Patent Grant
11593295
The disclosure relates generally to electronics, and, more specifically, an embodiment of the disclosure relates to circuitry to control unstructured data flow in a configurable spatial accelerator.
A processor, or set of processors, executes instructions from an instruction set, e.g., the instruction set architecture (ISA). The instruction set is the part of the computer architecture related to programming, and generally includes the native data types, instructions, register architecture, addressing modes, memory architecture, interrupt and exception handling, and external input and output (I/O). It should be noted that the term instruction herein may refer to a macro-instruction, e.g., an instruction that is provided to the processor for execution, or to a micro-instruction, e.g., an instruction that results from a processor's decoder decoding macro-instructions.
The present disclosure is illustrated by way of example and not limitation in the figures of the accompanying drawings, in which like references indicate similar elements and in which:
In the following description, numerous specific details are set forth. However, it is understood that embodiments of the disclosure may be practiced without these specific details. In other instances, well-known circuits, structures and techniques have not been shown in detail in order not to obscure the understanding of this description.
References in the specification to “one embodiment,” “an embodiment,” “an example embodiment,” etc., indicate that the embodiment described may include a particular feature, structure, or characteristic, but every embodiment may not necessarily include the particular feature, structure, or characteristic. Moreover, such phrases are not necessarily referring to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the knowledge of one skilled in the art to affect such feature, structure, or characteristic in connection with other embodiments whether or not explicitly described.
A processor (e.g., having one or more cores) may execute instructions (e.g., a thread of instructions) to operate on data, for example, to perform arithmetic, logic, or other functions. For example, software may request an operation and a hardware processor (e.g., a core or cores thereof) may perform the operation in response to the request. One non-limiting example of an operation is a blend operation to input a plurality of vectors elements and output a vector with a blended plurality of elements. In certain embodiments, multiple operations are accomplished with the execution of a single instruction.
Exascale performance, e.g., as defined by the Department of Energy, may require system-level floating point performance to exceed 10{circumflex over ( )}18 floating point operations per second (exaFLOPs) or more within a given (e.g., 20 MW) power budget. Certain embodiments herein are directed to a spatial array of processing elements (e.g., a configurable spatial accelerator (CSA)) that targets high performance computing (HPC), for example, of a processor. Certain embodiments herein of a spatial array of processing elements (e.g., a CSA) target the direct execution of a dataflow graph to yield a computationally dense yet energy-efficient spatial microarchitecture which far exceeds conventional roadmap architectures. Certain embodiments herein overlay (e.g., high-radix) dataflow operations on a communications network, e.g., in addition to the communications network's routing of data between the processing elements, memory, etc. and/or the communications network performing other communications (e.g., not data processing) operations. Certain embodiments herein are directed to a communications network (e.g., a packet switched network) of a (e.g., coupled to) spatial array of processing elements (e.g., a CSA) to perform certain dataflow operations, e.g., in addition to the communications network routing data between the processing elements, memory, etc. or the communications network performing other communications operations. Certain embodiments herein are directed to network dataflow endpoint circuits that (e.g., each) perform (e.g., a portion or all) a dataflow operation or operations, for example, a pick or switch dataflow operation, e.g., of a dataflow graph. Certain embodiments herein include augmented network endpoints (e.g., network dataflow endpoint circuits) to support the control for (e.g., a plurality of or a subset of) dataflow operation(s), e.g., utilizing the network endpoints to perform a (e.g., dataflow) operation instead of a processing element (e.g., core) or arithmetic-logic unit (e.g. to perform arithmetic and logic operations) performing that (e.g., dataflow) operation. In one embodiment, a network dataflow endpoint circuit is separate from a spatial array (e.g. an interconnect or fabric thereof) and/or processing elements.
Below also includes a description of the architectural philosophy of embodiments of a spatial array of processing elements (e.g., a CSA) and certain features thereof. As with any revolutionary architecture, programmability may be a risk. To mitigate this issue, embodiments of the CSA architecture have been co-designed with a compilation tool chain, which is also discussed below.
Exascale computing goals may require enormous system-level floating point performance (e.g., 1 ExaFLOPs) within an aggressive power budget (e.g., 20 MW). However, simultaneously improving the performance and energy efficiency of program execution with classical von Neumann architectures has become difficult: out-of-order scheduling, simultaneous multi-threading, complex register files, and other structures provide performance, but at high energy cost. Certain embodiments herein achieve performance and energy requirements simultaneously. Exascale computing power-performance targets may demand both high throughput and low energy consumption per operation. Certain embodiments herein provide this by providing for large numbers of low-complexity, energy-efficient processing (e.g., computational) elements which largely eliminate the control overheads of previous processor designs. Guided by this observation, certain embodiments herein include a spatial array of processing elements, for example, a configurable spatial accelerator (CSA), e.g., comprising an array of processing elements (PEs) connected by a set of light-weight, back-pressured (e.g., communication) networks. One example of a CSA tile is depicted in
The derivation of a dataflow graph from a sequential compilation flow allows embodiments of a CSA to support familiar programming models and to directly (e.g., without using a table of work) execute existing high performance computing (HPC) code. CSA processing elements (PEs) may be energy efficient. In
Certain embodiments herein provide for performance increases from parallel execution within a (e.g., dense) spatial array of processing elements (e.g., CSA) where each PE and/or network dataflow endpoint circuit utilized may perform its operations simultaneously, e.g., if input data is available. Efficiency increases may result from the efficiency of each PE and/or network dataflow endpoint circuit, e.g., where each PE's operation (e.g., behavior) is fixed once per configuration (e.g., mapping) step and execution occurs on local data arrival at the PE, e.g., without considering other fabric activity, and/or where each network dataflow endpoint circuit's operation (e.g., behavior) is variable (e.g., not fixed) when configured (e.g., mapped). In certain embodiments, a PE and/or network dataflow endpoint circuit is (e.g., each a single) dataflow operator, for example, a dataflow operator that only operates on input data when both (i) the input data has arrived at the dataflow operator and (ii) there is space available for storing the output data, e.g., otherwise no operation is occurring.
Certain embodiments herein include a spatial array of processing elements as an energy-efficient and high-performance way of accelerating user applications. In one embodiment, applications are mapped in an extremely parallel manner. For example, inner loops may be unrolled multiple times to improve parallelism. This approach may provide high performance, e.g., when the occupancy (e.g., use) of the unrolled code is high. However, if there are less used code paths in the loop body unrolled (for example, an exceptional code path like floating point de-normalized mode) then (e.g., fabric area of) the spatial array of processing elements may be wasted and throughput consequently lost.
One embodiment herein to reduce pressure on (e.g., fabric area of) the spatial array of processing elements (e.g., in the case of underutilized code segments) is time multiplexing. In this mode, a single instance of the less used (e.g., colder) code may be shared among several loop bodies, for example, analogous to a function call in a shared library. In one embodiment, spatial arrays (e.g., of processing elements) support the direct implementation of multiplexed codes. However, e.g., when multiplexing or demultiplexing in a spatial array involves choosing among many and distant targets (e.g., sharers), a direct implementation using dataflow operators (e.g., using the processing elements) may be inefficient in terms of latency, throughput, implementation area, and/or energy. Certain embodiments herein describe hardware mechanisms (e.g., network circuitry) supporting (e.g., high-radix) multiplexing or demultiplexing. Certain embodiments herein (e.g., of network dataflow endpoint circuits) permit the aggregation of many targets (e.g., sharers) with little hardware overhead or performance impact. Certain embodiments herein allow for compiling of (e.g., legacy) sequential codes to parallel architectures in a spatial array.
In one embodiment, a plurality of network dataflow endpoint circuits combine as a single dataflow operator, for example, as discussed in reference to
An embodiment of a “Pick” dataflow operator is to select data (e.g., a token) from a plurality of input channels and provide that data as its (e.g., single) output according to control data. Control data for a Pick may include an input selector value. In one embodiment, the selected input channel is to have its data (e.g., token) removed (e.g., discarded), for example, to complete the performance of that dataflow operation (or its portion of a dataflow operation). In one embodiment, additionally, those non-selected input channels are also to have their data (e.g., token) removed (e.g., discarded), for example, to complete the performance of that dataflow operation (or its portion of a dataflow operation).
An embodiment of a “PickSingleLeg” dataflow operator is to select data (e.g., a token) from a plurality of input channels and provide that data as its (e.g., single) output according to control data, but in certain embodiments, the non-selected input channels are ignored, e.g., those non-selected input channels are not to have their data (e.g., token) removed (e.g., discarded), for example, to complete the performance of that dataflow operation (or its portion of a dataflow operation). Control data for a PickSingleLeg may include an input selector value. In one embodiment, the selected input channel is also to have its data (e.g., token) removed (e.g., discarded), for example, to complete the performance of that dataflow operation (or its portion of a dataflow operation).
An embodiment of a “PickAny” dataflow operator is to select the first available (e.g., to the circuit performing the operation) data (e.g., a token) from a plurality of input channels and provide that data as its (e.g., single) output. In one embodiment, PickSingleLeg is also to output the index (e.g., indicating which of the plurality of input channels) had its data selected. In one embodiment, the selected input channel is to have its data (e.g., token) removed (e.g., discarded), for example, to complete the performance of that dataflow operation (or its portion of a dataflow operation). In certain embodiments, the non-selected input channels (e.g., with or without input data) are ignored, e.g., those non-selected input channels are not to have their data (e.g., token) removed (e.g., discarded), for example, to complete the performance of that dataflow operation (or its portion of a dataflow operation). Control data for a PickAny may include a value corresponding to the PickAny, e.g., without an input selector value.
An embodiment of a “Switch” dataflow operator is to steer (e.g., single) input data (e.g., a token) so as to provide that input data to one or a plurality of (e.g., less than all) outputs according to control data. Control data for a Switch may include an output(s) selector value or values. In one embodiment, the input data (e.g., from an input channel) is to have its data (e.g., token) removed (e.g., discarded), for example, to complete the performance of that dataflow operation (or its portion of a dataflow operation).
An embodiment of a “SwitchAny” dataflow operator is to steer (e.g., single) input data (e.g., a token) so as to provide that input data to one or a plurality of (e.g., less than all) outputs that may receive that data, e.g., according to control data. In one embodiment, SwitchAny may provide the input data to any coupled output channel that has availability (e.g., available storage space) in its ingress buffer, e.g., network ingress buffer in
Certain embodiments herein thus provide paradigm-shifting levels of performance and tremendous improvements in energy efficiency across a broad class of existing single-stream and parallel programs, e.g., all while preserving familiar HPC programming models. Certain embodiments herein may target HPC such that floating point energy efficiency is extremely important. Certain embodiments herein not only deliver compelling improvements in performance and reductions in energy, they also deliver these gains to existing HPC programs written in mainstream HPC languages and for mainstream HPC frameworks. Certain embodiments of the architecture herein (e.g., with compilation in mind) provide several extensions in direct support of the control-dataflow internal representations generated by modern compilers. Certain embodiments herein are direct to a CSA dataflow compiler, e.g., which can accept C, C++, and Fortran programming languages, to target a CSA architecture.
Section 1 below discloses embodiments of CSA architecture. In particular, novel embodiments of integrating memory within the dataflow execution model are disclosed. Section 2 delves into the microarchitectural details of embodiments of a CSA. In one embodiment, the main goal of a CSA is to support compiler produced programs. Section 3 discusses example operations of an Operation Set Architecture (OSA) for CSA. Section 4 below examines embodiments of a CSA compilation tool chain. The advantages of embodiments of a CSA are compared to other architectures in the execution of compiled codes in Section 5. Finally the performance of embodiments of a CSA microarchitecture is discussed in Section 6, further CSA details are discussed in Section 7, and a summary is provided in Section 8.
The goal of certain embodiments of a CSA is to rapidly and efficiently execute programs, e.g., programs produced by compilers. Certain embodiments of the CSA architecture provide programming abstractions that support the needs of compiler technologies and programming paradigms. Embodiments of the CSA execute dataflow graphs, e.g., a program manifestation that closely resembles the compiler's own internal representation (IR) of compiled programs. In this model, a program is represented as a dataflow graph comprised of nodes (e.g., vertices) drawn from a set of architecturally-defined dataflow operators (e.g., that encompass both computation and control operations) and edges which represent the transfer of data between dataflow operators. Execution may proceed by injecting dataflow tokens (e.g., that are or represent data values) into the dataflow graph. Tokens may flow between and be transformed at each node (e.g., vertex), for example, forming a complete computation. A sample dataflow graph and its derivation from high-level source code is shown in
Embodiments of the CSA are configured for dataflow graph execution by providing exactly those dataflow-graph-execution supports required by compilers. In one embodiment, the CSA is an accelerator (e.g., an accelerator in
Turning to embodiments of the CSA, dataflow operators are discussed next.
1.1 Dataflow Operators
The key architectural interface of embodiments of the accelerator (e.g., CSA) is the dataflow operator, e.g., as a direct representation of a node in a dataflow graph. From an operational perspective, dataflow operators behave in a streaming or data-driven fashion. Dataflow operators may execute as soon as their incoming operands become available. CSA dataflow execution may depend (e.g., only) on highly localized status, for example, resulting in a highly scalable architecture with a distributed, asynchronous execution model. Dataflow operators may include arithmetic dataflow operators, for example, one or more of floating point addition and multiplication, integer addition, subtraction, and multiplication, various forms of comparison, logical operators, and shift. However, embodiments of the CSA may also include a rich set of control operators which assist in the management of dataflow tokens in the program graph. Examples of these include a “pick” operator, e.g., which multiplexes two or more logical input channels into a single output channel, and a “switch” operator, e.g., which operates as a channel demultiplexor (e.g., outputting a single channel from two or more logical input channels). These operators may enable a compiler to implement control paradigms such as conditional expressions. Certain embodiments of a CSA may include a limited dataflow operator set (e.g., to relatively small number of operations) to yield dense and energy efficient PE microarchitectures. Certain embodiments may include dataflow operators for complex operations that are common in HPC code. The CSA dataflow operator architecture is highly amenable to deployment-specific extensions. For example, more complex mathematical dataflow operators, e.g., trigonometry functions, may be included in certain embodiments to accelerate certain mathematics-intensive HPC workloads. Similarly, a neural-network tuned extension may include dataflow operators for vectorized, low precision arithmetic.
In one embodiment, one or more of the processing elements in the array of processing elements 301 is to access memory through memory interface 302. In one embodiment, pick node 304 of dataflow graph 300 thus corresponds (e.g., is represented by) to pick operator 304A, switch node 306 of dataflow graph 300 thus corresponds (e.g., is represented by) to switch operator 306A, and multiplier node 308 of dataflow graph 300 thus corresponds (e.g., is represented by) to multiplier operator 308A. Another processing element and/or a flow control path network may provide the control values (e.g., control tokens) to the pick operator 304A and switch operator 306A to perform the operation in
1.2 Latency Insensitive Channels
Communications arcs are the second major component of the dataflow graph. Certain embodiments of a CSA describes these arcs as latency insensitive channels, for example, in-order, back-pressured (e.g., not producing or sending output until there is a place to store the output), point-to-point communications channels. As with dataflow operators, latency insensitive channels are fundamentally asynchronous, giving the freedom to compose many types of networks to implement the channels of a particular graph. Latency insensitive channels may have arbitrarily long latencies and still faithfully implement the CSA architecture. However, in certain embodiments there is strong incentive in terms of performance and energy to make latencies as small as possible. Section 2.2 herein discloses a network microarchitecture in which dataflow graph channels are implemented in a pipelined fashion with no more than one cycle of latency. Embodiments of latency-insensitive channels provide a critical abstraction layer which may be leveraged with the CSA architecture to provide a number of runtime services to the applications programmer. For example, a CSA may leverage latency-insensitive channels in the implementation of the CSA configuration (the loading of a program onto the CSA array).
1.3 Memory
Dataflow architectures generally focus on communication and data manipulation with less attention paid to state. However, enabling real software, especially programs written in legacy sequential languages, requires significant attention to interfacing with memory. Certain embodiments of a CSA use architectural memory operations as their primary interface to (e.g., large) stateful storage. From the perspective of the dataflow graph, memory operations are similar to other dataflow operations, except that they have the side effect of updating a shared store. In particular, memory operations of certain embodiments herein have the same semantics as every other dataflow operator, for example, they “execute” when their operands, e.g., an address, are available and, after some latency, a response is produced. Certain embodiments herein explicitly decouple the operand input and result output such that memory operators are naturally pipelined and have the potential to produce many simultaneous outstanding requests, e.g., making them exceptionally well suited to the latency and bandwidth characteristics of a memory subsystem. Embodiments of a CSA provide basic memory operations such as load, which takes an address channel and populates a response channel with the values corresponding to the addresses, and a store. Embodiments of a CSA may also provide more advanced operations such as in-memory atomics and consistency operators. These operations may have similar semantics to their von Neumann counterparts. Embodiments of a CSA may accelerate existing programs described using sequential languages such as C and Fortran. A consequence of supporting these language models is addressing program memory order, e.g., the serial ordering of memory operations typically prescribed by these languages.
1.4 Runtime Services
A primary architectural considerations of embodiments of the CSA involve the actual execution of user-level programs, but it may also be desirable to provide several support mechanisms which underpin this execution. Chief among these are configuration (in which a dataflow graph is loaded into the CSA), extraction (in which the state of an executing graph is moved to memory), and exceptions (in which mathematical, soft, and other types of errors in the fabric are detected and handled, possibly by an external entity). Section 2. below discusses the properties of a latency-insensitive dataflow architecture of an embodiment of a CSA to yield efficient, largely pipelined implementations of these functions. Conceptually, configuration may load the state of a dataflow graph into the interconnect (and/or communications network (e.g., a network dataflow endpoint circuit thereof)) and processing elements (e.g., fabric), e.g., generally from memory. During this step, all structures in the CSA may be loaded with a new dataflow graph and any dataflow tokens live in that graph, for example, as a consequence of a context switch. The latency-insensitive semantics of a CSA may permit a distributed, asynchronous initialization of the fabric, e.g., as soon as PEs are configured, they may begin execution immediately. Unconfigured PEs may backpressure their channels until they are configured, e.g., preventing communications between configured and unconfigured elements. The CSA configuration may be partitioned into privileged and user-level state. Such a two-level partitioning may enable primary configuration of the fabric to occur without invoking the operating system. During one embodiment of extraction, a logical view of the dataflow graph is captured and committed into memory, e.g., including all live control and dataflow tokens and state in the graph.
Extraction may also play a role in providing reliability guarantees through the creation of fabric checkpoints. Exceptions in a CSA may generally be caused by the same events that cause exceptions in processors, such as illegal operator arguments or reliability, availability, and serviceability (RAS) events. In certain embodiments, exceptions are detected at the level of dataflow operators, for example, checking argument values or through modular arithmetic schemes. Upon detecting an exception, a dataflow operator (e.g., circuit) may halt and emit an exception message, e.g., which contains both an operation identifier and some details of the nature of the problem that has occurred. In one embodiment, the dataflow operator will remain halted until it has been reconfigured. The exception message may then be communicated to an associated processor (e.g., core) for service, e.g., which may include extracting the graph for software analysis.
1.5 Tile-Level Architecture
Embodiments of the CSA computer architectures (e.g., targeting HPC and datacenter uses) are tiled.
In one embodiment, the goal of the CSA microarchitecture is to provide a high quality implementation of each dataflow operator specified by the CSA architecture. Embodiments of the CSA microarchitecture provide that each processing element (and/or communications network (e.g., a network dataflow endpoint circuit thereof)) of the microarchitecture corresponds to approximately one node (e.g., entity) in the architectural dataflow graph. In one embodiment, a node in the dataflow graph is distributed in multiple network dataflow endpoint circuits. In certain embodiments, this results in microarchitectural elements that are not only compact, resulting in a dense computation array, but also energy efficient, for example, where processing elements (PEs) are both simple and largely unmultiplexed, e.g., executing a single dataflow operator for a configuration (e.g., programming) of the CSA. To further reduce energy and implementation area, a CSA may include a configurable, heterogeneous fabric style in which each PE thereof implements only a subset of dataflow operators (e.g., with a separate subset of dataflow operators implemented with network dataflow endpoint circuit(s)). Peripheral and support subsystems, such as the CSA cache, may be provisioned to support the distributed parallelism incumbent in the main CSA processing fabric itself. Implementation of CSA microarchitectures may utilize dataflow and latency-insensitive communications abstractions present in the architecture. In certain embodiments, there is (e.g., substantially) a one-to-one correspondence between nodes in the compiler generated graph and the dataflow operators (e.g., dataflow operator compute elements) in a CSA.
Below is a discussion of an example CSA, followed by a more detailed discussion of the microarchitecture. Certain embodiments herein provide a CSA that allows for easy compilation, e.g., in contrast to an existing FPGA compilers that handle a small subset of a programming language (e.g., C or C++) and require many hours to compile even small programs.
Certain embodiments of a CSA architecture admits of heterogeneous coarse-grained operations, like double precision floating point. Programs may be expressed in fewer coarse grained operations, e.g., such that the disclosed compiler runs faster than traditional spatial compilers. Certain embodiments include a fabric with new processing elements to support sequential concepts like program ordered memory accesses. Certain embodiments implement hardware to support coarse-grained dataflow-style communication channels. This communication model is abstract, and very close to the control-dataflow representation used by the compiler. Certain embodiments herein include a network implementation that supports single-cycle latency communications, e.g., utilizing (e.g., small) PEs which support single control-dataflow operations. In certain embodiments, not only does this improve energy efficiency and performance, it simplifies compilation because the compiler makes a one-to-one mapping between high-level dataflow constructs and the fabric. Certain embodiments herein thus simplify the task of compiling existing (e.g., C, C++, or Fortran) programs to a CSA (e.g., fabric).
Energy efficiency may be a first order concern in modern computer systems. Certain embodiments herein provide a new schema of energy-efficient spatial architectures. In certain embodiments, these architectures form a fabric with a unique composition of a heterogeneous mix of small, energy-efficient, data-flow oriented processing elements (PEs) (and/or a packet switched communications network (e.g., a network dataflow endpoint circuit thereof)) with a lightweight circuit switched communications network (e.g., interconnect), e.g., with hardened support for flow control. Due to the energy advantages of each, the combination of these components may form a spatial accelerator (e.g., as part of a computer) suitable for executing compiler-generated parallel programs in an extremely energy efficient manner. Since this fabric is heterogeneous, certain embodiments may be customized for different application domains by introducing new domain-specific PEs. For example, a fabric for high-performance computing might include some customization for double-precision, fused multiply-add, while a fabric targeting deep neural networks might include low-precision floating point operations.
An embodiment of a spatial architecture schema, e.g., as exemplified in
Programs may be converted to dataflow graphs that are mapped onto the architecture by configuring PEs and the network to express the control-dataflow graph of the program. Communication channels may be flow-controlled and fully back-pressured, e.g., such that PEs will stall if either source communication channels have no data or destination communication channels are full. In one embodiment, at runtime, data flow through the PEs and channels that have been configured to implement the operation (e.g., an accelerated algorithm). For example, data may be streamed in from memory, through the fabric, and then back out to memory.
Embodiments of such an architecture may achieve remarkable performance efficiency relative to traditional multicore processors: compute (e.g., in the form of PEs) may be simpler, more energy efficient, and more plentiful than in larger cores, and communications may be direct and mostly short-haul, e.g., as opposed to occurring over a wide, full-chip network as in typical multicore processors. Moreover, because embodiments of the architecture are extremely parallel, a number of powerful circuit and device level optimizations are possible without seriously impacting throughput, e.g., low leakage devices and low operating voltage. These lower-level optimizations may enable even greater performance advantages relative to traditional cores. The combination of efficiency at the architectural, circuit, and device levels yields of these embodiments are compelling. Embodiments of this architecture may enable larger active areas as transistor density continues to increase.
Embodiments herein offer a unique combination of dataflow support and circuit switching to enable the fabric to be smaller, more energy-efficient, and provide higher aggregate performance as compared to previous architectures. FPGAs are generally tuned towards fine-grained bit manipulation, whereas embodiments herein are tuned toward the double-precision floating point operations found in HPC applications. Certain embodiments herein may include a FPGA in addition to a CSA according to this disclosure.
Certain embodiments herein combine a light-weight network with energy efficient dataflow processing elements (and/or communications network (e.g., a network dataflow endpoint circuit thereof)) to form a high-throughput, low-latency, energy-efficient HPC fabric. This low-latency network may enable the building of processing elements (and/or communications network (e.g., a network dataflow endpoint circuit thereof)) with fewer functionalities, for example, only one or two operations and perhaps one architecturally visible register, since it is efficient to gang multiple PEs together to form a complete program.
Relative to a processor core, CSA embodiments herein may provide for more computational density and energy efficiency. For example, when PEs are very small (e.g., compared to a core), the CSA may perform many more operations and have much more computational parallelism than a core, e.g., perhaps as many as 16 times the number of FMAs as a vector processing unit (VPU). To utilize all of these computational elements, the energy per operation is very low in certain embodiments.
The energy advantages our embodiments of this dataflow architecture are many. Parallelism is explicit in dataflow graphs and embodiments of the CSA architecture spend no or minimal energy to extract it, e.g., unlike out-of-order processors which must re-discover parallelism each time an operation is executed. Since each PE is responsible for a single operation in one embodiment, the register files and ports counts may be small, e.g., often only one, and therefore use less energy than their counterparts in core. Certain CSAs include many PEs, each of which holds live program values, giving the aggregate effect of a huge register file in a traditional architecture, which dramatically reduces memory accesses. In embodiments where the memory is multi-ported and distributed, a CSA may sustain many more outstanding memory requests and utilize more bandwidth than a core. These advantages may combine to yield an energy level per watt that is only a small percentage over the cost of the bare arithmetic circuitry. For example, in the case of an integer multiply, a CSA may consume no more than 25% more energy than the underlying multiplication circuit. Relative to one embodiment of a core, an integer operation in that CSA fabric consumes less than 1/30th of the energy per integer operation.
From a programming perspective, the application-specific malleability of embodiments of the CSA architecture yields significant advantages over a vector processing unit (VPU). In traditional, inflexible architectures, the number of functional units, like floating divide or the various transcendental mathematical functions, must be chosen at design time based on some expected use case. In embodiments of the CSA architecture, such functions may be configured (e.g., by a user and not a manufacturer) into the fabric based on the requirement of each application. Application throughput may thereby be further increased. Simultaneously, the compute density of embodiments of the CSA improves by avoiding hardening such functions, and instead provision more instances of primitive functions like floating multiplication. These advantages may be significant in HPC workloads, some of which spend 75% of floating execution time in transcendental functions.
Certain embodiments of the CSA represents a significant advance as a dataflow-oriented spatial architectures, e.g., the PEs of this disclosure may be smaller, but also more energy-efficient. These improvements may directly result from the combination of dataflow-oriented PEs with a lightweight, circuit switched interconnect, for example, which has single-cycle latency, e.g., in contrast to a packet switched network (e.g., with, at a minimum, a 300% higher latency). Certain embodiments of PEs support 32-bit or 64-bit operation. Certain embodiments herein permit the introduction of new application-specific PEs, for example, for machine learning or security, and not merely a homogeneous combination. Certain embodiments herein combine lightweight dataflow-oriented processing elements with a lightweight, low-latency network to form an energy efficient computational fabric.
In order for certain spatial architectures to be successful, programmers are to configure them with relatively little effort, e.g., while obtaining significant power and performance superiority over sequential cores. Certain embodiments herein provide for a CSA (e.g., spatial fabric) that is easily programmed (e.g., by a compiler), power efficient, and highly parallel. Certain embodiments herein provide for a (e.g., interconnect) network that achieves these three goals. From a programmability perspective, certain embodiments of the network provide flow controlled channels, e.g., which correspond to the control-dataflow graph (CDFG) model of execution used in compilers. Certain network embodiments utilize dedicated, circuit switched links, such that program performance is easier to reason about, both by a human and a compiler, because performance is predictable. Certain network embodiments offer both high bandwidth and low latency. Certain network embodiments (e.g., static, circuit switching) provides a latency of 0 to 1 cycle (e.g., depending on the transmission distance.) Certain network embodiments provide for a high bandwidth by laying out several networks in parallel, e.g., and in low-level metals. Certain network embodiments communicate in low-level metals and over short distances, and thus are very power efficient.
Certain embodiments of networks include architectural support for flow control. For example, in spatial accelerators composed of small processing elements (PEs), communications latency and bandwidth may be critical to overall program performance. Certain embodiments herein provide for a light-weight, circuit switched network which facilitates communication between PEs in spatial processing arrays, such as the spatial array shown in
Spatial architectures, such as the one shown in
Operations may be executed based on the availability of their inputs and the status of the PE. A PE may obtain operands from input channels and write results to output channels, although internal register state may also be used. Certain embodiments herein include a configurable dataflow-friendly PE.
Operation registers may be set during a special configuration step. During this step, auxiliary control wires and state, in addition to the inter-PE network, may be used to stream in configuration across the several PEs comprising the fabric. As result of parallelism, certain embodiments of such a network may provide for rapid reconfiguration, e.g., a tile sized fabric may be configured in less than about 10 microseconds.
Implementing distributed data channels may include two paths, illustrated in
The network may be statically configured, e.g., in addition to PEs being statically configured. During the configuration step, configuration bits may be set at each network component. These bits control, for example, the multiplexer selections and flow control functions. A network may comprise a plurality of networks, e.g., a data path network and a flow control path network. A network or plurality of networks may utilize paths of different widths (e.g., a first width, and a narrower or wider width). In one embodiment, a data path network has a wider (e.g., bit transport) width than the width of a flow control path network. In one embodiment, each of a first network and a second network includes their own data path network and flow control path network, e.g., data path network A and flow control path network A and wider data path network B and flow control path network B.
Certain embodiments of a network are bufferless, and data is to move between producer and consumer in a single cycle. Certain embodiments of a network are also boundless, that is, the network spans the entire fabric. In one embodiment, one PE is to communicate with any other PE in a single cycle. In one embodiment, to improve routing bandwidth, several networks may be laid out in parallel between rows of PEs.
Relative to FPGAs, certain embodiments of networks herein have three advantages: area, frequency, and program expression. Certain embodiments of networks herein operate at a coarse grain, e.g., which reduces the number configuration bits, and thereby the area of the network. Certain embodiments of networks also obtain area reduction by implementing flow control logic directly in circuitry (e.g., silicon). Certain embodiments of hardened network implementations also enjoys a frequency advantage over FPGA. Because of an area and frequency advantage, a power advantage may exist where a lower voltage is used at throughput parity. Finally, certain embodiments of networks provide better high-level semantics than FPGA wires, especially with respect to variable timing, and thus those certain embodiments are more easily targeted by compilers. Certain embodiments of networks herein may be thought of as a set of composable primitives for the construction of distributed, point-to-point data channels.
In certain embodiments, a multicast source may not assert its data valid unless it receives a ready signal from each sink. Therefore, an extra conjunction and control bit may be utilized in the multicast case.
Like certain PEs, the network may be statically configured. During this step, configuration bits are set at each network component. These bits control, for example, the multiplexer selection and flow control function. The forward path of our network requires some bits to swing its muxes. In the example shown in
For the third flow control box from the left in
2.1 Processing Elements
In certain embodiments, a CSA includes an array of heterogeneous PEs, in which the fabric is composed of several types of PEs each of which implement only a subset of the dataflow operators. By way of example,
PE execution may proceed in a dataflow style. Based on the configuration microcode, the scheduler may examine the status of the PE ingress and egress buffers, and, when all the inputs for the configured operation have arrived and the egress buffer of the operation is available, orchestrates the actual execution of the operation by a dataflow operator (e.g., on the ALU). The resulting value may be placed in the configured egress buffer. Transfers between the egress buffer of one PE and the ingress buffer of another PE may occur asynchronously as buffering becomes available. In certain embodiments, PEs are provisioned such that at least one dataflow operation completes per cycle. Section 2 discussed dataflow operator encompassing primitive operations, such as add, xor, or pick. Certain embodiments may provide advantages in energy, area, performance, and latency. In one embodiment, with an extension to a PE control path, more fused combinations may be enabled. In one embodiment, the width of the processing elements is 64 bits, e.g., for the heavy utilization of double-precision floating point computation in HPC and to support 64-bit memory addressing.
2.2 Communications Networks
Embodiments of the CSA microarchitecture provide a hierarchy of networks which together provide an implementation of the architectural abstraction of latency-insensitive channels across multiple communications scales. The lowest level of CSA communications hierarchy may be the local network. The local network may be statically circuit switched, e.g., using configuration registers to swing multiplexor(s) in the local network data-path to form fixed electrical paths between communicating PEs. In one embodiment, the configuration of the local network is set once per dataflow graph, e.g., at the same time as the PE configuration. In one embodiment, static, circuit switching optimizes for energy, e.g., where a large majority (perhaps greater than 95%) of CSA communications traffic will cross the local network. A program may include terms which are used in multiple expressions. To optimize for this case, embodiments herein provide for hardware support for multicast within the local network. Several local networks may be ganged together to form routing channels, e.g., which are interspersed (as a grid) between rows and columns of PEs. As an optimization, several local networks may be included to carry control tokens. In comparison to a FPGA interconnect, a CSA local network may be routed at the granularity of the data-path, and another difference may be a CSA's treatment of control. One embodiment of a CSA local network is explicitly flow controlled (e.g., back-pressured). For example, for each forward data-path and multiplexor set, a CSA is to provide a backward-flowing flow control path that is physically paired with the forward data-path. The combination of the two microarchitectural paths may provide a low-latency, low-energy, low-area, point-to-point implementation of the latency-insensitive channel abstraction. In one embodiment, a CSA's flow control lines are not visible to the user program, but they may be manipulated by the architecture in service of the user program. For example, the exception handling mechanisms described in Section 1.2 may be achieved by pulling flow control lines to a “not present” state upon the detection of an exceptional condition. This action may not only gracefully stalls those parts of the pipeline which are involved in the offending computation, but may also preserve the machine state leading up the exception, e.g., for diagnostic analysis. The second network layer, e.g., the mezzanine network, may be a shared, packet switched network. Mezzanine network may include a plurality of distributed network controllers, network dataflow endpoint circuits. The mezzanine network (e.g., the network schematically indicated by the dotted box in
The composability of channels across network layers may be extended to higher level network layers at the inter-tile, inter-die, and fabric granularities.
For example, suppose the operation of this processing (e.g., compute) element is (or includes) what is called call a pick in
For example, suppose the operation of this processing (e.g., compute) element is (or includes) what is called call a switch in
Multiple networks (e.g., interconnects) may be connected to a processing element, e.g., (input) networks 902, 904, 906 and (output) networks 908, 910, 912. The connections may be switches, e.g., as discussed in reference to
Data input buffer 924 and data input buffer 926 may perform similarly, e.g., local network 904 (e.g., set up as a data (as opposed to control) interconnect) is depicted as being switched (e.g., connected) to data input buffer 924. In this embodiment, a data path (e.g., network as in
A processing element 900 may be stalled from execution until its operands (e.g., a control input value and its corresponding data input value or values) are received and/or until there is room in the output buffer(s) of the processing element 900 for the data that is to be produced by the execution of the operation on those operands.
2.3 Memory Interface
The request address file (RAF) circuit, a simplified version of which is shown in
As an example for a load, an address arrives into queue 1022 which the scheduler 1012 matches up with a load in 1010. A completion buffer slot for this load is assigned in the order the address arrived. Assuming this particular load in the graph has no dependencies specified, the address and completion buffer slot are sent off to the memory system by the scheduler (e.g., via memory command 1042). When the result returns to multiplexer 1040 (shown schematically), it is stored into the completion buffer slot it specifies (e.g., as it carried the target slot all along though the memory system). The completion buffer sends results back into local network (e.g., local network 1002, 1004, 1006, or 1008) in the order the addresses arrived.
Stores may be similar except both address and data have to arrive before any operation is sent off to the memory system.
2.4 Cache
Dataflow graphs may be capable of generating a profusion of (e.g., word granularity) requests in parallel. Thus, certain embodiments of the CSA provide a cache subsystem with sufficient bandwidth to service the CSA. A heavily banked cache microarchitecture, e.g., as shown in
2.5 Network Resources, e.g., Circuitry, to Perform (e.g., Dataflow) Operations
In certain embodiments, processing elements (PEs) communicate using dedicated virtual circuits which are formed by statically configuring a (e.g., circuit switched) communications network. These virtual circuits may be flow controlled and fully back-pressured, e.g., such that a PE will stall if either the source has no data or its destination is full. At runtime, data may flow through the PEs implementing the mapped dataflow graph (e.g., mapped algorithm). For example, data may be streamed in from memory, through the (e.g., fabric area of a) spatial array of processing elements, and then back out to memory.
Such an architecture may achieve remarkable performance efficiency relative to traditional multicore processors: compute, e.g., in the form of PEs, may be simpler and more numerous than cores and communications may be direct, e.g., as opposed to an extension of the memory system. However, the (e.g., fabric area of) spatial array of processing elements may be tuned for the implementation of compiler-generated expression trees, which may feature little multiplexing or demultiplexing. Certain embodiments herein extend (for example, via network resources, such as, but not limited to, network dataflow endpoint circuits) the architecture to support (e.g., high-radix) multiplexing and/or demultiplexing, for example, especially in the context of function calls.
Spatial arrays, such as the spatial array of processing elements 101 in
In one embodiment, a circuit switched network between two points (e.g., between a producer and consumer of data) includes a dedicated communication line between those two points, for example, with (e.g., physical) switches between the two points set to create a (e.g., exclusive) physical circuit between the two points. In one embodiment, a circuit switched network between two points is set up at the beginning of use of the connection between the two points and maintained throughout the use of the connection. In another embodiment, a packet switched network includes a shared communication line (e.g., channel) between two (e.g., or more) points, for example, where packets from different connections share that communication line (for example, routed according to data of each packet, e.g., in the header of a packet including a header and a payload). An example of a packet switched network is discussed below, e.g., in reference to a mezzanine network.
