Patent Application
20220121954
This disclosure relates in general to the field of computer processing, and more particularly, though not exclusively, to performing matrix operations using a plurality of processing resources.
Matrix operations, such as matrix multiplication and convolutions, can be highly processor-intensive and memory-intensive operations, as they often involve complex operations on large, multi-dimensional matrix operands. Accordingly, the performance of complex matrix operations can be limited by the processing and/or memory latency. As matrix operations are increasingly utilized in a variety of applications and with ever-growing data sets (from graphics and image processing to machine learning and artificial intelligence), the demand for high-performance processing of matrix operations is increasing.
The present disclosure is best understood from the following detailed description when read with the accompanying figures. It is emphasized that, in accordance with the standard practice in the industry, various features are not necessarily drawn to scale, and are used for illustration purposes only. Where a scale is shown, explicitly or implicitly, it provides only one illustrative example. In other embodiments, the dimensions of the various features may be arbitrarily increased or reduced for clarity of discussion.
The following disclosure provides many different embodiments, or examples, for implementing different features of the present disclosure. Specific examples of components and arrangements are described below to simplify the present disclosure. These are, of course, merely examples and are not intended to be limiting. Further, the present disclosure may repeat reference numerals and/or letters in the various examples. This repetition is for the purpose of simplicity and clarity and does not in itself dictate a relationship between the various embodiments and/or configurations discussed. Different embodiments may have different advantages, and no particular advantage is necessarily required of any embodiment.
Matrix processing operations (e.g., linear algebra operations that involve matrix and/or vector operands) have a wide range of applications in computing systems, from graphics processing to machine learning and artificial intelligence, among other examples. For example, complex matrix operations may be used to implement artificial neural networks that provide artificial intelligence and machine learning capabilities, including computer vision, autonomous navigation, speech and audio recognition, and natural language processing, among other examples. These complex matrix operations (e.g., matrix multiplication and convolutions) may be used to implement the fundamental operations of neural networks, such as forward propagation, backward propagation, and weight updates. These matrix operations, however, can be highly processor and memory intensive, as they often involve complex operations on large, multi-dimensional matrix operands. Accordingly, the performance of these matrix operations can be limited by processing and/or memory latency. As matrix operations are increasingly utilized in a variety of applications with ever-growing data sets, such as artificial intelligence and machine learning, the demand for high-performance processing of matrix operations is increasing.
Existing matrix processing approaches suffer from various inefficiencies, particularly when used to implement artificial intelligence and machine learning in artificial neural networks. For example, while central processing units (CPUs) could be used to perform matrix operations, many CPU architectures are designed for low arithmetic intensity operations (i.e., a low ratio of arithmetic operations relative to memory operations), and thus are not designed for efficient execution of matrix operations. Moreover, many CPU architectures utilize complex local or cache memory management routines, which may increase processing overhead and execution complexity for operations involving large matrix operands. Graphics processing units (GPUs) could also be used to perform matrix operations. GPUs, however, are often designed for high precision computations and may provide a level of precision that is unnecessary for certain matrix operations, thus reducing the volume of matrix operations that can be performed. Accordingly, existing matrix processing approaches are inefficient for certain matrix operations, such as matrix multiplication or convolution operations involving large matrix operands and/or matrix operands with certain dimensions, among other examples. The existing approaches are unable to perform these matrix operations with 100% processing efficiency using all available processing resources. Moreover, existing approaches cannot be efficiently scaled to perform these matrix operations across additional processing resources in parallel. As an example, existing approaches are inefficient for matrix multiplication (e.g., general matrix multiplication or GEMM) on a large matrix operand which is neither square nor a single vector, such as a “thin” matrix with a much larger height than width. Existing approaches require more time to access and communicate the matrix operands than to perform the actual matrix computations, resulting in idle processing time while matrix operands are being obtained from memory and/or communicated to processing resources. Similarly, existing approaches are inefficient for convolution operations on large matrix operands, as they are unable to efficiently distribute or scale a convolution operation across a variable number of processing resources. Thus, existing approaches do not achieve 100% processing efficiency for these matrix operations.
The matrix processing functionality described throughout this disclosure performs matrix operations using a distributed approach that achieves 100% processing efficiency using the available processing resources. For example, this approach distributes matrix operations across multiple processing resources in a processing architecture that is optimized for performing matrix operations, thus enabling full utilization of the processing resources throughout the duration of the matrix operations. For example, the processing architecture may include multiple processing resources that are designed and optimized for performing matrix operations, and may support a higher volume of matrix operations than other architectures (e.g., GPUs). In some embodiments, these processing resources may be configured in a cyclical arrangement, with either unidirectional communication interfaces between neighboring processing resources (a “single-cyclical” configuration) or bi-directional communication interfaces between neighboring processing resources (a “dual-cyclical” configuration). In addition, the processing resources may be arranged hierarchically with multiple levels of processing resources. For example, in some embodiments, the processing resources may include multiple matrix processing chips, multiple high bandwidth memory (HBM) modules and matrix processing clusters on each matrix processing chip, and/or multiple matrix processing units (MPUs) on each matrix processing cluster. This processing architecture enables matrix operations to be distributed across multiple processing resources and/or processing hierarchies with 100% processing efficiency. In addition, this processing architecture enables matrix operations to be efficiently scaled across a variable number of processing resources operating in parallel, while still achieving 100% processing efficiency.
The distributed matrix processing functionality described throughout this disclosure provides numerous technical advantages, including alleviating the inefficiencies of existing approaches and enabling matrix operations to be executed efficiently, achieving 100% processing efficiency using the available processing resources, and efficiently scaling matrix operations across a variable number of processing resources operating in parallel. These advantages result in reduced processing time for matrix operations, which improves performance for applications that involve complex matrix operations, such as artificial intelligence and machine learning functionality implemented using artificial neural networks (e.g., convolutional neural networks, multilayer perceptrons (MLPs), restricted Boltzmann machines (RBM), and deep belief networks (DBN), among other examples).
Example embodiments that may be used to implement the matrix processing functionality of this disclosure will now be described with more particular reference to the attached FIGURES.
In some embodiments, the matrix processing functionality described throughout this disclosure may be implemented in system 100. Matrix processing functionality may be used in system 100 for a wide range of applications and/or use cases involving matrix operations, from graphics processing to machine learning and artificial intelligence, among other examples. For example, in some embodiments, matrix processing functionality may be used to implement artificial intelligence and machine learning in artificial neural networks. Moreover, matrix processing functionality may be implemented by any component of system 100. For example, in the illustrated embodiment, system 100 includes edge devices 110, cloud services 120, matrix processing nodes 130, and network 150. Matrix processing nodes 130 may include any component or device with matrix processing functionality, including any component of system 100. For example, matrix processing nodes 130 may include cloud services 120 and/or servers implemented with matrix processing functionality (e.g., application servers in a datacenter), edge devices 110 implemented with matrix processing functionality (e.g., end-user devices 112, Internet-of-Things devices 114, gateways 116), and so forth. These various components of system 100 are discussed further below.
Edge devices 110 may include any equipment and/or devices deployed or connected near the “edge” of a communication system 100. Edge devices 110 may communicate with each other and/or with other remote networks and services (e.g., cloud services 120) through one or more networks and/or communication protocols, such as network 150. In some embodiments, certain edge devices 110 may include the matrix processing functionality described throughout this disclosure, and thus may be used as matrix processing nodes 130. In the illustrated embodiment, edge devices 110 include end-user devices 112 (e.g., desktops, laptops, mobile devices), Internet-of-Things (IoT) devices 114, and gateways and/or routers 116, among other examples.
End-user devices 112 may include any device that enables or facilitates user interaction with computing system 100, including, for example, desktop computers, laptops, tablets, mobile phones and other mobile devices, and wearable devices (e.g., smart watches, smart glasses, headsets), among other examples.
IoT devices 114 may include any device capable of communicating and/or participating in an Internet-of-Things (IoT) system or network. IoT systems may refer to new or improved ad-hoc systems and networks composed of multiple different devices (e.g., IoT devices 114) interoperating and synergizing for a particular application or use case. Such ad-hoc systems are emerging as more and more products and equipment evolve to become “smart,” meaning they are controlled or monitored by computer processors and are capable of communicating with other devices. For example, an IoT device 114 may include a computer processor and/or communication interface to allow interoperation with other components of system 100, such as with cloud services 120 and/or other edge devices 110. IoT devices 114 may be “greenfield” devices that are developed with IoT capabilities from the ground-up, or “brownfield” devices that are created by integrating IoT capabilities into existing legacy devices that were initially developed without IoT capabilities. For example, in some cases, IoT devices 114 may be built from sensors and communication modules integrated in or attached to “things,” such as equipment, toys, tools, vehicles, living things (e.g., plants, animals, humans), and so forth. Alternatively, or additionally, certain IoT devices 114 may rely on intermediary components, such as edge gateways or routers 116, to communicate with the various components of system 100.
