a remote execution unit, and an adaptive remote atomic operation unit. The cache hierarchy includes a local cache at a first level and a shared cache at a second level. The local execution unit is to perform an atomic operation at the first level if the local cache is a storing a cache line including data for the atomic operation. The remote execution unit is to perform the atomic operation at the second level. The adaptive remote atomic operation unit is to determine whether to perform the first atomic operation at the first level or at the second level and whether to copy the cache line from the shared cache to the local cache.

Adaptive remote atomics

An embodiment of an integrated circuit may comprise circuitry communicatively coupled to two or more sub-non-uniform memory access clusters (SNCs) to allocate a specified memory space in the two or more SNCs in accordance with a SNC memory allocation policy indicated from a request to initialize the specified memory space. An embodiment of an apparatus may comprise decode circuitry to decode a single instruction, the single instruction to include a field for an opcode, and execution circuitry to execute the decoded instruction according to the opcode to provide an indicated SNC memory allocation policy (e.g., a SNC policy hint). Other embodiments are disclosed and claimed.

Inter-cluster shared data management in sub-NUMA cluster

Growing community of inventors

Santa Clara, CA, United States of America

Christopher J Hughes