Operations may be executed based on the availability of their inputs and the status of the PE. A PE may obtain operands from input channels and write results to output channels, although internal register state may also be used. Certain embodiments herein include a configurable dataflow-friendly PE.
Operation registers may be set during a special configuration step. During this step, auxiliary control wires and state, in addition to the inter-PE network, may be used to stream in configuration across the several PEs comprising the fabric. As result of parallelism, certain embodiments of such a network may provide for rapid reconfiguration, e.g., a tile sized fabric may be configured in less than about 10 microseconds.
Further, depicted accelerator tile 1300 includes packet switched communications network 1314, for example, as part of a mezzanine network, e.g., as described below. Certain embodiments herein allow for (e.g., a distributed) dataflow operations (e.g., operations that only route data) to be performed on (e.g., within) the communications network (e.g., and not in the processing element(s)). As an example, a distributed Pick dataflow operation of a dataflow graph is depicted in
As one example, a pick dataflow operation may have a plurality of inputs and steer (e.g., route) one of them as an output, e.g., as in
In the depicted embodiment, packet switched communications network 1314 may handle certain (e.g., configuration) communications, for example, to program the processing elements and/or circuit switched network (e.g., network 1313, which may include switches). In one embodiment, a circuit switched network is configured (e.g., programmed) to perform one or more operations (e.g., dataflow operations of a dataflow graph).
Packet switched communications network 1314 includes a plurality of endpoints (e.g., network dataflow endpoint circuits (1302, 1304, 1306). In one embodiment, each endpoint includes an address or other indicator value to allow data to be routed to and/or from that endpoint, e.g., according to (e.g., a header of) a data packet.
Additionally or alternatively to performing one or more of the above, packet switched communications network 1314 may perform dataflow operations. Network dataflow endpoint circuits (1302, 1304, 1306) may be configured (e.g., programmed) to perform a (e.g., distributed pick) operation of a dataflow graph. Programming of components (e.g., a circuit) are described herein. An embodiment of configuring a network dataflow endpoint circuit (e.g., an operation configuration register thereof) is discussed in reference to
As an example of a distributed pick dataflow operation, network dataflow endpoint circuits (1302, 1304, 1306) in
Network dataflow endpoint circuit 1302 may be configured to receive input data from a plurality of sources (e.g., network dataflow endpoint circuit 1304 and network dataflow endpoint circuit 1306), and to output resultant data, e.g., as in
When network dataflow endpoint circuit 1304 is to transmit input data to network dataflow endpoint circuit 1302 (e.g., when network dataflow endpoint circuit 1302 has available storage room for the data and/or network dataflow endpoint circuit 1304 has its input data), network dataflow endpoint circuit 1304 may generate a packet (e.g., including the input data and a header to steer that data to network dataflow endpoint circuit 1302 on the packet switched communications network 1314 (e.g., as a stop on that (e.g., ring) network 1314). This is illustrated schematically with dashed line 1326 in
When network dataflow endpoint circuit 1306 is to transmit input data to network dataflow endpoint circuit 1302 (e.g., when network dataflow endpoint circuit 1302 has available storage room for the data and/or network dataflow endpoint circuit 1306 has its input data), network dataflow endpoint circuit 1304 may generate a packet (e.g., including the input data and a header to steer that data to network dataflow endpoint circuit 1302 on the packet switched communications network 1314 (e.g., as a stop on that (e.g., ring) network 1314). This is illustrated schematically with dashed line 1318 in
Network dataflow endpoint circuit 1302 (e.g., on receipt of the Input 0 from network dataflow endpoint circuit 1304, Input 1 from network dataflow endpoint circuit 1306, and/or control data) may then perform the programmed dataflow operation (e.g., a Pick operation in this example). The network dataflow endpoint circuit 1302 may then output the according resultant data from the operation, e.g., to processing element 1308 in
In one embodiment, the control data to perform an operation (e.g., pick operation) comes from other components of the spatial array, e.g., a processing element or through network. An example of this is discussed below in reference to
In certain embodiments, a dataflow graph may have certain operations performed by a processing element and certain operations performed by a communication network (e.g., network dataflow endpoint circuit or circuits).
As one description of an embodiment of the microarchitecture, a pick dataflow operator may function to pick one output of resultant data from a plurality of inputs of input data, e.g., based on control data. A network dataflow endpoint circuit 1400 may be configured to consider one of the spatial array ingress buffer(s) 1402 of the circuit 1400 (e.g., data from the fabric being control data) as selecting among multiple input data elements stored in network ingress buffer(s) 1424 of the circuit 1400 to steer the resultant data to the spatial array egress buffer 1408 of the circuit 1400. Thus, the network ingress buffer(s) 1424 may be thought of as inputs to a virtual mux, the spatial array ingress buffer 1402 as the multiplexer select, and the spatial array egress buffer 1408 as the multiplexer output. In one embodiment, when a (e.g., control data) value is detected and/or arrives in the spatial array ingress buffer 1402, the scheduler 1428 (e.g., as programmed by an operation configuration in storage 1426) is sensitized to examine the corresponding network ingress channel. When data is available in that channel, it is removed from the network ingress buffer 1424 and moved to the spatial array egress buffer 1408. The control bits of both ingresses and egress may then be updated to reflect the transfer of data. This may result in control flow tokens or credits being propagated in the associated network. In certain embodiments, all inputs (e.g., control or data) may arise locally or over the network.
Initially, it may seem that the use of packet switched networks to implement the (e.g., high-radix staging) operators of multiplexed and/or demultiplexed codes hampers performance. For example, in one embodiment, a packet-switched network is generally shared and the caller and callee dataflow graphs may be distant from one another. Recall, however, that in certain embodiments, the intention of supporting multiplexing and/or demultiplexing is to reduce the area consumed by infrequent code paths within a dataflow operator (e.g., by the spatial array). Thus, certain embodiments herein reduce area and avoid the consumption of more expensive fabric resources, for example, like PEs, e.g., without (substantially) affecting the area and efficiency of individual PEs to supporting those (e.g., infrequent) operations.
Turning now to further detail of
Depicted network dataflow endpoint circuit 1400 includes a spatial array (e.g., fabric) egress buffer 1408, for example, to output data (e.g., control data) to a (e.g., circuit switched) network. As noted above, although a single spatial array (e.g., fabric) egress buffer 1408 is depicted, a plurality of spatial array (e.g., fabric) egress buffers may be in a network dataflow endpoint circuit. In one embodiment, spatial array (e.g., fabric) egress buffer 1408 is to send (e.g., transmit) data (e.g., control data) onto a communications network of a spatial array (e.g., a spatial array of processing elements), for example, onto one or more of network 1410 and network 1412. In one embodiment, network 1410 is part of network 1313 in
Additionally or alternatively, network dataflow endpoint circuit 1400 may be coupled to another network 1414, e.g., a packet switched network. Another network 1414, e.g., a packet switched network, may be used to transmit (e.g., send or receive) (e.g., input and/or resultant) data to processing elements or other components of a spatial array and/or to transmit one or more of input data or resultant data. In one embodiment, network 1414 is part of the packet switched communications network 1314 in
Network buffer 1418 (e.g., register(s)) may be a stop on (e.g., ring) network 1414, for example, to receive data from network 1414.
Depicted network dataflow endpoint circuit 1400 includes a network egress buffer 1422, for example, to output data (e.g., resultant data) to a (e.g., packet switched) network. As noted above, although a single network egress buffer 1422 is depicted, a plurality of network egress buffers may be in a network dataflow endpoint circuit. In one embodiment, network egress buffer 1422 is to send (e.g., transmit) data (e.g., resultant data) onto a communications network of a spatial array (e.g., a spatial array of processing elements), for example, onto network 1414. In one embodiment, network 1414 is part of packet switched network 1314 in
Depicted network dataflow endpoint circuit 1400 includes a network ingress buffer 1422, for example, to input data (e.g., inputted data) from a (e.g., packet switched) network. As noted above, although a single network ingress buffer 1424 is depicted, a plurality of network ingress buffers may be in a network dataflow endpoint circuit. In one embodiment, network ingress buffer 1424 is to receive (e.g., transmit) data (e.g., input data) from a communications network of a spatial array (e.g., a spatial array of processing elements), for example, from network 1414. In one embodiment, network 1414 is part of packet switched network 1314 in
In one embodiment, the data format (e.g., of the data on network 1414) includes a packet having data and a header (e.g., with the destination of that data). In one embodiment, the data format (e.g., of the data on network 1404 and/or 1406) includes only the data (e.g., not a packet having data and a header (e.g., with the destination of that data)). Network dataflow endpoint circuit 1400 may add (e.g., data output from circuit 1400) or remove (e.g., data input into circuit 1400) a header (or other data) to or from a packet. Coupling 1420 (e.g., wire) may send data received from network 1414 (e.g., from network buffer 1418) to network ingress buffer 1424 and/or multiplexer 1416. Multiplexer 1416 may (e.g., via a control signal from the scheduler 1428) output data from network buffer 1418 or from network egress buffer 1422. In one embodiment, one or more of multiplexer 1416 or network buffer 1418 are separate components from network dataflow endpoint circuit 1400. A buffer may include a plurality of (e.g., discrete) entries, for example, a plurality of registers.
In one embodiment, operation configuration storage 1426 (e.g., register or registers) is loaded during configuration (e.g., mapping) and specifies the particular operation (or operations) this network dataflow endpoint circuit 1400 (e.g., not a processing element of a spatial array) is to perform (e.g., data steering operations in contrast to logic and/or arithmetic operations). Buffer(s) (e.g., 1402, 1408, 1422, and/or 1424) activity may be controlled by that operation (e.g., controlled by the scheduler 1428). Scheduler 1428 may schedule an operation or operations of network dataflow endpoint circuit 1400, for example, when (e.g., all) input (e.g., payload) data and/or control data arrives. Dotted lines to and from scheduler 1428 indicate paths that may be utilized for control data, e.g., to and/or from scheduler 1428. Scheduler may also control multiplexer 1416, e.g., to steer data to and/or from network dataflow endpoint circuit 1400 and network 1414.
In reference to the distributed pick operation in
When network dataflow endpoint circuit 1304 is to transmit input data to network dataflow endpoint circuit 1302 (e.g., when network dataflow endpoint circuit 1302 has available storage room for the data and/or network dataflow endpoint circuit 1304 has its input data), network dataflow endpoint circuit 1304 may generate a packet (e.g., including the input data and a header to steer that data to network dataflow endpoint circuit 1302 on the packet switched communications network 1314 (e.g., as a stop on that (e.g., ring) network). This is illustrated schematically with dashed line 1326 in
When network dataflow endpoint circuit 1306 is to transmit input data to network dataflow endpoint circuit 1302 (e.g., when network dataflow endpoint circuit 1302 has available storage room for the data and/or network dataflow endpoint circuit 1306 has its input data), network dataflow endpoint circuit 1304 may generate a packet (e.g., including the input data and a header to steer that data to network dataflow endpoint circuit 1302 on the packet switched communications network 1314 (e.g., as a stop on that (e.g., ring) network). This is illustrated schematically with dashed line 1318 in
Network dataflow endpoint circuit 1302 (e.g., on receipt of the Input 0 from network dataflow endpoint circuit 1304 in circuit 1302's network ingress buffer(s), Input 1 from network dataflow endpoint circuit 1306 in circuit 1302's network ingress buffer(s), and/or control data from processing element 1308 in circuit 1302's spatial array ingress buffer) may then perform the programmed dataflow operation (e.g., a Pick operation in this example). The network dataflow endpoint circuit 1302 may then output the according resultant data from the operation, e.g., to processing element 1308 in
Depicted receive operation configuration data format 1704 includes an output field 1704A (e.g., indicating which component(s) in a network the (resultant) data is to be sent to), an input field 1704B (e.g., an identifier of the component(s) that is to send the input data), and an operation field 1704C (e.g., indicating which of a plurality of operations are to be performed). In one embodiment, the (e.g., inbound) operation is one of a Pick, PickSingleLeg, PickAny, or Merge dataflow operation, e.g., corresponding to a (e.g., same) dataflow operator of a dataflow graph. In one embodiment, a merge dataflow operation is a pick that requires and dequeues all operands (e.g., with the egress endpoint receiving control).
A configuration data format utilized herein may include one or more of the fields described herein, e.g., in any order.
In one embodiment, circuit 1900 (e.g., network dataflow endpoint circuit) is to receive packet of data in the data format of (e.g., send) operation configuration data format 1902, for example, with the input being the source(s) of the payload (e.g., input data) and the operation field indicating which operation is to be performed (e.g., shown schematically as Switch or SwitchAny). Depicted multiplexer 1904 may select the operation to be performed from a plurality of available operations, e.g., based on the value in operation field 1902D. In one embodiment, circuit 1900 is to perform that operation when both the input data is available and the credit status is a yes (for example, the dependency token indicates) indicating there is room for the output data to be stored, e.g., in a buffer of the destination.
In one embodiment, the send operation does not utilize control beyond checking its input(s) are available for sending. This may enable switch to perform the operation without credit on all legs. In one embodiment, the Switch and/or SwitchAny operation includes a multiplexer controlled by the value stored in the operation field 1902D to select the correct queue management circuitry.
Value stored in operation field 1902D may select among control options, e.g., with different control (e.g., logic) circuitry for each operation, for example, as in
In one embodiment, PickAny executes on the presence of any data and/or selection decoder creates multiplexer selection bits.
In one embodiment, (e.g., as with scheduling) the choice of dequeue is determined by the operation and its dynamic behavior, e.g., to dequeue the operation after performance. In one embodiment, a circuit is to use the operand selection bits to dequeue data (e.g., input, output and/or control data).
Network 2514 may be a circuit switched network, e.g., as discussed herein. Additionally or alternatively, a packet switched network (e.g., as discussed herein) may also be utilized, for example, coupled to network egress buffer 2522, network ingress buffer 2524, or other components herein. Argument queue 2502 may include a control buffer 2502A, for example, to indicate when a respective input queue (e.g., buffer) includes a (new) item of data, e.g., as a single bit. Turning now to
Referring again to
In certain embodiments, an accelerator (e.g., a PE thereof) couples to a RAF circuit or a plurality of RAF circuits through (i) a circuit switched network (for example, as discussed herein, e.g., in reference to
2.6 Floating Point Support
Certain HPC applications are characterized by their need for significant floating point bandwidth. To meet this need, embodiments of a CSA may be provisioned with multiple (e.g., between 128 and 256 each) of floating add and multiplication PEs, e.g., depending on tile configuration. A CSA may provide a few other extended precision modes, e.g., to simplify math library implementation. CSA floating point PEs may support both single and double precision, but lower precision PEs may support machine learning workloads. A CSA may provide an order of magnitude more floating point performance than a processor core. In one embodiment, in addition to increasing floating point bandwidth, in order to power all of the floating point units, the energy consumed in floating point operations is reduced. For example, to reduce energy, a CSA may selectively gate the low-order bits of the floating point multiplier array. In examining the behavior of floating point arithmetic, the low order bits of the multiplication array may often not influence the final, rounded product.
Given this maximum carry, if the result of the carry region is less than 2c-g, where the carry region is c bits wide, then the gated region may be ignored since it does not influence the result region. Increasing g means that it is more likely the gated region will be needed, while increasing c means that, under random assumption, the gated region will be unused and may be disabled to avoid energy consumption. In embodiments of a CSA floating multiplication PE, a two stage pipelined approach is utilized in which first the carry region is determined and then the gated region is determined if it is found to influence the result. If more information about the context of the multiplication is known, a CSA more aggressively tune the size of the gated region. In FMA, the multiplication result may be added to an accumulator, which is often much larger than either of the multiplicands. In this case, the addend exponent may be observed in advance of multiplication and the CSDA may adjust the gated region accordingly. One embodiment of the CSA includes a scheme in which a context value, which bounds the minimum result of a computation, is provided to related multipliers, in order to select minimum energy gating configurations.
2.7 Runtime Services
In certain embodiments, a CSA includes a heterogeneous and distributed fabric, and consequently, runtime service implementations are to accommodate several kinds of PEs in a parallel and distributed fashion. Although runtime services in a CSA may be critical, they may be infrequent relative to user-level computation. Certain implementations, therefore, focus on overlaying services on hardware resources. To meet these goals, CSA runtime services may be cast as a hierarchy, e.g., with each layer corresponding to a CSA network. At the tile level, a single external-facing controller may accepts or sends service commands to an associated core with the CSA tile. A tile-level controller may serve to coordinate regional controllers at the RAFs, e.g., using the ACI network. In turn, regional controllers may coordinate local controllers at certain mezzanine network stops (e.g., network dataflow endpoint circuits). At the lowest level, service specific micro-protocols may execute over the local network, e.g., during a special mode controlled through the mezzanine controllers. The micro-protocols may permit each PE (e.g., PE class by type) to interact with the runtime service according to its own needs. Parallelism is thus implicit in this hierarchical organization, and operations at the lowest levels may occur simultaneously. This parallelism may enables the configuration of a CSA tile in between hundreds of nanoseconds to a few microseconds, e.g., depending on the configuration size and its location in the memory hierarchy. Embodiments of the CSA thus leverage properties of dataflow graphs to improve implementation of each runtime service. One key observation is that runtime services may need only to preserve a legal logical view of the dataflow graph, e.g., a state that can be produced through some ordering of dataflow operator executions. Services may generally not need to guarantee a temporal view of the dataflow graph, e.g., the state of a dataflow graph in a CSA at a specific point in time. This may permit the CSA to conduct most runtime services in a distributed, pipelined, and parallel fashion, e.g., provided that the service is orchestrated to preserve the logical view of the dataflow graph. The local configuration micro-protocol may be a packet-based protocol overlaid on the local network. Configuration targets may be organized into a configuration chain, e.g., which is fixed in the microarchitecture. Fabric (e.g., PE) targets may be configured one at a time, e.g., using a single extra register per target to achieve distributed coordination. To start configuration, a controller may drive an out-of-band signal which places all fabric targets in its neighborhood into an unconfigured, paused state and swings multiplexors in the local network to a pre-defined conformation. As the fabric (e.g., PE) targets are configured, that is they completely receive their configuration packet, they may set their configuration microprotocol registers, notifying the immediately succeeding target (e.g., PE) that it may proceed to configure using the subsequent packet. There is no limitation to the size of a configuration packet, and packets may have dynamically variable length. For example, PEs configuring constant operands may have a configuration packet that is lengthened to include the constant field (e.g., X and Y in
The following section includes example operations of an operation set architecture (OSA) for a configurable spatial accelerator (CSA). A CSA may be programmed to perform one or more of the operations of the OSA, e.g., in contrast to an instruction that is decoded and the decoded instruction is executed. In certain embodiments, a CSA is a fabric comprised of various (e.g., small) processing elements connected by a configurable, statically circuit switched interconnection network. In certain embodiments, processing elements are configured to execute the dataflow operators present in a (e.g., control) dataflow graph, for example, with each processing element implementing approximately one dataflow operator. In certain embodiments, configuration occurs as a stage prior to execution and occurs only once for the life of the graph. As discussed above, dataflow operators may execute independently, e.g., whenever data is available locally at the processing element. Thus, parallelism may be achieved by the simultaneous execution of processing elements. For many forms of parallelism, high degrees of concurrent execution are achieved. As a purpose-built accelerator, a CSA may utilize a processor core (e.g., as discussed herein) to execute non-parallel or otherwise un-accelerable portions of programs.
The following includes a short description of certain concepts and terminology in section 3.1, some of which are described in more detail in other sections herein. Section 3 then discusses an example processing element with control lines in section 3.2, example communications (e.g., via a circuit switched network) in section 3.3, configuration of a CSA (e.g., configuration of the PEs and a circuit switched network) in section 3.4, an example operation format in section 3.5, and example operations in section 3.6.
3.1 Concepts and Terminology
An operation (e.g., which has input and output operands) may be configured on to some hardware component (e.g., a PE) at configuration time. Particularly, the hardware components (e.g., PEs) may be configured (e.g., programmed) as a dataflow operator (e.g., as a representation of a node in a dataflow graph) through the use of one of more of the operations of the OSA discussed herein. Operands may be sourced from and/or to latency insensitive channels (LICs), registers, or literal values. In certain embodiments, operations are initially triggered (e.g., able to start execution) by the availability of all required input operands and availability of a location for output. Operations may execute to produce an output directly when triggered, or may execute for an extended period generating multiple outputs, such as a sequence or stream. Operations that trigger and issue once without internal preserved state may be referred to as stateless operations. Operations that may perform extended processing, e.g., related to streams, may be referred to as stateful operations. Operations may be classified in several broad categories, such as integer logical and arithmetic, floating point arithmetic, comparisons, conversions, memory reference, fan-in/fan-out for dataflow (e.g., merge, copy, or switch), ordering, sequence generation, etc. Unlike other architectures, a stateful operation in a CSA may trigger and run for an extended duration. For example, a sequence generation operation might trigger on receiving the bounds of the sequence to generate, and it will be executing over an extended period as it sends out successive values in the sequence.
A latency insensitive channel (LIC) may refer to a point to point connection between operations (e.g., PEs) with exactly one head enqueuing values and one tail dequeuing them, e.g., first-in-first-out (FIFO) queues. In one embodiment, one or more LICs are formed between a single transmitting PE and a plurality of receiving PEs (e.g., multicast send). In one embodiment, one or more LICs are formed between a plurality of transmitting PE and a single receiving PE (e.g., multicast receive). In certain embodiments, ordering is preserved for values flowing through a LIC from the producer to the consumer. LICs may be characterized by a bit width, and a depth, e.g., the number of values that can be held. Note that in embodiments of an assembler, LICs are declared with a type. In one embodiment, only the bit size of the type matters for operating semantics. The operations for a latency insensitive channel may be: 1) check for empty (e.g., before reading), 2) check for full (e.g., before writing), 3) write a value at the tail (e.g., “put”), 4) read a value at the head, 5) remove the head (e.g., where 4 and 5 may be combined as a “get” operation).
A signal (e.g., a value of that signal) may refer to a LIC with no data width (e.g., the nil type—0 bits), for example, as only a presence signal, in which case only the fact that something has happened is conveyed.
A register may store state local to a unit that may be used to hold values. Registers are not a required part of hardware components (e.g., PEs), but may be available on some programmable hardware components. In certain embodiments, registers on one hardware component (e.g., PE) cannot be directly accessed by any other hardware components (e.g., PEs).
A CSA instance may include a network of processing elements (PEs), e.g., along with hardware to access memory. A hardware component (e.g., unit) may perform some set of operations of the OSA that are enabled by configuring them onto the component. Components (e.g., PEs) may be configured with one or more operations, e.g. to perform a variety of integer operations, and a particular instance of that component type may have multiple operations (e.g., add64 c1,c2,c3 and and64 c0,c1,1) loaded on it for a particular program, though certain embodiments may include only configuring a single operation per component (e.g., PE). Examples of kinds of hardware components (e.g., PEs) include ALU, floating multiply add, integer multiply, conversions, sequence generators, access, scratchpad, etc. Components may vary from having very little state (e.g. just operation descriptions for configuring) to small counts of latches combined with logic circuitry, to scratchpad components that are primarily storage, to (e.g., relatively complex) components for memory access. Some components may have multiple operations up to a small fixed limit (e.g., about 16), while others may only allow a single instance of an operation to be performed. Note that the exact concept of how large a CSA is flexible in certain embodiments, e.g., when loading a graph on aggregated CSA instances.
A (e.g., CSA) program may be a collection of operations and channels definitions that are configured (e.g., loaded) onto the hardware components (e.g., units) and network (e.g., interconnect) of a CSA instance. One CSA model expects that once configuration is complete, the program may be executed one or more times without reconfiguration, e.g., provided the CSA resources used for the program is loaded are not needed for another program between. In certain embodiments, routing of LICs is a property of configuration, e.g., and the configuration of the hardware (e.g., PE's and network) is not changed during the execution of a program. In certain embodiments, a CSA holds multiple programs at the same time, for example, and a given program may have multiple entry points (e.g. a CSA may hold code for several loop nests that are executed in a larger context to avoid repeated configuration steps).
Configuring may generally refer to when the program is loaded onto hardware, e.g. configuring a program onto the CSA, or configuring individual operations onto hardware components (e.g., PEs) during that load. In certain embodiments, configuring and transferring control to a CSA (e.g., from a processor core) has a reasonable configuration cost (e.g., dozens to hundreds of cycles to configure, not thousands or more), and the invocation of a CSA routine is relatively fast.
A sequence may generally refer to a sequence of values. The successive values in a given LIC may form a sequence.
A stream may generally refer to a set of channels including a stream control channel, e.g., as a single bit LIC, and one or more data channels. The values in the stream channel may be logical 1s until there is no more data, at which point there is a logical 0 to signal the end of the stream. E.g. a stream of the values 1-5 in a {control, data} formatted pair may look like {1,1}, {1,2}, {1,3}, {1,4}, {1,5}, {0} (note that no data value is included in the last data set {0} as the (first position) logical zero therein signals it is the end of the stream).
A CSA may utilize multiple data types. The types may be used in declaration of storage (e.g., including LICs, registers and static storage) and show up in the name of operations (e.g. add64, fmaf32, cvts64f32). In certain embodiments of an assembler, standalone type names are prefixed with a period (e.g., .lic .i64 achannel to declare a 64 bit LIC.)
In certain embodiments for storage, like LICs, only the bit size is semantically relevant to operations and the other properties are not (e.g., i32, s32, u32 and f32 are semantically equivalent for a LIC definition though they may affect the readability of output in a simulator dump).
In the example operations section, the s, u, or f types may be used for clarity of how the operation treats the bits in the operation, but not imply that hardware is doing sign extension beyond what is specified in the operation, or any type of implicit data conversion. Table 2 below indicates example types that may be used (e.g., in assembly).
Operations (e.g., CSA operations) may be the data values (e.g., including multiple fields) that are provided (e.g., as a plurality of set bits) to a hardware component (e.g., a PE) to program the PE to perform the desired operation (e.g., the PE performing that programmed operation when the input data arrives and there is storage available for the output data). A processing element may be any of the processing elements (or component or components thereof) discussed herein. The following discusses an embodiment of a processing element along with its example control lines.
3.2 Example Processing Element with Control Lines
In certain embodiments, the core architectural interface of the CSA is the dataflow operator, e.g., as a direct representation of a node in a dataflow graph. From an operational perspective, dataflow operators may behave in a streaming or data-driven fashion. Dataflow operators execute as soon as their incoming operands become available and there is space available to store the output (resultant) operand or operands. In certain embodiments, CSA dataflow execution depends only on highly localized status, e.g., resulting in a highly scalable architecture with a distributed, asynchronous execution model.
In certain embodiments, a CSA fabric architecture takes the position that each processing element of the microarchitecture corresponds to approximately one entity in the architectural dataflow graph. In certain embodiments, this results in processing elements that are not only compact, resulting in a dense computation array, but also energy efficient. To further reduce energy and implementation area, certain embodiments use a flexible, heterogeneous fabric style in which each PE implements only a (proper) subset of dataflow operators. For example, with floating point operations and integer operations mapped to separate processing element types, but both types support dataflow control operations discussed herein. In one embodiment, a CSA includes a dozen types of PEs, although the precise mix and allocation may vary in other embodiments.
In one embodiment, processing elements are organized as pipelines and support the injection of one pipelined dataflow operator per cycle. Processing elements may have a single-cycle latency. However, other pipelining choices may be used for other (e.g., more complicated) operations. For example, floating point operations may use multiple pipeline stages.
As discussed herein, in certain embodiments CSA PEs are configured (for example, as discussed in section 3.4 below, e.g., according to the operations discussed in section 3.6) before the beginning of graph execution to implement a particular dataflow operation from among the set that they support. A configuration value (e.g., stored in the configuration register of a PE) may consist of one or two control words (e.g., 32 or 64 bits) which specify an opcode controlling the operation circuitry (e.g., ALU), steer the various multiplexors within the PE, and actuate dataflow into and out of the PE channels. Dataflow operators may thus be implemented by micro coding these configurations bits. Once configured, in certain embodiments the PE operation is fixed for the life of the graph, e.g., although microcode may provide some (e.g., limited) flexibility to support dynamically controller operations.
To handle some of the more complex dataflow operators like floating-point fused-multiply add (FMA) and a loop-control sequencer operator, multiple PEs may be used rather than to provision a more complex single PE. In these cases, additional function-specific communications paths may be added between the combinable PEs. In the case of an embodiment of a sequencer (e.g., to implement loop control), combinational paths are established between (e.g., adjacent) PEs to carry control information related to the loop. Such PE combinations may maintain fully pipelined behavior while preserving the utility of a basic PE embodiment, e.g., in the case that the combined behavior is not used for a particular program graph.
Processing elements may implement a common interface, e.g., including the local network interfaces described herein. In addition to ports into the local network, a (e.g., every) processing element may implement a full complement of runtime services, e.g., including the micro-protocols associated with configuration, extraction, and exception. In certain embodiments, a common processing element perimeter enables the full parameterization of a particular hardware instance of a CSA with respect to processing element count, composition, and function, e.g., and the same properties make CSA processing element architecture highly amenable to deployment-specific extension. For example, CSA may include PEs tuned for the low-precision arithmetic machine learning applications.
In certain embodiments, a significant source of area and energy reduction is the customization of the dataflow operations supported by each type of processing element. In one embodiment, a proper subset (e.g., most) processing elements support only a few operations (e.g., one, two, three, or four operation types), for example, an implementation choice where a floating point PE only supports one of floating point multiply or floating point add, but not both.
Operation configuration value may be a (e.g., unique) value, for example, according to the format discussed in section 3.5 below, e.g., for the operations discussed in section 3.6 below. In certain embodiments, operation configuration value includes a plurality of bits that cause processing element 3300 to perform a desired (e.g., preselected) operation, for example, performing the desired (e.g., preselected) operation when the incoming operands become available (e.g., in input storage 3324 and/or input storage 3326) and when there is space available to store the output (resultant) operand or operands (e.g., in output storage 3334 and/or output storage 3336). The depicted processing element 3300 includes two sets of operation circuitry 3325 and 3327, for example, to each perform a different operation. In certain embodiments, a PE includes status (e.g., state) storage, for example, within operation circuitry or a status register. Status storage may be modified during the operation in the the course of execution. Status storage may be shared among several operations. See, for example, the status register 938 in
Depicted processing element 3300 includes an operation configuration storage 3319 (e.g., register(s)) to store an operation configuration value. In one embodiment, all of or a proper subset of a (e.g., single) operation configuration value is sent from the operation configuration storage 3319 (e.g., register(s)) to the multiplexers (e.g., multiplexer 3321 and multiplexer 3323) and/or demultiplexers (e.g., demultiplexer 3341 and demultiplexer 3343) of the processing element 3300 to steer the data according to the configuration.
Processing element 3300 includes a first input storage 3324 (e.g., input queue or buffer) coupled to (e.g., circuit switched) network 3302 and a second input storage 3326 (e.g., input queue or buffer) coupled to (e.g., circuit switched) network 3304. Network 3302 and network 3304 may be the same network (e.g., different circuit switched paths of the same network). Although two input storages are depicted, a single input storage or more than two input storages (e.g., any integer or proper subset of integers) may be utilized (e.g., with their own respective input controllers). Operation configuration value may be sent via the same network that the input storage 3324 and/or input storage 3326 are coupled to.
Depicted processing element 3300 includes input controller 3301, input controller 3303, output controller 3305, and output controller 3307 (e.g., together forming a scheduler for processing element 3300). Embodiments of input controllers are discussed in reference to
In
In certain embodiments, the input data (e.g., dataflow token or tokens) is sent to input storage 3324 and/or input storage 3326 by networks 3302 or networks 3302. In one embodiment, input data is stalled until there is available storage (e.g., in the targeted storage input storage 3324 or input storage 3326) in the storage that is to be utilized for that input data. In the depicted embodiment, operation configuration value (or a portion thereof) is sent to the multiplexers (e.g., multiplexer 3321 and multiplexer 3323) and/or demultiplexers (e.g., demultiplexer 3341 and demultiplexer 3343) of the processing element 3300 as control value(s) to steer the data according to the configuration. In certain embodiments, input operand selection switches 3321 and 3323 (e.g., multiplexers) allow data (e.g., dataflow tokens) from input storage 3324 and input storage 3326 as inputs to either of operation circuitry 3325 or operation circuitry 3327. In certain embodiments, result (e.g., output operand) selection switches 3337 and 3339 (e.g., multiplexers) allow data from either of operation circuitry 3325 or operation circuitry 3327 into output storage 3334 and/or output storage 3336. Storage may be a queue (e.g., FIFO queue). In certain embodiments, an operation takes one input operand (e.g., from either of input storage 3324 and input storage 3326) and produce two resultants (e.g., stored in output storage 3334 and output storage 3336). In certain embodiments, an operation takes two or more input operands (for example, one from each input storage queue, e.g., one from each of input storage 3324 and input storage 3326) and produces a single (or plurality of) resultant (for example, stored in output storage, e.g., output storage 3334 and/or output storage 3336).
In certain embodiments, processing element 3300 is stalled from execution until there is input data (e.g., dataflow token or tokens) in input storage and there is storage space for the resultant data available in the output storage (e.g., as indicated by a backpressure value sent that indicates the output storage is not full). In the depicted embodiment, the input storage (queue) status value from path 3309 indicates (e.g., by asserting a “not empty” indication value or an “empty” indication value) when input storage 3324 contains (e.g., new) input data (e.g., dataflow token or tokens) and the input storage (queue) status value from path 3311 indicates (e.g., by asserting a “not empty” indication value or an “empty” indication value) when input storage 3326 contains (e.g., new) input data (e.g., dataflow token or tokens). In one embodiment, the input storage (queue) status value from path 3309 for input storage 3324 and the input storage (queue) status value from path 3311 for input storage 3326 is steered to the operation circuitry 3325 and/or operation circuitry 3327 (e.g., along with the input data from the input storage(s) that is to be operated on) by multiplexer 3321 and multiplexer 3323.
In the depicted embodiment, the output storage (queue) status value from path 3313 indicates (e.g., by asserting a “not full” indication value or a “full” indication value) when output storage 3334 has available storage for (e.g., new) output data (e.g., as indicated by a backpressure token or tokens) and the output storage (queue) status value from path 3315 indicates (e.g., by asserting a “not full” indication value or a “full” indication value) when output storage 3336 has available storage for (e.g., new) output data (e.g., as indicated by a backpressure token or tokens). In the depicted embodiment, operation configuration value (or a portion thereof) is sent to both multiplexer 3341 and multiplexer 3343 to source the output storage (queue) status value(s) from the output controllers 3305 and/or 3307. In certain embodiments, operation configuration value includes a bit or bits to cause a first output storage status value to be asserted, where the first output storage status value indicates the output storage (queue) is not full or a second, different output storage status value to be asserted, where the second output storage status value indicates the output storage (queue) is full. The first output storage status value (e.g., “not full”) or second output storage status value (e.g., “full”) may be output from output controller 3305 and/or output controller 3307, e.g., as discussed below. In one embodiment, a first output storage status value (e.g., “not full”) is sent to the operation circuitry 3325 and/or operation circuitry 3327 to cause the operation circuitry 3325 and/or operation circuitry 3327, respectively, to perform the programmed operation when an input value is available in input storage (queue) and a second output storage status value (e.g., “full”) is sent to the operation circuitry 3325 and/or operation circuitry 3327 to cause the operation circuitry 3325 and/or operation circuitry 3327, respectively, to not perform the programmed operation even when an input value is available in input storage (queue).
In the depicted embodiment, dequeue (e.g., conditional dequeue) multiplexers 3329 and 3331 are included to cause a dequeue (e.g., removal) of a value (e.g., token) from a respective input storage (queue), e.g., based on operation completion by operation circuitry 3325 and/or operation circuitry 3327. The operation configuration value includes a bit or bits to cause the dequeue (e.g., conditional dequeue) multiplexers 3329 and 3331 to dequeue (e.g., remove) a value (e.g., token) from a respective input storage (queue). In the depicted embodiment, enqueue (e.g., conditional enqueue) multiplexers 3333 and 3335 are included to cause an enqueue (e.g., insertion) of a value (e.g., token) into a respective output storage (queue), e.g., based on operation completion by operation circuitry 3325 and/or operation circuitry 3327. The operation configuration value includes a bit or bits to cause the enqueue (e.g., conditional enqueue) multiplexers 3333 and 3335 to enqueue (e.g., insert) a value (e.g., token) into a respective output storage (queue).
Certain operations herein allow the manipulation of the control values sent to these queues, e.g., based on local values computed and/or stored in the PE.
In one embodiment, the dequeue multiplexers 3329 and 3331 are conditional dequeue multiplexers 3329 and 3331 that, when a programmed operation is performed, the consumption (e.g., dequeuing) of the input value from the input storage (queue) is conditionally performed. In one embodiment, the enqueue multiplexers 3333 and 3335 are conditional enqueue multiplexers 3333 and 3335 that, when a programmed operation is performed, the storing (e.g., enqueuing) of the output value for the programmed operation into the output storage (queue) is conditionally performed.
For example, as discussed herein, certain operations may make dequeuing (e.g., consumption) decisions for an input storage (queue) conditionally (e.g., based on token values) and/or enqueuing (e.g., output) decisions for an output storage (queue) conditionally (e.g., based on token values). An example of a conditional enqueue operation is a PredMerge operation that conditionally writes its outputs, so conditional enqueue multiplexer(s) will be swung, e.g., to store or not store the predmerge result into the appropriate output queue. An example of a conditional dequeue operation is a PredProp operation that conditionally reads its inputs, so conditional dequeue multiplexer(s) will be swung, e.g., to store or not store the predprop result into the appropriate input queue.