IoT devices 114 may include various types of sensors for monitoring, detecting, measuring, and generating sensor data and signals associated with characteristics of their environment. For instance, a given sensor may be configured to detect one or more respective characteristics, such as movement, weight, physical contact, temperature, wind, noise, light, position, humidity, radiation, liquid, specific chemical compounds, battery life, wireless signals, computer communications, and bandwidth, among other examples. Sensors can include physical sensors (e.g., physical monitoring components) and virtual sensors (e.g., software-based monitoring components). IoT devices 114 may also include actuators to perform various actions in their respective environments. For example, an actuator may be used to selectively activate certain functionality, such as toggling the power or operation of a security system (e.g., alarm, camera, locks) or household appliance (e.g., audio system, lighting, HVAC appliances, garage doors), among other examples.
Indeed, this disclosure contemplates use of a potentially limitless universe of IoT devices 114 and associated sensors/actuators. IoT devices 114 may include, for example, any type of equipment and/or devices associated with any type of system 100 and/or industry, including transportation (e.g., automobile, airlines), industrial manufacturing, energy (e.g., power plants), telecommunications (e.g., Internet, cellular, and television service providers), medical (e.g., healthcare, pharmaceutical), food processing, and/or retail industries, among others. In the transportation industry, for example, IoT devices 114 may include equipment and devices associated with aircrafts, automobiles, or vessels, such as navigation systems, autonomous flight or driving systems, traffic sensors and controllers, and/or any internal mechanical or electrical components that are monitored by sensors (e.g., engines). IoT devices 114 may also include equipment, devices, and/or infrastructure associated with industrial manufacturing and production, shipping (e.g., cargo tracking), communications networks (e.g., gateways, routers, servers, cellular towers), server farms, electrical power plants, wind farms, oil and gas pipelines, water treatment and distribution, wastewater collection and treatment, and weather monitoring (e.g., temperature, wind, and humidity sensors), among other examples. IoT devices 114 may also include, for example, any type of “smart” device or system, such as smart entertainment systems (e.g., televisions, audio systems, videogame systems), smart household or office appliances (e.g., heat-ventilation-air-conditioning (HVAC) appliances, refrigerators, washers and dryers, coffee brewers), power control systems (e.g., automatic electricity, light, and HVAC controls), security systems (e.g., alarms, locks, cameras, motion detectors, fingerprint scanners, facial recognition systems), and other home automation systems, among other examples. IoT devices 114 can be statically located, such as mounted on a building, wall, floor, ground, lamppost, sign, water tower, or any other fixed or static structure. IoT devices 114 can also be mobile, such as devices in vehicles or aircrafts, drones, packages (e.g., for tracking cargo), mobile devices, and wearable devices, among other examples. Moreover, an IoT device 114 can also be any type of edge device 110, including end-user devices 112 and edge gateways and routers 116.
Edge gateways and/or routers 116 may be used to facilitate communication to and from edge devices 110. For example, gateways 116 may provide communication capabilities to existing legacy devices that were initially developed without any such capabilities (e.g., “brownfield” IoT devices). Gateways 116 can also be utilized to extend the geographical reach of edge devices 110 with short-range, proprietary, or otherwise limited communication capabilities, such as IoT devices 114 with Bluetooth or ZigBee communication capabilities. For example, gateways 116 can serve as intermediaries between IoT devices 114 and remote networks or services, by providing a front-haul to the IoT devices 114 using their native communication capabilities (e.g., Bluetooth, ZigBee), and providing a back-haul to other networks 150 and/or cloud services 120 using another wired or wireless communication medium (e.g., Ethernet, Wi-Fi, cellular). In some embodiments, a gateway 116 may be implemented by a dedicated gateway device, or by a general purpose device, such as another IoT device 114, end-user device 112, or other type of edge device 110.
In some instances, gateways 116 may also implement certain network management and/or application functionality (e.g., IoT management and/or IoT application functionality for IoT devices 114), either separately or in conjunction with other components, such as cloud services 120 and/or other edge devices 110. For example, in some embodiments, configuration parameters and/or application logic may be pushed or pulled to or from a gateway device 116, allowing IoT devices 114 (or other edge devices 110) within range or proximity of the gateway 116 to be configured for a particular IoT application or use case.
Cloud services 120 may include services that are hosted remotely over a network 150, or in the “cloud.” In some embodiments, for example, cloud services 120 may be remotely hosted on servers in datacenter (e.g., application servers or database servers). Cloud services 120 may include any services that can be utilized by or for edge devices 110, including but not limited to, data storage, computational services (e.g., data analytics, searching, diagnostics and fault management), security services (e.g., surveillance, alarms, user authentication), mapping and navigation, geolocation services, network or infrastructure management, IoT application and management services, payment processing, audio and video streaming, messaging, social networking, news, and weather, among other examples. In some embodiments, certain cloud services 120 may include the matrix processing functionality described throughout this disclosure, and thus may be used as matrix processing nodes 130.
In general, edge devices 110 (and in particular IoT devices 114) may generate an extremely large volume and variety of data. IoT edge devices 114 typically offload this data to the cloud for processing and/or storage (e.g., by cloud services 120). Cloud services 120, however, may not necessarily be suited to handle the rapidly growing volume, variety, and velocity of data generated by IoT devices 114 and other edge devices 110. For example, cloud-based processing may not be ideal in certain circumstances, such as processing time-sensitive or highly confidential data, or when faced with network bandwidth constraints, among other examples. In some embodiments, cloud services 120 may leverage “edge” based processing using edge devices 110 to improve the performance of cloud services. Edge processing is an approach that involves processing certain data at the network edge (e.g., using edge devices 110), near where the data is generated, rather than simply funneling large volumes of data to the cloud for processing and storage. Certain data may still be sent to the cloud, as appropriate, such as for deeper analysis and/or long-term storage. Edge processing may be used to complement the shortcomings of cloud-based processing (e.g., when cloud-based processing is inefficient, ineffective, and/or unsecure), and thus improve the handling of the growing volume, variety, and velocity of data generated by IoT devices 114 and/or other edge devices 110. For example, in some cases, processing data near its source (e.g., in the network edge) rather than in the cloud may improve performance and/or avoid system failures or disasters. Edge processing may also conserve network bandwidth, which may be particularly beneficial when facing bandwidth constraints and/or limited network connectivity.
In some embodiments, edge devices 110 that provide edge-based processing for cloud services 120 may be collectively referred to as the “fog,” as they serve to extend the “cloud” to the edge of the network, thus creating a “fog” over the network edge. In some embodiments, devices 110 in the “fog” may connect and/or communicate with each other, for example, using an interconnection standard or protocol. For example, in some embodiments, device interconnection may be implemented using the open interconnect consortium (OIC) standard specification 1.0, released by the Open Connectivity Foundation™ (OCF) on Dec. 23, 2015, which enables devices to discover and connect with each other. Another interconnection protocol that may be used is Thread, a networking protocol for Internet-of-Things (IoT) devices used in “smart” home automation and similar deployments, which has been developed by an alliance of organizations named the “Thread Group.” Other interconnection protocols may also be used, including, for example, the optimized link state routing (OLSR) protocol, or the better approach to mobile ad-hoc networking (B.A.T.M.A.N.), among others.
Network 150 may be used to facilitate communication between the components of computing system 100. For example, edge devices 110, such as end-user devices 112 and IoT devices 114, may use network 150 to communicate with each other and/or access one or more remote cloud services 120. Network 150 may include any number or type of communication networks, including, for example, local area networks, wide area networks, public networks, the Internet, cellular networks, Wi-Fi networks, short-range networks (e.g., Bluetooth or ZigBee), and/or any other wired or wireless networks or communication mediums.
Any, all, or some of the computing devices of system 100 may be adapted to execute any operating system, including Linux or other UNIX-based operating systems, Microsoft Windows, Windows Server, MacOS, Apple iOS, Google Android, or any customized and/or proprietary operating system, along with virtual machines adapted to virtualize execution of a particular operating system.
While
Example Matrix Processing Architecture
In some embodiments, the matrix processing functionality described throughout this disclosure may be implemented using a matrix processing architecture, such as the matrix processing architecture of
Host processor 260 may be configured to control and/or manage matrix processing system 200. For example, in some embodiments, host processor 260 may use matrix processing resources 210 to perform complex matrix operations. Host processor 260 may be any processing resource capable of controlling and/or managing matrix processing functionality of matrix processing system 200. For example, in some embodiments, host processor 260 may be implemented using computer processors 300 or 400 of
Host memory 270 may include any type or combination of volatile and/or non-volatile memory. Examples of volatile memory include various types of random access memory (RAM), such as dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), and static random access memory (SRAM), among other examples. Examples of non-volatile memory include disk-based storage mediums (e.g., magnetic and/or optical storage mediums), solid-state storage (e.g., any form of persistent flash memory, including planar or three dimensional (3D) NAND flash memory or NOR flash memory), 3D crosspoint memory, electrically erasable programmable read-only memory (EEPROM), and/or other types of non-volatile random access memories (RAM), among other examples. Host memory 270 may be used, for example, to store information for host processor 260 during execution, such as code and/or data.
Interconnect bus 280 may be used, in some embodiments, to communicatively couple host processor 260 and host memory 270 to matrix processing resources 210. Interconnect bus 280 may use any interconnection protocol, such as Peripheral Component Interconnect express (PCIe), Universal Serial Bus (USB), or Small Computer Systems Interface (SCSI), among other examples.