In certain embodiments, control input value (e.g., bit or bits) (e.g., a control token) is input into a respective, input storage (e.g., queue), for example, into a control input buffer as discussed herein (e.g., control input buffer 922 in
Input Controllers
Depicted input controller circuitry 3400 includes a Status determiner 3404, a Not Full determiner 3406, and a Not Empty determiner 3408. A determiner may be implemented in software or hardware. A hardware determiner may be a circuit implementation, for example, a logic circuit programmed to produce an output based on the inputs into the state machine(s) discussed below. Depicted (e.g., new) Status determiner 3404 includes a port coupled to queue status register 3402 to read and/or write to input queue status register 3402.
Depicted Status determiner 3404 includes a first input to receive a Valid value (e.g., a value indicating valid) from a transmitting component (e.g., an upstream PE) that indicates if (e.g., when) there is data (valid data) to be sent to the PE that includes input controller circuitry 3400. The Valid value may be referred to as a dataflow token. Depicted Status determiner 3404 includes a second input to receive a value or values from queue status register 3402 that represents that current status of the input queue that input controller circuitry 3400 is controlling. Optionally, Status determiner 3404 includes a third input to receive a value (from within the PE that includes input controller circuitry 3400) that indicates if (when) there is a conditional dequeue, e.g., from operation circuitry 3325 and/or operation circuitry 3327 in
As discussed further below, the depicted Status determiner 3404 includes a first output to send a value on path 3410 that will cause input data (transmitted to the input queue that input controller circuitry 3400 is controlling) to be enqueued into the input queue or not enqueued into the input queue. Depicted Status determiner 3404 includes a second output to send an updated value to be stored in queue status register 3402, e.g., where the updated value represents the updated status (e.g., head value, tail value, count value, or any combination thereof) of the input queue that input controller circuitry 3400 is controlling.
Input controller circuitry 3400 includes a Not Full determiner 3406 that determines a Not Full (e.g., Ready) value and outputs the Not Full value to a transmitting component (e.g., an upstream PE) to indicate if (e.g., when) there is storage space available for input data in the input queue being controlled by input controller circuitry 3400. The Not Full (e.g., Ready) value may be referred to as a backpressure token, e.g., a backpressure token from a receiving PE sent to a transmitting PE.
Input controller circuitry 3400 includes a Not Empty determiner 3408 that determines an input storage (queue) status value and outputs (e.g., on path 3309 or path 3311 in
For example, assume that the operation that is to be performed is to source data from both input storage 3324 and input storage 3326 in
Queue status register 3502 may store any combination of a head value (e.g., pointer) that represents the head (beginning) of the data stored in the queue, a tail value (e.g., pointer) that represents the tail (ending) of the data stored in the queue, and a count value that represents the number of (e.g., valid) values stored in the queue). For example, a count value may be an integer (e.g., two) where the queue is storing the number of values indicated by the integer (e.g., storing two values in the queue). The capacity of data (e.g., storage slots for data, e.g., for data elements) in a queue may be preselected (e.g., during programming), for example, depending on the total bit capacity of the queue and the number of bits in each element. Queue status register 3502 may be updated with the initial values, e.g., during configuration time. Queue status register 3502 may be updated as discussed in reference to
Output Controllers
Depicted output controller circuitry 4400 includes a Status determiner 4404, a Not Full determiner 4406, and a Not Empty determiner 4408. A determiner may be implemented in software or hardware. A hardware determiner may be a circuit implementation, for example, a logic circuit programmed to produce an output based on the inputs into the state machine(s) discussed below. Depicted (e.g., new) Status determiner 4404 includes a port coupled to queue status register 4402 to read and/or write to output queue status register 4402.
Depicted Status determiner 4404 includes a first input to receive a Ready value from a receiving component (e.g., a downstream PE) that indicates if (e.g., when) there is space (e.g., in an input queue thereof) for new data to be sent to the PE. In certain embodiments, the Ready value from the receiving component is sent by an input controller that includes input controller circuitry 3400 in
As discussed further below, the depicted Status determiner 4404 includes a first output to send a value on path 4410 that will cause output data (sent to the output queue that output controller circuitry 4400 is controlling) to be enqueued into the output queue or not enqueued into the output queue. Depicted Status determiner 4404 includes a second output to send an updated value to be stored in queue status register 4402, e.g., where the updated value represents the updated status (e.g., head value, tail value, count value, or any combination thereof) of the output queue that output controller circuitry 4400 is controlling.
Output controller circuitry 4400 includes a Not Full determiner 4406 that determines a Not Full (e.g., Ready) value and outputs the Not Full value, e.g., within the PE that includes output controller circuitry 4400, to indicate if (e.g., when) there is storage space available for output data in the output queue being controlled by output controller circuitry 4400. In one embodiment, for an output queue of a PE, a Not Full value that indicates there is no storage space available in that output queue is to cause a stall of execution of the PE (e.g., stall execution that is to cause a resultant to be stored into the storage space) until storage space is available (e.g., and when there is available data in the input queue(s) being sourced from in that PE).
Output controller circuitry 4400 includes a Not Empty determiner 4408 that determines an output storage (queue) status value and outputs (e.g., on path 3345 or path 3347 in
For example, assume that the operation that is to be performed is to send (e.g., sink) data into both output storage 3334 and output storage 3336 in
Queue status register 4502 may store any combination of a head value (e.g., pointer) that represents the head (beginning) of the data stored in the queue, a tail value (e.g., pointer) that represents the tail (ending) of the data stored in the queue, and a count value that represents the number of (e.g., valid) values stored in the queue). For example, a count value may be an integer (e.g., two) where the queue is storing the number of values indicated by the integer (e.g., storing two values in the queue). The capacity of data (e.g., storage slots for data, e.g., for data elements) in a queue may be preselected (e.g., during programming), for example, depending on the total bit capacity of the queue and the number of bits in each element. Queue status register 4502 may be updated with the initial values, e.g., during configuration time. Queue status register 4502 may be updated as discussed in reference to
In certain embodiments, a state machine includes a plurality of single bit width input values (e.g., 0s or 1s), and produces a single output value that has a single bit width (e.g., a 0 or a 1).
In certain embodiments, a first LIC channel may be formed between an output of a first PE to an input of a second PE, and a second LIC channel may be formed between an output of the second PE and an input of a third PE. As an example, a ready value may be sent on a first path of a LIC channel by a receiving PE to a transmitting PE and a valid value may be sent on a second path of the LIC channel by the transmitting PE to the receiving PE. As an example, see
3.3 Example Communications (e.g., Circuit Switched Network)
In certain embodiments, multiple PEs are coupled together by a network to send data, e.g., data that includes ready values, valid values, and the payload data itself. As discussed herein, a dataflow graph is mapped directly to a CSA that includes multiple PEs coupled together by a circuit switched network in certain embodiments. In certain embodiments, the lowest level of the CSA communications hierarchy is the local network. In one embodiment, the local network is statically circuit switched, using configuration registers to swing multiplexor in the local network data-path, forming fixed electrical paths between communicating PEs. In one embodiment, the configuration of the local network is set once per dataflow graph at the same time as the PEs configuration. In one embodiment, a static, circuit switched network optimizes for energy, for example, where a large majority (e.g., greater than about 95%) of CSA communications traffic will cross the local network. As certain dataflow graphs include terms which are used in multiple expressions, certain embodiments herein include hardware support for multicast within the local network.
In certain embodiments, several local networks are ganged together to form routing channels which are interspersed between rows and columns of PEs. In one embodiment, several one-bit local networks are also included to carry control tokens. In contrast to a FPGA, embodiments of the CSA local network are routed at the granularity of the data path and the CSA architecture includes a novel treatment of control. In certain embodiments, the CSA local network is explicitly flow controlled (e.g., back pressured), that is, for each forward data path (e.g., and multiplexor) set, the CSA provides a backward-flowing flow control path that is physically paired with the forward data path. The combination of the two micro architectural paths provides a low-latency, low-energy, low-area, point-to-point implementation of the latency-insensitive channel abstraction in certain embodiments. In addition to point-to-point communications, certain embodiments of a CSA local network also support multicast, in which a single source sends a value to a plurality of downstream PEs. This functionality may be an extension of the point-to-point control logic combined with a multicast configuration state as discussed herein.
In certain embodiments, the CSA flow control lines are not visible to the user program, but they are manipulated by the architecture in service of the user program. For example, exception handling mechanisms may be achieved by pulling flow control lines to a “not present” state upon the detection of an exceptional condition. In one embodiment, this action not only gracefully stalls those parts of the pipeline which are involved in the offending computation, but also preserves the machine state leading up the exception for diagnostic analysis.
To enable a broad set of compiler-generated codes, certain embodiments of the CSA architecture support many control expressions. As a result, CSA dataflow graphs may often include a substantial number of Boolean values (e.g., a single bit zero for false and a single bit one for true), for example, the results of conditional or loop expressions. To decrease the overhead of these data flows, certain embodiments of the CSA provide a number of one-bit networks, e.g., in addition to the wider (number of bits) networks used to carry (e.g., arithmetic) multiple bit data types.
In certain embodiments, a CSA includes a second network layer (e.g., referred to as the mezzanine network) that is a shared, packet-switched network. In certain embodiments, the mezzanine provides more general, long range communications at the cost of latency, bandwidth, and energy. In well-routed programs, in certain embodiments, most communications will occur on the local network and the mezzanine network provisioning will be considerably reduced in comparison, e.g., where each PE connects to multiple local networks, but is provisioned with only one mezzanine endpoint per logical grouping (e.g., “neighborhood”) of PEs. Since the mezzanine is effectively a shared network, in certain embodiments each mezzanine network carries multiple logically independent channels, e.g., it is provisioned with multiple virtual channels. In certain embodiments, the main function of the mezzanine network is to provide long-range communications between PEs and between PEs and memory. The mezzanine may operate as a runtime support network, e.g., by which various services can access the complete fabric in a user-program-transparent manner. In this capacity, the mezzanine endpoint may function as a controller for its local neighborhood, for example, during CSA configuration.
To form channels spanning a CSA tile, as in the example shown in
In certain embodiments, the routing of data between components (e.g., PEs) is enabled by setting switches (e.g., multiplexers and/or demultiplexers) and/or logic gate circuits of a circuit switched network (e.g., a local network) to achieve a desired configuration, e.g., a configuration according to a dataflow graph.
The network(s) may be statically configured, e.g., in addition to PEs being statically configured during configuration for a dataflow graph. During the configuration step, configuration bits may be set at each network component. These bits may control, for example, the multiplexer selections to control the flow of a dataflow token (e.g., on a data path network) and its corresponding backpressure token (e.g., on a flow control path network). A network may comprise a plurality of networks, e.g., a data path network and a flow control path network. A network or plurality of networks may utilize paths of different widths (e.g., a first width, and a narrower or wider second width). In one embodiment, a data path network has a wider (e.g., bit transport) width than the width of a flow control path network. In one embodiment, each of a first network and a second network includes their own data paths and flow control paths, e.g., data path A and flow control path A and wider data path B and flow control path B. For example, a data path and flow control path for a single output buffer of a producer PE that couples to a plurality of input buffers of consumer PEs. In one embodiment, to improve routing bandwidth, several networks are laid out in parallel between rows of PEs. Like certain PEs, the network may be statically configured. During this step, configuration bits may be set at each network component. These bits control, for example, the data path (e.g., multiplexer created data path) and/or flow control path (e.g., multiplexer created flow control path). The forward (e.g., data) path may utilize control bits to swing its switches and/or logic gates.
In the zoomed in portion, multiplexer 5608 is provided with a configuration value from configuration storage (e.g., register) 5606 to cause the multiplexer 5608 to source data from one of more inputs (e.g., with those inputs being coupled to respective PEs or other CSA components). In one embodiment, an (e.g., each) input to multiplexer 5608 includes both (i) multiple bits of (e.g., payload) data as well as (ii) a (e.g., one bit) valid value, e.g., as discussed herein. In certain embodiments, the configuration value is stored into configuration storage locations (e.g., registers) to cause a transmitting PE or PEs to send data to receiving PE or PEs, e.g., according to a dataflow graph. Example configuration of a CSA is discussed further in Section 3.4 below.
In the zoomed in portion, OR logic gate 5710, OR logic gate 5712, and OR logic gate 5714 each include a first input coupled to configuration storage (e.g., register) 5706 to receive a configuration value (for example, where setting a logical one on that input effectively ignores the particular backpressure signal and a logical zero on that input cause the monitoring of that particular backpressure signal), and a second input coupled to a respective, receiving PE to provide a backpressure value that indicates when that receiving PE is not ready to receive a new data value (e.g., when a queue of that receiving PE is full). In the depicted embodiment, the output from each OR logic gate 5710, OR logic gate 5712, and OR logic gate 5714 is provided as a respective input to AND logic gate 5708 such that AND logic gate 5708 is to output a logical zero unless all of OR logic gate 5710, OR logic gate 5712, and OR logic gate 5714 are outputting a logical one, and AND logic gate 5708 will then output a logical one (e.g., to indicate that each of the monitored PEs are ready to receive a new data value). In one embodiment, an (e.g., each) input to OR logic gate 5710, OR logic gate 5712, and OR logic gate 5714 is a single bit. In certain embodiments, the configuration value is stored into configuration storage locations (e.g., registers) to cause a transmitting PE or PEs to send flow control (e.g., backpressure) data to transmitting PE or PEs, e.g., according to a dataflow graph. In one multicast embodiment, a (e.g., single) flow control (e.g., backpressure) value indicates that at least one of a plurality of receiving PEs does not have storage and thus is not ready to receive (e.g., payload) data that is to be transmitted, e.g., by ANDing the outputs from OR logic gate 5710, OR logic gate 5712, and OR logic gate 5714. Example configuration of a CSA is discussed further in Section 3.4 below.
3.4 Configuration of a CSA (e.g., PEs and Circuit Switched Network)
In certain embodiments, a CSA (e.g., PEs and a circuit switched network) is configured by setting one or more configuration values in one or more configuration storage locations (e.g., registers). For example, a (e.g., local) circuit switched network may be configured to provide path(s) to send and/or receive data between PEs (or between a PE and another CSA component(s)). In one embodiment, a compiler is to generate the configuration values (e.g., for PEs, for circuit switched networks, and/or for other CSA components) that overlay a dataflow graph to the dataflow architecture of a CSA. In certain embodiments, a (e.g., each) PE is a dataflow operator that is a direct representation of a node (e.g., or two nodes) in a dataflow graph. In certain embodiments, the circuit switched networks are configured with configuration values generated by a compiler to minimize the distance of paths between PEs that are transmitting data to receiving PEs. In certain embodiments, the circuit switched networks are configured with configuration values generated by a compiler to minimize the area utilized by a dataflow graph, e.g., by PEs that are transmitting data to receiving PEs according to that dataflow graph. In certain embodiments, the circuit switched networks are configured with configuration values generated by a compiler to minimize the data transfer latency between PEs that are transmitting data to receiving PEs. A circuit switched network may be a local network. A local network may further communicate via a packet switched network.
Section 7.1 discloses examples of how to configure a CSA (e.g., the PEs and the circuit switched network(s)). Embodiments of a CSA (e.g., fabric) may differ from traditional cores in that embodiments of a CSA use a configuration step in which the PEs and the (e.g., circuit switched) network are loaded with program configuration in advance of program execution.
In one embodiment, the CSA configuration protocol is for the PEs and the local, circuit switched network. In certain embodiments, a request for CSA configuration (e.g., the configuration code) arrives from a host (e.g., core of a processor that is coupled to the CSA). In one embodiment, the configuration (e.g., configuration values) are sent into the PEs and circuit switched network by configuration controllers, e.g., as discussed below. In certain embodiments, these controllers stream in graph configuration information and execute the local configuration protocol across their domains. Local configuration controllers may operate in parallel, e.g., decreasing the latency of the configuration operation.
The core of one embodiment of CSA configuration is the distributed protocol driven by the local configuration controller. In one embodiment, initially, configuration state (e.g., configuration values) resides in memory, and the local configuration controller receives a virtual pointer which points to a memory region containing the CSA dataflow graph. The PEs and network resources in the local neighborhood of the local configuration controller are put into an un-configured state in one embodiment. In certain embodiments of this state, all control signals associated with the local network in the local neighborhood are deactivated, effectively halting all communications within the local neighborhood and between the local neighborhood and other adjacent PEs. The local configuration controller then streams new configuration in to the PEs, initializing one at a time in a distributed fashion in one embodiment. As discussed further below,
In certain embodiments, the CSA program graph loaded at configuration time consists of both configuration values and data, either constants to be loaded in to the fabric or the prior execution state of a fabric, for example, as a result of an extraction operation described herein. In certain embodiments, CSA program state resides within the virtual memory space of the process associated with the CSA and may be resident within the CSA memory hierarchy or within die-level memory hierarchy. The performance of the configuration mechanism may be strongly influenced by the locality of the graph configuration.
3.5 Example Operation Format
The term “CSA program” may generally refer to a collection of operations and communication channels definitions that are configured (e.g., loaded) onto the components (e.g., PEs) and network (e.g., circuit switched network) of a CSA hardware instance. In one embodiment, once configuration is complete, the CSA program (e.g., representing a dataflow graph) is executed a plurality of times without reconfiguration, e.g., provided the CSA resources used for the program is loaded are not needed for another program between. In certain embodiments, the routing of communications (e.g., via setting up LICs) is a property of configuration and not changed during the execution of a program.
As discussed herein, in certain embodiments, a dataflow graph is overlaid on a CSA so that the CSA performs operations of the dataflow graph. The operations may include a format as discussed below. Data type(s) used in operations may be as discussed in reference to Table 2 herein.
In one embodiment, code may be written (e.g., by a programmer) that includes one or more of the operations discussed herein, e.g., according to the following format(s). In another embodiment, code is written in a first software language (e.g., C or C++ code), and then converted by an assembler into assembly code. In one embodiment, the assembly code includes operations written in the operation format(s) discussed herein. In certain embodiments, the operations correspond to configuration values (e.g., for PEs, for circuit switched networks, and/or for other CSA components) that overlay the dataflow graph on the dataflow architecture of a CSA. In one embodiment, the assembly code for a (e.g., proper subset of a) dataflow graph is further modified by a place and route tool that assigns an (e.g., each) operation to a particular hardware instance (e.g., a PE) of the CSA hardware.
Operands
In certain embodiments, there are 3 basic types of entities that may be (e.g., input and/or output) operands to a CSA operation: (i) latency insensitive channels (LICs), (ii) registers, and (iii) literal values. In one embodiment, the size of literals is the size of the operand supported on PEs or other dataflow units, e.g. a 64 bit (64b) operand having a full 64b literal.
The format (e.g., signatures) of operations in the descriptions that follow use the following form: [{name}.] {operand type} {uld}.{data type} [={default value}]. The first part is an optional operand name (e.g., “res.” for a resultant or “ctlseq.” for a control sequence). Next is the operand type, where characters C (Channel), R (Register) or L (Literal) specify what operand types are valid. If there is a d suffix, the operand is an output that is defined, while a u suffix means it is an input that is used. Next is a data type, which reflects the usage in the operation.
For example, res.CRd.s32 means that the operand is called res, it can either a channel (C) or register (R), it is defined (d) by the operation (e.g., it is an output), and uses 32 bits of input, which it treats inside the operation as being signed. Note that this does not mean that input channels smaller than 32 bits are sign extended, although sign extension may be optionally included.
Operands may have default values, denoted by ={default value}, allowing various trailing operands to be omitted in assembly code. This is shown for a given operand description by an=with a default value. Value can be: (i) a numeric value, which is that value (e.g. op2.CRLu.i1=1 means a default value of 1), (ii) the letter I means % ign—ignored/reads as 0, (iii) the letter N means % na—never available, either as input or output (e.g., % na in a field means that field is not utilized for that operation), (iv) the letter R means rounding mode literal ROUND NEAREST, and (v) the letter T means memory level literal MEMLEVEL+T0 (e.g., closest cache).
In the opcode description semantics, semicolons imply sequencing. If an operand appears by itself, the operation waits for the value to be available. e.g. for memrefs: op2; write(op0,op1); op3=0 means that the operation waits for op2 to be available, performs its access, and then defines op3. The following modifiers can appear for operands: non-consuming use (specified via a “*” prefix in the assembly code). This applies to any storage with empty/full semantics (e.g., LICs, and/or registers), and specifies that the operand is to be reused in the future.
Operation Naming Notes
In one embodiment, integer operations that do not care about signed-ness (e.g. and, add, cmpeq) are named based on the number of bits processed in the operation, and the corresponding output size (e.g. and32, add32). For cases where signed vs. unsigned matter, sN or uN specifies the signed (s) or unsigned (u) integer type (e.g. divu32, cmplts8). Floating point (f) data types are fN (e.g. f32/f64) (e.g. addf32). In certain embodiments, composite operations are named for the order of processing (e.g. fused multiply add=>fma, sll+add=>sladd). In certain embodiments, conversions (cvt) are named cvt{dsttype} {srctype} (e.g., “convert to xxx from yyy”, and the output size is the first type size).
Operand Ordering and Style
When there is a selector among operands (e.g. pick*, switch*), in certain embodiments, 0 is used for the 1st, 1 for the 2nd, etc. A selector may include 2 or 3 operands and a single bit of control, but there is a possibility of higher radix picks/switches (e.g., those with more or many more than two inputs or two outputs.
In certain embodiments, output operands precede input operands. In one embodiment, an exception is the memory ordering operands for memory references have both the output and input following the main operands. For memory references in this embodiment, the operands are ordered as if they were move operations that take more general operands, e.g., ld {target}, {memaddr} while store is st {memaddr}, {source}.) Further, in certain embodiments, displacement or index operands follow the base address operand, e.g., ldNx {target},{addr}, {index} vs. stNx {addr},{index},{source}. Note that (e.g., many) operations may allow defaulting of later operands.
Mixed Operation/Operand Size Semantics
In certain embodiments, a first rule is that a CSA operation's defined semantics require size consistency between operands and LICs, e.g., and use an explicit size conversion when a size change was involved. In one embodiment, if an input value (e.g. from an LIC) has a smaller number of bits than the corresponding input operand, it is automatically zero-extended to the width required, e.g., a comparison generates a single bit output, and using that as the input to an and64 operation will cause it to be zero-extended up through bit 63. Likewise, in certain embodiments, if an output value is a smaller number of bits than the consuming LIC, it is zero-extended, e.g., an add32 operation writing a 64b output only produces non-zero values in the low order 32 bits.
In certain embodiments, a second rule is that if an output value is larger than an output channel, the value is truncated to that many bits, e.g., it functions like a store to memory. For example, .lic .i32 c1; add64 c1 . . . ; add64, c1 would cause the 64b result from the first add to be truncated to 32b before being presented as the input to the second add.
In certain embodiments, a third rule is that the generated output is the size specified on the operation, e.g., an “add32” add operation generates 32b, and a “ld8” load operation generates 8 bits. In certain embodiments, CSA hardware detects when a smaller operation could be used because a smaller number of output bits are required, e.g., if an and32 is used to generate a 1 bit channel, only 1 bit is to be generated. Note that the first rule and the second rule mean that the bits semantics of a LIC matches a store x (“stx”) followed by a load x (“ldx”), where x is the bit size of the LIC. However, note that an arbitrary store/load would not provide ordering in certain embodiments, e.g., that would require memory with full/empty semantics.
Toolchain Modification of Code
In certain embodiments, optimization of hardware assignment happens in the compiler. However, in those embodiments, some decisions may be made after the assembly representation of the dataflow graph. Some examples of transformations are described below.
Expansion/Fission
Some single operations may be expanded to a sequence of two or more operations, for example, large (e.g., greater than 64 bit input operands) integer multiply, integer and floating point division, math functions such as square root, displacement and indexing for memory references, some variations of streaming memory references, etc. Implementations may also have operations inserted for handling of cases like mismatched sizes. For example, some implementations may not allow different sizes of network connections to operands, so performing an add using the result of a comparison may involve an operation to change networks.
Fusion
In certain embodiments, there are a number of cases where dataflow operations, particularly including pick, switch and repeat, are implemented in the underlying hardware without requiring use of an entire PE, for example, a switch as an output operand, a repeat as an input operand, and a pick as an input operand.
The CSA operations may each include a (e.g., unique for each operation type) configuration value, that when loaded into a PE or other CSA component (e.g., registers that control a circuit switched network), causes the PE or other CSA component to perform the desired CSA operation. As a non-limiting example, an add operation may include the format of:
add {8-64} res.Ld.iN, opl.LCu.iN, op2L.Cu.iN
such that the resultant (res) is equal to the first operand (op1) added to the second operand (op2). In reference to
3.6 Example CSA Operations
The following are examples of CSA operations. Hardware (e.g., a CSA) may perform one or more of the following operations, e.g., via a processing element. CSA operations may include arithmetic and/or logical operations, e.g., with one or a plurality of (e.g., 0 to 3) inputs and one or a plurality of (e.g., 0 to 1) outputs. In contrast to other architectures, the operands in certain embodiments of CSA are channels, registers, or literals. CSA operations may also include families of operations related to dataflow, sequence processing, reductions, etc. In certain embodiments, conversion operations are provided between floating types, and between a first size (e.g., 32b) and a second size (e.g., 64b) signed or unsigned integer and/or floating point types of data. In one embodiment, (e.g., most) integer operations are provided in 8, 16, 32, and 64b widths, and single bit (e.g., control data) as well. Note that although certain buffers are discussed as being used to provide input values and to stored output values, those buffers are merely examples and the particular buffer or buffers used for an operation may be selected (e.g., via setting the configuration value accordingly).
In certain embodiments, each (e.g., single) operation is performed by a single PE configured via a configuration value being set, e.g., in a register of that PE, to a value corresponding to that operation. In certain embodiments, a CSA (e.g., a PE thereof) does not change its function each clock cycle. In certain embodiments, a CSA (e.g., a PE thereof) does not receive bits of instruction from a centralized memory (e.g., an element instruction stream memory) during execution. In certain embodiments, a CSA (e.g., a PE thereof) does not change its function based upon bits of instruction received from a centralized memory (e.g., an element instruction stream memory) during execution. In certain embodiments, a CSA (e.g., a PE thereof) does not receive programming (e.g., configuration values) each execution cycle. In certain embodiments, a CSA does not utilize algorithms stored in a centralized memory and access them before each operation. In certain embodiments, a CSA (e.g., a PE thereof) executes (e.g., only) when input data is available and storage for a resultant(s) is available, e.g., in certain embodiments a CSA does not execute based only on a clock cycling (e.g., for a predetermined number of cycles). In certain embodiments, a CSA (e.g., a PE thereof) stores state information locally (for example, in queues and/or registers of the CSA element (e.g., PE)), and not in a centralized repository of state memory.
The following discusses examples of certain CSA operations, including certain streaming operations, Boolean control operations, dataflow operations, storage (buffer) operations, and fountain operations, and then includes a table of other CSA operations. The following operations are discussed in reference to a PE having one or more (e.g., all) of the components of PE 5800 in
Note that in certain PEs herein, a configuration register includes storage for multiple operation configuration values. In any of these embodiments, a PE may only include storage for a single operation configuration value, for example, with the operation configuration value controlling which operation circuitry is used. See, for example,
Input (e.g., control) queues 5804, 5806, and 5822 are coupled to local network(s) 5802 (e.g., and local network 5802 may include a data path network as in
In certain embodiments, status register 5838 is loaded whenever the ALU (or other operations circuitry) 5818 executes (also controlled by output of multiplexer 5816). In one embodiment, data in control input queues 5804, 5806, 5822, and/or control output queue 5832 is a single bit. In the depicted embodiment, multiplexer 5821 (e.g., operand A) and multiplexer 5823 (e.g., operand B) sources inputs, e.g., according to the configuration value.
For example, suppose the operation of this processing (e.g., compute) element is (or includes) what is called call a pick in
For example, suppose the operation of this processing (e.g., compute) element is (or includes) what is called call a switch in
In certain embodiments, multiple networks (e.g., LICs thereof) are connected to a processing element, e.g., (input) network(s) 5802 and (output) network(s) 5812. The connections may be switches, e.g., as discussed in reference to
Note that certain operations of this disclosure include a combination of inputs (e.g., from queues of a PE performing the operation), but in certain embodiments, a PE only stalls when certain proper subset of the inputs is available instead of requiring all of the inputs be available. The proper subset of inputs determining the stall may be chosen based on the combination of the value of particular inputs to the operation, the value of status storage associated with the operation, and the PE configuration. In one embodiment, a pick operation that is to pick data from a first input queue or a second input queue is not to stall when the second input queue is empty if the pick operation is currently picking from the first input queue that includes at least one value.
Note that certain operations of this disclosure include a combination of outputs (e.g., from queues of a PE performing the operation), but in certain embodiments, a PE only stalls when certain proper subset of the outputs are not full (e.g. available to accept new data) instead of requiring all of the outputs. The proper subset of outputs determining the stall may be chosen based on the combination of the value of particular inputs to the operation, the value of status storage associated with the operation, and the PE configuration. In one embodiment, a switch operation that is to steer data from a first input queue to a first output queue or a second output queue is not to stall when the second output queue is full (e.g. not available to accept new data) if the switch operation is currently steering (e.g., sourcing) from the first input queue to the first output queue and the first output queue is not full (e.g. available to accept new data).
Data input queue 5824 and data input queue 5826 may perform similarly, e.g., local network 5802 (e.g., set up as a data (as opposed to control) interconnect) being switched (e.g., connected) to data input queue 5824. In this embodiment, a data path (e.g., network as in
A processing element 5800 may be stalled from execution until its operands (e.g., a control input value and its corresponding data input value or values) are received and/or until there is room in the output queue(s) of the processing element 5800 for the data that is to be produced by the execution of the operation on those operands.
Streaming Operations
In certain embodiments, dataflow architectures operate on scalar values. However, in some cases, it may be more efficient to process streams of data (e.g., aggregations of dataflow tokens). This allows natural management of irregular data and complex control, for example, when sorting lists or performing sparse matrix arithmetic. The section below describes several dataflow operations (e.g., and their dataflow operators in embodiments of a CSA) which facilitate the manipulation of streams. In one embodiment, Stream Compare (“stcmp”) allows the comparison of two streams of values (e.g., data values). This operation permits the merging of partially ordered lists, e.g., in merge sort and also in sparse matrix multiplication (e.g., where it is used to calculate the unions and intersections of sparse matrix rows and columns). Stream Pick (“stpick”) and Stream Switch (“stswitch”) allow for the steering of stream-based data. Is Null (“snull”) assists in controlling stream operations by checking the length of a stream object. In one embodiment, one or more (e.g., all) of these operations are sufficient to implement a large number of streaming algorithms.
In certain embodiments, streaming operations transform input streams into other, output stream(s).
Further, control data 5906 may be used to select which of the two input streams is to be output from the stream pick operation 5900, for example, a first value (e.g., zero) of control data 5906 to cause the stream pick operation 5900 to output the (e.g., entire) first input stream of data 5902, and a second value (e.g., a one) of control data 5906 to cause the stream pick operation 5900 to output the (e.g., entire) second input stream of data 5904. In the depicted embodiment, the first value received for control data 5906 is a one, which is to cause the stream pick operation 5900 to output 5908 the entire stream from second input stream of data 5904, and the second value received for control data 5906 is a zero, which is to then (e.g., after the completion of outputting the entire second input stream 5904) cause the stream pick operation 5900 to output 5908 the entire stream of the first input stream of data 5902. Like scalar dataflow operators, certain embodiments of stream operations (e.g., operators) execute when all operands are available, but operate on entire streams (e.g., the entire stream need not be available at the commencement of the streaming operation). In one embodiment, stream operations accept multiple streams and, optionally, additional control tokens which indicate an action to be taken on an entire stream. The following discusses examples of stream operations, e.g., where a PE is configured to perform a stream operation when its configuration value is set accordingly.
Stream Compare
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Stream Compare operation according to the following (e.g., semantics and/or description).
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein.
In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc.
In
In
In
In
In certain embodiments, PE 6100 is stalled from performing the comparison operation until there is both (i) space available in the output queues that are to be used for storing resultant data, and (ii) input data for each stream (e.g., control data of a Boolean one and the associated payload data for a stream, or control data of a Boolean zero for the end of a stream).
In certain embodiments, PE 6100 removes input data (e.g., tokens) subject to a conditional comparison, e.g., where stream compare allows merging of ordered streams. In one embodiment, PE 6100 emits optional Boolean control (e.g., to queue 6132) to be used as control values for other PEs.
In the depicted embodiment, PE 6100 includes the components of PE 5800 from
Strewn Pick
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Stream Pick operation according to the following (e.g., semantics and/or description).
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein.
In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc.
In
In
In
In
In
In
In
In
In certain embodiments, PE 6200 is stalled from performing the pick operation until there is both (i) space available in the output queues that are to be used for storing resultant data, and (ii) pick control value (e.g., selection control bit) and input data for the picked stream.
In certain embodiments, PE 6200 selects a single stream from a pair of streams and copies the entire, single stream to the output using a predicate (e.g., a selection control value) to control the selection.
In the depicted embodiment, PE 6200 includes the components of PE 5800 from
Stream Switch
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Stream Switch operation according to the following (e.g., semantics and/or description).
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein.
In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc.
In
In
In
In
In
In
In
In
In
In certain embodiments, PE 6300 is stalled from performing the switch operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) the switch control value (e.g., selection control bit) and input data for the switched stream.
In certain embodiments, PE 6300 steers a single stream to one of a plurality of outputs by using a predicate (e.g., a selection control value) to control the selection.
In the depicted embodiment, PE 6300 includes the components of PE 5800 from
Is Null
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform an Is Null (snull) operation according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In certain embodiments, PE 6400 is stalled from performing the IsNull operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) an input control value in input queue 6422.
In the depicted embodiment, PE 6400 includes the components of PE 5800 from
Stream Split
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Stream Split operation according to the following (e.g., semantics and/or description).
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein.
In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc.
In
In
In
The data value from the output queue 6534 and the associated control data from the control queue 6544 may be consumed from the output queues, e.g., by a downstream PE or PEs.
In
In
In
In
In
In certain embodiments, PE 6500 is stalled from performing the split operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) the split control value (e.g., selection control bit) and input data for the split stream.
In certain embodiments, PE 6500 steers a single stream to one of a plurality of outputs by using a predicate (e.g., a split control value) (and optionally, a count) to control the selection, for example, moving the first two elements of an input stream to a first output and the remainder of the input stream to a second output. In certain embodiments, the data stream is shorter than the predicate stream and the extra predicates are discarded.
In the depicted embodiment, PE 6500 includes the components of PE 5800 from
In one embodiment, a PE performing a stream split operation is to send outputs (e.g., tokens) (e.g., store data into the PEs output queues) only at termination of both predicate and input stream. In another embodiment, e.g., to avoid stalling while waiting for these values, state storage is added to track when a stream termination value (e.g., Boolean zero) has been sent. In one embodiment, only one of a predicate stream and a data stream terminate first, so a PE utilized a single bit of state storage, although a plurality of bits of state storage may be used.
As one example, input queue (e.g., having a single bit width) 6604 is provided to receive a stream control value (e.g., token) for one of (i) input queue 6624 (for example, having a multiple bit width, e.g., 8, 16, 32, or 64 as shown above in the example stream comparison opcodes) or (ii) input queue (e.g., having a single bit width) 6606 is provided to receive a stream control value (e.g., token) for input queue 6626 (for example, having a multiple bit width, e.g., 8, 16, 32, or 64 as shown above in the example stream comparison opcodes). In
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein.
In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc.
In
In
In
The data value from the output queue 6634 and the associated control data from the control queue 6644 may be consumed from the output queues, e.g., by a downstream PE or PEs.
In
In
In
In the depicted embodiment, the Stream Split operation causes PE 6400 to produce a Boolean value (e.g., zero) internally in state storage 3.6605 to track that the end of predicate stream has been encountered (e.g., the zero in the first slot of (e.g., split control) input queue 6622) and a stream termination value (e.g., zero) for first output queue 6634 has been sent (e.g., stored) to control output queue 6644. In this embodiment, the zero in the first slot of (e.g., split control) input queue 6622 is cleared. In certain embodiments, the state storage is set to a Boolean one before each new operation is performed.
In
In
In certain embodiments, PE 6600 is stalled from performing the split operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) the split control value (e.g., selection control bit) and input data for the split stream.
In certain embodiments, PE 6600 steers a single stream to one of a plurality of outputs by using a predicate (e.g., a split control value) (and optionally, a count) to control the selection, for example, moving the first two elements of an input stream to a first output and the remainder of the input stream to a second output. In certain embodiments, the data stream is shorter than the predicate stream and the extra predicates are discarded.
In the depicted embodiment, PE 6600 includes the components of PE 5800 from
Stream Combine (SCMB)
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Stream Combine (SCMB) operation according to the following (e.g., semantics and/or description).
In
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein.
In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc.
In
As the data value of (e.g., integer) two in the first slot of input queue 6726 is dequeued (e.g., deleted) from the first slot of input queue 6726 and no data value is stored in a second slot of input queue 6726, then no data value is stored into the first slot of input queue 6726 but a Boolean zero is moved into the first slot from the second slot of the associated (e.g., control) input queue 6706 to indicate that data value two in the first slot of input queue 6726 is the end of that stream.
As the data value of (e.g., integer) two in the first slot of input queue 6724 is dequeued (e.g., deleted) from the first slot of input queue 6724, the data value of (e.g., integer) four is moved into the first slot from the second slot of input queue 6724 along with the Boolean one moved into the first slot from the second slot of the associated (e.g., control) input queue 6704 to indicate that data value is a valid value of the stream. In
The stream equal (or not equal) value from the (e.g., control) output queue 6732 and the control data from the control queues (e.g., 6744 and/or 6746) may be consumed from the output queues, e.g., by a downstream PE or PEs.
In
As the data value of (e.g., integer) four in the first slot of input queue 6724 is dequeued (e.g., deleted) from the first slot of input queue 6724 and no data value is stored in a second slot of input queue 6724, then no data value is stored into the first slot of input queue 6724 but a Boolean zero is moved into the first slot from the second slot of the associated (e.g., control) input queue 6704 to indicate that data value four from the first slot of input queue 6724 was the end of that stream.