Matrix processing resources 210 may include any processing resources configured to perform matrix operations. For example, matrix processing resources 210 may be configured to perform matrix multiplication operations, convolution operations, element-wise matrix operations (e.g., +, *, /<, >, ==), dimension shuffle operations, and/or any combination thereof. In some embodiments, matrix processing resources 210 may include processing resources that are designed and optimized for performing matrix operations. In some embodiments, matrix processing resources 210 may also be arranged hierarchically with multiple levels of processing resources. For example, in the illustrated embodiment, matrix processing resources 210 include a plurality of matrix processing chips 220, and may also include any processing resources within each matrix processing chip 220. For example, as discussed below in connection with
Matrix processing chips 220 may be, for example, any chips or other components configured to perform matrix operations. For example, in some embodiments, a matrix processing chip 220 may be a peripheral card or chip connected to host processor 260 using any type of interconnect interface, such as a PCIe interface. In some embodiments, a matrix processing chip 220 may be implemented using an integrated circuit, such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), and/or any other type of circuitry. In the illustrated embodiment, matrix processing chips 220 are configured in a cyclical arrangement, with communication channels 215 between neighboring matrix processing chips 220. In some embodiments, communication channels 215 may provide one-way communication between neighboring matrix processing chips 220. In other embodiments, however, communication channels 215 may provide bi-directional communication between neighboring matrix processing chips 220. A cyclical arrangement with one-way communication between neighboring processing resources may be referred to as a “single-cyclical” configuration, while a cyclical arrangement with bi-directional communication between neighboring processing resources may be referred to as a “dual-cyclical” configuration.
Moreover, although not illustrated, in some embodiments matrix processing system 200 may include a communication interface to communicate over a communication network. For example, in some embodiments, matrix processing system 200 may communicate over a network with one or more remote matrix processing chips to perform distributed matrix operations.
Controller 222 may be configured to control and/or manage matrix operations performed by matrix processing chip 220. In some embodiments, controller 222 may control and/or manage matrix operations in conjunction with host processor 260 of
Host interface 224 may be a communication interface that enables a matrix processing chip 220 to communicate with host processor 260 of
Inter-chip links (ICLs) 225 may enable a matrix processing chip 220 to communicate with other matrix processing chips. For example, inter-chip links 225 may be used to implement the communication channels 215 between matrix processing chips 220 in
High bandwidth memory (HBM) modules 240 may be memory components associated with matrix processing chip 220 that are used to store matrix operands and other matrix data. In some embodiments, high bandwidth memory (HBM) modules 240 may be designed to efficiently store and retrieve matrix data. In some embodiments, high bandwidth memory (HBM) modules 240 may be multi-dimensional memory components configured to store and retrieve data in multiple dimensions. For example, in some embodiments, high bandwidth memory (HBM) modules 240 may be memory components configured to store and retrieve data in two dimensions, such as rows and columns. Other embodiments, however, may use memory components configured to store and retrieve data using any other number of dimensions (e.g., one dimension, three dimensions, four dimensions, and so forth). In the illustrated embodiment, matrix processing chip 220 includes four high bandwidth memory (HBM) modules 240a-d. In some embodiments, high bandwidth memory (HBM) modules 240 may be shared by the matrix processing clusters 230 of a matrix processing chip 220.
Matrix processing clusters 230 may include processing resources configured to perform matrix operations, such as matrix multiplication, convolutions, and/or dimension shuffling, among other examples. In some embodiments, matrix processing clusters 230 may be collectively used to execute a particular matrix operation by performing matrix processing in parallel. In the illustrated embodiment, matrix processing chip 220 includes twelve matrix processing clusters 230a-l. Moreover, in the illustrated embodiment, matrix processing clusters 230 are configured or arranged using a two-dimensional mesh interconnection topology. The interconnection topology of matrix processing clusters 230 may facilitate cyclical communication among the matrix processing clusters 230. Moreover, other embodiments may include any number and/or arrangement of matrix processing clusters 230.
Master control CPU (MCC) 232 may be configured to control and/or manage matrix operations performed by a matrix processing cluster 230. In some embodiments, master control CPU 232 may be a microprocessor, an integrated circuit, and/or any other type of circuitry and/or processing logic. In some embodiments, master control CPU 232 may receive instructions from another component, such as host processor 260 of
Matrix processing units (MPUs) 234 may be configured to perform matrix operations, such as matrix multiplication, convolutions, and/or dimension shuffling. In some embodiments, matrix processing units (MPUs) 234 perform matrix operations based on commands received from master control CPU (MCC) 232. Moreover, in some embodiments, each matrix processing cluster 230 may include multiple matrix processing units (MPUs) 234. For example, in the illustrated embodiment, matrix processing cluster 230 includes two matrix processing units (MPUs) 234. A matrix processing unit (MPU) 234 may be capable of performing matrix operations, such as matrix multiplication, on small matrices (e.g., 32×32 matrices). In some cases, a matrix processing unit (MPU) 234 may be designed and/or optimized to perform matrix multiplication operations. A matrix processing unit (MPU) 234 may load matrix operands from memory resource blocks (MRBs) 238. In some embodiments, a matrix processing unit (MPU) 234 may support the following arithmetic operations: matrix multiplication; unary matrix operations; binary matrix operations, such as addition (+), subtraction (−), multiplication (*), division (/), bitwise XOR, AND, OR, logical and arithmetic left and right shift, comparison (>, <, >=, <=, ==, !=); and column-wise, row-wise, and matrix-wide operations, such as sum, max value, and min value.
Slicing engine 236 may be configured to slice the matrix operands of a particular matrix operation into smaller partial matrices. For example, in some embodiments, master control CPU (MCC) 232 may use slicing engine 236 to break up matrix operands into smaller partial matrices for matrix processing units (MPUs) 234. In some embodiments, slicing engine 236 may include a convolution slicing engine (CSE) to perform matrix slicing for convolution operations. For example, in some embodiments, a convolution slicing engine (CSE) may slice matrix operands in a manner that enables a convolution operation to be cast as a matrix multiplication operation, thus enabling the same processing logic to perform both matrix multiplication and convolution operations. Moreover, in some embodiments, slicing engine 236 and/or the associated convolution slicing engine (CSE) may be used to perform the dimension shuffle operations to reorder the dimensions of a matrix.
Memory resource blocks (MRBs) 238 may be memory components on matrix processing cluster 230 used to store matrix operands and other matrix data. In some embodiments, memory resource blocks (MRBs) 238 may be designed to store and retrieve matrix data efficiently. In some embodiments, memory resource blocks (MRBs) 238 may be multi-dimensional memory components configured to store and retrieve data in multiple dimensions. For example, in some embodiments, memory resource blocks (MRBs) 238 may be memory components configured to store and retrieve data in two dimensions, such as rows and columns. In the illustrated embodiment, matrix processing cluster 230 includes ten memory resource blocks (MRBs) 238. Other embodiments, however, may include a different number of memory resource blocks (MRBs) 238 on a matrix processing cluster 230. In some embodiments, each memory resource block (MRB) 238 may be capable of storing a matrix of a certain size (e.g., a 256×512 matrix). In some embodiments, memory resource blocks (MRBs) 238 may be shared by the matrix processing units (MPUs) 234 of a particular matrix processing cluster 230.
In some embodiments, the matrix processing architecture of
As an example, the matrix processing architecture of
As an example, when a matrix operation or command is received, the matrix operation may be distributed across the processing resources 210 of matrix processing system 200. For example, the matrix operands (or input matrices) may be partitioned based on the number of available processing resources 210. Moreover, in some embodiments, the partitions may be across the rows of the matrix operands, and/or across any other dimension of the matrix operands. Each partition may then be distributed to a particular processing resource 210. Each processing resource 210 may then perform a plurality of partial matrix operations. In some embodiments, the plurality of partial matrix operations is performed in a plurality of stages. For example, each processing resource 210 may perform a particular stage of partial matrix operations while simultaneously sending and receiving partial matrix data to and from its neighboring processing resources 210. For example, in a single-cyclical configuration of processing resources 210, each processing resource 210 either sends or receives partial matrix data to or from each neighbor processing resource. Similarly, in a dual-cyclical configuration of processing resources 210, each processing resource 210 may send and receive partial matrix data to and from each neighboring processing resource 210.
Each processing resource 210 may then use the partial matrix data for subsequent partial matrix operations. The result of the matrix operation may then be determined based on the partial matrix operations collectively performed by the processing resources 210.
Moreover, if the processing resources 210 are arranged hierarchically, the matrix operation may be distributed in a hierarchical manner. For example, the matrix operands (or input matrices) may initially be partitioned based on the number of available matrix processing chips 220. Each partition, and the associated partial matrix operations, may then be distributed to a particular matrix processing chip 220. The partition and partial matrix operations distributed to a particular matrix processing chip 220 may then be similarly partitioned and distributed across the matrix processing clusters 230 and/or high bandwidth memory (HBM) modules 240 of the particular matrix processing chip 220. For example, for certain matrix operations, partial matrix operations may be distributed to each matrix processing cluster 230. Alternatively, for certain matrix operations, partial matrix operations may be distributed across various “logical processing nodes” (e.g., groups of matrix processing clusters 230 associated with a high-bandwidth memory (HBM) module 240), and may then be distributed to each matrix processing cluster 230 of a particular logical processing node. In some embodiments, the matrix processing clusters 230 (and/or the logical processing nodes) may be cyclically configured similar to the matrix processing chips 220. The partition and partial matrix operations distributed to a particular matrix processing cluster 230 may then be similarly partitioned and distributed across the matrix processing units (MPUs) 234 of the particular matrix processing cluster 230.