The stream not equal (or equal) value from the (e.g., control) output queue 6732 and the control data from the control queues (e.g., 6744 and/or 6746) may be consumed from the output queues, e.g., by a downstream PE or PEs.
In
In certain embodiments, PE 6700 is stalled from performing the stream combine operation until there is both (i) space available in the output queues that are to be used for storing resultant data, and (ii) input data for each stream (e.g., control data of a Boolean one and the associated payload data for a stream, or control data of a Boolean zero for the end of a stream).
In certain embodiments, PE 6700 prepares “streams are equal” value or “streams are not equal” value and/or control values for combination of streams (e.g., by the Union operation or Inter operation discussed next). In one embodiment, both the first slot of (e.g., control) output queue 6744 and the first slot of (e.g., control) output queue 6746 storing a zero indicates the end of both streams (e.g., the end of the stream combine operation).
In the depicted embodiment, PE 6700 includes the components of PE 5800 from
Union
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Union operation according to the following (e.g., semantics and/or description).
In the depicted embodiment, input queue (e.g., having a single bit width) 6804 is provided to receive stream combine (SCMB) values (e.g., tokens) for input queue 6824 (for example, having a multiple bit width, e.g., 8, 16, 32, or 64) and input queue (e.g., having a single bit width) 6806 is provided to receive stream combine (SCMB) values (e.g., tokens) for input queue 6826 (for example, having a multiple bit width, e.g., 8, 16, 32, or 64). In
In
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein.
In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc.
In
The control value from the (e.g., control) output queue 6844 and the data value from output queue 6834 may be consumed from the output queues, e.g., by a downstream PE or PEs.
In
In
In certain embodiments, PE 6800 is stalled from performing the union operation until there is both (i) space available in the output queues that are to be used for storing resultant data, and (ii) input data for each stream (e.g., stream combine (SCMB) value of a Boolean one and the associated payload data for a stream, or stream combine (SCMB) value of a Boolean zero).
In certain embodiments, PE 6800 determines the union of two streams using SCMB generated control data and using the data values of the streams.
In the depicted embodiment, PE 6800 includes the components of PE 5800 from
Inter
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform an Inter operation according to the following (e.g., semantics and/or description).
In the depicted embodiment, input queue (e.g., having a single bit width) 6904 is provided to receive stream combine (SCMB) values (e.g., tokens) for input queue 6924 (for example, having a multiple bit width, e.g., 8, 16, 32, or 64) and input queue (e.g., having a single bit width) 6906 is provided to receive stream combine (SCMB) values (e.g., tokens) for input queue 6926 (for example, having a multiple bit width, e.g., 8, 16, 32, or 64). In
In
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein.
In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc.
In
The control value from the (e.g., control) output queue 6944 and the data value from output queue 6934 may be consumed from the output queues, e.g., by a downstream PE or PEs.
In
In
In certain embodiments, PE 6900 is stalled from performing the inter operation until there is both (i) space available in the output queues that are to be used for storing resultant data, and (ii) input data for each stream (e.g., stream combine (SCMB) value of a Boolean one and the associated payload data for a stream, or stream combine (SCMB) value of a Boolean zero).
In certain embodiments, PE 6900 determines the inter of two streams using SCMB generated control data and using the data values of the streams.
In the depicted embodiment, PE 6900 includes the components of PE 5800 from
Boolean Control Operations
As noted herein, one type of data is the data value (e.g., payload) and another type of data is control values. In certain embodiments, data values are transmitted on LICs (e.g., between PEs). Additionally, in certain embodiments, control values are transmitted on LICs (e.g., between PEs). The following discusses a plurality of Boolean control operations that may utilize control values.
NetAll0
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a NetAll0 operation according to the following (e.g., semantics and/or description).
In certain embodiments, NetAll0 ensures that a value is sent from each of a plurality of transmitting PEs to a single receiving PE. In one embodiment, the receiving PE outputs a control value (e.g., a one) when all corresponding values (e.g., both instances labeled 0s, 1s, or 2s, respectively) are collected in the transmitting PEs. The use of matched labels (e.g., a pair of 0s) is for explanation only, e.g., matching values (e.g., an integer zero and an integer zero) are not required during actual execution, only the presence of some value.
In one embodiment, a circuit switched network 7010 includes (i) a data path to send data from first PE 7000A to third PE 7000C and a data path from second PE 7000B to third PE 7000C, and (ii) a flow control path to send control values that controls (or is used to control) the sending of that data from first PE 7000A and second PE 7000B to third PE 7000C. Data path may send a data (e.g., valid) value when data is in an output queue (e.g., buffer) (e.g., when data is in control output buffer 7032A, first data output buffer 7034A, or second data output queue (e.g., buffer) 7036A of first PE 7000A and when data is in control output buffer 7032B, first data output buffer 7034B, or second data output queue (e.g., buffer) 7036B of second PE 7000B). In one embodiment, each output buffer includes its own data path, e.g., for its own data value from producer PE to consumer PE. Components in PE are examples, for example, a PE may include only a single (e.g., data) input buffer and/or a single (e.g., data) output buffer. Flow control path may send control data that controls (or is used to control) the sending of corresponding data from first PE 7000A and second PE 7000B to third PE 7000C. Flow control data may include a backpressure value from each consumer PE (or aggregated from all consumer PEs, e.g., with an AND logic gate). Flow control data may include a backpressure value, for example, indicating a buffer of the third PE 7000C that is to receive an input value is full. Flow control may include a value that indicates a netall0 operation has completed at third PE 7000C in a prior cycle.
Turning to the depicted PEs, processing elements 7000A-C include operation configuration registers 7019A-C that may be loaded during configuration (e.g., mapping) and specify the particular operation or operations (for example, to indicate whether to enable NetAll0 mode or not. In one embodiment, the operation configuration registers 7019A of the transmitting PE 7000A, 7019B of the transmitting PE 7000B, and 7019C of the receiving PE 7000C are loaded with the operation configuration values for NetAll0. It should be understood that operation configuration registers 7019A of the transmitting PE 7000A, 7019B of the transmitting PE 7000B, and 7019C of the receiving PE 7000C may be loaded with other configuration values, in addition to those associated with NetAll0, that may enable PEs 7000A, 7000B, and 7000C to execute other operations concurrently with NetAll0.
Multiple networks (e.g., interconnects) may be connected to a processing element, e.g., networks 7002, 7004, 7006, and 7010. The connections may be switches, e.g., as discussed in reference to
First processing element (PE) 7000A includes storage (e.g., a register) 7005A to store a transmitted last value (transmitted last, indicating that this transmitter 7000A has already sent a value for this NetAll0 execution), second processing element (PE) 7000B includes storage (e.g., a register) 7005B to store a transmitted last value (transmitted last, indicating that this transmitter 7000B has already sent a value for this NetAll0 execution), and third processing element (PE) 7000C includes storage (e.g., a register) 7005C to store a value (AllCompleteReg) that when set to a first value, causes the receiving PE to read the transmittedLast line and that when set to a second value, cause the receiving PE to read the valid line and not the transmittedLast line.
The following discussion sometimes refers to a cycle or cycles. It should be understood that the steps (e.g., instances in time) outlined herein may occur as a sequence of timesteps independent of the oscillation of a particular cycle value in certain embodiments.
In
In
In
In
In
In certain embodiments, the control indications (e.g., from input and/or output controllers of scheduler) are used to indicate presence of zero-bit all0 tokens and leverages control programmability to do this
In certain embodiments, a NetAll0 operation reduces a set of 0-bit inputs to a single output, for example, to aggregate counting values coming from memory operations (e.g., one NetAll0 operation occurring per store operation). In certain memory-heavy dataflow graphs, the use of the NetAll operation accounts for about 8% of the total operations. In certain embodiments, a plurality of transmitting PEs send an indication to a receiver PE that they have data, and these indications are combined in the network by NetAll0 to form a single value representative of the indications from the plurality of transmitting PEs.
In certain embodiments, no modifications are required to the PE-to-PE network because control is fanned out and fanned in using programmable state machines, the control can be steered to or from any number of transmitters (transmitter PEs) to a receiver (receiver PE) by correctly configuring the network. In one embodiment where all transmitters must send a value simultaneously, the control fan-in network into the receiver will allow the receiver to accept data only when all transmitters are sending, and all transmitters will be dequeued by the control signals sent by the receiver. In some embodiments, a receiver will assert that it has room to receive tokens (a “ready” signal), and transmitters will observe this and dequeue their tokens. Unfortunately it may be the case that not all transmitters were ready to send. To correct this, NetAll0 may utilize the following mechanisms, e.g., as a configuration extension at the receiver and/or transmitter as explained in reference to
An example combinational implementation is for the receiver to use uses the incoming valid indication to decide if it will send an indication to its buffer (e.g., queue) to accept new values (e.g., tokens). In certain embodiments, if valid is driven, the receiver may set its ready signal to ensure that transmitters are only dequeued once all transmitters are signaling values. To limit combinational path scope, simultaneous multicast at a transmitter is disallowed i certain embodiments.
An example multiple (e.g., two) cycle implementation uses a bit at the receiver to track whether all transmitters attempted to send a value in this cycle, and the receiver ready indication is driven from a register representing whether the NetAll0 operation will complete in the present cycle. In certain embodiments, this eliminates the combinational loop described above.
An example protocol for obtaining distributed agreement is a modification of the existing four-wire protocol, e.g., as shown in
Transmitter (e.g. 7000A, 7000B):
Deq=(output.notEmpty && transmitter.ready && (!transmittedLastReg∥allComplete))
valid=output.notEmpty
transmittedLast=transmittedLastReg∥output.notEmpty
transmittedLastReg<=deq∥(transmittedLastReg && !allComplete)
Receiver (e.g. 7000C):
Enq: input.notFull && ((!allCompleteReg && transmittedLast)∥(allCompleteReg && valid))
allCompleteReg<=input.notFull && ((!allCompleteReg && (transmittedLast))∥(allCompleteReg&& valid))
ready:input.notFull
allComplete: allCompleteReg
This implementation allows the NetAll0 to slip in case not all transmitters are ready to send in a single cycle. However, if all transmitters and the receiver are ready in every cycle, full throughput it maintained. In certain embodiments, configuration bits are used to select this mode, which adjusts the control from the normal multicast to the all reduction (NetAll0).
Because the behavior of the protocol is modified in this case, transmitters participating in the NetAll0 do not simultaneously participate in a multicast in certain embodiments.
Land
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a logical AND (land) operation according to the following (e.g., semantics and/or description).
In
In
In
Note that certain embodiments herein discuss moving a value between slots (e.g., from a first slot to a second slot). In one embodiment, the value physically moves from one slot to another in a same queue. In another embodiment, the physical storage slot that is used is the same slot, but it is a logical (not physical) move of data. For example, the head/tail pointer in
In
In
In certain embodiments, land is the logical AND of successive operands, e.g., where a zero is to stop the land from examining successive operands and instead immediately output a zero. This may be used for nested combinational statements.
In certain embodiments, PE 7100 is stalled from performing the land operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) an input value in each of the source input queues (e.g., but absence of a value in the low priority input queue will not block the execution of the operation if that value is not needed).
In the depicted embodiment, PE 7100 includes the components of PE 5800 from
Lor
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a logical OR (Lor) operation according to the following (e.g., semantics and/or description).
In
In
In
Additionally, in the depicted embodiment, the value of one is dequeued from the first slot of input queue 7206, the value of zero is moved from the second slot into the first slot of input queue 7206, the value of one is not dequeued from the first slot of input queue 7204, and the value of zero is not moved from the second slot into the first slot of input queue 7204. The value from the output queue 7234 may be consumed, e.g., by a downstream PE or PEs.
In
In
In certain embodiments, lor is the logical OR of successive operands, e.g., where a one is to stop the lor from examining successive operands and instead immediately output a one. This may be used for nested combinational statements.
In certain embodiments, PE 7200 is stalled from performing the lor operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) an input value in each of the source input queues if the value contained by each input queue is a logical false value or a subset of source input queues defined by the priority order of lor with precedence greater than the first queue containing a logical true value, including the queue containing the logical true value. That is, the absence of a value in a low priority queue will not stall execution if that value is not needed.
In the depicted embodiment, PE 7200 includes the components of PE 5800 from
First
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a First operation according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In certain embodiments, PE 7300 is to first convert a stream of “k” number of ones, followed by a single zero marking the end of the stream to a k length sequence in which the first value is one and the remaining k−1 values are zero (e.g., corresponding to the first output of a sequencer operator).
In certain embodiments, PE 7300 is stalled from performing the First operation until there is both (i) space available in the output queue that is to be used for storing resultant data (e.g. if the operation is producing output data), and (ii) an input control value in input queue 7322.
In the depicted embodiment, PE 7300 includes the components of PE 5800 from
Last
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Last operation according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In certain embodiments, PE 7400 is to first convert a stream of “k” number of ones, followed by a single zero marking the end of the stream to a k length sequence in which the last value is one and the remaining k−1 values are zero (e.g., corresponding to the last output of a sequencer operator).
In certain embodiments, PE 7400 is stalled from performing the Last operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) an input control value in input queue 7422.
In the depicted embodiment, PE 7400 includes the components of PE 5800 from
Countbuffer0 (cntbuffer0)
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Countbuffer0 (cntbuffer0) operation according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In certain embodiments, PE 7500 is to implement a storage structure for zero-bit data values by maintaining a counter tracked in a PE register (e.g., to enable a large number of tokens to be stored).
In certain embodiments, PE 7500 is not stalled from performing the Countbuffer0 operation because space is not available in the output queue that is to be used for storing resultant data (e.g., assuming the counter value has space available for the counter value).
In the depicted embodiment, PE 7500 includes the components of PE 5800 from
Countbuffer1 (cntbuffer1)
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Countbuffer1 (cntbuffer1) operation according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In
In certain embodiments, PE 7600 is to implement a storage structure for one-bit data values by maintaining a counter tracked in a PE register (e.g., to enable a large number of a “same value” tokens to be stored).
In certain embodiments, PE 7600 is not stalled from performing the Countbuffer1 operation because space is not available in the output queue that is to be used for storing resultant data (e.g., assuming the counter value has space available for the counter value).
In the depicted embodiment, PE 7600 includes the components of PE 5800 from
OnCount0
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform an OnCount0 operation according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In
In certain embodiments, PE 7700 is stalled from performing the OnCount0 operation until there is space available in the output queue that is to be used for storing resultant data.
In the depicted embodiment, PE 7700 includes the components of PE 5800 from
OnEnd
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform an OnEnd operation according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In one embodiment of the
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein.
In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc. The data value from the output queue 7832 may be consumed from the output queues, e.g., by a downstream PE or PEs.
In certain embodiments, PE 7800 is stalled from performing the switch operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) the stream control value (e.g., and input data value) are available. In one embodiment, PE 7800 is not stalled if no output is to be produced.
In the depicted embodiment, PE 7800 includes the components of PE 5800 from
Replace1
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Replace 1 operation according to the following (e.g., semantics and/or description).
In
In certain embodiments, “emit state” value in storage 7905 tracks the state of the “replace” state machine. In one embodiment, that state machine has the following three states: (i) ACCUMULATE (state value of 0 in
In
In
In
In
In
As the to-be-replaced pattern did match the input values in the previous comparisons, the replacement pattern of a 0, followed by a 1, and followed by another 1 will be sent to output queue 7932. In one embodiment, had the to-be-replaced pattern not matched the input values in the previous comparisons (e.g., 0, 1, 0), the accumulated bits would have been copied to the output rather than replaced.
In
In
In
In certain embodiments, PE 7900 is stalled from performing the Replace1 operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) an input value in input queue 7904.
In the depicted embodiment, PE 7900 includes the components of PE 5800 from
Replicate1
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Replicate1 operation according to the following (e.g., semantics and/or description).
In certain embodiments, the input stream of values is received on one of input queues (e.g., having a single bit width) 8017, 8021, 8022, 8004, or 8006). In the depicted embodiment, the input stream of values is received on input queue 8004. In certain embodiments, one or more of the match value, and the count value are received on one of (e.g., wider) input queue or queues 8024 or 8026 (for example, having a multiple bit width, e.g., 8, 16, 32, or 64).
In
In
In
In
Here, the input value is a one, and thus does not match the match value of zero, so without a match, no replication occurs.
As noted throughout, the output values may be consumed between points in time indicated by the Figures, e.g., by a downstream PE. In one embodiment, each Figure illustrates a separate cycle. In another embodiment, each figure illustrates a distinct moment in time, but not necessarily an entire cycle passing between two figures.
In
In
In
e.g., In
In certain embodiments, PE 8000 is stalled from performing the Replicate1 operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) an input value in input queue 8004.
In the depicted embodiment, PE 8000 includes the components of PE 5800 from
Dataflow Operations
NetUnpack
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a NetUnpack operation according to the following (e.g., semantics and/or description).
In addition to point-to-point communications, certain networks herein also support multicast communications, e.g., sending data from a single, transmitting PE to a plurality of receiving PEs. Communication channels may be formed by statically configuring the network to from virtual circuits (e.g., LICs) between PEs.
Thus, each receiving PE may select a high portion or a low portion of the value in its input queue. In one embodiment, the selection of the high portion (e.g., upper half), the lower portion (e.g., lower half), or the entirety of a value is selected by setting the configuration value in that PE to a value to select one of those three options.
In one embodiment, a circuit switched network 8110 includes (i) a data path to send data from first PE 8100A to both second PE 8100B and third PE 8100C, e.g., for operations to be performed on that data by second PE 8100B and third PE 8100C, and (ii) a flow control path to send control data that controls (or is used to control) the sending of that data from first PE 8100A to both second PE 8100B and third PE 8100C. Data path may send a data (e.g., valid) value when data is in an output buffer (e.g., when data is in control output buffer 8132A, first data output buffer 8134A, or second data output buffer 8136A of first PE 8100A). In one embodiment, each output buffer includes its own data path, e.g., for its own data value from producer PE to consumer PE. Components in PE are examples, for example, a PE may include only a single (e.g., data) input buffer and/or a single (e.g., data) output buffer. Flow control path may send control data that controls (or is used to control) the sending of corresponding data from first PE 8100A (e.g., control output buffer 8132A, first data output buffer 8134A, or second data output buffer 8136A thereof) to both second PE 8100B and third PE 8100C. Flow control data may include a backpressure value from each consumer PE (or aggregated from all consumer PEs, e.g., with an AND logic gate). Flow control data may include a backpressure value, for example, indicating the buffer of the second PE 8100B (e.g., control input buffer 8122B, first data input buffer 8124B, or second data input buffer 8126B) and/or the buffer of the third PE 8100B (e.g., control input buffer 8122C, first data input buffer 8124C, or second data input buffer 8126C) where the data (e.g., from control output buffer 8132A, first data output buffer 8134A, or second data output buffer 8136A of first PE 8100A) is to-be-stored is (e.g., in the current cycle) full or has an empty slot (e.g., empty in the current cycle or next cycle) (e.g., transmission attempt). Flow control data may include a speculation value and/or success value. Network 8110 may include a speculation path (e.g., to transport a speculation value) and/or success path (e.g., to transport a success value). In one embodiment, a success path follows (e.g., is parallel to) the data path, e.g., is sent from the producer PE to the consumer PEs. In one embodiment, a speculation path follows (e.g., is parallel to) the backpressure path, e.g., is sent from a consumer PE to the producer PE. In one embodiment, each consumer PE has its own flow control path, e.g., in a circuit switched network 8110, to its producer PE. In one embodiment, each consumer PEs flow control path is combined into an aggregated flow control path for its producer PE.
Turning to the depicted PEs, processing elements 8100A-C include operation configuration registers 8119A-C that may be loaded during configuration (e.g., mapping) and specify the particular operation or operations (for example, and indicate whether to enable non-blocking (e.g., reduced critical path) multicast mode or not (e.g., enable multicast mode that blocks transmission from producer PE until all consumer PEs are ready) that processing (e.g., compute) element is to perform. Register 8120A-C activity may be controlled by that operation (an output of multiplexer 8116A-C, e.g., controlled by the scheduler 8114A-C). Scheduler 8114A-C may schedule an operation or operations of processing element 8100A-C, respectively, for example, when a dataflow token arrives (e.g., input data and/or control input). Control input buffer 8122A, first data input buffer 8124A, and second data input buffer 8126A are connected to local network 8102 for first PE 8100A. In one embodiment, control output buffer 8132A is connected to network 8110 for first PE 8100A, control input buffer 8122B is connected to local network 8110 for second PE 8100B, and control input buffer 8122C is connected to local network 8110 for third PE 8100C (e.g., and each local network may include a data path as in
For example, suppose the operation of first processing (e.g., compute) element 8100A is (or includes) what is called call a pick in
For example, suppose the operation of first processing (e.g., compute) element 8100A is (or includes) what is called call a switch in
Multiple networks (e.g., interconnects) may be connected to a processing element, e.g., networks 8102, 8104, 8106, and 8110. The connections may be switches, e.g., as discussed in reference to
In one embodiment, the backpressure value and the speculation value (e.g., and the success value) allow the PEs and network (e.g., cumulatively the system) to handle the distributed coordination case, e.g., where all consumer PEs (e.g., receivers) must receive the multicast data item before it may be dequeued (e.g., discarded) by the producer PE (e.g., transmitter). Certain embodiments herein allow the target receivers to speculatively receive data, e.g., even if it is not known that all receivers will receive (e.g., store) the data (e.g., in that cycle). Thus, in certain embodiments the data itself is not speculative and it will eventually be sent. Here speculation may generally refer to the producer PE (e.g., transmitter) assuming that (e.g., at least some of) the consumer PEs (e.g., receivers) might receive the transmitted data (e.g., in that cycle). For example, in contrast to waiting for the backpressure value from all multicast consumer PEs to indicate they have storage available for that data. In one embodiment, if any receivers are unready, then the backpressure (e.g., ready) value will be pulled to a value (e.g., binary low) indicating there is no storage available in the consumer PE, for example, by the flow control function, e.g., and the producer PE (e.g., transmitter) would also pull its data flow (e.g., transmit valid) value to a value (e.g., binary low) so that no data would be transmitted.
In a reduced multicast critical path embodiment, the producer PE (e.g., transmitter) may drive its dataflow (e.g., valid) signal to a value (e.g., binary high) to indicate it has data to-be-transmitted. The speculation value(s) and/or a success value may resolve the case in which not all consumer PEs (e.g., receivers) were ready to receive data (e.g., have storage available for that data) (e.g., in that cycle). In one embodiment, the success signal (e.g., a single bit) is driven to a value that indicates success (e.g., binary high) by the producer PE (e.g., transmitter) when the producer PE (e.g., transmitter) was able to successfully complete a transmission in the previous cycle for a dataflow token (e.g., the dataflow token is stored in all of the multicast consumer PEs), e.g., as noted by the success value (e.g., success bit) being set as discussed herein. In one embodiment, the producer PE (e.g., transmitter) determines that it was able to complete a transmission of a dataflow token in the previous cycle when the producer PE (e.g., transmitter) observed for all of the multicast receiver PEs that either a speculation value was set to the value (e.g., binary high) to indicate the dataflow token was stored in the buffer (e.g., as indicated by a reception value (e.g., bit)) or the backpressure value (e.g., ready value) was set to the value (e.g., binary high) to indicate that storage is to be available in the buffer of the consumer PE (e.g., in the next cycle (e.g., transmission attempt)) for the dataflow token. In certain embodiments, when a producer PE (e.g., transmitter) determines that the success value is already at a value (e.g., binary high) that indicates the producer PE was able to successfully complete a transmission in the previous cycle to the multicast consumer PEs, then the producer PE (e.g., transmitter) ignores the speculation value(s) (e.g., a single bit), e.g., since it is known to refer to a completed transaction. In one embodiment, in all cycles where success is driven high, the producer PE (e.g., transmitter) also dequeues its data, e.g., dequeued from its output buffer (e.g., removed from control output buffer 8132A, first data output buffer 8134A, or second data output buffer 8136A of first PE 8100A). In certain embodiments, the success value being set in storage of a producer PE (to indicate success) causes a success value to be sent (e.g., in the next cycle after the success value was set or in the same cycle the success value was set) to the consumer PEs to clear their reception values (e.g., bits) (e.g., in the same cycle the success value is sent). In certain embodiments, the success value is set following any cycle in which a multicast transmission is completed and cleared otherwise, e.g., and success may happen in back-to-back cycles. In one embodiment, the reception bit(s) are cleared in the cycle following the dequeue of the dataflow token from the output buffer.
In one embodiment, the speculation value (e.g., a single bit) is driven to a value by a consumer PE (e.g., receiver) that indicates if that consumer PE (e.g., receiver) has accepted the data sent by the producer PE (e.g., transmitter), e.g., as noted by the reception value (e.g., reception bit) being set (e.g., in that cycle) as discussed herein or if the receiver was ready to receive anyway (for example, the backpressure value indicates that storage is available or is to be available on the next cycle, e.g., that PE is consuming a dataflow token that is to be cleared from the buffer at the end of the current cycle). In one embodiment, the backpressure value (e.g., ready value) and the reception value are logically OR'd (e.g., returns the Boolean value true (e.g., binary high, e.g., 1) if either or both input operands are true and returns false (e.g., binary low, e.g., 0) otherwise) together to form the speculation value. In one embodiment, the reception value (e.g., value) is cleared when (e.g., following any cycle in which) the success value (e.g., value) is observed, e.g., indicating the producer PE was able to successfully complete a transmission in the previous cycle to the multicast consumer PEs. Certain embodiments herein permit speculatively transmitted data to proceed through the pipeline. In one embodiment, once a dataflow token (e.g., value) has been obtained it may be used, e.g., it is not stalled. In one embodiment, each consumer PE (e.g., receiver) is to drive its speculation value until the cycle after it observes the producer PE (e.g., transmitter) driving its success value to indicate success. This may improve the performance of some dataflow graphs. In one embodiment, having both backpressure values (e.g., ready) and speculation values enables the transmittal of data in a fully pipelined fashion for multicast. Backpressure (e.g., ready) value may be used in cycles in which the speculation value is unusable due to a previous transmission having completed in a previous cycle. In one embodiment, PEs are provisioned with at least two input buffer slots in each input buffer to allow for full pipelining to be obtained.
In certain embodiments, distributed agreement of the consumers (e.g., PEs) allows for a reduced multicast critical path, for example, where success is checked in the next cycle after a transmission attempt, e.g., instead of a producer (e.g., PE) waiting for all the backpressure to be clear (e.g., ready) values from consumers. In one embodiment, the producer sends the data (e.g., at the beginning of a first cycle), then the consumers check if they received that data (e.g., simultaneously, at the end of the first cycle, or the beginning of a second cycle), e.g., if the data was stored in the target buffer of that consumer. If all the transmissions were successful, in one embodiment (e.g., at the clock edge), the producer is to set the success bit and then drive the success value to the consumers (e.g., in the next cycle). If not, then data may be sent for another cycle until all the consumers pass the check that the data was received. In one embodiment, a first value (e.g., from a first wire between a consumer and a producer) indicates whether data is ready (e.g., in its output buffer) and a second value (e.g., from a second wire between the consumer to the producer) indicates that data is ready, but it is a retransmission (e.g., not new data). The second value (e.g., from second wire) may thus keep from having two of the same data in a consumer, e.g., to avoid having two or more copies in an input buffer of a consumer PE for the same instance of an output value from a producer PE that was transmitted multiple times. Certain embodiments herein add a state element at each consumer, e.g., a reception bit. Flow control may indicate full or empty (e.g., backpressure) and indicate if a consumer took the data in a previous cycle. Producer may use knowledge of (i) if the consumer took the data, and (ii) whether the consumer may take more data, to control its output of data. Consumer PEs may send a speculation value back to a producer. Consumer PE may indicate that its target buffer is full, but producer PE may utilize the embodiments herein to determine if that target buffer is full for a consumer PE, and that consumer PE took the data (versus not taking the data and being full from a previous transmission for a different instance of an output value from the producer PE). In certain embodiments, one or more of the following aggregated values are utilized: (1) whether all the consumer PEs are full or empty, and (2) whether a consumer PE (e.g., all multicast consumer PEs) took data in the prior cycle e.g., so the backpressure value indicates no storage is available because it took the current data in that cycle or because there was and/or is no room for the data).
In one embodiment, first PE 8100A includes first storage 8101 for a success value (e.g., bit) for control output buffer 8132A, second storage 8103 for a success value (e.g., bit) for first data output buffer 8134A, and third storage 8105 for a success value (e.g., bit) for second data output buffer 8136A. Depicted scheduler 8114A is coupled to first storage 8101 to set or clear a success value (e.g., bit) therein for control output buffer 8132A, coupled to second storage 8103 to set or clear a success value (e.g., bit) therein for first data output buffer 8134A, and coupled to third storage 8105 to set or clear a success value (e.g., bit) therein for second data output buffer 8136A. In one embodiment, the scheduler 8114A sets the success value based on flow control data from the second PE 8100B and flow control data from the third PE 8100C. Some or all of the flow control data may be aggregated into a single value, e.g., sent to the first (e.g., as producer) PE 8100A. First (e.g., as producer) PE 8100A includes a (e.g., input) port 8108A(1-3) coupled to network 8110, e.g., to receive a backpressure value from second (e.g., as consumer) PE 8100B and/or third (e.g., as consumer) PE 8100C. In one circuit switched configuration, (e.g., input) port 8108A(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3)) is to receive a respective backpressure value from each one of control input buffer 8122B, first data input buffer 8124B, and second data input buffer 8126B and/or control input buffer 8122C, first data input buffer 8124C, and second data input buffer 8126C. In one embodiment, (e.g., input) port 8108A(1-3) is to receive an aggregated (e.g., single) respective backpressure value of each of (i) a backpressure value from control input buffer 8122B logically AND'd (e.g., it returns the Boolean value true (e.g., binary high, e.g., binary 1) if both input operands are true and returns false (e.g., binary 0) otherwise) with a backpressure value from control input buffer 8122C (e.g., on input 8108A(1)), (ii) a backpressure value from first data input buffer 8124B logically AND'd with a backpressure value from first data input buffer 8124C (e.g., on input 8108A(2)), and (iii) a backpressure value from second data input buffer 8126B logically AND'd with a backpressure value from second data input buffer 8126C (e.g., on input 8108A(3)). In one embodiment, an input or output marked as a (1), (2), or (3) is its own respective wire or other coupling.
First (e.g., as producer) PE 8100A includes a (e.g., input) port 8112A(1-3) coupled to network 8110, e.g., to receive a speculation value from second (e.g., as consumer) PE 8100B and/or third (e.g., as consumer) PE 8100C. In one circuit switched configuration, (e.g., input) port 8112A(1-3) (e.g., having a plurality of parallel inputs(1), (2), and (3)) is to receive a respective speculation value for each one of control input buffer 8122B, first data input buffer 8124B, and second data input buffer 8126B and/or control input buffer 8122C, first data input buffer 8124C, and second data input buffer 8126C. In one embodiment, (e.g., input) port 8112A(1-3) is to receive an aggregated (e.g., single) speculation value for each of (i) speculation value for control input buffer 8122B logically AND'd with speculation value for control input buffer 8122C (e.g., on input 8108A(1)), (ii) speculation value for first data input buffer 8124B logically AND'd with speculation value for first data input buffer 8124C (e.g., on input 8108A(2)), and (iii) speculation value for second data input buffer 8126B logically AND'd with speculation value for second data input buffer 8126C (e.g., on input 8108A(3)).
In one circuit switched configuration, a multicast data path is formed from (i) control output buffer 8132A to control input buffer 8122B and control input buffer 8122C, (ii) first data output buffer 8134A to first data input buffer 8124B and first data input buffer 8124C, (iii) second data output buffer 8136A to second data input buffer 8126B and second data input buffer 8126C, or any combination thereof. A data path may be used to send a data token from the producer PE to the consumer PEs. In the depicted embodiment, second PE 8100B includes first storage 8107 for a reception value (e.g., bit) for control input buffer 8122B, second storage 8109 for a reception value (e.g., bit) for first data input buffer 8124B, and third storage 8111 for a reception value (e.g., bit) for second data input buffer 8126B, e.g., set by scheduler 8114B. In the depicted embodiment, second (e.g., as consumer) PE 8100B includes an (e.g., output) port 8108B(1-3) coupled to network 8110, e.g., to send a backpressure value from second (e.g., as consumer) PE 8100B to first (e.g., as producer) PE 8100A. In one circuit switched configuration, (e.g., output) port 8108B(1-3) is to send a respective backpressure value for each one of control input buffer 8122B (e.g., on output 8108B(1)), first data input buffer 8124B (e.g., on output 8108B(2)), and second data input buffer 8126B (e.g., on output 8108B(3)), e.g., by scheduler 8114B. Second (e.g., as consumer) PE 8100B includes a (e.g., input) port 8112B(1-3) coupled to network 8110, e.g., to receive a success value from first (e.g., as producer) PE 8100A. In one circuit switched configuration, (e.g., input) port 8112B(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3))) is to receive a respective success value for each one of control input buffer 8122B (e.g., on input 8112B(1)), first data input buffer 8124B (e.g., on input 8112B(2)), and second data input buffer 8126B (e.g., on input 8112B(3)).
In the depicted embodiment, third PE 8100C includes first storage 8113 for a reception value (e.g., bit) for control input buffer 8122C, second storage 8115 for a reception value (e.g., bit) for first data input buffer 8124C, and third storage 8117 for a reception value (e.g., bit) for second data input buffer 8126C, e.g., set by scheduler 8114C. Third (e.g., as consumer) PE 8100C includes an (e.g., output) port 8108C(1-3) coupled to network 8110, e.g., to send a backpressure value from third (e.g., as consumer) PE 8100C to first (e.g., as producer) PE 8100A. In one circuit switched configuration, (e.g., output) port 8108C(1-3) is to send a respective backpressure value for each one of control input buffer 8122C (e.g., on output 8108C(1)), first data input buffer 8124C (e.g., on output 8108C(2)), and second data input buffer 8126C (e.g., on output 8108C(3)), e.g., by scheduler 8114C. Second (e.g., as consumer) PE 8100B includes a (e.g., input) port 8112C(1-3) coupled to network 8110, e.g., to receive a success value from first (e.g., as producer) PE 8100A. In one circuit switched configuration, (e.g., input) port 8112C(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3)) is to receive a respective success value for each one of control input buffer 8122C (e.g., on input 8112C(1)), first data input buffer 8124C (e.g., on input 8112C(2)), and second data input buffer 8126C (e.g., on input 8112C(3)).
As noted herein, speculation value may be formed by logically OR'ing the reception bit (for example, where a binary low value indicates the buffer did not take an input since it was last cleared and a binary high value indicates the buffer did take an input since it was last cleared, e.g., by the success value) and a backpressure bit (e.g., where a binary low value indicates there is no backpressure and a binary high value indicates there is backpressure). A port may include a plurality of inputs and/or outputs. A processing element may include a single port into network 8110, or any plurality of ports. Although
First PE 8100A may include first storage 8129 for a reception value (e.g., bit) for control input buffer 8122A, second storage 8131 for a reception value (e.g., bit) for first data input buffer 8124A, and third storage 8133 for a reception value (e.g., bit) for second data input buffer 8126A, e.g., set by scheduler 8114A. First (e.g., as consumer) PE 8100A may include an (e.g., output) port 8125(1-3) coupled to network 8102, e.g., to send a backpressure value from first (e.g., as consumer) PE 8100A to an upstream (e.g., as producer) PE. In one circuit switched configuration, (e.g., output) port 8125(1-3) is to send a respective backpressure value for each one of control input buffer 8122A (e.g., on output 8125(1)), first data input buffer 8124A (e.g., on output 8125(2)), and second data input buffer 8126A (e.g., on output 8125(3)), e.g., by scheduler 8114A. First (e.g., as consumer) PE 8100A includes a (e.g., input) port 8127(1-3) coupled to network 8102, e.g., to receive a success value from an upstream (e.g., as producer) PE. In one circuit switched configuration, (e.g., input) port 8127(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3))) is to receive a respective success value for each one of control input buffer 8122A (e.g., on input 8127(1)), first data input buffer 8124A (e.g., on input 8127(2)), and second data input buffer 8126A (e.g., on input 8127(3)).
Second (e.g., as producer) PE 8100B may include a (e.g., input) port 8135(1-3) coupled to network 8104 (e.g., which may be the same network as network 8106), e.g., to receive a backpressure value from a downstream (e.g., as consumer) PE or PEs. In one circuit switched configuration, (e.g., input) port 8135(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3)) is to receive a respective backpressure value from each one of control input buffer, first data input buffer, and second data input buffer of a first downstream PE and/or control input buffer, first data input buffer, and second data input buffer of a second downstream PE. In one embodiment, (e.g., input) port 8135(1-3) is to receive an aggregated (e.g., single) respective backpressure value of each of (i) a backpressure value from control input buffer for first downstream PE logically AND'd (e.g., it returns the Boolean value true (e.g., binary high, e.g., binary 1) if both input operands are true and returns false (e.g., binary 0) otherwise) with a backpressure value from control input buffer for second downstream PE (e.g., on input 8135(1)), (ii) a backpressure value from first data input buffer for first downstream PE logically AND'd with a backpressure value from first data input buffer for first downstream PE (e.g., on input 8135(2)), and (iii) a backpressure value from second data input buffer for first downstream PE logically AND'd with a backpressure value from second data input buffer for first downstream PE (e.g., on input 8135(3)). In one embodiment, an input or output marked as a (1), (2), or (3) is its own respective wire or other coupling. In one embodiment, each PE includes the same circuitry and/or components.