Example Computer Processor Architectures
Processor 300 can execute any type of instructions associated with algorithms, processes, or operations detailed herein. Generally, processor 300 can transform an element or an article (e.g., data) from one state or thing to another state or thing.
Code 304, which may be one or more instructions to be executed by processor 300, may be stored in memory 302, or may be stored in software, hardware, firmware, or any suitable combination thereof, or in any other internal or external component, device, element, or object where appropriate and based on particular needs. In one example, processor 300 can follow a program sequence of instructions indicated by code 304. Each instruction enters a front-end logic 306 and is processed by one or more decoders 308. The decoder may generate, as its output, a micro operation such as a fixed width micro operation in a predefined format, or may generate other instructions, microinstructions, or control signals that reflect the original code instruction. Front-end logic 306 may also include register renaming logic and scheduling logic, which generally allocate resources and queue the operation corresponding to the instruction for execution.
Processor 300 can also include execution logic 314 having a set of execution units 316a, 316b, 316n, etc. Some embodiments may include a number of execution units dedicated to specific functions or sets of functions. Other embodiments may include only one execution unit or one execution unit that can perform a particular function. Execution logic 314 performs the operations specified by code instructions.
After completion of execution of the operations specified by the code instructions, back-end logic 318 can retire the instructions of code 304. In one embodiment, processor 300 allows out of order execution but requires in order retirement of instructions. Retirement logic 320 may take a variety of known forms (e.g., re-order buffers or the like). In this manner, processor 300 is transformed during execution of code 304, at least in terms of the output generated by the decoder, hardware registers and tables utilized by register renaming logic 310, and any registers (not shown) modified by execution logic 314.
Although not shown in
Processors 470 and 480 are shown including integrated memory controller (IMC) units 472 and 482, respectively. Processor 470 also includes as part of its bus controller units point-to-point (P-P) interfaces 476 and 478; similarly, second processor 480 includes P-P interfaces 486 and 488. Processors 470, 480 may exchange information via a point-to-point (P-P) interface 450 using P-P interface circuits 478, 488. As shown in
Processors 470, 480 may each exchange information with a chipset 490 via individual P-P interfaces 452, 454 using point to point interface circuits 476, 494, 486, 498. Chipset 490 may optionally exchange information with the coprocessor 438 via a high-performance interface 439. In one embodiment, the coprocessor 438 is a special-purpose processor, such as, for example, a high-throughput MIC processor, a network or communication processor, compression engine, graphics processor, GPGPU, embedded processor, matrix processor, or the like.
A shared cache (not shown) may be included in either processor or outside of both processors, yet connected with the processors via P-P interconnect, such that either or both processors' local cache information may be stored in the shared cache if a processor is placed into a low power mode.
Chipset 490 may be coupled to a first bus 416 via an interface 496. In one embodiment, first bus 416 may be a Peripheral Component Interconnect (PCI) bus, or a bus such as a PCI Express bus or another third generation I/O interconnect bus, although the scope of this disclosure is not so limited.
As shown in
All or part of any component of
Embodiments of the mechanisms disclosed herein may be implemented in hardware, software, firmware, or a combination of such implementation approaches. Certain embodiments may be implemented as computer programs or program code executing on programmable systems comprising at least one processor, a storage system (including volatile and non-volatile memory and/or storage elements), at least one input device, and at least one output device.
Program code, such as code 430 illustrated in
The program code may be implemented in a high level procedural or object oriented programming language to communicate with a processing system. The program code may also be implemented in assembly or machine language, if desired. In fact, the mechanisms described herein are not limited in scope to any particular programming language. In any case, the language may be a compiled or interpreted language.
One or more aspects of at least one embodiment may be implemented by representative instructions stored on a machine-readable medium which represents various logic within the processor, which when read by a machine causes the machine to fabricate logic to perform the techniques described herein. Such representations, known as “IP cores” may be stored on a tangible, machine readable medium and supplied to various customers or manufacturing facilities to load into the fabrication machines that actually make the logic or processor.
Such machine-readable storage media may include, without limitation, non-transitory, tangible arrangements of articles manufactured or formed by a machine or device, including storage media such as hard disks, any other type of disk including floppy disks, optical disks, compact disk read-only memories (CD-ROMs), compact disk rewritable's (CD-RWs), and magneto-optical disks, semiconductor devices such as read-only memories (ROMs), random access memories (RAMs) such as dynamic random access memories (DRAMs), static random access memories (SRAMs), erasable programmable read-only memories (EPROMs), flash memories, electrically erasable programmable read-only memories (EEPROMs), phase change memory (PCM), magnetic or optical cards, or any other type of media suitable for storing electronic instructions.
Accordingly, embodiments of this disclosure also include non-transitory, tangible machine-readable media containing instructions or containing design data, such as Hardware Description Language (HDL), which defines structures, circuits, apparatuses, processors and/or system features described herein. Such embodiments may also be referred to as program products.
Distributed Convolution Operations
In the illustrated embodiment, for example, a convolution operation 500 is performed on an image 502 using a filter 504, resulting in a convolution result 506. In some embodiments, the image 502, filter 504, and result 506 of a convolution operation 500 may be represented as matrices. For example, an image may be represented by a matrix whose elements correspond to each pixel of the image. For example, in some embodiments, a black and white image may be represented by a two-dimensional (2D) matrix with dimensions corresponding to the height (H) and width (W) of the image. As another example, a color image (e.g., image 502) may be represented by a three-dimensional (3D) matrix with dimensions corresponding to the number of channels (C), height (H), and width (W) of the image. The channels of the image, for example, may be used to define the color of each pixel in the image. For example, in some embodiments, a color image may include three RGB channels to define the color of each pixel (e.g., a red channel, a green channel, and a blue channel). Similarly, a filter (e.g., filter 504) for a color image may be represented by a 3D matrix with dimensions corresponding to the number of channels (C), filter height (R), and filter width (S). In these embodiments, a convolution operation 500 may be performed by moving the filter 504 throughout the image 502 and computing the dot product between the filter 504 and the various portions of the image 502. For example, in some embodiments, the filter 504 may be moved along the height and width of the image 502 using a certain stride or interval, the dot product may be computed at each location, and the result may be stored in the corresponding location of a result matrix 506. The result matrix 506, for example, may be represented by a 3D matrix with dimensions corresponding to the number of channels (C), output height (P), and output width (Q).
Moreover, in some embodiments, when convolution operations need to be performed on multiple images and using multiple filters, the images may be processed together to increase processing efficiency. Thus, in some embodiments, a collection of images may be represented using a four-dimensional (4D) matrix, with dimensions corresponding to the number of channels (C), image height (H), image width (W), and number of images (N). Similarly, a collection of filters may be represented using a 4D matrix, with dimensions corresponding to the number of filters (K), number of channels (C), filter height (R), and filter width (S). Accordingly, multiple convolution operations may then be performed using the 4D matrices that are used to represent the images and filters. The collective results of the convolution operations may be represented using a 4D matrix, with dimensions corresponding to number of filters (K), output height (P), output width (Q), and number of images (N).
forward propagation: IFM[CHW×N]*FILTER[CRS×K]=OFM[KPQ×N]
backward propagation: IFM[KPQ×N]*FILTER[KRS×C]=OFM[CHW×N]
weight update: IFM1[CHW×N]*IFM2[KPQ×N]=OFM[CRS×K]
In some embodiments, the example operations illustrated by
In some embodiments, forward propagation can be implemented for convolution operations. For example, the illustrated forward propagation operation is implemented by performing convolutions using an input feature matrix (IFM) 602 and a filter 604 to generate an output feature matrix (OFM) 606. In some embodiments, IFM 602 may correspond to a collection of images represented using a four-dimensional (4D) matrix, with dimensions corresponding to the number of channels (C), image height (H), image width (W), and number of images (N). In some embodiments, however, IFM 602 may need to be stored in 2D memory (e.g., when using matrix processing architectures that store matrix operands using 2D memory components). Accordingly, because IFM 602 is a 4D matrix, its dimensions may need to be rearranged in order to store it in 2D memory. For example, in some embodiments, IFM 602 could be stored in 2D memory by treating the channels (C), height (H), and width (W) as one dimension, and treating the number of images (N) as another dimension (e.g., arranging the dimensions of IFM 602 as CHW×N).
Similarly, filter 604 may correspond to a collection of filters represented using a four-dimensional (4D) matrix, with dimensions corresponding to the number of filters (K), number of channels (C), filter height (R), and filter width (S). Moreover, because filter 604 is a 4D matrix, its dimensions may need to be rearranged in order to store it in 2D memory. For example, in some embodiments, filter 604 could be stored in 2D memory by treating the channels (C), filter height (R), and filter width (S) as one dimension, and treating the number of filters (K) as another dimension (e.g., arranging the dimensions of filter 604 as CRS×K).
OFM 606 may correspond to the result of the convolution operation, and may be represented using a four-dimensional (4D) matrix with dimensions corresponding to the number of filters (K), output height (P), output width (Q), and number of images (N). Moreover, because OFM 606 is a 4D matrix, its dimensions may need to be rearranged in order to store it in 2D memory. For example, in some embodiments, OFM 606 could be stored in 2D memory by treating the number of filters (K), output height (P), and output width (Q) as one dimension, and treating the number of images (N) as another dimension (e.g., arranging the dimensions of OFM 606 as KPQ×N).
In some embodiments, the operands of a forward propagation operation may be partitioned into a number of partitions corresponding to the number of available processing resources. For example, if there are P processing resources available to perform the operation, the input feature matrix (IFM) 602 and filter 604 may be partitioned into P partitions. Moreover, in some embodiments, the P partitions may be across any of the various dimensions of the matrix operands, such as the channels (C), images (N), and/or filters (K). Each partition may then be distributed to a particular processing resource, as described throughout this disclosure.
In the illustrated example, the matrix operands are partitioned for distribution across four processing resources. Accordingly, IFM 602 and filter 604 are partitioned into four partitions p1-p4 (designated by label 608) corresponding to the four processing resources. Moreover, in some embodiments, partitions p1-p4 of filter 604 may be further partitioned into virtual partitions v1-v4 (designated by label 609). Virtual partitions v1-v4, for example, may simply be further sub-partitions of partitions p1-p4, which are respectively used in different stages of a forward propagation operation. Moreover, the partitions of IFM 602 and filter 604 are across the channels (C) and filters (K) dimensions. Each partition p1-p4 may then be distributed to a particular matrix processing resource. For example, partition p1 of IFM 602 and filter 604 may be distributed to a first matrix processing resource, partition p2 of IFM 602 and filter 604 may be distributed to a second matrix processing resource, and so forth.
Moreover, in some embodiments, matrix operands may be partitioned hierarchically based on the hierarchical arrangement of processing resources. For example, as described above regarding the matrix processing architecture of
At the multi-chip level, the matrix operands may initially be partitioned based on the number of available matrix processing chips (e.g., matrix processing chips 220 of
At the multi-HBM level, the partition and partial matrix operations distributed to a particular matrix processing chip may then be similarly partitioned and distributed across various “logical processing nodes” of that matrix processing chip. A “logical processing node” of a matrix processing chip may include, for example, a group of matrix processing clusters and an associated high-bandwidth memory (HBM) module. For example, matrix processing chip 220 of
At the multi-cluster level, the partition and partial matrix operations distributed to a particular logical processing node may then be similarly partitioned and distributed across the matrix processing clusters of that logical processing node (e.g., a group of three matrix processing clusters 230 in
The forward propagation operation may then be performed as described in connection with
The partial results of OFM 606 are determined using the following partial calculations:
partial result ΔK0PQ×N (partition p1 of OFM 606)=convolution of IFM 602 and virtual partition v1 of filter 604; 1st partial calculation:
partial result ΔK1PQ×N (partition p2 of OFM 606)=convolution of IFM 602 and virtual partition v2 of filter 604; 2nd partial calculation:
partial result ΔK2PQ×N (partition p3 of OFM 606)=convolution of IFM 602 and virtual partition v3 of filter 604; and 3rd partial calculation:
partial result ΔK3PQ×N (partition p4 of OFM 606)=convolution of IFM 602 and virtual partition v4 of filter 604. 4th partial calculation:
In the illustrated example, the 1st partial calculation (e.g., for the partial result corresponding to partition p1 of OFM 606) is broken down as the sum of four sub-convolutions s1-s4 (where the asterisks (*) represent a convolution operation):
s
1
: ΔC
0
HW×N*ΔC
0
RS×ΔK
0
s
2
: ΔC
1
HW×N*ΔC
1
RS×ΔK
0
s
3
: ΔC
2
HW×N*ΔC
2
RS×ΔK
0
s
4
: ΔC
3
HW×N*ΔC
3
RS×ΔK
0
These four sub-convolutions s1-s4 are then performed in four stages, as described further in connection with
For example, in the first stage, sub-convolution s4 of the 1st partial calculation is performed by a first processing resource. Moreover, during this stage, the other processing resources may be performing sub-convolutions associated with the other partial calculations identified above. In the second stage, sub-convolution s3 of the 1st partial calculation is performed by a second processing resource, and while that sub-convolution is being performed, the result of sub-convolution s4 is transmitted from the first processing resource to the second processing resource. When the second processing resource completes sub-convolution s3, it calculates the sum of s4 and s3. Moreover, during this stage, the other processing resources may be performing similar operations associated with the other partial calculations identified above. In the third stage, sub-convolution s2 of the 1st partial calculation is performed by a third processing resource, and while that sub-convolution is being performed, the sum of s4 and s3 is transmitted from the second processing resource to the third processing resource. When the third processing resource completes sub-convolution 52, it calculates the sum of s4, s3, and s2. Moreover, during this stage, the other processing resources may be performing similar operations associated with the other partial calculations identified above. In the fourth stage, sub-convolution s1 of the 1st partial calculation is performed by a fourth processing resource, and while that sub-convolution is being performed, the sum of s4, s3, and s2 is transmitted from the third processing resource to the fourth processing resource. When the fourth processing resource completes sub-convolution s1, it calculates the sum of s4, s3, s2, and s1, which is the final result of the 1st partial calculation (e.g., the partial result corresponding to partition p1 of OFM 606). Similarly, during this stage, the other processing resources may be performing similar operations associated with the other partial calculations identified above, and thus may obtain the partial results corresponding to partitions p2-p4 of OFM 606.
Although not illustrated, partitions p1-p4 of the matrix operands have been distributed across processing resources 610-1 to 610-4 (e.g., p1 to processing resource 610-1, p2 to processing resource 610-2, and so forth), as described above in connection with
In the illustrated example of stage 1, processing resource 610-4 performs sub-convolution s4 (e.g., ΔC3HW×N*ΔC3RS×ΔK0) of the 1st partial calculation. Moreover, although not illustrated, the processing resources may be performing similar stage 1 operations associated with the other partial calculations identified above in connection with
In the illustrated example of stage 2, processing resource 610-3 performs sub-convolution s3 (e.g., ΔC2HW×N*ΔC2RS×ΔK0) of the 1st partial calculation. Moreover, while processing resource 610-3 is performing sub-convolution s3, processing resource 610-4 transmits the result of sub-convolution s4 (which was computed during the previous stage) to processing resource 610-3 (denoted by label 615a). Moreover, although not illustrated, the processing resources may be performing similar stage 2 operations associated with the other partial calculations identified above in connection with
In the illustrated example of stage 3, processing resource 610-2 performs sub-convolution s2 (e.g., ΔC1HW×N*ΔC1RS×ΔK0) of the 1st partial calculation. Moreover, while processing resource 610-2 is performing sub-convolution s2, processing resource 610-3 transmits the sum of sub-convolutions s4 and s3 (or s4+s3) to processing resource 610-2 (denoted by label 615b). Moreover, although not illustrated, the processing resources may be performing similar stage 3 operations associated with the other partial calculations identified above in connection with
In the illustrated example of stage 4, processing resource 610-1 performs sub-convolution s1 (e.g., ΔC0HW×N*ΔC0RS×ΔK0) of the 1st partial calculation. Moreover, while processing resource 610-1 is performing sub-convolution s1, processing resource 610-2 transmits the sum of sub-convolutions s4, s3, and s2 (or s4+s3+s2) to processing resource 610-1 (denoted by label 615c). Moreover, although not illustrated, the processing resources may be performing similar stage 4 operations associated with the other partial calculations identified above in connection with
After all stages are complete, processing resource 610-1 contains the partial result of the 1st partial calculation from
Accordingly, when all stages have been completed, each processing resource 610 contains one of the partial results of OFM 606 from
In some embodiments, the operands of a weight update operation may be partitioned into a number of partitions corresponding to the number of available processing resources. For example, if there are P processing resources available to perform the weight update operation, IFM1701 and IFM2702 may be partitioned into P partitions. Moreover, in some embodiments, the P partitions may be across any of the various dimensions of the matrix operands, such as the channels (C), images (N), and/or filters (K) dimensions. Each partition may then be distributed to a particular processing resource, as described throughout this disclosure.
In the illustrated example, the matrix operands are partitioned for distribution across four processing resources. Accordingly, IFM1701 and IFM2702 are partitioned into four partitions p1-p4 (designated by label 708) corresponding to the four processing resources. Moreover, the partitions of IFM1701 and IFM2702 are across the channels (C) and filters (K) dimensions. Each partition p1-p4 may then be distributed to a particular matrix processing resource. For example, partition p1 of IFM1701 and IFM2702 may be distributed to a first matrix processing resource, partition p2 of IFM1701 and IFM2702 may be distributed to a second matrix processing resource, and so forth.