Second PE 8100B includes first storage 8139 for a success value (e.g., bit) for control output buffer 8132B, second storage 8141 for a success value (e.g., bit) for first data output buffer 8134B, and third storage 8143 for a success value (e.g., bit) for second data output buffer 8136B. Depicted scheduler 8114B is coupled to first storage 8139 to set or clear a success value (e.g., bit) therein for control output buffer 8132B, coupled to second storage 8141 to set or clear a success value (e.g., bit) therein for first data output buffer 8134B, and coupled to third storage 8143 to set or clear a success value (e.g., bit) therein for second data output buffer 8136B. In one embodiment, the setting of the success value in storage 8139 causes a success value to be sent on a path from storage 8139 through network 8104 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from second PE 8100B (e.g., from storage 8139 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, the setting of the success value in storage 8141 causes a success value to be sent on a path from storage 8141 through network 8104 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from second PE 8100B (e.g., from storage 8141 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, receipt of success value from second PE 8100B (e.g., from storage 8143 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, the setting of the success value in storage 8143 causes a success value to be sent on a path from storage 8143 through network 8104 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from second PE 8100B (e.g., from storage 8143 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer.
Second (e.g., as producer) PE 8100B may include a (e.g., input) port 8137(1-3) coupled to network 8104, e.g., to receive a speculation value from a first downstream (e.g., as consumer) PE and/or second downstream (e.g., as consumer) PE. In one circuit switched configuration, (e.g., input) port 8137(1-3) (e.g., having a plurality of parallel inputs(1), (2), and (3)) is to receive a respective speculation value for each one of control input buffer for first downstream PE, first data input buffer for first downstream PE, and second data input buffer for first downstream PE and/or control input buffer for second downstream PE, first data input buffer for second downstream PE, and second data input buffer for second downstream PE. In one embodiment, (e.g., input) port 8137(1-3) is to receive an aggregated (e.g., single) speculation value for each of (i) speculation value for control input buffer for first downstream PE logically AND'd with speculation value for control input buffer for second downstream PE (e.g., on input 8137(1)), (ii) speculation value for first data input buffer for first downstream PE logically AND'd with speculation value for first data input buffer for second downstream PE (e.g., on input 8137(2)), and (iii) speculation value for second data input buffer for first downstream PE logically AND'd with speculation value for second data input buffer for second downstream PE (e.g., on input 8137(3)).
Second (e.g., as producer) PE 8100B may include a (e.g., input) port 8135(1-3) coupled to network 8104 (e.g., which may be the same network as network 8106), e.g., to receive a backpressure value from a downstream (e.g., as consumer) PE or PEs. In one circuit switched configuration, (e.g., input) port 8135(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3)) is to receive a respective backpressure value from each one of control input buffer, first data input buffer, and second data input buffer of a first downstream PE and/or control input buffer, first data input buffer, and second data input buffer of a second downstream PE. In one embodiment, (e.g., input) port 8135(1-3) is to receive an aggregated (e.g., single) respective backpressure value of each of (i) a backpressure value from control input buffer for first downstream PE logically AND'd (e.g., it returns the Boolean value true (e.g., binary high, e.g., binary 1) if both input operands are true and returns false (e.g., binary 0) otherwise) with a backpressure value from control input buffer for second downstream PE (e.g., on input 8135(1)), (ii) a backpressure value from first data input buffer for first downstream PE logically AND'd with a backpressure value from first data input buffer for first downstream PE (e.g., on input 8135(2)), and (iii) a backpressure value from second data input buffer for first downstream PE logically AND'd with a backpressure value from second data input buffer for first downstream PE (e.g., on input 8135(3)). In one embodiment, an input or output marked as a (1), (2), or (3) is its own respective wire or other coupling. In one embodiment, each PE includes the same circuitry and/or components.
Second (e.g., as producer) PE 8100B may include a (e.g., input) port 8137(1-3) coupled to network 8104, e.g., to receive a speculation value from a first downstream (e.g., as consumer) PE and/or second downstream (e.g., as consumer) PE. In one circuit switched configuration, (e.g., input) port 8137(1-3) (e.g., having a plurality of parallel inputs(1), (2), and (3)) is to receive a respective speculation value for each one of control input buffer for first downstream PE, first data input buffer for first downstream PE, and second data input buffer for first downstream PE and/or control input buffer for second downstream PE, first data input buffer for second downstream PE, and second data input buffer for second downstream PE. In one embodiment, (e.g., input) port 8137(1-3) is to receive an aggregated (e.g., single) speculation value for each of (i) speculation value for control input buffer for first downstream PE logically AND'd with speculation value for control input buffer for second downstream PE (e.g., on input 8137(1)), (ii) speculation value for first data input buffer for first downstream PE logically AND'd with speculation value for first data input buffer for second downstream PE (e.g., on input 8137(2)), and (iii) speculation value for second data input buffer for first downstream PE logically AND'd with speculation value for second data input buffer for second downstream PE (e.g., on input 8137(3)).
Third (e.g., as producer) PE 8100C may include a (e.g., input) port 8145(1-3) coupled to network 8106 (e.g., which may be the same network as network 8104), e.g., to receive a backpressure value from a downstream (e.g., as consumer) PE or PEs. In one circuit switched configuration, (e.g., input) port 8145(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3)) is to receive a respective backpressure value from each one of control input buffer, first data input buffer, and second data input buffer of a first downstream PE and/or control input buffer, first data input buffer, and second data input buffer of a second downstream PE. In one embodiment, (e.g., input) port 8145(1-3) is to receive an aggregated (e.g., single) respective backpressure value of each of (i) a backpressure value from control input buffer for first downstream PE logically AND'd (e.g., it returns the Boolean value true (e.g., binary high, e.g., binary 1) if both input operands are true and returns false (e.g., binary 0) otherwise) with a backpressure value from control input buffer for second downstream PE (e.g., on input 8145(1)), (ii) a backpressure value from first data input buffer for first downstream PE logically AND'd with a backpressure value from first data input buffer for first downstream PE (e.g., on input 8145(2)), and (iii) a backpressure value from second data input buffer for first downstream PE logically AND'd with a backpressure value from second data input buffer for first downstream PE (e.g., on input 8145(3)). In one embodiment, an input or output marked as a (1), (2), or (3) is its own respective wire or other coupling. In one embodiment, each PE includes the same circuitry and/or components.
Third (e.g., as producer) PE 8100C may include a (e.g., input) port 8147(1-3) coupled to network 8106, e.g., to receive a speculation value from a first downstream (e.g., as consumer) PE and/or second downstream (e.g., as consumer) PE. In one circuit switched configuration, (e.g., input) port 8147(1-3) (e.g., having a plurality of parallel inputs(1), (2), and (3)) is to receive a respective speculation value for each one of control input buffer for first downstream PE, first data input buffer for first downstream PE, and second data input buffer for first downstream PE and/or control input buffer for second downstream PE, first data input buffer for second downstream PE, and second data input buffer for second downstream PE. In one embodiment, (e.g., input) port 8147(1-3) is to receive an aggregated (e.g., single) speculation value for each of (i) speculation value for control input buffer for first downstream PE logically AND'd with speculation value for control input buffer for second downstream PE (e.g., on input 8147(1)), (ii) speculation value for first data input buffer for first downstream PE logically AND'd with speculation value for first data input buffer for second downstream PE (e.g., on input 8147(2)), and (iii) speculation value for second data input buffer for first downstream PE logically AND'd with speculation value for second data input buffer for second downstream PE (e.g., on input 8147(3)).
Third PE 8100C includes first storage 8149 for a success value (e.g., bit) for control output buffer 8132C, second storage 815 for a success value (e.g., bit) for first data output buffer 8134C, and third storage 8153 for a success value (e.g., bit) for second data output buffer 8136C. Depicted scheduler 8114C is coupled to first storage 8149 to set or clear a success value (e.g., bit) therein for control output buffer 8132C, coupled to second storage 8151 to set or clear a success value (e.g., bit) therein for first data output buffer 8134C, and coupled to third storage 8153 to set or clear a success value (e.g., bit) therein for second data output buffer 8136C. In one embodiment, the setting of the success value in storage 8149 causes a success value to be sent on a path from storage 8149 through network 8104 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from third PE 8100C (e.g., from storage 8149 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, the setting of the success value in storage 8151 causes a success value to be sent on a path from storage 8151 through network 8104 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from third PE 8100C (e.g., from storage 8151 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, receipt of success value from third PE 8100C (e.g., from storage 8153 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, the setting of the success value in storage 8153 causes a success value to be sent on a path from storage 8153 through network 8104 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from third PE 8100C (e.g., from storage 8143 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer.
A processing element may include two sub-networks (or two channels on the network), e.g., one for a data path and one for a flow control path. A processing element (e.g., PE 8100A, PE 8100B, and PE 8100C) may function and/or include the components as in any of the disclosure herein. A processing element may be stalled from execution until its operands (e.g., in its input buffer(s)) are received and/or until there is room in the output buffer(s) of the processing element for the data that is to be produced by the execution of the operation on those operands. Next, three reduced multicast critical path embodiments are discussed.
As a first example,
Scheduler 8114A is coupled to first storage 8101 to set or clear a success value (e.g., bit) for control output buffer 8132A. In one embodiment, the scheduler 8114A sets the success value based on flow control data from the second PE 8100B and flow control data the second PE 8100C. Some or all of the flow control data may be aggregated into a single value, e.g., sent to the first (e.g., as producer) PE 8100A. First (e.g., as producer) PE 8100A includes a (e.g., input) port 8108A(1) coupled to network 8110, e.g., to receive a backpressure value from second (e.g., as consumer) PE 8100B and/or third (e.g., as consumer) PE 8100C. In one circuit switched configuration, (e.g., input) port 8108A(1) is to receive a respective backpressure value from each one of control input buffer 8122B and control input buffer 8122C. In the depicted embodiment, (e.g., input) port 8108A(1) is to receive an aggregated (e.g., single) respective backpressure value of a backpressure value from control input buffer 8122B logically AND'd (e.g., it returns the Boolean value true (e.g., binary high, e.g., binary 1) if both input operands are true and returns false (e.g., binary low, e.g., binary 0) otherwise) with a backpressure value from control input buffer 8122C by AND logic gate 8152.
First (e.g., as producer) PE 8100A includes a (e.g., input) port 8112A(1) coupled to network 8110, e.g., to receive a speculation value from second (e.g., as consumer) PE 8100B and/or third (e.g., as consumer) PE 8100C. In one circuit switched configuration, (e.g., input) port 8112A(1) is to receive a respective speculation value for each one of control input buffer 8122B and control input buffer 8122C. In the depicted embodiment, (e.g., input) port 8112A(1) is to receive an aggregated (e.g., single) speculation value for speculation value for control input buffer 8122B logically AND'd with speculation value for control input buffer 8122C by AND logic gate 8150. In the depicted embodiment, the speculation value for control input buffer 8122B is formed by OR'ing the reception bit for the speculative path (e.g., reception bit from storage 8107) (e.g., where a binary low value indicates the buffer did not store an input since it was last cleared) and a backpressure bit from backpressure path (e.g., from port 8108B(1)) (e.g., where a binary low value indicates there is no backpressure) by OR logic gate 8154. In the depicted embodiment, the speculation value for control input buffer 8122C is formed by OR'ing the reception bit for the speculative path (e.g., reception bit from storage 8113) (e.g., where a binary low value indicates the buffer did not store an input since it was last cleared) and a backpressure bit from backpressure path (e.g., from port 8108C(1)) (e.g., where a binary low value indicates there is no backpressure) by OR logic gate 8156. In one embodiment, a PE (e.g., scheduler thereof) is to set (e.g., to binary high) a reception value (e.g., reception bit) to indicate a value was stored in that buffer (e.g., second PE 8100B setting a reception bit in storage 8107 to indicate a dataflow token was stored (e.g., since the reception bit was last cleared) in the control input buffer 8122B and/or third PE 8100C setting a reception bit in storage 8113 to indicate a dataflow token was stored (e.g., since the reception bit was last cleared) in the control input buffer 8122C). In certain embodiments herein, logic gate functionality is achieved by using NAND/NOR circuit designs.
In one circuit switched configuration, a multicast data path is formed from control output buffer 8132A to control input buffer 8122B and control input buffer 8122C. A data path may be used to send a data token from the producer PE to the consumer PEs. In the depicted embodiment, second PE 8100B includes first storage 8107 for a reception value (e.g., bit) for control input buffer 8122B. Second (e.g., as consumer) PE 8100B includes an (e.g., output) port 8108B(1) coupled to network 8110, e.g., to send a backpressure value from second (e.g., as consumer) PE 8100B to first (e.g., as producer) PE 8100A. In one circuit switched configuration, (e.g., output) port 8108B(1) is to send a respective backpressure value for control input buffer 8122B. Second (e.g., as consumer) PE 8100B includes a (e.g., input) port 8112B(1) coupled to network 8110, e.g., to receive a success value from first (e.g., as producer) PE 8100A. In one circuit switched configuration, (e.g., input) port 8112B(1) is to receive a respective success value for control input buffer 8122B.
In the depicted embodiment, third PE 8100C includes first storage 8113 for a reception value (e.g., bit) for control input buffer 8122C. Third (e.g., as consumer) PE 8100C includes an (e.g., output) port 8108C(1) coupled to network 8110, e.g., to send a backpressure value from third (e.g., as consumer) PE 8100C to first (e.g., as producer) PE 8100A. In one circuit switched configuration, (e.g., output) port 8108C(1) is to send a respective backpressure value for control input buffer 8122C. Second (e.g., as consumer) PE 8100B includes a (e.g., input) port 8112C(1) coupled to network 8110, e.g., to receive a success value from first (e.g., as producer) PE 8100A. In one circuit switched configuration, (e.g., input) port 8112C(1) is to receive a respective success value for control input buffer 8122C.
In one embodiment, a data token is received in control output buffer 8132A which causes the reduced multicast critical path of the first example to begin operation. In one embodiment, the data token's reception therein causes the producer PE 8100A (e.g., transmitter) to drive its dataflow (e.g., valid) value (e.g., on the path from control output buffer 8132A to control input buffer 8122B (e.g., through network 8110) and the path from control output buffer 8132A to control input buffer 8122C (e.g., through network 8110)) to a value (e.g., binary high) to indicate it has data to-be-transmitted. In one embodiment, the dataflow value (e.g., valid) is the transmittal of the dataflow token (e.g., payload data) itself. In one embodiment, a first path is included from producer PE to (e.g., each) consumer PE through network 8110 for the dataflow token and a second path is included from producer PE to (e.g., each) consumer PE through network 8110 for a dataflow value to indicate if that dataflow token (e.g., in storage coupled to the first path) is valid or invalid. The speculation value(s) and/or a success value may resolve the case in which not all consumer PEs (e.g., receivers) were ready to receive the dataflow token (e.g., have storage available for that dataflow token).
In the first transmission attempt for this dataflow token, if the backpressure value (e.g., ready value) on the path from port 8108B(1) of second PE 8100B to port 8108A(1) of first PE 8100A and the backpressure value (e.g., ready value) on the path from port 8108C(1) of third PE 8100C to port 8108A(1) of first PE 8100A both indicate (e.g., as the output from AND logic gate 8152) there is no backpressure (e.g., there is storage available in each of control input buffer 8122B and control input buffer 8122C), then the first PE (e.g., scheduler 8114A) determines that this transmission attempt will be successful, for example, and the dataflow token is to be dequeued (e.g., in the next cycle) from the control output buffer 8132A of the first PE 8100A and/or the success value (e.g., success bit) in first storage 8101 is set (e.g., in the next cycle) to indicate a successful transmission. In the first transmission attempt for this data token, if the backpressure value (e.g., ready value) on the path from port 8108B(1) of second PE 8100B to port 8108A(1) of first PE 8100A or the backpressure value (e.g., ready value) on the path from port 8108C(1) of third PE 8100C to port 8108A(1) of first PE 8100A indicate (e.g., as the output from AND logic gate 8152) there is backpressure (e.g., there is not storage available in both (e.g., all) of control input buffer 8122B and control input buffer 8122C, respectively), then one or more retransmissions of that dataflow token will occur until the speculation value from each of second (e.g., as consumer) PE 8100B and third (e.g., as consumer) PE 8100C indicates speculation is true, for example, until the speculation value is driven to a value by each of second (e.g., as consumer) PE 8100B and third (e.g., as consumer) PE 8100C that indicates that consumer PE (e.g., receiver) has either (i) accepted the data sent by the producer PE 8100A, e.g., as noted by the reception value (e.g., reception bit) being set (e.g., in a previous cycle) (e.g., in storage 8107 or storage 8113, respectively) or (ii) that the consumer is ready (e.g., by the next cycle) to receive the dataflow token (e.g., the backpressure value indicates that storage is currently available). For example, where the speculation value for control input buffer 8122B is formed by OR'ing the reception bit for the speculative path (e.g., reception bit from storage 8107) (e.g., where a binary low value indicates the buffer did not store an input since it was last cleared) and a backpressure bit from backpressure path (e.g., from port 8108B(1)) (e.g., where a binary low value indicates there is no backpressure) by OR logic gate 8154. In one embodiment, once the speculation values (e.g., from speculation paths) indicate the dataflow token is to be stored (e.g., in the next cycle) in control input buffer 8122B and control input buffer 8122C, the success value (e.g., a single bit) is driven by the producer PE 8100A to a value that the producer PE was able to successfully complete a transmission in the previous cycle (e.g., the value is stored in all of the multicast consumer PEs), e.g., as noted by the success value (e.g., success bit) (e.g., binary high, e.g., binary 1) being set in storage 8101. In one embodiment, the setting of the success value in storage 8101 causes a success value to be sent on a path from storage 8101 through network 8110 to (e.g., input) port 8112B(1) of (e.g., as consumer) second PE 8100B and to (e.g., input) port 8112C(1) of (e.g., as consumer) third PE 8100C. In one embodiment, receipt of success value from first PE 8100A (e.g., from storage 8101 thereof) by second PE 8100B is to cause the clearing of the reception bit in storage 8107, e.g., by scheduler 8114B. In one embodiment, receipt of success value from first PE 8100A (e.g., from storage 8101 thereof) by third PE 8100C is to cause the clearing of the reception bit in storage 8113, e.g., by scheduler 8114C.
In
In
In the depicted embodiment, PEs 8100A, 8100B, or 8100C include the components of PE 5800 from
In certain embodiments, a single line on a figure may illustrate one wire, or a plurality of wires. Note that a two wire protocol is discussed above, however, network may use a four wire protocol. In one embodiment, network 8110 uses the reduced multicast critical path discussed below (e.g., and adding the high-low muxes, etc.) Certain embodiments of a reduced multicast critical path utilize a speculation path (e.g., to transport a speculation value). Additionally or alternatively, certain embodiments of a reduced multicast critical path utilize a success path (e.g., to transport a success value). In one embodiment, a success path follows (e.g., is parallel to) the data path, e.g., is sent from the producer PE to the consumer PEs. In one embodiment, a speculation path follows (e.g., is parallel to) the flow control (e.g., backpressure) path, e.g., is sent from the consumer PEs to the producer PE. In one embodiment, the speculation value reflects the behavior in the current and previous cycle of the PEs and network(s) transmitting the data. In one embodiment, the success value reflects the behavior in the previous cycle of the PEs and network(s) transmitting the data. A cycle may be defined by a (e.g., rising or falling) clock edge. In one embodiment, a new cycle begins with (e.g., and includes) the rising clock edge. In one embodiment, a value is locked in from its asserted value on a (e.g., rising) clock edge. In one embodiment, a value is set in a first cycle, and an action caused by that value being set is begun in the second (e.g., next) cycle. Certain embodiments herein include storage (e.g., a register) in a PE and/or network to store a reception value, e.g., in storage in each consumer PE. Certain embodiments herein include storage (e.g., a register) in a PE and/or network to store a success value (e.g., from a success path), e.g., storage in the producer PE. In one embodiment, the storage is a one bit register in each PE, for example, for each set of buffers.
In one embodiment, a circuit switched network 8210 includes (i) a data path to send data from first PE 8200A to both second PE 8200B and third PE 8200C, e.g., for operations to be performed on that data by second PE 8200B and third PE 8200C, and (ii) a flow control path to send control data that controls (or is used to control) the sending of that data from first PE 8200A to both second PE 8200B and third PE 8200C. Data path may send a data (e.g., valid) value when data is in an output buffer (e.g., when data is in control output buffer 8232A, first data output buffer 8234A, or second data output buffer 8236A of first PE 8200A). In one embodiment, each output buffer includes its own data path, e.g., for its own data value from producer PE to consumer PE. Components in PE are examples, for example, a PE may include only a single (e.g., data) input buffer and/or a single (e.g., data) output buffer. Flow control path may send control data that controls (or is used to control) the sending of corresponding data from first PE 8200A (e.g., control output buffer 8232A, first data output buffer 8234A, or second data output buffer 8236A thereof) to both second PE 8200B and third PE 8200C. Flow control data may include a backpressure value from each consumer PE (or aggregated from all consumer PEs, e.g., with an AND logic gate). Flow control data may include a backpressure value, for example, indicating the buffer of the second PE 8200B (e.g., control input buffer 8222B, first data input buffer 8224B, or second data input buffer 8226B) and/or the buffer of the third PE 8200B (e.g., control input buffer 8222C, first data input buffer 8224C, or second data input buffer 8226C) where the data (e.g., from control output buffer 8232A, first data output buffer 8234A, or second data output buffer 8236A of first PE 8200A) is to-be-stored is (e.g., in the current cycle) full or has an empty slot (e.g., empty in the current cycle or next cycle) (e.g., transmission attempt). Flow control data may include a speculation value and/or success value. Network 8210 may include a speculation path (e.g., to transport a speculation value) and/or success path (e.g., to transport a success value). In one embodiment, a success path follows (e.g., is parallel to) the data path, e.g., is sent from the producer PE to the consumer PEs. In one embodiment, a speculation path follows (e.g., is parallel to) the backpressure path, e.g., is sent from a consumer PE to the producer PE. In one embodiment, each consumer PE has its own flow control path, e.g., in a circuit switched network 8210, to its producer PE. In one embodiment, each consumer PEs flow control path is combined into an aggregated flow control path for its producer PE.
Turning to the depicted PEs, processing elements 8200A-C include operation configuration registers 8219A-C that may be loaded during configuration (e.g., mapping) and specify the particular operation or operations (for example, and indicate whether to enable non-blocking (e.g., reduced critical path) multicast mode or not (e.g., enable multicast mode that blocks transmission from producer PE until all consumer PEs are ready) that processing (e.g., compute) element is to perform. Register 8220A-C activity may be controlled by that operation (an output of multiplexer 8216A-C, e.g., controlled by the scheduler 8214A-C). Scheduler 8214A-C may schedule an operation or operations of processing element 8200A-C, respectively, for example, when a dataflow token arrives (e.g., input data and/or control input). Control input buffer 8222A, first data input buffer 8224A, and second data input buffer 8226A are connected to local network 8202 for first PE 8200A. In one embodiment, control output buffer 8232A is connected to network 8210 for first PE 8200A, control input buffer 8222B is connected to local network 8210 for second PE 8200B, and control input buffer 8222C is connected to local network 8210 for third PE 8200C (e.g., and each local network may include a data path as in
For example, suppose the operation of first processing (e.g., compute) element 8200A is (or includes) what is called call a pick in
For example, suppose the operation of first processing (e.g., compute) element 8200A is (or includes) what is called call a switch in
Multiple networks (e.g., interconnects) may be connected to a processing element, e.g., networks 8202, 8204, 8206, and 8210. The connections may be switches, e.g., as discussed in reference to
In one embodiment, the backpressure value and the speculation value (e.g., and the success value) allow the PEs and network (e.g., cumulatively the system) to handle the distributed coordination case, e.g., where all consumer PEs (e.g., receivers) must receive the multicast data item before it may be dequeued (e.g., discarded) by the producer PE (e.g., transmitter). Certain embodiments herein allow the target receivers to speculatively receive data, e.g., even if it is not known that all receivers will receive (e.g., store) the data (e.g., in that cycle). Thus, in certain embodiments the data itself is not speculative and it will eventually be sent. Here speculation may generally refer to the producer PE (e.g., transmitter) assuming that (e.g., at least some of) the consumer PEs (e.g., receivers) might receive the transmitted data (e.g., in that cycle). For example, in contrast to waiting for the backpressure value from all multicast consumer PEs to indicate they have storage available for that data. In one embodiment, if any receivers are unready, then the backpressure (e.g., ready) value will be pulled to a value (e.g., binary low) indicating there is no storage available in the consumer PE, for example, by the flow control function, e.g., and the producer PE (e.g., transmitter) would also pull its data flow (e.g., transmit valid) value to a value (e.g., binary low) so that no data would be transmitted.
In a reduced multicast critical path embodiment, the producer PE (e.g., transmitter) may drive its dataflow (e.g., valid) signal to a value (e.g., binary high) to indicate it has data to-be-transmitted. The speculation value(s) and/or a success value may resolve the case in which not all consumer PEs (e.g., receivers) were ready to receive data (e.g., have storage available for that data) (e.g., in that cycle). In one embodiment, the success signal (e.g., a single bit) is driven to a value that indicates success (e.g., binary high) by the producer PE (e.g., transmitter) when the producer PE (e.g., transmitter) was able to successfully complete a transmission in the previous cycle for a dataflow token (e.g., the dataflow token is stored in all of the multicast consumer PEs), e.g., as noted by the success value (e.g., success bit) being set as discussed herein. In one embodiment, the producer PE (e.g., transmitter) determines that it was able to complete a transmission of a dataflow token in the previous cycle when the producer PE (e.g., transmitter) observed for all of the multicast receiver PEs that either a speculation value was set to the value (e.g., binary high) to indicate the dataflow token was stored in the buffer (e.g., as indicated by a reception value (e.g., bit)) or the backpressure value (e.g., ready value) was set to the value (e.g., binary high) to indicate that storage is to be available in the buffer of the consumer PE (e.g., in the next cycle (e.g., transmission attempt)) for the dataflow token. In certain embodiments, when a producer PE (e.g., transmitter) determines that the success value is already at a value (e.g., binary high) that indicates the producer PE was able to successfully complete a transmission in the previous cycle to the multicast consumer PEs, then the producer PE (e.g., transmitter) ignores the speculation value(s) (e.g., a single bit), e.g., since it is known to refer to a completed transaction. In one embodiment, in all cycles where success is driven high, the producer PE (e.g., transmitter) also dequeues its data, e.g., dequeued from its output buffer (e.g., removed from control output buffer 8232A, first data output buffer 8234A, or second data output buffer 8236A of first PE 8200A). In certain embodiments, the success value being set in storage of a producer PE (to indicate success) causes a success value to be sent (e.g., in the next cycle after the success value was set or in the same cycle the success value was set) to the consumer PEs to clear their reception values (e.g., bits) (e.g., in the same cycle the success value is sent). In certain embodiments, the success value is set following any cycle in which a multicast transmission is completed and cleared otherwise, e.g., and success may happen in back-to-back cycles. In one embodiment, the reception bit(s) are cleared in the cycle following the dequeue of the dataflow token from the output buffer.
In one embodiment, the speculation value (e.g., a single bit) is driven to a value by a consumer PE (e.g., receiver) that indicates if that consumer PE (e.g., receiver) has accepted the data sent by the producer PE (e.g., transmitter), e.g., as noted by the reception value (e.g., reception bit) being set (e.g., in that cycle) as discussed herein or if the receiver was ready to receive anyway (for example, the backpressure value indicates that storage is available or is to be available on the next cycle, e.g., that PE is consuming a dataflow token that is to be cleared from the buffer at the end of the current cycle). In one embodiment, the backpressure value (e.g., ready value) and the reception value are logically OR'd (e.g., returns the Boolean value true (e.g., binary high, e.g., 1) if either or both input operands are true and returns false (e.g., binary low, e.g., 0) otherwise) together to form the speculation value. In one embodiment, the reception value (e.g., value) is cleared when (e.g., following any cycle in which) the success value (e.g., value) is observed, e.g., indicating the producer PE was able to successfully complete a transmission in the previous cycle to the multicast consumer PEs. Certain embodiments herein permit speculatively transmitted data to proceed through the pipeline. In one embodiment, once a dataflow token (e.g., value) has been obtained it may be used, e.g., it is not stalled. In one embodiment, each consumer PE (e.g., receiver) is to drive its speculation value until the cycle after it observes the producer PE (e.g., transmitter) driving its success value to indicate success. This may improve the performance of some dataflow graphs. In one embodiment, having both backpressure values (e.g., ready) and speculation values enables the transmittal of data in a fully pipelined fashion for multicast. Backpressure (e.g., ready) value may be used in cycles in which the speculation value is unusable due to a previous transmission having completed in a previous cycle. In one embodiment, PEs are provisioned with at least two input buffer slots in each input buffer to allow for full pipelining to be obtained.
In certain embodiments, distributed agreement of the consumers (e.g., PEs) allows for a reduced multicast critical path, for example, where success is checked in the next cycle after a transmission attempt, e.g., instead of a producer (e.g., PE) waiting for all the backpressure to be clear (e.g., ready) values from consumers. In one embodiment, the producer sends the data (e.g., at the beginning of a first cycle), then the consumers check if they received that data (e.g., simultaneously, at the end of the first cycle, or the beginning of a second cycle), e.g., if the data was stored in the target buffer of that consumer. If all the transmissions were successful, in one embodiment (e.g., at the clock edge), the producer is to set the success bit and then drive the success value to the consumers (e.g., in the next cycle). If not, then data may be sent for another cycle until all the consumers pass the check that the data was received. In one embodiment, a first value (e.g., from a first wire between a consumer and a producer) indicates whether data is ready (e.g., in its output buffer) and a second value (e.g., from a second wire between the consumer to the producer) indicates that data is ready, but it is a retransmission (e.g., not new data). The second value (e.g., from second wire) may thus keep from having two of the same data in a consumer, e.g., to avoid having two or more copies in an input buffer of a consumer PE for the same instance of an output value from a producer PE that was transmitted multiple times. Certain embodiments herein add a state element at each consumer, e.g., a reception bit. Flow control may indicate full or empty (e.g., backpressure) and indicate if a consumer took the data in a previous cycle. Producer may use knowledge of (i) if the consumer took the data, and (ii) whether the consumer may take more data, to control its output of data. Consumer PEs may send a speculation value back to a producer. Consumer PE may indicate that its target buffer is full, but producer PE may utilize the embodiments herein to determine if that target buffer is full for a consumer PE, and that consumer PE took the data (versus not taking the data and being full from a previous transmission for a different instance of an output value from the producer PE). In certain embodiments, one or more of the following aggregated values are utilized: (1) whether all the consumer PEs are full or empty, and (2) whether a consumer PE (e.g., all multicast consumer PEs) took data in the prior cycle e.g., so the backpressure value indicates no storage is available because it took the current data in that cycle or because there was and/or is no room for the data).
In one embodiment, first PE 8200A includes first storage 8201 for a success value (e.g., bit) for control output buffer 8232A, second storage 8203 for a success value (e.g., bit) for first data output buffer 8234A, and third storage 8205 for a success value (e.g., bit) for second data output buffer 8236A. Depicted scheduler 8214A is coupled to first storage 8201 to set or clear a success value (e.g., bit) therein for control output buffer 8232A, coupled to second storage 8203 to set or clear a success value (e.g., bit) therein for first data output buffer 8234A, and coupled to third storage 8205 to set or clear a success value (e.g., bit) therein for second data output buffer 8236A. In one embodiment, the scheduler 8214A sets the success value based on flow control data from the second PE 8200B and flow control data from the third PE 8200C. Some or all of the flow control data may be aggregated into a single value, e.g., sent to the first (e.g., as producer) PE 8200A. First (e.g., as producer) PE 8200A includes a (e.g., input) port 8208A(1-3) coupled to network 8210, e.g., to receive a backpressure value from second (e.g., as consumer) PE 8200B and/or third (e.g., as consumer) PE 8200C. In one circuit switched configuration, (e.g., input) port 8208A(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3)) is to receive a respective backpressure value from each one of control input buffer 8222B, first data input buffer 8224B, and second data input buffer 8226B and/or control input buffer 8222C, first data input buffer 8224C, and second data input buffer 8226C. In one embodiment, (e.g., input) port 8208A(1-3) is to receive an aggregated (e.g., single) respective backpressure value of each of (i) a backpressure value from control input buffer 8222B logically AND'd (e.g., it returns the Boolean value true (e.g., binary high, e.g., binary 1) if both input operands are true and returns false (e.g., binary 0) otherwise) with a backpressure value from control input buffer 8222C (e.g., on input 8208A(1)), (ii) a backpressure value from first data input buffer 8224B logically AND'd with a backpressure value from first data input buffer 8224C (e.g., on input 8208A(2)), and (iii) a backpressure value from second data input buffer 8226B logically AND'd with a backpressure value from second data input buffer 8226C (e.g., on input 8208A(3)). In one embodiment, an input or output marked as a (1), (2), or (3) is its own respective wire or other coupling.
First (e.g., as producer) PE 8200A includes a (e.g., input) port 8212A(1-3) coupled to network 8210, e.g., to receive a speculation value from second (e.g., as consumer) PE 8200B and/or third (e.g., as consumer) PE 8200C. In one circuit switched configuration, (e.g., input) port 8212A(1-3) (e.g., having a plurality of parallel inputs(1), (2), and (3)) is to receive a respective speculation value for each one of control input buffer 8222B, first data input buffer 8224B, and second data input buffer 8226B and/or control input buffer 8222C, first data input buffer 8224C, and second data input buffer 8226C. In one embodiment, (e.g., input) port 8212A(1-3) is to receive an aggregated (e.g., single) speculation value for each of (i) speculation value for control input buffer 8222B logically AND'd with speculation value for control input buffer 8222C (e.g., on input 8208A(1)), (ii) speculation value for first data input buffer 8224B logically AND'd with speculation value for first data input buffer 8224C (e.g., on input 8208A(2)), and (iii) speculation value for second data input buffer 8226B logically AND'd with speculation value for second data input buffer 8226C (e.g., on input 8208A(3)).
In one circuit switched configuration, a multicast data path is formed from (i) control output buffer 8232A to control input buffer 8222B and control input buffer 8222C, (ii) first data output buffer 8234A to first data input buffer 8224B and first data input buffer 8224C, (iii) second data output buffer 8236A to second data input buffer 8226B and second data input buffer 8226C, or any combination thereof. A data path may be used to send a data token from the producer PE to the consumer PEs. In the depicted embodiment, second PE 8200B includes first storage 8207 for a reception value (e.g., bit) for control input buffer 8222B, second storage 8209 for a reception value (e.g., bit) for first data input buffer 8224B, and third storage 8211 for a reception value (e.g., bit) for second data input buffer 8226B, e.g., set by scheduler 8214B. In the depicted embodiment, second (e.g., as consumer) PE 8200B includes an (e.g., output) port 8208B(1-3) coupled to network 8210, e.g., to send a backpressure value from second (e.g., as consumer) PE 8200B to first (e.g., as producer) PE 8200A. In one circuit switched configuration, (e.g., output) port 8208B(1-3) is to send a respective backpressure value for each one of control input buffer 8222B (e.g., on output 8208B(1)), first data input buffer 8224B (e.g., on output 8208B(2)), and second data input buffer 8226B (e.g., on output 8208B(3)), e.g., by scheduler 8214B. Second (e.g., as consumer) PE 8200B includes a (e.g., input) port 8212B(1-3) coupled to network 8210, e.g., to receive a success value from first (e.g., as producer) PE 8200A. In one circuit switched configuration, (e.g., input) port 8212B(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3))) is to receive a respective success value for each one of control input buffer 8222B (e.g., on input 8212B(1)), first data input buffer 8224B (e.g., on input 8212B(2)), and second data input buffer 8226B (e.g., on input 8212B(3)).
In the depicted embodiment, third PE 8200C includes first storage 8213 for a reception value (e.g., bit) for control input buffer 8222C, second storage 8215 for a reception value (e.g., bit) for first data input buffer 8224C, and third storage 8217 for a reception value (e.g., bit) for second data input buffer 8226C, e.g., set by scheduler 8214C. Third (e.g., as consumer) PE 8200C includes an (e.g., output) port 8208C(1-3) coupled to network 8210, e.g., to send a backpressure value from third (e.g., as consumer) PE 8200C to first (e.g., as producer) PE 8200A. In one circuit switched configuration, (e.g., output) port 8208C(1-3) is to send a respective backpressure value for each one of control input buffer 8222C (e.g., on output 8208C(1)), first data input buffer 8224C (e.g., on output 8208C(2)), and second data input buffer 8226C (e.g., on output 8208C(3)), e.g., by scheduler 8214C. Second (e.g., as consumer) PE 8200B includes a (e.g., input) port 8212C(1-3) coupled to network 8210, e.g., to receive a success value from first (e.g., as producer) PE 8200A. In one circuit switched configuration, (e.g., input) port 8212C(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3)) is to receive a respective success value for each one of control input buffer 8222C (e.g., on input 8212C(1)), first data input buffer 8224C (e.g., on input 8212C(2)), and second data input buffer 8226C (e.g., on input 8212C(3)).
As noted herein, speculation value may be formed by logically OR'ing the reception bit (for example, where a binary low value indicates the buffer did not take an input since it was last cleared and a binary high value indicates the buffer did take an input since it was last cleared, e.g., by the success value) and a backpressure bit (e.g., where a binary low value indicates there is no backpressure and a binary high value indicates there is backpressure). A port may include a plurality of inputs and/or outputs. A processing element may include a single port into network 8210, or any plurality of ports. Although
First PE 8200A may include first storage 8229 for a reception value (e.g., bit) for control input buffer 8222A, second storage 8231 for a reception value (e.g., bit) for first data input buffer 8224A, and third storage 8233 for a reception value (e.g., bit) for second data input buffer 8226A, e.g., set by scheduler 8214A. First (e.g., as consumer) PE 8200A may include an (e.g., output) port 8225(1-3) coupled to network 8202, e.g., to send a backpressure value from first (e.g., as consumer) PE 8200A to an upstream (e.g., as producer) PE. In one circuit switched configuration, (e.g., output) port 8225(1-3) is to send a respective backpressure value for each one of control input buffer 8222A (e.g., on output 8225(1)), first data input buffer 8224A (e.g., on output 8225(2)), and second data input buffer 8226A (e.g., on output 8225(3)), e.g., by scheduler 8214A. First (e.g., as consumer) PE 8200A includes a (e.g., input) port 8227(1-3) coupled to network 8202, e.g., to receive a success value from an upstream (e.g., as producer) PE. In one circuit switched configuration, (e.g., input) port 8227(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3))) is to receive a respective success value for each one of control input buffer 8222A (e.g., on input 8227(1)), first data input buffer 8224A (e.g., on input 8227(2)), and second data input buffer 8226A (e.g., on input 8227(3)).