Moreover, each processing resource then performs various stages of partial matrix operations to compute various partial results of the output feature matrix (OFM) 703. The partial results from these partial matrix operations are stored in partitions p1-p4 (and virtual partitions v1-v4) of OFM 703, as illustrated in
Moreover, in some embodiments, matrix operands may be partitioned and distributed hierarchically based on the hierarchical arrangement of processing resources, as described above in connection with
The weight update operation may then be performed as described in connection with
The result of the weight update operation is determined using the following partial weight update operations (where an asterisk (*) represents a partial weight update operation):
ΔC0HW×N(p1 of IFM1 701)*KPQ×N(IFM2 702)=ΔC0RS×K(p1 of OFM 703); 1st partial weight update operation:
ΔC1HW×N(p2 of IFM1 701)*KPQ×N(IFM2 702)=ΔC1RS×K(p2 of OFM 703); 2nd partial weight update operation:
ΔC2HW×N(p3 of IFM1 701)*KPQ×N(IFM2 702)=ΔC2RS×K(p3 of OFM 703); and 3rd partial weight update operation:
ΔC3HW×N(p4 of IFM1 701)*KPQ×N(IFM2 702)=ΔC3RS×K(p4 of OFM 703). 4th partial weight update operation:
The entire weight update operation (including all associated partial weight update operations and partial matrix calculations) is then performed in four stages, as described further in connection with
Accordingly, in stage 1 of the weight update operation, each processing resource performs a partial matrix calculation using its respective partitions of IFM1701 and IFM2702, as illustrated in
The first processing resource performs the following partial matrix calculation using the matrix operands on partition p1 of IFM1701 and IFM2702: ΔC0HW×N*ΔK0PQ×N=ΔC0RS×ΔK0. The result is stored in the corresponding location of OFM 703 (e.g., on partition p1,v1 of OFM 703).
The second processing resource performs the following partial matrix calculation using the matrix operands on partition p2 of IFM1701 and IFM2702: ΔC1HW×N*ΔK1PQ×N=ΔC1RS×ΔK1. The result is stored in the corresponding location of OFM 703 (e.g., on partition p2,v2 of OFM 703).
The third processing resource performs the following partial matrix calculation using the matrix operands on partition p3 of IFM1701 and IFM2702: ΔC2HW×N*ΔK2PQ×N=ΔC2RS×ΔK2. The result is stored in the corresponding location of OFM 703 (e.g., on partition p3,v3 of OFM 703).
Finally, the fourth processing resource performs the following partial matrix calculation using the matrix operands on partition p4 of IFM1701 and IFM2702: ΔC3HW×N*ΔK3PQ×N=ΔC3RS×ΔK3. The result is stored in the corresponding location of OFM 703 (e.g., on partition p4,v4 of OFM 703).
Moreover, while the partial matrix calculations are being performed by the processing resources in stage 1, each processing resource simultaneously sends and receives partial matrix operands to and from its neighbor processing resources. For example, in some embodiments, the processing resources may be configured in a single-cyclical arrangement (e.g., with unidirectional communication between neighboring processing resources) or a dual-cyclical arrangement (e.g., with bi-directional communication between neighboring processing resources). In a single-cyclical configuration, each processing resource either sends to, or receives from, each neighbor processing resource. In a dual-cyclical configuration, each processing resource may send and receive to and from each neighbor processing resource. Accordingly, for some operations, a dual-cyclical configuration may reduce the latency for communicating matrix operands, thus avoiding any idle processing time.
Accordingly, using a cyclical configuration (e.g., single-cyclical, dual-cyclical, or any other cyclical-based configuration), the partitions p1-p4 of IFM2702 are shifted across the processing resources during each stage of the weight update operation. In the illustrated examples of
In this manner, during each stage of the weight update operation, partial matrix operands (e.g., partitions p1-p4 of IFM2702) are shifted across the processing resources to and from their respective neighboring processing resource(s). For example, in
Each processing resource may then use the partial matrix operands received from neighboring processing resource(s) in subsequent stages, as described below in connection with
The first processing resource performs the following partial matrix calculation using the matrix operands on partition p1 of IFM1701 and IFM2702: ΔC0HW×N*ΔK1PQ×N=ΔC0RS×ΔK1. The result is stored in the corresponding location of OFM 703 (e.g., on partition p1,v2 of OFM 703).
The second processing resource performs the following partial matrix calculation using the matrix operands on partition p2 of IFM1701 and IFM2702: ΔC1HW×N*ΔK2PQ×N=ΔC1RS×ΔK2. The result is stored in the corresponding location of OFM 703 (e.g., on partition p2,v3 of OFM 703).
The third processing resource performs the following partial matrix calculation using the matrix operands on partition p3 of IFM1701 and IFM2702: ΔC2HW×N*ΔK3PQ×N=ΔC2RS×ΔK3. The result is stored in the corresponding location of OFM 703 (e.g., on partition p3,v4 of OFM 703).
Finally, the fourth processing resource performs the following partial matrix calculation using the matrix operands on partition p4 of IFM1701 and IFM2702: ΔC3HW×N*ΔK0PQ×N=ΔC3RS×ΔK0. The result is stored in the corresponding location of OFM 703 (e.g., on partition p4,v1 of OFM 703).
Moreover, while the processing resources are performing the partial matrix calculations in stage 2, each processing resource simultaneously sends and receives partitions of IFM2702 to and from its neighbor processing resources. Accordingly, during stage 2 of the weight update operation, the partitions of IFM2702 are shifted across the processing resources, in a similar manner as described for stage 1 of
Each processing resource may then use the partial matrix operands received from neighboring processing resource(s) in subsequent stages, as described below in connection with
The first processing resource performs the following partial matrix calculation using the matrix operands on partition p1 of IFM1701 and IFM2702: ΔC0HW×N*ΔK2PQ×N=ΔC0RS×ΔK2. The result is stored in the corresponding location of OFM 703 (e.g., on partition p1,v3 of OFM 703).
The second processing resource performs the following partial matrix calculation using the matrix operands on partition p2 of IFM1701 and IFM2702: ΔC1HW×N*ΔK3PQ×N=ΔC1RS×ΔK3. The result is stored in the corresponding location of OFM 703 (e.g., on partition p2,v4 of OFM 703).
The third processing resource performs the following partial matrix calculation using the matrix operands on partition p3 of IFM1701 and IFM2702: ΔC2HW×N*ΔK0PQ×N=ΔC2RS×ΔK0. The result is stored in the corresponding location of OFM 703 (e.g., on partition p3,v1 of OFM 703).
Finally, the fourth processing resource performs the following partial matrix calculation using the matrix operands on partition p4 of IFM1701 and IFM2702: ΔC3HW×N*ΔK1PQ×N=ΔC3RS×ΔK1. The result is stored in the corresponding location of OFM 703 (e.g., on partition p4,v2 of OFM 703).
Moreover, while the processing resources are performing the partial matrix calculations in stage 3, each processing resource simultaneously sends and receives partitions of IFM2702 to and from its neighbor processing resources. Accordingly, during stage 3 of the weight update operation, the partitions of IFM2702 are shifted across the processing resources, in a similar manner as described for stages 1 and 2 of
Each processing resource may then use the partial matrix operands received from neighboring processing resource(s) in subsequent stages, as described below in connection with
The first processing resource performs the following partial matrix calculation using the matrix operands on partition p1 of IFM1701 and IFM2702: ΔC0HW×N*ΔK3PQ×N=ΔC0RS×ΔK3. The result is stored in the corresponding location of OFM 703 (e.g., on partition p1,v4 of OFM 703).
The second processing resource performs the following partial matrix calculation using the matrix operands on partition p2 of IFM1701 and IFM2702: ΔC0HW×N*ΔK0PQ×N=ΔC1RS×ΔK0. The result is stored in the corresponding location of OFM 703 (e.g., on partition p2,v1 of OFM 703).
The third processing resource performs the following partial matrix calculation using the matrix operands on partition p3 of IFM1701 and IFM2702: ΔC2HW×N*ΔK1PQ×N=ΔC2RS×ΔK1. The result is stored in the corresponding location of OFM 703 (e.g., on partition p3,v2 of OFM 703).
Finally, the fourth processing resource performs the following partial matrix calculation using the matrix operands on partition p4 of IFM1701 and IFM2702: ΔC3HW×N*ΔK2PQ×N=ΔC3RS×ΔK2. The result is stored in the corresponding location of OFM 703 (e.g., on partition p4,v3 of OFM 703).
After the partial matrix calculations of stage 4 are complete, all partial results of the weight update operation have been calculated and are stored in OFM 703. Accordingly, the weight update operation is complete.
The flowchart may begin at block 802 by receiving a command to perform a matrix operation. The matrix operation, for example, may comprise an operation on a plurality of input matrices (e.g., matrix operands). Moreover, the matrix operation may be associated with one or more convolution operations.
Convolution operations, for example, may be used for image processing. For example, a convolution operation may involve applying a filter to an image. The filter may be used to modify the image (e.g., applying visual effects), detect certain features in the image (e.g., facial recognition or identification of other objects), and so forth. Moreover, in some embodiments, the image and filter of a convolution operation may be represented using matrices. For example, a color image may be represented by a three-dimensional (3D) matrix with dimensions corresponding to the number of channels (C), height (H), and width (W) of the image. Similarly, a filter may be represented by a 3D matrix with dimensions corresponding to the number of channels (C), height (R), and width (S) of the filter. In these embodiments, a convolution operation may be performed by moving the filter throughout the image and computing the dot product between the filter and the various portions of the image. Moreover, in some embodiments, when convolution operations need to be performed on multiple images and using multiple filters, the images may be processed together to increase processing efficiency. Thus, in some embodiments, a collection of images may be represented using a four-dimensional (4D) matrix, with dimensions corresponding to the number of channels (C), image height (H), image width (W), and number of images (N). Similarly, a collection of filters may be represented using a 4D matrix, with dimensions corresponding to the number of filters (K), number of channels (C), filter height (R), and filter width (S). Accordingly, a convolution operation may then be performed using the 4D matrices that are used to represent the images and filters.