Second (e.g., as producer) PE 8200B may include a (e.g., input) port 8235(1-3) coupled to network 8204 (e.g., which may be the same network as network 8206), e.g., to receive a backpressure value from a downstream (e.g., as consumer) PE or PEs. In one circuit switched configuration, (e.g., input) port 8235(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3)) is to receive a respective backpressure value from each one of control input buffer, first data input buffer, and second data input buffer of a first downstream PE and/or control input buffer, first data input buffer, and second data input buffer of a second downstream PE. In one embodiment, (e.g., input) port 8235(1-3) is to receive an aggregated (e.g., single) respective backpressure value of each of (i) a backpressure value from control input buffer for first downstream PE logically AND'd (e.g., it returns the Boolean value true (e.g., binary high, e.g., binary 1) if both input operands are true and returns false (e.g., binary 0) otherwise) with a backpressure value from control input buffer for second downstream PE (e.g., on input 8235(1)), (ii) a backpressure value from first data input buffer for first downstream PE logically AND'd with a backpressure value from first data input buffer for first downstream PE (e.g., on input 8235(2)), and (iii) a backpressure value from second data input buffer for first downstream PE logically AND'd with a backpressure value from second data input buffer for first downstream PE (e.g., on input 8235(3)). In one embodiment, an input or output marked as a (1), (2), or (3) is its own respective wire or other coupling. In one embodiment, each PE includes the same circuitry and/or components.
Second PE 8200B includes first storage 8239 for a success value (e.g., bit) for control output buffer 8232B, second storage 8241 for a success value (e.g., bit) for first data output buffer 8234B, and third storage 8243 for a success value (e.g., bit) for second data output buffer 8236B. Depicted scheduler 8214B is coupled to first storage 8239 to set or clear a success value (e.g., bit) therein for control output buffer 8232B, coupled to second storage 8241 to set or clear a success value (e.g., bit) therein for first data output buffer 8234B, and coupled to third storage 8243 to set or clear a success value (e.g., bit) therein for second data output buffer 8236B. In one embodiment, the setting of the success value in storage 8239 causes a success value to be sent on a path from storage 8239 through network 8204 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from second PE 8200B (e.g., from storage 8239 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, the setting of the success value in storage 8241 causes a success value to be sent on a path from storage 8241 through network 8204 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from second PE 8200B (e.g., from storage 8241 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, receipt of success value from second PE 8200B (e.g., from storage 8243 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, the setting of the success value in storage 8243 causes a success value to be sent on a path from storage 8243 through network 8204 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from second PE 8200B (e.g., from storage 8243 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer.
Second (e.g., as producer) PE 8200B may include a (e.g., input) port 8237(1-3) coupled to network 8204, e.g., to receive a speculation value from a first downstream (e.g., as consumer) PE and/or second downstream (e.g., as consumer) PE. In one circuit switched configuration, (e.g., input) port 8237(1-3) (e.g., having a plurality of parallel inputs(1), (2), and (3)) is to receive a respective speculation value for each one of control input buffer for first downstream PE, first data input buffer for first downstream PE, and second data input buffer for first downstream PE and/or control input buffer for second downstream PE, first data input buffer for second downstream PE, and second data input buffer for second downstream PE. In one embodiment, (e.g., input) port 8237(1-3) is to receive an aggregated (e.g., single) speculation value for each of (i) speculation value for control input buffer for first downstream PE logically AND'd with speculation value for control input buffer for second downstream PE (e.g., on input 8237(1)), (ii) speculation value for first data input buffer for first downstream PE logically AND'd with speculation value for first data input buffer for second downstream PE (e.g., on input 8237(2)), and (iii) speculation value for second data input buffer for first downstream PE logically AND'd with speculation value for second data input buffer for second downstream PE (e.g., on input 8237(3)).
Second (e.g., as producer) PE 8200B may include a (e.g., input) port 8235(1-3) coupled to network 8204 (e.g., which may be the same network as network 8206), e.g., to receive a backpressure value from a downstream (e.g., as consumer) PE or PEs. In one circuit switched configuration, (e.g., input) port 8235(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3)) is to receive a respective backpressure value from each one of control input buffer, first data input buffer, and second data input buffer of a first downstream PE and/or control input buffer, first data input buffer, and second data input buffer of a second downstream PE. In one embodiment, (e.g., input) port 8235(1-3) is to receive an aggregated (e.g., single) respective backpressure value of each of (i) a backpressure value from control input buffer for first downstream PE logically AND'd (e.g., it returns the Boolean value true (e.g., binary high, e.g., binary 1) if both input operands are true and returns false (e.g., binary 0) otherwise) with a backpressure value from control input buffer for second downstream PE (e.g., on input 8235(1)), (ii) a backpressure value from first data input buffer for first downstream PE logically AND'd with a backpressure value from first data input buffer for first downstream PE (e.g., on input 8235(2)), and (iii) a backpressure value from second data input buffer for first downstream PE logically AND'd with a backpressure value from second data input buffer for first downstream PE (e.g., on input 8235(3)). In one embodiment, an input or output marked as a (1), (2), or (3) is its own respective wire or other coupling. In one embodiment, each PE includes the same circuitry and/or components.
Second (e.g., as producer) PE 8200B may include a (e.g., input) port 8237(1-3) coupled to network 8204, e.g., to receive a speculation value from a first downstream (e.g., as consumer) PE and/or second downstream (e.g., as consumer) PE. In one circuit switched configuration, (e.g., input) port 8237(1-3) (e.g., having a plurality of parallel inputs(1), (2), and (3)) is to receive a respective speculation value for each one of control input buffer for first downstream PE, first data input buffer for first downstream PE, and second data input buffer for first downstream PE and/or control input buffer for second downstream PE, first data input buffer for second downstream PE, and second data input buffer for second downstream PE. In one embodiment, (e.g., input) port 8237(1-3) is to receive an aggregated (e.g., single) speculation value for each of (i) speculation value for control input buffer for first downstream PE logically AND'd with speculation value for control input buffer for second downstream PE (e.g., on input 8237(1)), (ii) speculation value for first data input buffer for first downstream PE logically AND'd with speculation value for first data input buffer for second downstream PE (e.g., on input 8237(2)), and (iii) speculation value for second data input buffer for first downstream PE logically AND'd with speculation value for second data input buffer for second downstream PE (e.g., on input 8237(3)).
Third (e.g., as producer) PE 8200C may include a (e.g., input) port 8245(1-3) coupled to network 8206 (e.g., which may be the same network as network 8204), e.g., to receive a backpressure value from a downstream (e.g., as consumer) PE or PEs. In one circuit switched configuration, (e.g., input) port 8245(1-3) (e.g., having a plurality of parallel inputs (1), (2), and (3)) is to receive a respective backpressure value from each one of control input buffer, first data input buffer, and second data input buffer of a first downstream PE and/or control input buffer, first data input buffer, and second data input buffer of a second downstream PE. In one embodiment, (e.g., input) port 8245(1-3) is to receive an aggregated (e.g., single) respective backpressure value of each of (i) a backpressure value from control input buffer for first downstream PE logically AND'd (e.g., it returns the Boolean value true (e.g., binary high, e.g., binary 1) if both input operands are true and returns false (e.g., binary 0) otherwise) with a backpressure value from control input buffer for second downstream PE (e.g., on input 8245(1)), (ii) a backpressure value from first data input buffer for first downstream PE logically AND'd with a backpressure value from first data input buffer for first downstream PE (e.g., on input 8245(2)), and (iii) a backpressure value from second data input buffer for first downstream PE logically AND'd with a backpressure value from second data input buffer for first downstream PE (e.g., on input 8245(3)). In one embodiment, an input or output marked as a (1), (2), or (3) is its own respective wire or other coupling. In one embodiment, each PE includes the same circuitry and/or components.
Third (e.g., as producer) PE 8200C may include a (e.g., input) port 8247(1-3) coupled to network 8206, e.g., to receive a speculation value from a first downstream (e.g., as consumer) PE and/or second downstream (e.g., as consumer) PE. In one circuit switched configuration, (e.g., input) port 8247(1-3) (e.g., having a plurality of parallel inputs(1), (2), and (3)) is to receive a respective speculation value for each one of control input buffer for first downstream PE, first data input buffer for first downstream PE, and second data input buffer for first downstream PE and/or control input buffer for second downstream PE, first data input buffer for second downstream PE, and second data input buffer for second downstream PE. In one embodiment, (e.g., input) port 8247(1-3) is to receive an aggregated (e.g., single) speculation value for each of (i) speculation value for control input buffer for first downstream PE logically AND'd with speculation value for control input buffer for second downstream PE (e.g., on input 8247(1)), (ii) speculation value for first data input buffer for first downstream PE logically AND'd with speculation value for first data input buffer for second downstream PE (e.g., on input 8247(2)), and (iii) speculation value for second data input buffer for first downstream PE logically AND'd with speculation value for second data input buffer for second downstream PE (e.g., on input 8247(3)).
Third PE 8200C includes first storage 8249 for a success value (e.g., bit) for control output buffer 8232C, second storage 825 for a success value (e.g., bit) for first data output buffer 8234C, and third storage 8253 for a success value (e.g., bit) for second data output buffer 8236C. Depicted scheduler 8214C is coupled to first storage 8249 to set or clear a success value (e.g., bit) therein for control output buffer 8232C, coupled to second storage 8251 to set or clear a success value (e.g., bit) therein for first data output buffer 8234C, and coupled to third storage 8253 to set or clear a success value (e.g., bit) therein for second data output buffer 8236C. In one embodiment, the setting of the success value in storage 8249 causes a success value to be sent on a path from storage 8249 through network 8204 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from third PE 8200C (e.g., from storage 8249 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, the setting of the success value in storage 8251 causes a success value to be sent on a path from storage 8251 through network 8204 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from third PE 8200C (e.g., from storage 8251 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, receipt of success value from third PE 8200C (e.g., from storage 8253 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer. In one embodiment, the setting of the success value in storage 8253 causes a success value to be sent on a path from storage 8253 through network 8204 to (e.g., input) port of (e.g., as consumer) a first downstream PE and to (e.g., input) port of (e.g., as consumer) a second downstream PE. In one embodiment, receipt of success value from third PE 8200C (e.g., from storage 8243 thereof) by first downstream PE or second downstream PE is to cause the clearing of their reception bit in storage for that input buffer.
A processing element may include two sub-networks (or two channels on the network), e.g., one for a data path and one for a flow control path. A processing element (e.g., PE 8200A, PE 8200B, and PE 8200C) may function and/or include the components as in any of the disclosure herein. A processing element may be stalled from execution until its operands (e.g., in its input buffer(s)) are received and/or until there is room in the output buffer(s) of the processing element for the data that is to be produced by the execution of the operation on those operands. Next, three reduced multicast critical path embodiments are discussed.
As a first example,
First (e.g., as producer) PE 8200A includes a (e.g., input) port 8212A(1) coupled to network 8210, e.g., to receive a speculation value from second (e.g., as consumer) PE 8200B and/or third (e.g., as consumer) PE 8200C. In one circuit switched configuration, (e.g., input) port 8212A(1) is to receive a respective speculation value for each one of control input buffer 8222B and control input buffer 8222C. In the depicted embodiment, (e.g., input) port 8212A(1) is to receive an aggregated (e.g., single) speculation value for speculation value for control input buffer 8222B logically AND'd with speculation value for control input buffer 8222C by AND logic gate 8250. In the depicted embodiment, the speculation value for control input buffer 8222B is formed by OR'ing the reception bit for the speculative path (e.g., reception bit from storage 8207) (e.g., where a binary low value indicates the buffer did not store an input since it was last cleared) and a backpressure bit from backpressure path (e.g., from port 8208B(1)) (e.g., where a binary low value indicates there is no backpressure) by OR logic gate 8254. In the depicted embodiment, the speculation value for control input buffer 8222C is formed by OR'ing the reception bit for the speculative path (e.g., reception bit from storage 8213) (e.g., where a binary low value indicates the buffer did not store an input since it was last cleared) and a backpressure bit from backpressure path (e.g., from port 8208C(1)) (e.g., where a binary low value indicates there is no backpressure) by OR logic gate 8256. In one embodiment, a PE (e.g., scheduler thereof) is to set (e.g., to binary high) a reception value (e.g., reception bit) to indicate a value was stored in that buffer (e.g., second PE 8200B setting a reception bit in storage 8207 to indicate a dataflow token was stored (e.g., since the reception bit was last cleared) in the control input buffer 8222B and/or third PE 8200C setting a reception bit in storage 8213 to indicate a dataflow token was stored (e.g., since the reception bit was last cleared) in the control input buffer 8222C). In certain embodiments herein, logic gate functionality is achieved by using NAND/NOR circuit designs.
In one circuit switched configuration, a multicast data path is formed from control output buffer 8232A to control input buffer 8222B and control input buffer 8222C. A data path may be used to send a data token from the producer PE to the consumer PEs. In the depicted embodiment, second PE 8200B includes first storage 8207 for a reception value (e.g., bit) for control input buffer 8222B. Second (e.g., as consumer) PE 8200B includes an (e.g., output) port 8208B(1) coupled to network 8210, e.g., to send a backpressure value from second (e.g., as consumer) PE 8200B to first (e.g., as producer) PE 8200A. In one circuit switched configuration, (e.g., output) port 8208B(1) is to send a respective backpressure value for control input buffer 8222B. Second (e.g., as consumer) PE 8200B includes a (e.g., input) port 8212B(1) coupled to network 8210, e.g., to receive a success value from first (e.g., as producer) PE 8200A. In one circuit switched configuration, (e.g., input) port 8212B(1) is to receive a respective success value for control input buffer 8222B.
In the depicted embodiment, third PE 8200C includes first storage 8213 for a reception value (e.g., bit) for control input buffer 8222C. Third (e.g., as consumer) PE 8200C includes an (e.g., output) port 8208C(1) coupled to network 8210, e.g., to send a backpressure value from third (e.g., as consumer) PE 8200C to first (e.g., as producer) PE 8200A. In one circuit switched configuration, (e.g., output) port 8208C(1) is to send a respective backpressure value for control input buffer 8222C. Second (e.g., as consumer) PE 8200B includes a (e.g., input) port 8212C(1) coupled to network 8210, e.g., to receive a success value from first (e.g., as producer) PE 8200A. In one circuit switched configuration, (e.g., input) port 8212C(1) is to receive a respective success value for control input buffer 8222C.
In one embodiment, a data token is received in control output buffer 8232A which causes the reduced multicast critical path of the first example to begin operation, e.g., as discussed below in reference to
In the first transmission attempt for this dataflow token, if the backpressure value (e.g., ready value) on the path from port 8208B(1) of second PE 8200B to port 8208A(1) of first PE 8200A and the backpressure value (e.g., ready value) on the path from port 8208C(1) of third PE 8200C to port 8208A(1) of first PE 8200A both indicate (e.g., as the output from AND logic gate 8252) there is no backpressure (e.g., there is storage available in each of control input buffer 8222B and control input buffer 8222C), then the first PE (e.g., scheduler 8214A) determines that this transmission attempt will be successful, for example, and the dataflow token is to be dequeued (e.g., in the next cycle) from the control output buffer 8232A of the first PE 8200A and/or the success value (e.g., success bit) in first storage 8201 is set (e.g., in the next cycle) to indicate a successful transmission. In the first transmission attempt for this data token, if the backpressure value (e.g., ready value) on the path from port 8208B(1) of second PE 8200B to port 8208A(1) of first PE 8200A or the backpressure value (e.g., ready value) on the path from port 8208C(l) of third PE 8200C to port 8208A(1) of first PE 8200A indicate (e.g., as the output from AND logic gate 8252) there is backpressure (e.g., there is not storage available in both (e.g., all) of control input buffer 8222B and control input buffer 8222C, respectively), then one or more retransmissions of that dataflow token will occur until the speculation value from each of second (e.g., as consumer) PE 8200B and third (e.g., as consumer) PE 8200C indicates speculation is true, for example, until the speculation value is driven to a value by each of second (e.g., as consumer) PE 8200B and third (e.g., as consumer) PE 8200C that indicates that consumer PE (e.g., receiver) has either (i) accepted the data sent by the producer PE 8200A, e.g., as noted by the reception value (e.g., reception bit) being set (e.g., in a previous cycle) (e.g., in storage 8207 or storage 8213, respectively) or (ii) that the consumer is ready (e.g., by the next cycle) to receive the dataflow token (e.g., the backpressure value indicates that storage is currently available). For example, where the speculation value for control input buffer 8222B is formed by OR'ing the reception bit for the speculative path (e.g., reception bit from storage 8207) (e.g., where a binary low value indicates the buffer did not store an input since it was last cleared) and a backpressure bit from backpressure path (e.g., from port 8208B(1)) (e.g., where a binary low value indicates there is no backpressure) by OR logic gate 8254. In one embodiment, once the speculation values (e.g., from speculation paths) indicate the dataflow token is to be stored (e.g., in the next cycle) in control input buffer 8222B and control input buffer 8222C, the success value (e.g., a single bit) is driven by the producer PE 8200A to a value that the producer PE was able to successfully complete a transmission in the previous cycle (e.g., the value is stored in all of the multicast consumer PEs), e.g., as noted by the success value (e.g., success bit) (e.g., binary high, e.g., binary 1) being set in storage 8201. In one embodiment, the setting of the success value in storage 8201 causes a success value to be sent on a path from storage 8201 through network 8210 to (e.g., input) port 8212B(1) of (e.g., as consumer) second PE 8200B and to (e.g., input) port 8212C(1) of (e.g., as consumer) third PE 8200C. In one embodiment, receipt of success value from first PE 8200A (e.g., from storage 8201 thereof) by second PE 8200B is to cause the clearing of the reception bit in storage 8207, e.g., by scheduler 8214B. In one embodiment, receipt of success value from first PE 8200A (e.g., from storage 8201 thereof) by third PE 8200C is to cause the clearing of the reception bit in storage 8213, e.g., by scheduler 8214C.
NetPack
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a NetPack operation according to the following (e.g., semantics and/or description).
Coarse-grained reconfigurable arrays (e.g., a CSA) may support multiple data widths, yet some structures may operate at a finer or coarser granularity. For example, it may be more efficient from an area perspective to build a larger (e.g., 64 bit wide) communications networks or memory interfaces, but also to support smaller (e.g., 32 bit and smaller) data types. Due to properties of the PE interconnection in a circuit-switched network in a CSA, is possible to implement some data manipulations in parallel with network transit.
In certain embodiments, NetPack creates a single value from two smaller values sent from each of a plurality of transmitting PEs to the single receiving PE.
In certain embodiments, each local network endpoint is a mux, and the network includes AND logic gates and OR logic gates (or other logic gates) to combine signals from PEs, and NetPack utilizes the network to pack several data elements into one packed data value.
Unlike an embodiment of a pick operation, both transmitting PEs are to view the flow control line and both are to dequeue values for a NetPack operation. Unlike an embodiment of an All operation, the receiver PE is to input data from all (e.g., both) transmitting PEs, e.g., and this data is be latched in the cycle that it is ready. In certain embodiments, one transmitting PE may have a value to transmit, but the other transmitting PE does not, so stalling may be utilized, for example, by using the “opComplete” values from the receiver and state bit to prevent too many dequeues (e.g., state bit is to indicate (remember) the dequeue and prevent additional dequeues). In certain embodiments, the receiver PE is to observe flow control of both transmitter PEs to determine completion of the pack operation and set the pack complete indication (e.g., register) accordingly. In certain embodiments, the enqueue indication is modified to allow for the enqueue of multiple data elements into different portions of the receiver input queue. In one embodiment, the transmitter FIFO status state machine is augmented with bit tracking whether it has dequeued (e.g., queue status not updated if already updated) and the bit is cleared on ‘complete’ indication.
Depicted output controller circuitry 8300 includes a Status determiner 8304, a Not Full determiner 8306, and an Out determiner 8308. A determiner may be implemented in software or hardware. A hardware determiner may be a circuit implementation, for example, a logic circuit programmed to produce an output based on the inputs into the state machine(s) discussed below. Depicted (e.g., new) Status determiner 8304 includes a port coupled to queue status register 8302 to read and/or write to output queue status register 8302.
Depicted Status determiner 8304 includes a first input to receive a Ready value from a receiving component (e.g., a downstream PE) that indicates if (e.g., when) there is space (e.g., in an input queue thereof) for new data to be sent to the PE and a second input to receive a Complete value from the receiving component (e.g., a downstream PE) that indicates if (e.g., when) the NetPack operation is complete. In certain embodiments, the Ready value from the receiving component is sent by an input controller that includes input controller circuitry 3400 in
As discussed further below, the depicted Status determiner 8304 includes a first output to send a value on path 8310 that will cause output data (sent to the output queue that output controller circuitry 8300 is controlling) to be enqueued into the output queue or not enqueued into the output queue. Depicted Status determiner 8304 includes a second output to send an updated value to be stored in queue status register 8302, e.g., where the updated value represents the updated status (e.g., head value, tail value, count value, or any combination thereof) of the output queue that output controller circuitry 8300 is controlling.
Output controller circuitry 8300 includes a Not Full determiner 8306 that determines a Not Full (e.g., Ready) value and outputs the Not Full value, e.g., within the PE that includes output controller circuitry 8300, to indicate if (e.g., when) there is storage space available for output data in the output queue being controlled by output controller circuitry 8300. In one embodiment, for an output queue of a PE, a Not Full value that indicates there is no storage space available in that output queue is to cause a stall of execution of the PE (e.g., stall execution that is to cause a resultant to be stored into the storage space) until storage space is available (e.g., and when there is available data in the input queue(s) being sourced from in that PE).
Output controller circuitry 8300 includes an Out (e.g., logic) determiner 8308 that determines an output storage (queue) status value and outputs (e.g., on path 3345 or path 3347 in
For example, assume that the operation that is to be performed is to send (e.g., sink) data into both output storage 3334 and output storage 3336 in
In comparison to
State machine in
The && symbol indicates a logical AND operation. The ∥ symbol indicates a logical OR operation. The ! symbol indicates a logical NOT operation.
State machine in
State machine in
The state machine in
In one embodiment, a circuit switched network 9710 includes (i) a data path to send data from first PE 9700A to third PE 9700C and a data path from second PE 9700B to third PE 9700C, and (ii) a flow control path to send control values that controls (or is used to control) the sending of that data from first PE 9700A and second PE 9700B to third PE 9700C. Data path may send a data (e.g., valid) value when data is in an output queue (e.g., buffer) (e.g., when data is in control output buffer 9732A, first data output buffer 9734A, or second data output queue (e.g., buffer) 9736A of first PE 9700A and when data is in control output buffer 9732B, first data output buffer 9734B, or second data output queue (e.g., buffer) 9736B of second PE 9700B). In one embodiment, each output buffer includes its own data path, e.g., for its own data value from producer PE to consumer PE. Components in PE are examples, for example, a PE may include only a single (e.g., data) input buffer and/or a single (e.g., data) output buffer. Flow control path may send control data that controls (or is used to control) the sending of corresponding data from first PE 9700A and second PE 9700B to third PE 9700C. Flow control data may include a backpressure value from each consumer PE (or aggregated from all consumer PEs, e.g., with an AND logic gate). Flow control data may include a backpressure value, for example, indicating a buffer of the third PE 9700C that is to receive an input value is full.
Turning to the depicted PEs, processing elements 9700A-C include operation configuration registers 9719A-C that may be loaded during configuration (e.g., mapping) and specify the particular operation or operations (for example, to indicate whether to enable NetPack mode or not). In one embodiment, all the operation configuration registers for transmitter PEs and the receiver PE are loaded with the operation configuration value for NetAll0 (e.g., a first configuration value for a PE to be in receiver NetAll mode and a second configuration value for a PE to be in transmitter NetAll mode).
Multiple networks (e.g., interconnects) may be connected to a processing element, e.g., networks 9702, 9704, 9706, and 9710. The connections may be switches, e.g., as discussed in reference to
Depicted network 9710 includes a dataflow path and a flow control (e.g., backpressure) path with the paths as indicated. First processing element (PE) 9700A includes storage (e.g., a register) 9705A to store a dequeue completed (done) value for its output queue(s), second processing element (PE) 9700B includes storage (e.g., a register) 9705B to store a to store a dequeue completed (done) value for its output queue(s), and third processing element (PE) 7000C includes storage (e.g., a register) 7005C to store an operation complete ready (e.g., en12ready[0]) value when third PE 9700C has received the first value from the first PE 9700A and storage (e.g., a register) 7007C to store an operation complete ready (e.g., en12ready[1]) value when third PE 9700C has received the second value from the second PE 9700A.
In
In
In
In
In
In
In one embodiment, all transmitters are required to send a value simultaneously, so the control fan-in into the receiver PE from the transmitter PEs allows the receiver to accept data only when all transmitters have data in their output queues (e.g., and are all sending a Valid indication).
Repeato
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Repeato operation according to the following (e.g., semantics and/or description).
In certain embodiments, the Repeato operation causes PE 9900 to produce a Boolean value (e.g., zero) internally in state storage 9901 when a data value is output into an output queue, but no associated control value (e.g., one or zero) has been received in input queue of the PE. In one embodiment, when the state bit in state storage 9901 is set to zero, no new data value it output unless an associated control value of one is received.
In
In
In
In
In
In
In
In certain embodiments, Repeato produces copies on an input, consumes input data values (e.g., tokens) when a control value (e.g., token) is a zero, and produces an output value for each instance of the control input (including all the 1s followed by an ending 0). In certain embodiments, since at least one copy is always produced, a value is sent speculatively, e.g., if that speculative value is sent, no value will be sent on 0.
In certain embodiments, PE 9900 is stalled from performing the Repeato operation until there is both (i) space available in the output queue that is to be used for storing resultant data, in the case that repeato would need to produce output data, for example if a control value 1 is present on the control input, and (ii) an input control value in input queue 9922, except in the case that the first output data value has not yet been produced.
In the depicted embodiment, PE 9900 includes the components of PE 5800 from
Strideo
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Strideo operation according to the following (e.g., semantics and/or description).
In certain embodiments, the Strideo operation causes PE 10000 to produce a Boolean value (e.g., zero) internally in state storage 10001 when the (e.g., strided or base) data value is output into an output queue. In one embodiment, when the state bit in state storage 10001 is set to zero, no new data value it output unless an associated control value of one is received.
In
In
In
In
In
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein. In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc. For all of the embodiments herein, the data value from the output queues may be consumed from the output queues, e.g., by a downstream PE or PEs.
In certain embodiments, Strideo produces a set of data values (e.g., strided from a base value), where a zero for a control value indicates the stride should be consumed and the updated stride value is discarded. In certain embodiments, since at least one copy is always produced, a data value (e.g., base value) is sent speculatively, e.g., if sent, the state bit is marked (e.g., to a zero) to indicate the data value has been sent.
In certain embodiments, PE 10000 is stalled from performing the Strideo operation until there is both (i) space available in the output queue that is to be used for storing resultant data in the case that strideo would need to produce an output value, for example if a control value 1 is present on the control input, and (ii) an input control value in input queue 10022 after emission of the first output to an output queue.
In the depicted embodiment, PE 10000 includes the components of PE 5800 from
Nestrepeat
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Nestrepeat operation according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein. In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc.
In certain embodiments, PE 10100 is stalled from performing the Nestrepeat operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) the inner loop control values and outer loop control values (e.g., selection control bit) are available.
In certain embodiments, PE 10100 constructs control patterns for nested loops by taking control (e.g., iteration) values from two control input streams (e.g., for the inner loop and outer loop, respectively). In certain embodiments, the Nestrepeat operation produces an output that is used to control inner loop repeats using a single repeat, e.g., according to the outer, inner, and output (out) table in the above description of Nestrepeat.
In the depicted embodiment, PE 10100 includes the components of PE 5800 from
Predfilter
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Predfilter operation according to the following (e.g., semantics and/or description).
In
In
In
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein. In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc.
In certain embodiments, PE 10200 is stalled from performing the Predfilter operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) a stream control value and a predicate stream value are available.
In certain embodiments, PE 10200, when given an iteration stream and a control stream, removes some iterations of the control stream based on the iteration stream. In certain embodiments, PE 10200 produces truncated control stream (with a 0) as an output, e.g., according to the control (ctl), predicate (pred), and output (ctlout) table in the above description of Predfilter.
In the depicted embodiment, PE 10200 includes the components of PE 5800 from
Reduction (Red*)
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a reduction (Red*) operation according to the following (e.g., semantics and/or description), e.g., where the * indicates an arithmetic or logical operation such as, but not limited to, add, subtract, AND, OR, and XOR.
In
In
In
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein. In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc. For all of the embodiments herein, the data value from the output queues may be consumed from the output queues, e.g., by a downstream PE or PEs.
In certain embodiments, Red* computes a reduction operation and carries a store value (e.g., in register 6120) and uses it as an operand. In certain embodiments, the predicate stream is used to control the operation (e.g., a summation in this example), and when the stream element is false (e.g., a Boolean zero), the stored value is output.
In certain embodiments, PE 10300 is stalled from performing the Red* operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) an input control value is in input queue 10322 and a first set of input operands (e.g., elements) for the * operation are available.
In the depicted embodiment, PE 10300 includes the components of PE 5800 from
Sequence Reduction (Sred*)
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a sequence reduction (Sred*) operation according to the following (e.g., semantics and/or description), e.g., where the * indicates an arithmetic or logical operation such as, but not limited to, add, subtract, AND, OR, and XOR.
In
In
In
In
In
The input data that is queued may be sent from another component of a CSA, e.g., from a plurality of other PEs as discussed herein. In certain embodiments, the data is read from the first slot of a queue, and when consumed (e.g., removed), any data from other slots of the queue are advanced such that data from the second slot is moved into the first slot, etc. For all of the embodiments herein, the data value from the output queues may be consumed from the output queues, e.g., by a downstream PE or PEs.
In certain embodiments, Sred* computes a reduction operation, carries the store value (e.g., in register 10420) and uses it as an operand, but also produces an output for each intermediate result. The intermediate result may be sent to second output, but may also be combined in first output (e.g. in this example all values would be written to 10434). In certain embodiments, the predicate stream is used to control the operation (e.g., a summation in this example), and when the stream element is false (e.g., a Boolean zero), the stored value is output.
In certain embodiments, PE 10400 is stalled from performing the Sred* operation until there is both (i) space available in the output queue(s) that is to be used for storing resultant data, and (ii) an input control value is in input queue 10422 and a first set of input operands (e.g., elements) for the * operation are available.
In the depicted embodiment, PE 10400 includes the components of PE 5800 from
Pack
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Pack operation according to the following (e.g., semantics and/or description).
In
In
In
Also, PE 10500 moved the second data value (labeled as circled 2) from the first slot into the second slot of input queue 10524, and the second data value (labeled as circled 3) from the first slot into the second slot of input queue 10526. A third data value (labeled as circled 4) has been stored (e.g., by an upstream PE) into the second slot of input queue 10524, and a third data value (labeled as circled 5) has been stored (e.g., by an upstream PE) into the second slot of input queue 10526. In one embodiment, the intermediate packed data value is stored into register 10520.
In
Also, PE 10500 moved the third data value (labeled as circled 4) from the first slot into the second slot of input queue 10524, and the third data value (labeled as circled 5) from the first slot into the second slot of input queue 10526. A fourth data value (labeled as circled 6) has been stored (e.g., by an upstream PE) into the second slot of input queue 10524, and a fourth data value (labeled as circled 7) has been stored (e.g., by an upstream PE) into the second slot of input queue 10526. In one embodiment, the updated, intermediate packed data value is stored into register 10520.
In
Also, PE 10500 moved the fourth data value (labeled as circled 6) from the first slot into the second slot of input queue 10524, and the fourth data value (labeled as circled 7) from the first slot into the second slot of input queue 10526. In one embodiment, the updated, intermediate packed data value is stored into register 10520.
In
In
In one embodiment, a pack operation sources data values from two (e.g., wide) input queues, and the components to be packed are streaming in serially. In another embodiment, a PE includes more than two input queues (e.g., eight input queues) to support the packing of all of the data elements into the resultant packed data value in parallel (e.g., in one cycle of that PE).
In certain embodiments, PE 10500 is stalled from performing the Pack operation until there is both (i) space available in the output queue that is to be used for storing resultant data, and (ii) a control input value in input queue (e.g., 10517) (for example, using a control stream to control (e.g., start and/or end) the packing operation), a data input value in input queue 10524, and a data input value in input queue 10526.
In the depicted embodiment, PE 10500 includes the components of PE 5800 from
Unpack
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform an Unpack operation according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In
In
In
In
In
In certain embodiments, PE 10600 is stalled from performing the Unpack operation until there is both (i) space available in any output queues that are to be used for storing resultant data, and (ii) a packed data value is available in an input queue.
In the depicted embodiment, PE 10600 includes the components of PE 5800 from
Gate
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to perform a Gate operation according to the following (e.g., semantics and/or description).
In
In
In
In
In certain embodiments, PE 10700 is stalled from performing the Gate operation when there is an input data value and a gate control value until there is space available in the output queue that is to be used for storing resultant data.
In certain embodiments, PE 10700 configured to perform a gate operation is used to synchronize values.
In the depicted embodiment, PE 10700 includes the components of PE 5800 from
Storage (Buffer Box Element) Operations
In certain embodiments of a CSA architecture, local storage mechanisms are used to store temporary data, implement read-only-memory (ROM), and/or add buffering to certain portions (e.g., legs) of a dataflow graph executing on the CSA. A buffer box element is discussed below to provide storage. In one embodiment, the buffer box element uses the same communications protocols as a processing element, for example, such that one or more processing elements in a CSA are replaced by a corresponding buffer box element. One embodiment of a buffer box element supports multiple type of data storage mechanisms, e.g., those used by most dataflow graphs. In certain embodiments, a buffer box element is used for storage of certain data instead of sending/receiving that data in main memory (e.g., memory that is accessed via a RAF circuit as discussed herein). Accessing data from a (e.g., local) buffer box element is faster and more flexible (e.g., the buffer box element(s) may be placed within the CSA in any desired location).
In certain embodiments, a buffer box element is a configurable storage element of a CSA that implements multiple data storage types of modes. In certain embodiments, the buffer box element(s) allow a CSA to support data storage mechanisms within the mapped dataflow graphs in a flexible and reusable manner. In one embodiment, a buffer box element is a CSA compute circuit that includes a (e.g., small) shared memory that can act as a RAM, ROM, or buffer (first-in, first-out (FIFO) buffer) connected to the dataflow graph instantiated in the CSA array. In one embodiment, the primary purpose of the RAM mode is to serve as a small scratchpad memory, e.g., without being part of a CSA's main coherent memory space. In one embodiment, the primary purpose of the ROM mode is to supply constants, e.g., without being part of a CSA's main coherent memory space. A buffer box element uses less power and communication (e.g., network) bandwidth to access (e.g., store and/or load) data than main memory. In one embodiment, the primary purpose of the FIFO mode is to provide additional buffering in the dataflow graph (e.g., along the links thereof) to achieve optimal throughput.
In the depicted embodiment, input queues 10804 and 10806 (e.g., narrower queues), and input queues 10824 and 10826 (e.g., wider queues) are coupled to local network(s) 10802 (e.g., and local network 10802 may include a data path network as in
In certain embodiments, multiple networks (e.g., LICs thereof) are connected to a buffer box element, e.g., (input) network(s) 10802 and (output) network(s) 10812. The connections may be switches, e.g., as discussed in reference to
In
In certain embodiments, a buffer box element or elements support a variety of operational modes within the dataflow paradigm. In order to have basic functionality, in one embodiment, the smallest realization of the buffer box element has two wide local network inputs, two wide local network outputs, one narrow (1-bit, control input (ctl_in)) local network input, and one narrow (1-bit, control output (ctl_out)) local network output.
As one embodiment, the FIFO Buffer mode (e.g., the default mode) being selected for a buffer box element causes the buffer (e.g., buffers 10840 and 10850 in
As one embodiment, the FIFO Buffer preload mode being selected for a buffer box element allows a buffer (e.g., buffers 10840 and 10850 in
As one embodiment, the repeat mode being selected for a buffer box element allows a buffer (e.g., buffers 10840 and 10850 in
As one embodiment, the random access memory (RAM) mode being selected for a buffer box element allows a buffer (e.g., buffers 10840 and 10850 in
As one embodiment, the read only memory (ROM) mode being selected for a buffer box element allows a buffer (e.g., buffers 10840 and 10850 in
As one embodiment, the random access memory (RAM) mode being selected for a buffer box element uses a buffer (e.g., buffers 10840 and 10850 in
As one embodiment, the Fast-clear RAM mode (e.g., Clearing RAM between configurations mode) being selected for a buffer box element allows a buffer (e.g., buffers 10840 and 10850 in
As one embodiment, the Fast-clear RAM mode (e.g., Clearing RAM in the course of a calculation mode) being selected for a buffer box element allows a buffer (e.g., buffers 10840 and 10850 in
In one embodiment, the RAM mode (number 4 above) of a buffer box element uses two wide inputs (addr and data_in) (e.g., two wide input queues) and one wide output (data_out) (e.g., one wide output queue). As one embodiment, the Streaming-unload RAM mode being selected for a buffer box element allows a buffer (e.g., buffers 10840 and 10850 in
As one embodiment, the completion buffer mode being selected for a buffer box element allows a buffer (e.g., buffers 10840 and 10850 in
As one embodiment, the Fast-clear FIFO Buffer mode (e.g., the default mode) being selected for a buffer box element causes the buffer (e.g., buffers 10840 and 10850 in
In certain embodiments, any of these modes is expanded upon by allowing overflow of values to go to memory (e.g., main memory) to increase the effective buffer size to be bigger than the physical buffer in the buffer box element. In one embodiment, the memory used for the overflow of the physical buffer is not forced to be coherent with system memory, thus still acting as a scratchpad implementation.