In some embodiments, convolution operations may be used to implement computer vision artificial intelligence and machine learning capabilities in an artificial neural network. For example, in some embodiments, the matrix operation of block 802 may be associated with operations in an artificial neural network, such as forward propagation, backward propagation, and/or weight update operations.
The flowchart may then proceed to block 804 to partition the input matrices into a plurality of partitions based on the number of available processing elements. In some embodiments, the input matrices may be partitioned based on the hierarchical arrangement of processing resources, as described further in connection with block 806. In some embodiments, the partitions may be across any of the various dimensions of the input matrices, such as the channels (C), images (N), and/or filters (K) dimensions.
The flowchart may then proceed to block 806 to distribute the partitions to the available processing elements. For example, in some embodiments, each partition may be distributed to a particular processing element. Moreover, in some embodiments, the processing elements may be configured in a hierarchical arrangement with a plurality of processing levels, and the matrix operation may be distributed across the hierarchy of processing levels. For example, in some embodiments, the processing elements may include multiple matrix processing chips (e.g., matrix processing chips 220 of
Moreover, in some cases, a partial matrix operation distributed to a particular processing element may be distributed across the height (P) and width (Q) of the corresponding output of the partial matrix operation. For example, a partial matrix operation distributed to a logical processing node may be distributed to the associated matrix processing clusters across the height (P) of the output matrix, and the partial matrix operation distributed to each matrix processing cluster may then be distributed to the matrix processing units (MPUs) across the width (Q) of the output matrix.
The flowchart may then proceed to block 808 to perform partial matrix operations using the processing elements. For example, each processing element may perform a partial matrix operation based on the matrix data distributed to that processing element.
The flowchart may then proceed to block 810 to transmit partial matrix data between processing elements while performing the partial matrix operations. For example, in some embodiments, the processing elements may be configured in a cyclical arrangement such that each processing element is communicatively coupled to multiple neighbor processing elements. Moreover, the partial matrix operations may be performed in a plurality of stages, and each processing element may transmit partial matrix data to its neighbor processing elements while performing a particular stage of the partial matrix operations. For example, in some embodiments, each processing element may transmit partial matrix data to one of its neighbor processing elements (e.g., using a single-cyclical approach) or to both of its neighbor processing elements (e.g., using a dual-cyclical approach) during each stage of partial matrix operations. For example, a first processing element may use or calculate partial matrix data in a particular stage of the partial matrix operations, the first processing element may transmit the partial matrix data to a second processing element, and the second processing element may then use the partial matrix data in a subsequent stage of the partial matrix operations. For some matrix operations, the partial matrix data may include a partial input matrix, while for other matrix operations, the partial matrix data may include a partial result matrix.
The flowchart may then proceed to block 812 to determine a result of the matrix operation. For example, the result of the matrix operation may be determined based on the partial results collectively computed by the processing elements.
At this point, the flowchart may be complete. In some embodiments, however, the flowchart may restart and/or certain blocks may be repeated. For example, in some embodiments, the flowchart may restart at block 802 to continue receiving and processing commands to perform matrix operations.
The flowcharts and block diagrams in the FIGURES illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various aspects of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order or alternative orders, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.
The foregoing disclosure outlines features of several embodiments so that those skilled in the art may better understand various aspects of the present disclosure. Those skilled in the art should appreciate that they may readily use the present disclosure as a basis for designing or modifying other processes and structures for carrying out the same purposes and/or achieving the same advantages of the embodiments introduced herein. Those skilled in the art should also realize that such equivalent constructions do not depart from the spirit and scope of the present disclosure, and that they may make various changes, substitutions, and alterations herein without departing from the spirit and scope of the present disclosure.
All or part of any hardware element disclosed herein may readily be provided in a system-on-a-chip (SoC), including a central processing unit (CPU) package. An SoC represents an integrated circuit (IC) that integrates components of a computer or other electronic system into a single chip. The SoC may contain digital, analog, mixed-signal, and radio frequency functions, all of which may be provided on a single chip substrate. Other embodiments may include a multi-chip-module (MCM), with a plurality of chips located within a single electronic package and configured to interact closely with each other through the electronic package. In various other embodiments, the computing functionalities disclosed herein may be implemented in one or more silicon cores in Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs), and other semiconductor chips.
As used throughout this specification, the term “processor” or “microprocessor” should be understood to include not only a traditional microprocessor (such as Intel's® industry-leading x86 and x64 architectures), but also matrix processors, graphics processors, and any ASIC, FPGA, microcontroller, digital signal processor (DSP), programmable logic device, programmable logic array (PLA), microcode, instruction set, emulated or virtual machine processor, or any similar “Turing-complete” device, combination of devices, or logic elements (hardware or software) that permit the execution of instructions.
Note also that in certain embodiments, some of the components may be omitted or consolidated. In a general sense, the arrangements depicted in the figures should be understood as logical divisions, whereas a physical architecture may include various permutations, combinations, and/or hybrids of these elements. It is imperative to note that countless possible design configurations can be used to achieve the operational objectives outlined herein. Accordingly, the associated infrastructure has a myriad of substitute arrangements, design choices, device possibilities, hardware configurations, software implementations, and equipment options.
In a general sense, any suitably-configured processor can execute instructions associated with data or microcode to achieve the operations detailed herein. Any processor disclosed herein could transform an element or an article (for example, data) from one state or thing to another state or thing. In another example, some activities outlined herein may be implemented with fixed logic or programmable logic (for example, software and/or computer instructions executed by a processor) and the elements identified herein could be some type of a programmable processor, programmable digital logic (for example, a field programmable gate array (FPGA), an erasable programmable read only memory (EPROM), an electrically erasable programmable read only memory (EEPROM)), an ASIC that includes digital logic, software, code, electronic instructions, flash memory, optical disks, CD-ROMs, DVD ROMs, magnetic or optical cards, other types of machine-readable mediums suitable for storing electronic instructions, or any suitable combination thereof.
In operation, a storage may store information in any suitable type of tangible, non-transitory storage medium (for example, random access memory (RAM), read only memory (ROM), field programmable gate array (FPGA), erasable programmable read only memory (EPROM), electrically erasable programmable ROM (EEPROM), or microcode), software, hardware (for example, processor instructions or microcode), or in any other suitable component, device, element, or object where appropriate and based on particular needs. Furthermore, the information being tracked, sent, received, or stored in a processor could be provided in any database, register, table, cache, queue, control list, or storage structure, based on particular needs and implementations, all of which could be referenced in any suitable timeframe. Any of the memory or storage elements disclosed herein should be construed as being encompassed within the broad terms ‘memory’ and ‘storage,’ as appropriate. A non-transitory storage medium herein is expressly intended to include any non-transitory special-purpose or programmable hardware configured to provide the disclosed operations, or to cause a processor to perform the disclosed operations. A non-transitory storage medium also expressly includes a processor having stored thereon hardware-coded instructions, and optionally microcode instructions or sequences encoded in hardware, firmware, or software.
Computer program logic implementing all or part of the functionality described herein is embodied in various forms, including, but in no way limited to, hardware description language, a source code form, a computer executable form, machine instructions or microcode, programmable hardware, and various intermediate forms (for example, forms generated by an HDL processor, assembler, compiler, linker, or locator). In an example, source code includes a series of computer program instructions implemented in various programming languages, such as an object code, an assembly language, or a high-level language such as OpenCL, FORTRAN, C, C++, JAVA, or HTML for use with various operating systems or operating environments, or in hardware description languages such as Spice, Verilog, and VHDL. The source code may define and use various data structures and communication messages. The source code may be in a computer executable form (e.g., via an interpreter), or the source code may be converted (e.g., via a translator, assembler, or compiler) into a computer executable form, or converted to an intermediate form such as byte code. Where appropriate, any of the foregoing may be used to build or describe appropriate discrete or integrated circuits, whether sequential, combinatorial, state machines, or otherwise.
In one example, any number of electrical circuits of the FIGURES may be implemented on a board of an associated electronic device. The board can be a general circuit board that can hold various components of the internal electronic system of the electronic device and, further, provide connectors for other peripherals. More specifically, the board can provide the electrical connections by which the other components of the system can communicate electrically. Any suitable processor and memory can be suitably coupled to the board based on particular configuration needs, processing demands, and computing designs. Other components such as external storage, additional sensors, controllers for audio/video display, and peripheral devices may be attached to the board as plug-in cards, via cables, or integrated into the board itself. In another example, the electrical circuits of the FIGURES may be implemented as stand-alone modules (e.g., a device with associated components and circuitry configured to perform a specific application or function) or implemented as plug-in modules into application specific hardware of electronic devices.
Note that with the numerous examples provided herein, interaction may be described in terms of two, three, four, or more electrical components. However, this has been done for purposes of clarity and example only. It should be appreciated that the system can be consolidated or reconfigured in any suitable manner. Along similar design alternatives, any of the illustrated components, modules, and elements of the FIGURES may be combined in various possible configurations, all of which are within the broad scope of this specification. In certain cases, it may be easier to describe one or more of the functionalities of a given set of flows by only referencing a limited number of electrical elements. It should be appreciated that the electrical circuits of the FIGURES and its teachings are readily scalable and can accommodate a large number of components, as well as more complicated/sophisticated arrangements and configurations. Accordingly, the examples provided should not limit the scope or inhibit the broad teachings of the electrical circuits as potentially applied to a myriad of other architectures.
Numerous other changes, substitutions, variations, alterations, and modifications may be ascertained to one skilled in the art and it is intended that the present disclosure encompass all such changes, substitutions, variations, alterations, and modifications as falling within the scope of the appended claims.
The following examples pertain to embodiments described throughout this disclosure.
One or more embodiments may include an apparatus, comprising: a plurality of memory elements to store matrix data; and a plurality of processing elements to perform a matrix operation using a plurality of input matrices, wherein the matrix operation is associated with one or more convolution operations; wherein the plurality of processing elements is configured to: partition the plurality of input matrices into a plurality of input partitions, wherein the plurality of input matrices is partitioned based on a number of available processing elements; distribute the plurality of input partitions among the plurality of processing elements, wherein each input partition is distributed to a particular processing element of the plurality of processing elements; perform a plurality of partial matrix operations using the plurality of processing elements; transmit partial matrix data between the plurality of processing elements while performing the plurality of partial matrix operations; and determine a result of the matrix operation based on the plurality of partial matrix operations.
In one example embodiment of an apparatus: the plurality of processing elements is configured in a hierarchical arrangement comprising a plurality of processing levels; and the plurality of processing elements is further configured to distribute the matrix operation across the plurality of processing levels.
In one example embodiment of an apparatus, the plurality of input matrices comprises matrix data associated with one or more images and one or more filters, wherein the one or more images are associated with one or more channels.
In one example embodiment of an apparatus, the plurality of processing elements is further configured to partition the plurality of input matrices based on one or more of: a number of channels associated with the one or more images; a number of filters; and a number of images.
In one example embodiment of an apparatus, the plurality of processing elements is further configured to distribute the plurality of partial matrix operations among the plurality of processing elements based on a height and a width of the result of the matrix operation.
In one example embodiment of an apparatus: the plurality of processing elements is configured in a cyclic arrangement such that each processing element is communicatively coupled to a plurality of neighbor processing elements; and the plurality of processing elements is further configured to: perform the plurality of partial matrix operations in a plurality of stages; and transmit a portion of the partial matrix data from each processing element to one or more of the neighbor processing elements while performing a particular stage of the partial matrix operations.
In one example embodiment of an apparatus, the partial matrix data comprises a partial input matrix, wherein the partial input matrix is to be used by a first processing element in a particular stage of the partial matrix operations, and wherein the partial input matrix is to be used by a second processing element in a subsequent stage of the partial matrix operations.
In one example embodiment of an apparatus, the partial matrix data comprises a partial result matrix determined by a first processing element in a particular stage of the partial matrix operations, and wherein the partial result matrix is to be used by a second processing element in a subsequent stage of the partial matrix operations.
In one example embodiment of an apparatus, the plurality of processing elements comprises: a plurality of matrix processing chips; a plurality of logical processing nodes associated with each matrix processing chip; and a memory element and a plurality of matrix processing clusters associated with each logical processing node.
In one example embodiment of an apparatus, the matrix operation is associated with a weight update operation in a neural network.
In one example embodiment of an apparatus, the matrix operation is associated with a forward propagation operation in a neural network.
In one example embodiment of an apparatus, the matrix operation is associated with a backward propagation operation in a neural network.
One or more embodiments may include a method, comprising: performing a matrix operation using a plurality of input matrices, wherein the matrix operation is associated with one or more convolution operations, and wherein performing the matrix operation comprises: partitioning the plurality of input matrices into a plurality of input partitions, wherein the plurality of input matrices is partitioned based on a number of available processing elements; distributing the plurality of input partitions among a plurality of processing elements, wherein each input partition is distributed to a particular processing element of the plurality of processing elements; performing a plurality of partial matrix operations using the plurality of processing elements; transmitting partial matrix data between the plurality of processing elements while performing the plurality of partial matrix operations; and determining a result of the matrix operation based on the plurality of partial matrix operations.
In one example embodiment of a method: the plurality of processing elements is configured in a hierarchical arrangement comprising a plurality of processing levels; and the matrix operation is distributed across the plurality of processing levels.
In one example embodiment of a method, the plurality of processing elements comprises: a plurality of matrix processing chips; a plurality of logical processing nodes associated with each matrix processing chip; and a memory element and a plurality of matrix processing clusters associated with each logical processing node.
In one example embodiment of a method, the plurality of input matrices comprises matrix data associated with one or more images and one or more filters, wherein the one or more images are associated with one or more channels.
In one example embodiment of a method, the plurality of input matrices is further partitioned based on one or more of: a number of channels associated with the one or more images; a number of filters; and a number of images.
In one example embodiment of a method, the method further comprises distributing the plurality of partial matrix operations to the plurality of processing elements based on a height and a width of the result of the matrix operation.
In one example embodiment of a method, the plurality of processing elements is configured in a cyclic arrangement such that each processing element is communicatively coupled to a plurality of neighbor processing elements.
In one example embodiment of a method, the plurality of partial matrix operations is performed in a plurality of stages, and wherein each processing element transmits a portion of the partial matrix data to one or more of the neighbor processing elements while performing a particular stage of the partial matrix operations.
In one example embodiment of a method, the partial matrix data comprises a partial input matrix, wherein the partial input matrix is used by a first processing element in a particular stage of the partial matrix operations, and wherein the partial input matrix is used by a second processing element in a subsequent stage of the partial matrix operations.
In one example embodiment of a method, the matrix operation is associated with a weight update operation in a neural network.
In one example embodiment of a method, the partial matrix data comprises a partial result matrix determined by a first processing element in a particular stage of the partial matrix operations, and wherein the partial result matrix is used by a second processing element in a subsequent stage of the partial matrix operations.
In one example embodiment of a method, the matrix operation is associated with a forward propagation operation in a neural network.
In one example embodiment of a method, the matrix operation is associated with a backward propagation operation in a neural network.
One or more embodiments may include a system, comprising: a plurality of memory elements to store matrix data; a plurality of processing elements to perform a matrix operation using a plurality of input matrices, wherein the matrix operation is associated with one or more convolution operations, and wherein the plurality of processing elements comprises: a host processor; one or more matrix processing chips; a plurality of matrix processors associated with the one or more matrix processing chips; wherein the plurality of processing elements is configured to: partition the plurality of input matrices into a plurality of input partitions, wherein the plurality of input matrices is partitioned based on a number of available processing elements; distribute the plurality of input partitions among the plurality of processing elements, wherein each input partition is distributed to a particular processing element of the plurality of processing elements; perform a plurality of partial matrix operations using the plurality of processing elements; transmit partial matrix data between the plurality of processing elements while performing the plurality of partial matrix operations; and determine a result of the matrix operation based on the plurality of partial matrix operations.
In one example embodiment of a system, the system further comprises a communication interface to communicate with one or more remote matrix processing chips over a communication network.
One or more embodiments may include at least one machine accessible storage medium having instructions stored thereon, the instructions, when executed on a machine, cause the machine to: perform a matrix operation using a plurality of input matrices, wherein the matrix operation is associated with one or more convolution operations, and wherein the instructions that cause the machine to perform the matrix operation further cause the machine to: partition the plurality of input matrices into a plurality of input partitions, wherein the plurality of input matrices is partitioned based on a number of available processing elements; distribute the plurality of input partitions among a plurality of processing elements, wherein each input partition is distributed to a particular processing element of the plurality of processing elements; perform a plurality of partial matrix operations using the plurality of processing elements; transmit partial matrix data between the plurality of processing elements while performing the plurality of partial matrix operations; and determine a result of the matrix operation based on the plurality of partial matrix operations.
In one example embodiment of a storage medium, the plurality of input matrices comprises matrix data associated with one or more images and one or more filters, wherein the one or more images are associated with one or more channels.
In one example embodiment of a storage medium, the instructions further cause the machine to partition the plurality of input matrices based on one or more of: a number of channels associated with the one or more images; a number of filters; and a number of images.
In one example embodiment of a storage medium, the instructions further cause the machine to distribute the plurality of partial matrix operations to the plurality of processing elements based on a height and a width of the result of the matrix operation.
In one example embodiment of a storage medium, the plurality of processing elements is configured in a cyclic arrangement such that each processing element is communicatively coupled to a plurality of neighbor processing elements.
In one example embodiment of a storage medium, the instructions further cause the machine to: perform the plurality of partial matrix operations in a plurality of stages; and transmit a portion of the partial matrix data from each processing element to one or more of the neighbor processing elements while performing a particular stage of the partial matrix operations.
This patent arises from a continuation of U.S. patent application Ser. No. 15/395,675, filed Dec. 30, 2016, which is hereby incorporated herein in its entirety.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 15395675 | Dec 2016 | US |
| Child | 17564098 | US |