FIFO Mode
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in a FIFO mode according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In
In certain embodiments, buffer box element 11100 is stalled from sending data stored in its buffer storage until there is room in the target storage location (e.g., the output queue that is coupled to an input queue of a receiving PE or PEs).
In the depicted embodiment, buffer box element 11100 includes the components of buffer box element 10800 from
FIFO Mode with Retention
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in a FIFO mode with retention according to the following (e.g., semantics and/or description).
Thus, “FIFO with retention” mode allows the specification (e.g., by a programmer or compiler) that a particular buffer box element (e.g., buffer thereof) will retain some predetermined number of prior values (e.g., dataflow tokens). In one embodiment, this is achieved by limiting the number of values that can be alive in the buffer at a given point in time. In certain embodiments, retained values are visible to software flows, e.g., and can be used to recover machine state in case of an error. In certain embodiments, a buffer box element (e.g., buffer thereof) includes a programmable threshold which guarantees the retention of a proper subset (e.g., a plurality of) values. In the case of error, these retained dataflow tokens are then used to reconstruct or replay erroneous values, e.g., allowing the CSA to self-recover in an automated fashion. Moreover, since buffer box elements are programmable, they can be disabled during executions of certain dataflow graphs, e.g., those which do not require a high level of reliability.
In
In contrast to
Preload Mode
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in a Preload mode according to the following (e.g., semantics and/or description).
In the depicted embodiment, a preload has occurred that has stored the (e.g., data) value labeled circled 1 into the (e.g., first storage location of) buffer storage 11440, the (e.g., data) value labeled circled 2 into the (e.g., second storage location of) buffer storage 11440, the (e.g., configuration) value labeled circled 3 into the (e.g., first storage location of) buffer storage 11450, and the (e.g., configuration) value labeled circled 4 into the (e.g., second storage location of) buffer storage 11450.
In certain embodiments, during configuration, a bit is set indicating the next values received by buffer box element 11400 are to be preloaded into the buffer storage.
In the depicted embodiment, buffer box element 11400 includes the components of buffer box element 10800 from
Repeat (Using Buffer Box Element) Mode
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in a repeat mode according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In
In certain embodiments, buffer box element 11500 is stalled from sending data stored in its buffer storage until there is room in the target storage location (e.g., the output queue that is coupled to an input queue of a receiving PE or PEs).
In the depicted embodiment, buffer box element 11500 includes the components of buffer box element 10800 from
Repeat—Controlled Mode
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in a Repeat—controlled mode according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In
In
In
In certain embodiments, buffer box element 11600 is stalled from sending data stored in its buffer storage until there is room in the target storage location (e.g., the output queue that is coupled to an input queue of a receiving PE or PEs) even when the control value is set to one.
In the depicted embodiment, buffer box element 11600 includes the components of buffer box element 10800 from
RAM Mode
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in a RAM mode according to the following (e.g., semantics and/or description).
In
In
In
Also, scheduler 11714 dequeues the address value labeled circled 1 from the input queue 11726, and the corresponding data value (for the address value) from the input queue 11724. An address value labeled circled 2 is stored into the first slot of input queue 11726, and the corresponding data value (for the address value) labeled circled 2 is stored into the first slot of input queue 11724, e.g., by an upstream component that is requesting that store.
In
Also, scheduler 11714 dequeues the address value labeled circled 2 from the input queue 11726, and the corresponding data value (for the address value) labeled circled 2 from the input queue 11724.
In
In
In
In
In certain embodiments, buffer box element 11700 is stalled from sending data stored in its buffer storage until there is room in the target storage location (e.g., the output queue that is coupled to an input queue of a receiving PE or PEs).
In the depicted embodiment, buffer box element 11700 includes the components of buffer box element 10800 from
In certain embodiments, a CSA includes in network storage (e.g., as scratchpads). The in-network storage (e.g., a buffer box element) may be treated as part of the general memory (storage) access paradigm, which uses loads and stores, and which can be ordered by using ordering operands.
Each memory access operation instance can be configured to access a single disjoint address space. In one embodiment, there is not a provision for a unified address space that allows access to both standard memory and scratchpads transparently to a single operation. By default, in certain embodiments, memory operations (e.g., load and store) access the standard process address space. However, it is possible to declare storage that should be allocated to scratchpad, if available, and memory operations which can access that specific storage. In that embodiment, addresses that are outside the bounds of a given address space are an error and there is no assurance that addresses in different address spaces are disjoint (e.g. scratchpad addresses may simply be an index relative to the scratchpad, which might appear to be address 0 for each independent scratchpad). In certain embodiments, accesses use the scratchpad address as the base for a displacement or indexed memory access. However, this is not a requirement, to allow other kinds of pointer accesses. Unlike general memory, references to scratchpad storage are naturally aligned in certain embodiments.
Random Access Assembly Language Handling
In one embodiment to specify storage should be allocated to a scratchpad, it should be placed in a section that starts with the string “.csa.sp.”, for example, where each individual scratchpad section specifies a separable allocation—a separate “address space”.
Note: In C/C++, variables can be allocated to sections using attributes, e.g.:
Small scratchpads (e.g., less than or equal to 64×64 bit elements in size) may only allow a single static load and single static store operation. This may make them very difficult to use from high level code.
In assembly code, the .addrspace [{section}] directive can be used to specify the address space that applies to all following references which do not have an explicit base. If no section is specified, the address space that applies reverts to the standard address space. e.g.
In certain embodiments, code does not make assumptions about whether a given section will actually be allocated to scratchpad, e.g., the runtime and toolchain (e.g., place and route, etc.) is to decide whether the address space (.addrspace) is to be implemented in a hardware buffer box element, e.g., or instead implemented in shared virtual memory. In particular, it is possible that for the above program, .csa.sp.foo will reside in memory, with a standard memory address. On the other hand, it may be allocated to scratchpad, in which case the effective address of sym might be 0, some small integer value, or some other bitpattern entirely.
Streaming-Unload RAM Mode
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in a Streaming-unload RAM mode according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In
In
In certain embodiments, buffer box element 11800 is stalled from sending data stored in its buffer storage until there is room in the target storage location (e.g., the output queue that is coupled to an input queue of a receiving PE or PEs) even when the control value is set to one.
In the depicted embodiment, buffer box element 11800 includes the components of buffer box element 10800 from
ROM Mode
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in a ROM mode according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In certain embodiments, buffer box element 11900 is stalled from sending data stored in its buffer storage until there is room in the target storage location (e.g., the output queue that is coupled to an input queue of a receiving PE or PEs).
In the depicted embodiment, buffer box element 11900 includes the components of buffer box element 10800 from
Stack (e.g., Stack Mode)
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a CSA (e.g., a PE thereof) to implement a stack according to the following (e.g., semantics and/or description).
A CSA component (e.g., a PE) supplies a value to be stored on the stack (e.g., pushed) for the “valin” in the above operation, and a (e.g., different or same) CSA component (e.g., a PE) loads a value from the stack (e.g., popped) as the “pop” in the above operation.
The stack itself may be implemented in a buffer box element.
Stack (Using Buffer Box Element) Mode
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in a stack mode according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In
In
In
In certain embodiments, buffer box element 12100 also uses buffer storage 12150 to push and/or pop values, e.g., as indicated by a configuration value.
In certain embodiments, buffer box element 12100 stalls the popping data stored in its buffer storage until there is room in the configured target storage location (e.g., the output queue of the buffer box element 12100 that is coupled to an input queue of a receiving PE or PEs).
In the depicted embodiment, buffer box element 12100 includes the components of buffer box element 10800 from
Completion Buffer Mode
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in a Completion Buffer mode according to the following (e.g., semantics and/or description).
Another mode of operation for in-fabric storage (e.g., buffer box element) is as a completion buffer, e.g., where values are inserted in a sliding window, and values can be retrieved in order.
The completion operation supports this functionality. Note that the example completion operation below can be viewed as these three operations bound together:
In
In
In
Thus, in
In
Also in
In
Also in
In
In
In
In
In certain embodiments, buffer box element 12200 is stalled from sending data stored in its buffer storage until there is room in the target storage location (e.g., the output queue that is coupled to an input queue of a receiving PE or PEs).
In the depicted embodiment, buffer box element 12200 includes the components of buffer box element 10800 from
Overflow Buffer Mode
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in an overflow buffer mode according to the following (e.g., semantics and/or description).
In
In
In
In
In
In
In
In
In certain embodiments, buffer box element 12300 uses external memory 12301 to extend the size of the storage available for buffering data.
In certain embodiments, buffer box element 12300 is not stalled from storing data until there is no room available in the external memory 12301.
In the depicted embodiment, buffer box element 12300 includes the components of buffer box element 10800 from
Depicted accelerator tile 12408 includes in-fabric storage 12406 (e.g., a buffer box element according to this disclosure) that uses the overflow mode discussed above. As one example, in-fabric storage 12406 is used as a stack (e.g., according to the stack operation) by a CSA component (e.g., PE 12410 and PE 12416). At the times indicated by a circled number, time 1 indicates a push of data from PE 12410 onto the stack implemented in in-fabric storage 12406. When the in-fabric storage uses the overflow mode (e.g., when the data to be pushed will not fit in the in-fabric storage or the new data being pushed would also push the oldest data out of the stack), (e.g., the oldest) data from the stack (e.g., data that is being pushed out of the stack from the new data value being pushed onto the stack) is sent by the in-fabric storage 12406 (e.g., as discussed above in reference to the buffer box element 12300) to external storage (depicted here as cache bank (1)).
As one example, the request to store the (e.g., oldest) data value from the stack (e.g., along with that data value) is sent at circled 2 to RAF circuit 12404, which receives (e.g., queues) the request (and data value) in the RAF circuit 12404 at circled 3.
In one embodiment, RAF circuit 12404 is to send a memory request (e.g., generated by in-fabric storage 12406) into accelerator-cache network 12414 (e.g., ACI as described herein), for example, to be serviced by one of the cache banks, and the corresponding data for a memory request of that data (e.g., payload data for a load request) is steered back to RAF circuit 12410 (and then in-fabric storage 12406), e.g., when the in-fabric storage has storage space for that value again.
In certain embodiments, accelerator-cache network 12414 is further coupled to cache home agent and/or next level cache 12412. In certain embodiments, accelerator-cache network 12414 is separate from any (for example, circuit switched or packet switched) network of an accelerator (e.g., accelerator tile), e.g., RAF is the interface between the accelerator (e.g., accelerator tile) and the cache. In one embodiment, a cache home agent is to connect to a memory (e.g., separate from the cache banks) to access data from that memory (e.g., memory 202 in
At time circled 5, PE 12416 pops a (e.g., most recent) value from the stack implements in the in-fabric storage 12406. Note that time 5 is just an example, and data may be popped from the stack at any time. In one embodiment, PEBY16 is coupled to the in-fabric storage via a circuit switched network as discussed herein.
In one embodiment, a component of accelerator (e.g., an accelerator tile) 12408 (e.g., one of a distributed set of processing elements) sends a memory request (e.g., via a packet switched network and/or circuit switched network of the accelerator) to access a memory location (e.g., virtual address), for example, a memory location that is separate from the accelerator (e.g., accelerator tile). The request may be sent to a RAF circuit (e.g., RAF circuit 12404). Although the number of RAF circuits and cache banks are depicted as six, any one or plurality of RAF circuits and/or any one or plurality of cache banks may be utilized in certain embodiments herein. The number of RAF circuits may be the same or different than the number of cache banks. A RAF circuit may be coupled to any of the cache banks, e.g., via accelerator-cache network 12414. The depicted arrangement of components is on example and other arrangements are possible, for example, a next level cache and/or cache home agent (CHA) may be omitted in certain embodiments.
Fast Clear Mode
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes a buffer box element thereof to operate in a Fast Clear mode according to the following (e.g., semantics and/or description).
In certain embodiments, fast clear mode is used in addition to FIFO mode, RAM mode (e.g., clearing RAM in the course of a calculation), and ROM mode. As one example, a (e.g., control) input queue is set (e.g., with a LIC) such that a first value on the input queue is to clear certain (e.g., all) of the data in the buffer box element. For example, in FIFO mode, the head and tail pointer are reset to clear the data in one embodiment.
In
In
In
In
In
In the depicted embodiment, buffer box element 12500 includes the components of buffer box element 10800 from
Fountain Operations
In one embodiment of a CSA, there are two aspects to a dataflow graph: (i) data compute nodes and (ii) nodes which do support work (e.g., control). In certain embodiments, it is desirable to keep the control logic and circuitry as small and not impactful to the overall performance of the graph as possible. In one embodiment, control in dataflow (e.g., spatial) graphs is a stream of single bit values, e.g., which are formed into sequences to control data flow through the graph. As one example, a CSA architecture includes a “fountain” operation used to generate streams of patterns used for control (e.g., for a loop operation) and/or used to generate data patterns for specific workload calculations with a single PE (e.g., not multiple PEs). In one embodiment, the number of PEs generating control values is less than one times, less than two times, or less than three times the number of PEs generating data values.
In one embodiment, a CSA architecture includes a configuration value that, when stored into the configuration storage (e.g., register), causes the CSA (e.g., a PE thereof) to operate as a fountain according to the following (e.g., semantics and/or description).
An operation that produces results continuously is generally referred to as a “fountain”. There are some fountains, such as mov64 lic, 10, which would generate the value 10 constantly filling the output. The fountain operations below are designed to provide repeating sequences of values, not just a single literal. These may be useful in constructing patterns to feed various control structures.
These operations may be used with in-fabric CSA storage (e.g., buffer box elements) where the address feeding the fountains holding the pattern is in the in-fabric CSA storage (e.g., scratchpad), and not in memory (e.g., cache or system memory).
This fountain functionality may be added in as micro-architecture onto all processing elements, e.g., allowing them to generate (e.g., small) “N” (e.g., 1 to 14 elements) of bit patterns as inputs into the processing element's operation circuitry (e.g., ALU). For example, where each processing element contains multiple input queues (FIFO queues) for each operand sent to the PE, storing literals (e.g., values which are preprogrammed into the PE during graph configuration) allows the PE to cycle through the entries creating a small pattern (e.g., the size of the number of total slots in the queues.
In one embodiment, during normal operation the shifter in the PE will implement a normal shifter calculation when needed, but when the PE is configured to implement a fountain, the input queues (FIFOs) are preloaded during configuration with literal values defining the bit pattern to generate. In certain embodiments, an N bit counter (e.g., the size of the counter is based on the number of pattern bits to generate) is used to control the shift amount. In one embodiment, a first value out will be shifted by 0, then shifted by 1, etc., and the output will be pulled from the last bit shifted out. This fountain functionality using a shifter circuit allows the programmer to generate a complex pattern using a single PE (e.g., where if created by an equation might take 3-5 or more PEs).
In one embodiment, when the PE is configured to implement a fountain, a (e.g., preloaded before execution of a dataflow graph) literal value(s) is (are) shifted in a desired pattern to produce an output pattern from the literal value(s). In one embodiment, the output of the desired pattern is sent on a bit by bit basis to a (e.g., single bit) output queue (e.g., 12732). In certain embodiments, the (e.g., narrow) input queue receives a value that is used to select start, stop, reset, or do not use a current value (e.g., bit or bits) of the literal value.
In certain embodiments, a (e.g., narrow) input queue (e.g., 12722) receives a (e.g., single bit) value used as predicate to control the mode of the fountain. Example modes are:
In the fountain implementation mode, the sequencer dataflow operator 12801 is to load the base value and stride value as literals in input queue(s) of PE 12800A (e.g., shown with the base value stored in input buffer 12824A and the stride value stored in input buffer 12826A). In certain embodiments, this forces the counter 12842 to start at 0 and count by 1 every time it starts. Literals (labeled X, Y, and Z) are loaded into the input queues of PE 12800B and specifying the length of X, Y, and Z. The input queues of “compare” PE 12800B are used to let the “sequence” PE 12800A know that it has completed in this embodiment. For example, when the sequencer PE 12800A starts, it counts from 0 with a stride of 1 to value X, and once value X is reached, a restart signal is sent to the SEQ Stride counter (e.g., in register 12844) to restart the counter, at the same time the sequence compare controller 12840 changes the pointer to point to the Y value. The counting process continues, starting at 0 and counting by 1 each time until the total reaches the value Y. The counter in 12800A is then reset again to zero, and counting continues until the total reaches the value Z. While counting up to X, a constant pattern value (e.g. 1) The bit values for X, Y, Z are determined as part of the command and are stored within the PE's control registers changing the output value generated for each section, in one embodiment. In another embodiment, this is expanded to N number of sub patterns consisting of some length of 0 and 1 values, by increasing the size and number of input queues for the compare values (e.g. in this example X, Y, and Z and 12824A, 12826A), as well as the state machine states in the controller. In one embodiment, a controller is to use dedicated restart line instead of the last_token narrow channel loop.
In one baseline (e.g., non-fountain) example, processing element 12800A of sequencer dataflow operator 12801 is to perform an add or subtract (e.g., to increment or decrement a counter for the number of iterations) and processing element 12820B is to perform a compare (e.g., to compare the current iteration to the number of iterations to either continue or stop the operation being performed by the sequencer dataflow operator 12801). The left part (e.g., left half) (e.g., processing element 12800A) of the sequencer dataflow operator 12801 has a (e.g., single) (e.g., 64 bit) register(s) 12844, for example, which is used to accumulate the stride data (e.g., stride data token) repeatedly into the base data (e.g., base data token). This may be referred to as the sequencer stride PE (seqstr). The right part (e.g., right half) (e.g., processing element 12800B) of the sequencer dataflow operator 12801 has a (e.g., single) (e.g., 64 bit) register(s) 12844, for example, which is used to do comparison operations. This may be referred to as the sequencer compare PE (seqcmp). The compare result may be passed back (e.g., on datapath 12841) from from sequencer compare PE (seqcmp) (e.g., processing element 12800B) to the sequencer stride PE (seqstr) (e.g., processing element 12800A), for example, so both PEs together decide when the sequence is done (e.g., the sequencer compare PE (seqcmp) (e.g., processing element 12800B) updates the sequencer stride PE (seqstr) (e.g., processing element 12800A) when the end (e.g., limit or bound) is reached).
In one embodiment, data passed into the sequencer dataflow operator 12801 includes a new strided length, e.g., where processing element 12800A is performing the add (or subtract) of the strided length to the total number of strides (e.g., iterations) thus far and processing element 12800B is performing the compare of that total number of strides (e.g., iterations) thus far to the total number of strides (e.g., iterations) to be performed.
Sequencer dataflow operator 12801 (e.g., processing element 12800B) may include a sequencer compare controller 12840. Sequencer compare controller 12840 may cause the processing element 12800B to perform the compare of that total number of strides (e.g., iterations) thus far (e.g., stored in register(s) 12844) to the total number of strides (e.g., iterations) to be performed (e.g., stored in register(s) 12844). Sequencer dataflow operator 12801 (e.g., processing element 12800A) may include a sequencer stride controller 12842. Sequencer stride controller 12842 may cause the processing element 12800A to performing the add (or subtract) of the strided length (e.g., increment for each iteration) (e.g., in one embodiment, the strided length is one unit (e.g., a numerical one)) to the total number of strides (e.g., iterations) thus far (e.g., “res” in
Another possible implementation of a sequencer dataflow operator is to use a single integer PE that contains two ALUs (e.g., one is used for accumulation and the other is used for comparison). The two ALUs may be pipelined (e.g., with additional pipeline hazard control circuitry) to maximize circuit frequency and/or the two ALUs may be put in series in a single clock cycle, e.g., to simplify the controller. In one embodiment, data passed into the sequencer dataflow operator 12801 includes a new strided length, e.g., where processing element 12800A is performing the add (or subtract) of the strided length to the total number of strides (e.g., iterations) thus far and processing element 12800B is performing the compare of that total number of strides (e.g., iterations) thus far to the total number of strides (e.g., iterations) to be performed.
Additionally or alternatively to forming a sequencer dataflow operator, each of processing elements 12800A and 12800B may perform as an integer PE.
In one embodiment, operation configuration register 2109A is loaded during configuration (e.g., mapping) and specifies the particular operation (or operations) this processing (e.g., compute) element is to perform. Scheduler 2114A (e.g., operations selector) may schedule an operation or operations of processing element 2100A, for example, when input data and control input arrives. Input and outputs (e.g., via queue(s)) may be sent via a network, e.g., any network discussed herein. Control input queue 2122A may be connected to local network (e.g., and local network may include a data path network as in
For example, suppose the operation of this processing (e.g., compute) element is (or includes) what is called call a pick in
For example, suppose the operation of this processing (e.g., compute) element is (or includes) what is called a switch in
Multiple networks (e.g., interconnects) may be connected to a processing element, e.g., (input) networks (e.g., networks 902, 904, 906 and (output) networks 908, 910, 912 in
Data input queue 12824A and data input queue 12826A may perform similarly, e.g., local network (e.g., set up as a data (as opposed to control) interconnect) may be switched (e.g., connected) to couple to data input queue 12824A. In this embodiment, a data path (e.g., network as in
A processing element 12800A may be stalled from execution until its operands (e.g., a control input value and its corresponding data input value or values) are received and/or until there is room in the output queue(s) of the processing element 12800A for the data that is to be produced by the execution of the operation on those operands.
In one embodiment, operation configuration register 12809B is loaded during configuration (e.g., mapping) and specifies the particular operation (or operations) this processing (e.g., compute) element is to perform. Scheduler 12814B (e.g., operations selector) may schedule an operation or operations of processing element 12800A, for example, when input data and control input arrives. Input and outputs (e.g., via queue(s)) may be sent via a network, e.g., any network discussed herein. Control input queue 12822B may be connected to local network (e.g., and local network may include a data path network as in
For example, suppose the operation of this processing (e.g., compute) element is (or includes) what is called call a pick in
For example, suppose the operation of this processing (e.g., compute) element is (or includes) what is called call a switch in
Multiple networks (e.g., interconnects) may be connected to a processing element, e.g., (input) networks (e.g., networks 902, 904, 906 and (output) networks 908, 910, 912 in
Data input queue 12824B and data input queue 12826B may perform similarly, e.g., local network (e.g., set up as a data (as opposed to control) interconnect) may be switched (e.g., connected) to couple to data input queue 12824B. In this embodiment, a data path (e.g., network as in
A processing element 12800B may be stalled from execution until its operands (e.g., a control input value and its corresponding data input value or values) are received and/or until there is room in the output queue(s) of the processing element 12800B for the data that is to be produced by the execution of the operation on those operands.
In certain embodiments, a processing element (PE) has one or a plurality of (e.g., two or three) operations that it may perform, e.g., the PE may be configured based on the input of the operation (e.g., operation value) into a PE.
Although certain embodiments herein illustrate an output queue with a single slot, other embodiments utilize an output queue with a plurality of slots, for example, such that stalling based on the output queue occurs when all of the plurality of the slots are used (e.g., are full).
Other Operations
Other operations (e.g., logical operation and/or arithmetic operation) of a CSA architecture may be part of a set of operations. In certain embodiments, one or more of the following is loaded (e.g., by a corresponding configuration value) in a CSA component (e.g., PE) to cause the following semantics and/or description to be performed:
Floating Point Comparison
Floating point comparisons are like their integer counterparts in that they generate 1 bit results in this embodiment.
The specific opcodes are:
Floating Point Conversion
Generally there are conversions between f32, f64, u32 and u64. If f16 were added, the expectation is that would need to be converted through f32 to get to integer types.
Prefetching
While certain embodiments of the CSA are designed to allow a large number of memory operations in flight with sufficient buffering, there are some circumstances in which prefetching may be beneficial.
Local Memory Reference Ordering
One major difference from Von Neumann architectures is memory ops are not ordered relative to other memory ops unless they are on the sequential unit. For operations that access state (load, store, atomics, send/recv, etc.) on dataflow units, there are 3 additional operands, an input ordering channel, a local output ordering channel, and an external output ordering channel. An operation cannot proceed until all inputs, including the ordering channel, are available. After a memory operation is done, the local output ordering channel is defined as soon as the reference is known to be ordered relative to local references—e.g., in the same graph. The external ordering channel is not defined until the effects are known to be ordered relative to any external observer. Memory ordering channels are defined to be 0 bit—e.g., no data, only presence/absence. The most common consuming operations are those that take 0 bit inputs—e.g. onend, all, any, etc.
To give a flavor of memory reference ordering, consider a store, two loads that depend on it, and a store that must follow, because a compiler cannot determine address relationships in a way that allows disambiguation. On the sequential unit, this would just be:
In order to place these on dataflow units while preserving semantic ordering, we use channels to fan out from the store to the loads, and fan in from the loads to the store.
The order operands are designed to allow compilers to express as much—or preferably, as little—ordering as is required by the program specification. The more relaxed the semantic ordering requirements, the less constrained the graph will be.
Other operations may include, but are not limited to add, subtract, multiply, divide, move (to move data from a PE to memory (e.g., cache), logical NOT, logical AND, logical OR, logical XOR, comparison (e.g., less than, greater than, less than or equal, and greater than or equal), copy, etc.
The ability to compile programs written in high-level languages onto a CSA may be essential for industry adoption. This section gives a high-level overview of compilation strategies for embodiments of a CSA. First is a proposal for a CSA software framework that illustrates the desired properties of an ideal production-quality toolchain. Next, a prototype compiler framework is discussed. A “control-to-dataflow conversion” is then discussed, e.g., to converts ordinary sequential control-flow code into CSA dataflow assembly code.
4.1 Example Production Framework
4.2 Prototype Compiler
4.3 Control to Dataflow Conversion
A key portion of the compiler may be implemented in the control-to-dataflow conversion pass, or dataflow conversion pass for short. This pass takes in a function represented in control flow form, e.g., a control-flow graph (CFG) with sequential machine operations operating on virtual registers, and converts it into a dataflow function that is conceptually a graph of dataflow operations (e.g., not instructions) connected by latency-insensitive channels (LICs). This section gives a high-level description of this pass, describing how it conceptually deals with memory operations, branches, and loops in certain embodiments.
Straight-Line Code
First, consider the simple case of converting straight-line sequential code to dataflow. The dataflow conversion pass may convert a basic block of sequential code, such as the code shown in
Branches
To convert programs with multiple basic blocks and conditionals to dataflow, the compiler generates special dataflow operators to replace the branches. More specifically, the compiler uses switch operators to steer outgoing data at the end of a basic block in the original CFG, and pick operators to select values from the appropriate incoming channel at the beginning of a basic block. As a concrete example, consider the code and corresponding dataflow graph in
Control Equivalence: Consider a single-entry-single-exit control flow graph G with two basic blocks A and B. A and B are control-equivalent if all complete control flow paths through G visit A and B the same number of times.
LIC Replacement: In a control flow graph G, suppose an operation in basic block A defines a virtual register x, and an operation in basic block B that uses x. Then a correct control-to-dataflow transformation can replace x with a latency-insensitive channel only if A and B are control equivalent. The control-equivalence relation partitions the basic blocks of a CFG into strong control-dependence regions.
Loops
Another important class of CFGs in dataflow conversion are CFGs for single-entry-single-exit loops, a common form of loop generated in (LLVM) IR. These loops may be almost acyclic, except for a single back edge from the end of the loop back to a loop header block. The dataflow conversion pass may use same high-level strategy to convert loops as for branches, e.g., it inserts switches at the end of the loop to direct values out of the loop (either out the loop exit or around the back-edge to the beginning of the loop), and inserts picks at the beginning of the loop to choose between initial values entering the loop and values coming through the back edge.
In one embodiment, the core writes a command into a memory queue and a CSA (e.g., the plurality of processing elements) monitors the memory queue and begins executing when the command is read. In one embodiment, the core executes a first part of a program and a CSA (e.g., the plurality of processing elements) executes a second part of the program. In one embodiment, the core does other work while the CSA is executing its operations.
In certain embodiments, the CSA architecture and microarchitecture provides profound energy, performance, and usability advantages over roadmap processor architectures and FPGAs. In this section, these architectures are compared to embodiments of the CSA and highlights the superiority of CSA in accelerating parallel dataflow graphs relative to each.
5.1 Processors
5.2 Comparison of CSA Embodiments and FPGAs
The choice of dataflow operators as the fundamental architecture of embodiments of a CSA differentiates those CSAs from a FPGA, and particularly the CSA is as superior accelerator for HPC dataflow graphs arising from traditional programming languages. Dataflow operators are fundamentally asynchronous. This enables embodiments of a CSA not only to have great freedom of implementation in the microarchitecture, but it also enables them to simply and succinctly accommodate abstract architectural concepts. For example, embodiments of a CSA naturally accommodate many memory microarchitectures, which are essentially asynchronous, with a simple load-store interface. One need only examine an FPGA DRAM controller to appreciate the difference in complexity. Embodiments of a CSA also leverage asynchrony to provide faster and more-fully-featured runtime services like configuration and extraction, which are believed to be four to six orders of magnitude faster than an FPGA. By narrowing the architectural interface, embodiments of a CSA provide control over most timing paths at the microarchitectural level. This allows embodiments of a CSA to operate at a much higher frequency than the more general control mechanism offered in a FPGA. Similarly, clock and reset, which may be architecturally fundamental to FPGAs, are microarchitectural in the CSA, e.g., obviating the need to support them as programmable entities. Dataflow operators may be, for the most part, coarse-grained. By only dealing in coarse operators, embodiments of a CSA improve both the density of the fabric and its energy consumption: CSA executes operations directly rather than emulating them with look-up tables. A second consequence of coarseness is a simplification of the place and route problem. CSA dataflow graphs are many orders of magnitude smaller than FPGA net-lists and place and route time are commensurately reduced in embodiments of a CSA. The significant differences between embodiments of a CSA and a FPGA make the CSA superior as an accelerator, e.g., for dataflow graphs arising from traditional programming languages.
The CSA is a novel computer architecture with the potential to provide enormous performance and energy advantages relative to roadmap processors. Consider the case of computing a single strided address for walking across an array. This case may be important in HPC applications, e.g., which spend significant integer effort in computing address offsets. In address computation, and especially strided address computation, one argument is constant and the other varies only slightly per computation. Thus, only a handful of bits per cycle toggle in the majority of cases. Indeed, it may be shown, using a derivation similar to the bound on floating point carry bits described in Section 2.6, that less than two bits of input toggle per computation in average for a stride calculation, reducing energy by 50% over a random toggle distribution. Were a time-multiplexed approach used, much of this energy savings may be lost. In one embodiment, the CSA achieves approximately 3× energy efficiency over a core while delivering an 8× performance gain. The parallelism gains achieved by embodiments of a CSA may result in reduced program run times, yielding a proportionate, substantial reduction in leakage energy. At the PE level, embodiments of a CSA are extremely energy efficient. A second important question for the CSA is whether the CSA consumes a reasonable amount of energy at the tile level. Since embodiments of a CSA are capable of exercising every floating point PE in the fabric at every cycle, it serves as a reasonable upper bound for energy and power consumption, e.g., such that most of the energy goes into floating point multiply and add.
This section discusses further details for configuration and exception handling.
7.1 Microarchitecture for Configuring a CSA
This section discloses examples of how to configure a CSA (e.g., fabric), how to achieve this configuration quickly, and how to minimize the resource overhead of configuration. Configuring the fabric quickly may be of preeminent importance in accelerating small portions of a larger algorithm, and consequently in broadening the applicability of a CSA. The section further discloses features that allow embodiments of a CSA to be programmed with configurations of different length.
Embodiments of a CSA (e.g., fabric) may differ from traditional cores in that they make use of a configuration step in which (e.g., large) parts of the fabric are loaded with program configuration in advance of program execution. An advantage of static configuration may be that very little energy is spent at runtime on the configuration, e.g., as opposed to sequential cores which spend energy fetching configuration information (an instruction) nearly every cycle. The previous disadvantage of configuration is that it was a coarse-grained step with a potentially large latency, which places an under-bound on the size of program that can be accelerated in the fabric due to the cost of context switching. This disclosure describes a scalable microarchitecture for rapidly configuring a spatial array in a distributed fashion, e.g., that avoids the previous disadvantages.
As discussed above, a CSA may include light-weight processing elements connected by an inter-PE network. Programs, viewed as control-dataflow graphs, are then mapped onto the architecture by configuring the configurable fabric elements (CFEs), for example PEs and the interconnect (fabric) networks. Generally, PEs may be configured as dataflow operators and once all input operands arrive at the PE, some operation occurs, and the results are forwarded to another PE or PEs for consumption or output. PEs may communicate over dedicated virtual circuits which are formed by statically configuring the circuit switched communications network. These virtual circuits may be flow controlled and fully back-pressured, e.g., such that PEs will stall if either the source has no data or destination is full. At runtime, data may flow through the PEs implementing the mapped algorithm. For example, data may be streamed in from memory, through the fabric, and then back out to memory. Such a spatial architecture may achieve remarkable performance efficiency relative to traditional multicore processors: compute, in the form of PEs, may be simpler and more numerous than larger cores and communications may be direct, as opposed to an extension of the memory system.
Embodiments of a CSA may not utilize (e.g., software controlled) packet switching, e.g., packet switching that requires significant software assistance to realize, which slows configuration. Embodiments of a CSA include out-of-band signaling in the network (e.g., of only 2-3 bits, depending on the feature set supported) and a fixed configuration topology to avoid the need for significant software support.
One key difference between embodiments of a CSA and the approach used in FPGAs is that a CSA approach may use a wide data word, is distributed, and includes mechanisms to fetch program data directly from memory. Embodiments of a CSA may not utilize JTAG-style single bit communications in the interest of area efficiency, e.g., as that may require milliseconds to completely configure a large FPGA fabric.
Embodiments of a CSA include a distributed configuration protocol and microarchitecture to support this protocol. Initially, configuration state may reside in memory. Multiple (e.g., distributed) local configuration controllers (boxes) (LCCs) may stream portions of the overall program into their local region of the spatial fabric, e.g., using a combination of a small set of control signals and the fabric-provided network. State elements may be used at each CFE to form configuration chains, e.g., allowing individual CFEs to self-program without global addressing.
Embodiments of a CSA include specific hardware support for the formation of configuration chains, e.g., not software establishing these chains dynamically at the cost of increasing configuration time. Embodiments of a CSA are not purely packet switched and do include extra out-of-band control wires (e.g., control is not sent through the data path requiring extra cycles to strobe this information and reserialize this information). Embodiments of a CSA decreases configuration latency by fixing the configuration ordering and by providing explicit out-of-band control (e.g., by at least a factor of two), while not significantly increasing network complexity.
Embodiments of a CSA do not use a serial mechanism for configuration in which data is streamed bit by bit into the fabric using a JTAG-like protocol. Embodiments of a CSA utilize a coarse-grained fabric approach. In certain embodiments, adding a few control wires or state elements to a 64 or 32-bit-oriented CSA fabric has a lower cost relative to adding those same control mechanisms to a 4 or 6 bit fabric.
Embodiments of a CSA include hardware that provides for efficient, distributed, low-latency configuration of a heterogeneous spatial fabric. This may be achieved according to four techniques. First, a hardware entity, the local configuration controller (LCC) is utilized, for example, as in
Local Configuration Controller
LCC operation may begin when it receives a pointer to a code segment. Depending on the LCB microarchitecture, this pointer (e.g., stored in pointer register 13906) may come either over a network (e.g., from within the CSA (fabric) itself) or through a memory system access to the LCC. When it receives such a pointer, the LCC optionally drains relevant state from its portion of the fabric for context storage, and then proceeds to immediately reconfigure the portion of the fabric for which it is responsible. The program loaded by the LCC may be a combination of configuration data for the fabric and control commands for the LCC, e.g., which are lightly encoded. As the LCC streams in the program portion, it may interprets the program as a command stream and perform the appropriate encoded action to configure (e.g., load) the fabric.
Two different microarchitectures for the LCC are shown in
Extra Out-of-Band Control Channels (e.g., Wires)
In certain embodiments, configuration relies on 2-8 extra, out-of-band control channels to improve configuration speed, as defined below. For example, configuration controller 13902 may include the following control channels, e.g., CFG_START control channel 13908, CFG_VALID control channel 13910, and CFG_DONE control channel 13912, with examples of each discussed in Table 3 below.
Generally, the handling of configuration information may be left to the implementer of a particular CFE. For example, a selectable function CFE may have a provision for setting registers using an existing data path, while a fixed function CFE might simply set a configuration register.
Due to long wire delays when programming a large set of CFEs, the CFG_VALID signal may be treated as a clock/latch enable for CFE components. Since this signal is used as a clock, in one embodiment the duty cycle of the line is at most 50%. As a result, configuration throughput is approximately halved. Optionally, a second CFG_VALID signal may be added to enable continuous programming.
In one embodiment, only CFG_START is strictly communicated on an independent coupling (e.g., wire), for example, CFG_VALID and CFG_DONE may be overlaid on top of other network couplings.
Reuse of Network Resources
To reduce the overhead of configuration, certain embodiments of a CSA make use of existing network infrastructure to communicate configuration data. A LCC may make use of both a chip-level memory hierarchy and a fabric-level communications networks to move data from storage into the fabric. As a result, in certain embodiments of a CSA, the configuration infrastructure adds no more than 2% to the overall fabric area and power.
Reuse of network resources in certain embodiments of a CSA may cause a network to have some hardware support for a configuration mechanism. Circuit switched networks of embodiments of a CSA cause an LCC to set their multiplexors in a specific way for configuration when the ‘CFG_START’ signal is asserted. Packet switched networks do not require extension, although LCC endpoints (e.g., configuration terminators) use a specific address in the packet switched network. Network reuse is optional, and some embodiments may find dedicated configuration buses to be more convenient.
Per CFE State
Each CFE may maintain a bit denoting whether or not it has been configured (see, e.g.,
Internal to the CFE, this bit may be used to drive flow control ready signals. For example, when the configuration bit is de-asserted, network control signals may automatically be clamped to a values that prevent data from flowing, while, within PEs, no operations or other actions will be scheduled.
Dealing with High-Delay Configuration Paths
One embodiment of an LCC may drive a signal over a long distance, e.g., through many multiplexors and with many loads. Thus, it may be difficult for a signal to arrive at a distant CFE within a short clock cycle. In certain embodiments, configuration signals are at some division (e.g., fraction of) of the main (e.g., CSA) clock frequency to ensure digital timing discipline at configuration. Clock division may be utilized in an out-of-band signaling protocol, and does not require any modification of the main clock tree.
Ensuring Consistent Fabric Behavior During Configuration
Since certain configuration schemes are distributed and have non-deterministic timing due to program and memory effects, different portions of the fabric may be configured at different times. As a result, certain embodiments of a CSA provide mechanisms to prevent inconsistent operation among configured and unconfigured CFEs. Generally, consistency is viewed as a property required of and maintained by CFEs themselves, e.g., using the internal CFE state. For example, when a CFE is in an unconfigured state, it may claim that its input buffers are full, and that its output is invalid. When configured, these values will be set to the true state of the buffers. As enough of the fabric comes out of configuration, these techniques may permit it to begin operation. This has the effect of further reducing context switching latency, e.g., if long-latency memory requests are issued early.
Variable-Width Configuration
Different CFEs may have different configuration word widths. For smaller CFE configuration words, implementers may balance delay by equitably assigning CFE configuration loads across the network wires. To balance loading on network wires, one option is to assign configuration bits to different portions of network wires to limit the net delay on any one wire. Wide data words may be handled by using serialization/deserialization techniques. These decisions may be taken on a per-fabric basis to optimize the behavior of a specific CSA (e.g., fabric). Network controller (e.g., one or more of network controller 13710 and network controller 13712 may communicate with each domain (e.g., subset) of the CSA (e.g., fabric), for example, to send configuration information to one or more LCCs. Network controller may be part of a communications network (e.g., separate from circuit switched network). Network controller may include a network dataflow endpoint circuit.
7.2 Microarchitecture for Low Latency Configuration of a CSA and for Timely Fetching of Configuration Data for a CSA
Embodiments of a CSA may be an energy-efficient and high-performance means of accelerating user applications. When considering whether a program (e.g., a dataflow graph thereof) may be successfully accelerated by an accelerator, both the time to configure the accelerator and the time to run the program may be considered. If the run time is short, then the configuration time may play a large role in determining successful acceleration. Therefore, to maximize the domain of accelerable programs, in some embodiments the configuration time is made as short as possible. One or more configuration caches may be includes in a CSA, e.g., such that the high bandwidth, low-latency store enables rapid reconfiguration. Next is a description of several embodiments of a configuration cache.
In one embodiment, during configuration, the configuration hardware (e.g., LCC) optionally accesses the configuration cache to obtain new configuration information. The configuration cache may operate either as a traditional address based cache, or in an OS managed mode, in which configurations are stored in the local address space and addressed by reference to that address space. If configuration state is located in the cache, then no requests to the backing store are to be made in certain embodiments. In certain embodiments, this configuration cache is separate from any (e.g., lower level) shared cache in the memory hierarchy.
Caching Modes
In certain embodiments, a configuration cache may have the configuration data pre-loaded into it, e.g., either by external direction or internal direction. This may allow reduction in the latency to load programs. Certain embodiments herein provide for an interface to a configuration cache which permits the loading of new configuration state into the cache, e.g., even if a configuration is running in the fabric already. The initiation of this load may occur from either an internal or external source. Embodiments of a pre-loading mechanism further reduce latency by removing the latency of cache loading from the configuration path.
Pre Fetching Modes
Certain embodiments of a CSA (e.g., a spatial fabric) include large amounts of operation and configuration state, e.g., which is largely static during the operation of the CSA. Thus, the configuration state may be vulnerable to soft errors. Rapid and error-free recovery of these soft errors may be critical to the long-term reliability and performance of spatial systems.
Certain embodiments herein provide for a rapid configuration recovery loop, e.g., in which configuration errors are detected and portions of the fabric immediately reconfigured. Certain embodiments herein include a configuration controller, e.g., with reliability, availability, and serviceability (RAS) reprogramming features. Certain embodiments of CSA include circuitry for high-speed configuration, error reporting, and parity checking within the spatial fabric. Using a combination of these three features, and optionally, a configuration cache, a configuration/exception handling circuit may recover from soft errors in configuration. When detected, soft errors may be conveyed to a configuration cache which initiates an immediate reconfiguration of (e.g., that portion of) the fabric. Certain embodiments provide for a dedicated reconfiguration circuit, e.g., which is faster than any solution that would be indirectly implemented in the fabric. In certain embodiments, co-located exception and configuration circuit cooperates to reload the fabric on configuration error detection.
7.4 Hardware for Fabric-Initiated Reconfiguration of a CSA
Some portions of an application targeting a CSA (e.g., spatial array) may be run infrequently or may be mutually exclusive with other parts of the program. To save area, to improve performance, and/or reduce power, it may be useful to time multiplex portions of the spatial fabric among several different parts of the program dataflow graph. Certain embodiments herein include an interface by which a CSA (e.g., via the spatial program) may request that part of the fabric be reprogrammed. This may enable the CSA to dynamically change itself according to dynamic control flow. Certain embodiments herein allow for fabric initiated reconfiguration (e.g., reprogramming). Certain embodiments herein provide for a set of interfaces for triggering configuration from within the fabric. In some embodiments, a PE issues a reconfiguration request based on some decision in the program dataflow graph. This request may travel a network to our new configuration interface, where it triggers reconfiguration. Once reconfiguration is completed, a message may optionally be returned notifying of the completion. Certain embodiments of a CSA thus provide for a program (e.g., dataflow graph) directed reconfiguration capability.
Configuration Modes
Configure-by-address—In this mode, the fabric makes a direct request to load configuration data from a particular address.
Configure-by-reference—In this mode the fabric makes a request to load a new configuration, e.g., by a pre-determined reference ID. This may simplify the determination of the code to load, since the location of the code has been abstracted.
Configuring Multiple Domains
A CSA may include a higher level configuration controller to support a multicast mechanism to cast (e.g., via network indicated by the dotted box) configuration requests to multiple (e.g., distributed or local) configuration controllers. This may enable a single configuration request to be replicated across larger portions of the fabric, e.g., triggering a broad reconfiguration.
6.5 Exception Aggregators
Certain embodiments of a CSA may also experience an exception (e.g., exceptional condition), for example, floating point underflow. When these conditions occur, a special handlers may be invoked to either correct the program or to terminate it. Certain embodiments herein provide for a system-level architecture for handling exceptions in spatial fabrics. Since certain spatial fabrics emphasize area efficiency, embodiments herein minimize total area while providing a general exception mechanism. Certain embodiments herein provides a low area means of signaling exceptional conditions occurring in within a CSA (e.g., a spatial array). Certain embodiments herein provide an interface and signaling protocol for conveying such exceptions, as well as a PE-level exception semantics. Certain embodiments herein are dedicated exception handling capabilities, e.g., and do not require explicit handling by the programmer.
One embodiments of a CSA exception architecture consists of four portions, e.g., shown in
1. PE Exception Generator
2. Local Exception Network
3. Mezzanine Exception Aggregator
4. Tile-Level Exception Aggregator
PE Exception Generator
Processing element 14500 may include processing element 900 from
The initiation of the exception may either occur explicitly, by the execution of a programmer supplied command, or implicitly when a hardened error condition (e.g., a floating point underflow) is detected. Upon an exception, the PE 14500 may enter a waiting state, in which it waits to be serviced by the eventual exception handler, e.g., external to the PE 14500. The contents of the exception packet depend on the implementation of the particular PE, as described below.
Local Exception Network
A (e.g., local) exception network steers exception packets from PE 14500 to the mezzanine exception network. Exception network (e.g., 14513) may be a serial, packet switched network consisting of a (e.g., single) control wire and one or more data wires, e.g., organized in a ring or tree topology, e.g., for a subset of PEs. Each PE may have a (e.g., ring) stop in the (e.g., local) exception network, e.g., where it can arbitrate to inject messages into the exception network.
PE endpoints needing to inject an exception packet may observe their local exception network egress point. If the control signal indicates busy, the PE is to wait to commence inject its packet. If the network is not busy, that is, the downstream stop has no packet to forward, then the PE will proceed commence injection.
Network packets may be of variable or fixed length. Each packet may begin with a fixed length header field identifying the source PE of the packet. This may be followed by a variable number of PE-specific field containing information, for example, including error codes, data values, or other useful status information.
Mezzanine Exception Aggregator
The mezzanine exception aggregator 14404 is responsible for assembling local exception network into larger packets and sending them to the tile-level exception aggregator 14402. The mezzanine exception aggregator 14404 may pre-pend the local exception packet with its own unique ID, e.g., ensuring that exception messages are unambiguous. The mezzanine exception aggregator 14404 may interface to a special exception-only virtual channel in the mezzanine network, e.g., ensuring the deadlock-freedom of exceptions.
The mezzanine exception aggregator 14404 may also be able to directly service certain classes of exception. For example, a configuration request from the fabric may be served out of the mezzanine network using caches local to the mezzanine network stop.
Tile-Level Exception Aggregator
The final stage of the exception system is the tile-level exception aggregator 14402. The tile-level exception aggregator 14402 is responsible for collecting exceptions from the various mezzanine-level exception aggregators (e.g., 14404) and forwarding them to the appropriate servicing hardware (e.g., core). As such, the tile-level exception aggregator 14402 may include some internal tables and controller to associate particular messages with handler routines. These tables may be indexed either directly or with a small state machine in order to steer particular exceptions.
Like the mezzanine exception aggregator, the tile-level exception aggregator may service some exception requests. For example, it may initiate the reprogramming of a large portion of the PE fabric in response to a specific exception.
6.6 Extraction Controllers
Certain embodiments of a CSA include an extraction controller(s) to extract data from the fabric. The below discusses embodiments of how to achieve this extraction quickly and how to minimize the resource overhead of data extraction. Data extraction may be utilized for such critical tasks as exception handling and context switching. Certain embodiments herein extract data from a heterogeneous spatial fabric by introducing features that allow extractable fabric elements (EFEs) (for example, PEs, network controllers, and/or switches) with variable and dynamically variable amounts of state to be extracted.
Embodiments of a CSA include a distributed data extraction protocol and microarchitecture to support this protocol. Certain embodiments of a CSA include multiple local extraction controllers (LECs) which stream program data out of their local region of the spatial fabric using a combination of a (e.g., small) set of control signals and the fabric-provided network. State elements may be used at each extractable fabric element (EFE) to form extraction chains, e.g., allowing individual EFEs to self-extract without global addressing.
Embodiments of a CSA do not use a local network to extract program data. Embodiments of a CSA include specific hardware support (e.g., an extraction controller) for the formation of extraction chains, for example, and do not rely on software to establish these chains dynamically, e.g., at the cost of increasing extraction time. Embodiments of a CSA are not purely packet switched and do include extra out-of-band control wires (e.g., control is not sent through the data path requiring extra cycles to strobe and reserialize this information). Embodiments of a CSA decrease extraction latency by fixing the extraction ordering and by providing explicit out-of-band control (e.g., by at least a factor of two), while not significantly increasing network complexity.
Embodiments of a CSA do not use a serial mechanism for data extraction, in which data is streamed bit by bit from the fabric using a JTAG-like protocol. Embodiments of a CSA utilize a coarse-grained fabric approach. In certain embodiments, adding a few control wires or state elements to a 64 or 32-bit-oriented CSA fabric has a lower cost relative to adding those same control mechanisms to a 4 or 6 bit fabric.
Embodiments of a CSA include hardware that provides for efficient, distributed, low-latency extraction from a heterogeneous spatial fabric. This may be achieved according to four techniques. First, a hardware entity, the local extraction controller (LEC) is utilized, for example, as in
The following sections describe the operation of the various components of embodiments of an extraction network.
Local Extraction Controller
LEC operation may begin when it receives a pointer to a buffer (e.g., in virtual memory) where fabric state will be written, and, optionally, a command controlling how much of the fabric will be extracted. Depending on the LEC microarchitecture, this pointer (e.g., stored in pointer register 14804) may come either over a network or through a memory system access to the LEC. When it receives such a pointer (e.g., command), the LEC proceeds to extract state from the portion of the fabric for which it is responsible. The LEC may stream this extracted data out of the fabric into the buffer provided by the external caller.
Two different microarchitectures for the LEC are shown in
Extra Out-of-Band Control Channels (e.g., Wires)
In certain embodiments, extraction relies on 2-8 extra, out-of-band signals to improve configuration speed, as defined below. Signals driven by the LEC may be labelled LEC. Signals driven by the EFE (e.g., PE) may be labelled EFE. Configuration controller 14802 may include the following control channels, e.g., LEC EXTRACT control channel 14906, LEC START control channel 14808, LEC_STROBE control channel 14810, and EFE COMPLETE control channel 14812, with examples of each discussed in Table 4 below.
Generally, the handling of extraction may be left to the implementer of a particular EFE. For example, selectable function EFE may have a provision for dumping registers using an existing data path, while a fixed function EFE might simply have a multiplexor.
Due to long wire delays when programming a large set of EFEs, the LEC_STROBE signal may be treated as a clock/latch enable for EFE components. Since this signal is used as a clock, in one embodiment the duty cycle of the line is at most 50%. As a result, extraction throughput is approximately halved. Optionally, a second LEC_STROBE signal may be added to enable continuous extraction.
In one embodiment, only LEC_START is strictly communicated on an independent coupling (e.g., wire), for example, other control channels may be overlayed on existing network (e.g., wires).
Reuse of Network Resources
To reduce the overhead of data extraction, certain embodiments of a CSA make use of existing network infrastructure to communicate extraction data. A LEC may make use of both a chip-level memory hierarchy and a fabric-level communications networks to move data from the fabric into storage. As a result, in certain embodiments of a CSA, the extraction infrastructure adds no more than 2% to the overall fabric area and power.
Reuse of network resources in certain embodiments of a CSA may cause a network to have some hardware support for an extraction protocol. Circuit switched networks require of certain embodiments of a CSA cause a LEC to set their multiplexors in a specific way for configuration when the ‘LEC_START’ signal is asserted. Packet switched networks may not require extension, although LEC endpoints (e.g., extraction terminators) use a specific address in the packet switched network. Network reuse is optional, and some embodiments may find dedicated configuration buses to be more convenient.
Per EFE State
Each EFE may maintain a bit denoting whether or not it has exported its state. This bit may de-asserted when the extraction start signal is driven, and then asserted once the particular EFE finished extraction. In one extraction protocol, EFEs are arranged to form chains with the EFE extraction state bit determining the topology of the chain. A EFE may read the extraction state bit of the immediately adjacent EFE. If this adjacent EFE has its extraction bit set and the current EFE does not, the EFE may determine that it owns the extraction bus. When an EFE dumps its last data value, it may drives the ‘EFE_DONE’ signal and sets its extraction bit, e.g., enabling upstream EFEs to configure for extraction. The network adjacent to the EFE may observe this signal and also adjust its state to handle the transition. As a base case to the extraction process, an extraction terminator (e.g., extraction terminator 14604 for LEC 14602 or extraction terminator 14608 for LEC 14606 in
Internal to the EFE, this bit may be used to drive flow control ready signals. For example, when the extraction bit is de-asserted, network control signals may automatically be clamped to a values that prevent data from flowing, while, within PEs, no operations or actions will be scheduled.
Dealing with High-delay Paths
One embodiment of a LEC may drive a signal over a long distance, e.g., through many multiplexors and with many loads. Thus, it may be difficult for a signal to arrive at a distant EFE within a short clock cycle. In certain embodiments, extraction signals are at some division (e.g., fraction of) of the main (e.g., CSA) clock frequency to ensure digital timing discipline at extraction. Clock division may be utilized in an out-of-band signaling protocol, and does not require any modification of the main clock tree.
Ensuring Consistent Fabric Behavior During Extraction
Since certain extraction scheme are distributed and have non-deterministic timing due to program and memory effects, different members of the fabric may be under extraction at different times. While LEC_EXTRACT is driven, all network flow control signals may be driven logically low, e.g., thus freezing the operation of a particular segment of the fabric.
An extraction process may be non-destructive. Therefore a set of PEs may be considered operational once extraction has completed. An extension to an extraction protocol may allow PEs to optionally be disabled post extraction. Alternatively, beginning configuration during the extraction process will have similar effect in embodiments.
Single PE Extraction
In some cases, it may be expedient to extract a single PE. In this case, an optional address signal may be driven as part of the commencement of the extraction process. This may enable the PE targeted for extraction to be directly enabled. Once this PE has been extracted, the extraction process may cease with the lowering of the LEC_EXTRACT signal. In this way, a single PE may be selectively extracted, e.g., by the local extraction controller.
Handling Extraction Backpressure
In an embodiment where the LEC writes extracted data to memory (for example, for post-processing, e.g., in software), it may be subject to limited memory bandwidth. In the case that the LEC exhausts its buffering capacity, or expects that it will exhaust its buffering capacity, it may stops strobing the LEC_STROBE signal until the buffering issue has resolved.
Note that in certain figures (e.g.,
6.7 Flow Diagrams
6.8 Memory
In one embodiment, programs, viewed as control data flow graphs, are mapped onto the spatial architecture by configuring PEs and a communications network. Generally, PEs are configured as dataflow operators, similar to functional units in a processor: once the input operands arrive at the PE, some operation occurs, and results are forwarded to downstream PEs in a pipelined fashion. Dataflow operators (or other types of operators) may choose to consume incoming data on a per-operator basis. Simple operators, like those handling the unconditional evaluation of arithmetic expressions often consume all incoming data. It is sometimes useful, however, for operators to maintain state, for example, in accumulation.
The PEs communicate using dedicated virtual circuits, which are formed by statically configuring a circuit switched communications network. These virtual circuits are flow controlled and fully back pressured, such that PEs will stall if either the source has no data or the destination is full. At runtime, data flows through the PEs implementing a mapped algorithm according to a dataflow graph, also referred to as a subprogram herein. For example, data may be streamed in from memory, through the acceleration hardware 15102, and then back out to memory. Such an architecture can achieve remarkable performance efficiency relative to traditional multicore processors: compute, in the form of PEs, is simpler and more numerous than larger cores and communication is direct, as opposed to an extension of the memory subsystem 15110. Memory system parallelism, however, helps to support parallel PE computation. If memory accesses are serialized, high parallelism is likely unachievable. To facilitate parallelism of memory accesses, the disclosed memory ordering circuit 15105 includes memory ordering architecture and microarchitecture, as will be explained in detail. In one embodiment, the memory ordering circuit 15105 is a request address file circuit (or “RAF”) or other memory request circuitry.
Each memory ordering circuit 15105 may accept read and write requests to the memory subsystem 15110. The requests from the acceleration hardware 15102 arrive at the memory ordering circuit 15105 in a separate channel for each node of the dataflow graph that initiates read or write accesses, also referred to as load or store accesses herein. Buffering is provided so that the processing of loads will return the requested data to the acceleration hardware 15102 in the order it was requested. In other words, iteration six data is returned before iteration seven data, and so forth. Furthermore, note that the request channel from a memory ordering circuit 15105 to a particular cache bank may be implemented as an ordered channel and any first request that leaves before a second request will arrive at the cache bank before the second request.
By considering this sequence of operations, it may be evident that spatial arrays more naturally map to channels. Furthermore, the acceleration hardware 15102 is latency-insensitive in terms of the request and response channels, and inherent parallel processing that may occur. The acceleration hardware may also decouple execution of a program from implementation of the memory subsystem 15110 (
The memory ordering circuit 15105 may further include, but not be limited to, a memory interface 15410, an operations queue 15412, input queue(s) 15416, a completion queue 15420, an operation configuration data structure 15424, and an operations manager circuit 15430 that may further include a scheduler circuit 15432 and an execution circuit 15434. In one embodiment, the memory interface 15410 may be circuit switched, and in another embodiment, the memory interface 15410 may be packet-switched, or both may exist simultaneously. The operations queue 15412 may buffer memory operations (with corresponding arguments) that are being processed for request, and may, therefore, correspond to addresses and data coming into the input queues 15416.
More specifically, the input queues 15416 may be an aggregation of at least the following: a load address queue, a store address queue, a store data queue, and a dependency queue. When implementing the input queue 15416 as aggregated, the memory ordering circuit 15105 may provide for sharing of logical queues, with additional control logic to logically separate the queues, which are individual channels with the memory ordering circuit. This may maximize input queue usage, but may also require additional complexity and space for the logic circuitry to manage the logical separation of the aggregated queue. Alternatively, as will be discussed with reference to
When shared, the input queues 15416 and the completion queue 15420 may be implemented as ring buffers of a fixed size. A ring buffer is an efficient implementation of a circular queue that has a first-in-first-out (FIFO) data characteristic. These queues may, therefore, enforce a semantical order of a program for which the memory operations are being requested. In one embodiment, a ring buffer (such as for the store address queue) may have entries corresponding to entries flowing through an associated queue (such as the store data queue or the dependency queue) at the same rate. In this way, a store address may remain associated with corresponding store data.
More specifically, the load address queue may buffer an incoming address of the memory 18 from which to retrieve data. The store address queue may buffer an incoming address of the memory 18 to which to write data, which is buffered in the store data queue. The dependency queue may buffer dependency tokens in association with the addresses of the load address queue and the store address queue. Each queue, representing a separate channel, may be implemented with a fixed or dynamic number of entries. When fixed, the more entries that are available, the more efficient complicated loop processing may be made. But, having too many entries costs more area and energy to implement. In some cases, e.g., with the aggregated architecture, the disclosed input queue 15416 may share queue slots. Use of the slots in a queue may be statically allocated.
The completion queue 15420 may be a separate set of queues to buffer data received from memory in response to memory commands issued by load operations. The completion queue 15420 may be used to hold a load operation that has been scheduled but for which data has not yet been received (and thus has not yet completed). The completion queue 15420, may therefore, be used to reorder data and operation flow.
The operations manager circuit 15430, which will be explained in more detail with reference to
From an architectural perspective, there are at least two goals: first, to run general sequential codes correctly, and second, to obtain high performance in the memory operations performed by the microarchitecture 15500. To ensure program correctness, the compiler expresses the dependency between the store operation and the load operation to an array, p, in some fashion, which are expressed via dependency tokens as will be explained. To improve performance, the microarchitecture 15500 finds and issues as many load commands of an array in parallel as is legal with respect to program order.
In one embodiment, the microarchitecture 15500 may include the operations queue 15412, the input queues 15416, the completion queues 15420, and the operations manager circuit 15430 discussed with reference to
The input queues 15416, as mentioned, may include a load address queue 15522, a store address queue 15524, and a store data queue 15526. (The small numbers 0, 1, 2 are channel labels and will be referred to later in
In one embodiment, the completion queues 15420 may include a set of output buffers 15544 and 15546 for receipt of load data from the memory subsystem 15110 and a completion queue 15542 to buffer addresses and data for load operations according to an index maintained by the operations manager circuit 15430. The operations manager circuit 15430 can manage the index to ensure in-order execution of the load operations, and to identify data received into the output buffers 15544 and 15546 that may be moved to scheduled load operations in the completion queue 15542.
More specifically, because the memory subsystem 15110 is out of order, but the acceleration hardware 15102 completes operations in order, the microarchitecture 15500 may re-order memory operations with use of the completion queue 15542. Three different sub-operations may be performed in relation to the completion queue 15542, namely to allocate, enqueue, and dequeue. For allocation, the operations manager circuit 15430 may allocate an index into the completion queue 15542 in an in-order next slot of the completion queue. The operations manager circuit may provide this index to the memory subsystem 15110, which may then know the slot to which to write data for a load operation. To enqueue, the memory subsystem 15110 may write data as an entry to the indexed, in-order next slot in the completion queue 15542 like random access memory (RAM), setting a status bit of the entry to valid. To dequeue, the operations manager circuit 15430 may present the data stored in this in-order next slot to complete the load operation, setting the status bit of the entry to invalid. Invalid entries may then be available for a new allocation.
In one embodiment, the status signals 15448 may refer to statuses of the input queues 15416, the completion queues 15420, the dependency queues 15518, and the dependency token counters 15514. These statuses, for example, may include an input status, an output status, and a control status, which may refer to the presence or absence of a dependency token in association with an input or an output. The input status may include the presence or absence of addresses and the output status may include the presence or absence of store values and available completion buffer slots. The dependency token counters 15514 may be a compact representation of a queue and track a number of dependency tokens used for any given input queue. If the dependency token counters 15514 saturate, no additional dependency tokens may be generated for new memory operations. Accordingly, the memory ordering circuit 15105 may stall scheduling new memory operations until the dependency token counters 15514 becomes unsaturated.
With additional reference to
ldNo[d,x] result.outN, addr.in64, order.in0, order.out0
stNo[d,x] addr.in64, data.inN, order.in0, order.out0
The executable determiner circuit 15600 may be integrated as a part of the scheduler circuit 15432 and which may perform a logical operation to determine whether a given memory operation is executable, and thus ready to be issued to memory. A memory operation may be executed when the queues corresponding to its memory arguments have data and an associated dependency token is present. These memory arguments may include, for example, an input queue identifier 15610 (indicative of a channel of the input queue 15416), an output queue identifier 15620 (indicative of a channel of the completion queues 15420), a dependency queue identifier 15630 (e.g., what dependency queue or counter should be referenced), and an operation type indicator 15640 (e.g., load operation or store operation). A field (e.g., of a memory request) may be included, e.g., in the above format, that stores a bit or bits to indicate to use the hazard checking hardware.
These memory arguments may be queued within the operations queue 15412, and used to schedule issuance of memory operations in association with incoming addresses and data from memory and the acceleration hardware 15102. (See
For a load operation, and by way of example, the memory ordering circuit 15105 may issue a load command when the load operation has an address (input status) and room to buffer the load result in the completion queue 15542 (output status). Similarly, the memory ordering circuit 15105 may issue a store command for a store operation when the store operation has both an address and data value (input status). Accordingly, the status signals 15448 may communicate a level of emptiness (or fullness) of the queues to which the status signals pertain. The operation type may then dictate whether the logic results in an executable signal depending on what address and data should be available.
To implement dependency ordering, the scheduler circuit 15432 may extend memory operations to include dependency tokens as underlined above in the example load and store operations. The control status 15632 may indicate whether a dependency token is available within the dependency queue identified by the dependency queue identifier 15630, which could be one of the dependency queues 15518 (for an incoming memory operation) or a dependency token counter 15514 (for a completed memory operation). Under this formulation, a dependent memory operation requires an additional ordering token to execute and generates an additional ordering token upon completion of the memory operation, where completion means that data from the result of the memory operation has become available to program-subsequent memory operations.
In one embodiment, with further reference to
The priority encoder 15706, for example, may be a circuit (such as a state machine or a simpler converter) that compresses multiple binary inputs into a smaller number of outputs, including possibly just one output. The output of a priority encoder is the binary representation of the original number starting from zero of the most significant input bit. So, in one example, when memory operation 0 (“zero”), memory operation one (“1”), and memory operation two (“2”) are executable and scheduled, corresponding to 15704A, 15704B, and 15704C, respectively. The priority encoder 15706 may be configured to output the selector signal 15707 to the selection circuitry 15708 indicating the memory operation zero as the memory operation that has highest priority. The selection circuitry 15708 may be a multiplexer in one embodiment, and be configured to output its selection (e.g., of memory operation zero) onto the control lines 15710, as a control signal, in response to the selector signal from the priority encoder 15706 (and indicative of selection of memory operation of highest priority). This control signal may go to the multiplexers 15532, 15534, 15536, and/or 15538, as discussed with reference to
An example of memory ordering by the memory ordering circuit 15105 will be illustrated with a simplified example for purposes of explanation with relation to
Assume, for this example, that array p contains 0,1,2,3,4,5,6, and at the end of loop execution, array p will contain 0,1,0,1,0,1,0. This code may be transformed by unrolling the loop, as illustrated in
The way the microarchitecture may perform this reordering is discussed with reference to
In
In
In
Note that the address p[2] for the newest load operation is dependent on the value that first needs to be stored by the store operation for address p[2], which is at the top of the store address queue. Later, the indexed entry in the completion queue for the load operation from address p[2] may remain buffered until the data from the store operation to the address p[2] is completed (see
In
In
In
In
In
In the present embodiment, the process of executing the code of
More specifically, referring to
The method 16200 may continue with the memory ordering circuit scheduling issuance of the second memory operation to the memory in response to receiving the dependency token and the address associated with the dependency token (16240). For example, when the load address queue receives the address for an address argument of a load operation and the dependency queue receives the dependency token for a control argument of the load operation, the memory ordering circuit may schedule issuance of the second memory operation as a load operation. The method 16200 may continue with the memory ordering circuit issuing the second memory operation (e.g., in a command) to the memory in response to completion of the first memory operation (16250). For example, if the first memory operation is a store, completion may be verified by acknowledgement that the data in a store data queue of the set of input queues has been written to the address in the memory. Similarly, if the first memory operation is a load operation, completion may be verified by receipt of data from the memory for the load operation.
Supercomputing at the ExaFLOP scale may be a challenge in high-performance computing, a challenge which is not likely to be met by conventional von Neumann architectures. To achieve ExaFLOPs, embodiments of a CSA provide a heterogeneous spatial array that targets direct execution of (e.g., compiler-produced) dataflow graphs. In addition to laying out the architectural principles of embodiments of a CSA, the above also describes and evaluates embodiments of a CSA which showed performance and energy of larger than 10× over existing products. Compiler-generated code may have significant performance and energy gains over roadmap architectures. As a heterogeneous, parametric architecture, embodiments of a CSA may be readily adapted to all computing uses. For example, a mobile version of CSA might be tuned to 32-bits, while a machine-learning focused array might feature significant numbers of vectorized 8-bit multiplication units. The main advantages of embodiments of a CSA are high performance and extreme energy efficiency, characteristics relevant to all forms of computing ranging from supercomputing and datacenter to the internet-of-things.
In one embodiment, a processor includes a spatial array of processing elements; and a packet switched communications network to route data within the spatial array between processing elements according to a dataflow graph to perform a first dataflow operation of the dataflow graph, wherein the packet switched communications network further comprises a plurality of network dataflow endpoint circuits to perform a second dataflow operation of the dataflow graph. A network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits may include a network ingress buffer to receive input data from the packet switched communications network; and a spatial array egress buffer to output resultant data to the spatial array of processing elements according to the second dataflow operation on the input data. The spatial array egress buffer may output the resultant data based on a scheduler within the network dataflow endpoint circuit monitoring the packet switched communications network. The spatial array egress buffer may output the resultant data based on the scheduler within the network dataflow endpoint circuit monitoring a selected channel of multiple network virtual channels of the packet switched communications network. A network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits may include a spatial array ingress buffer to receive control data from the spatial array that causes a network ingress buffer of the network dataflow endpoint circuit that received input data from the packet switched communications network to output resultant data to the spatial array of processing elements according to the second dataflow operation on the input data and the control data. A network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits may stall an output of resultant data of the second dataflow operation from a spatial array egress buffer of the network dataflow endpoint circuit when a backpressure signal from a downstream processing element of the spatial array of processing elements indicates that storage in the downstream processing element is not available for the output of the network dataflow endpoint circuit. A network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits may send a backpressure signal to stall a source from sending input data on the packet switched communications network into a network ingress buffer of the network dataflow endpoint circuit when the network ingress buffer is not available. The spatial array of processing elements may include a plurality of processing elements; and an interconnect network between the plurality of processing elements to receive an input of the dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the interconnect network, the plurality of processing elements, and the plurality of network dataflow endpoint circuits with each node represented as a dataflow operator in either of the plurality of processing elements and the plurality of network dataflow endpoint circuits, and the plurality of processing elements and the plurality of network dataflow endpoint circuits are to perform an operation by an incoming operand set arriving at each of the dataflow operators of the plurality of processing elements and the plurality of network dataflow endpoint circuits. The spatial array of processing elements may include a circuit switched network to transport the data within the spatial array between processing elements according to the dataflow graph.
In another embodiment, a method includes providing a spatial array of processing elements; routing, with a packet switched communications network, data within the spatial array between processing elements according to a dataflow graph; performing a first dataflow operation of the dataflow graph with the processing elements; and performing a second dataflow operation of the dataflow graph with a plurality of network dataflow endpoint circuits of the packet switched communications network. The performing the second dataflow operation may include receiving input data from the packet switched communications network with a network ingress buffer of a network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits; and outputting resultant data from a spatial array egress buffer of the network dataflow endpoint circuit to the spatial array of processing elements according to the second dataflow operation on the input data. The outputting may include outputting the resultant data based on a scheduler within the network dataflow endpoint circuit monitoring the packet switched communications network. The outputting may include outputting the resultant data based on the scheduler within the network dataflow endpoint circuit monitoring a selected channel of multiple network virtual channels of the packet switched communications network. The performing the second dataflow operation may include receiving control data, with a spatial array ingress buffer of a network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits, from the spatial array; and configuring the network dataflow endpoint circuit to cause a network ingress buffer of the network dataflow endpoint circuit that received input data from the packet switched communications network to output resultant data to the spatial array of processing elements according to the second dataflow operation on the input data and the control data. The performing the second dataflow operation may include stalling an output of the second dataflow operation from a spatial array egress buffer of a network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits when a backpressure signal from a downstream processing element of the spatial array of processing elements indicates that storage in the downstream processing element is not available for the output of the network dataflow endpoint circuit. The performing the second dataflow operation may include sending a backpressure signal from a network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits to stall a source from sending input data on the packet switched communications network into a network ingress buffer of the network dataflow endpoint circuit when the network ingress buffer is not available. The routing, performing the first dataflow operation, and performing the second dataflow operation may include receiving an input of a dataflow graph comprising a plurality of nodes; overlaying the dataflow graph into the spatial array of processing elements and the plurality of network dataflow endpoint circuits with each node represented as a dataflow operator in either of the processing elements and the plurality of network dataflow endpoint circuits; and performing the first dataflow operation with the processing elements and performing the second dataflow operation with the plurality of network dataflow endpoint circuits when an incoming operand set arrives at each of the dataflow operators of the processing elements and the plurality of network dataflow endpoint circuits. The method may include transporting the data within the spatial array between processing elements according to the dataflow graph with a circuit switched network of the spatial array.
In yet another embodiment, a non-transitory machine readable medium that stores code that when executed by a machine causes the machine to perform a method including providing a spatial array of processing elements; routing, with a packet switched communications network, data within the spatial array between processing elements according to a dataflow graph; performing a first dataflow operation of the dataflow graph with the processing elements; and performing a second dataflow operation of the dataflow graph with a plurality of network dataflow endpoint circuits of the packet switched communications network. The performing the second dataflow operation may include receiving input data from the packet switched communications network with a network ingress buffer of a network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits; and outputting resultant data from a spatial array egress buffer of the network dataflow endpoint circuit to the spatial array of processing elements according to the second dataflow operation on the input data. The outputting may include outputting the resultant data based on a scheduler within the network dataflow endpoint circuit monitoring the packet switched communications network. The outputting may include outputting the resultant data based on the scheduler within the network dataflow endpoint circuit monitoring a selected channel of multiple network virtual channels of the packet switched communications network. The performing the second dataflow operation may include receiving control data, with a spatial array ingress buffer of a network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits, from the spatial array; and configuring the network dataflow endpoint circuit to cause a network ingress buffer of the network dataflow endpoint circuit that received input data from the packet switched communications network to output resultant data to the spatial array of processing elements according to the second dataflow operation on the input data and the control data. The performing the second dataflow operation may include stalling an output of the second dataflow operation from a spatial array egress buffer of a network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits when a backpressure signal from a downstream processing element of the spatial array of processing elements indicates that storage in the downstream processing element is not available for the output of the network dataflow endpoint circuit. The performing the second dataflow operation may include sending a backpressure signal from a network dataflow endpoint circuit of the plurality of network dataflow endpoint circuits to stall a source from sending input data on the packet switched communications network into a network ingress buffer of the network dataflow endpoint circuit when the network ingress buffer is not available. The routing, performing the first dataflow operation, and performing the second dataflow operation may include receiving an input of a dataflow graph comprising a plurality of nodes; overlaying the dataflow graph into the spatial array of processing elements and the plurality of network dataflow endpoint circuits with each node represented as a dataflow operator in either of the processing elements and the plurality of network dataflow endpoint circuits; and performing the first dataflow operation with the processing elements and performing the second dataflow operation with the plurality of network dataflow endpoint circuits when an incoming operand set arrives at each of the dataflow operators of the processing elements and the plurality of network dataflow endpoint circuits. The method may include transporting the data within the spatial array between processing elements according to the dataflow graph with a circuit switched network of the spatial array.
In another embodiment, a processor includes a spatial array of processing elements; and a packet switched communications network to route data within the spatial array between processing elements according to a dataflow graph to perform a first dataflow operation of the dataflow graph, wherein the packet switched communications network further comprises means to perform a second dataflow operation of the dataflow graph.
In one embodiment, a processor includes a core with a decoder to decode an instruction into a decoded instruction and an execution unit to execute the decoded instruction to perform a first operation; a plurality of processing elements; and an interconnect network between the plurality of processing elements to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the interconnect network and the plurality of processing elements with each node represented as a dataflow operator in the plurality of processing elements, and the plurality of processing elements are to perform a second operation by a respective, incoming operand set arriving at each of the dataflow operators of the plurality of processing elements. A processing element of the plurality of processing elements may stall execution when a backpressure signal from a downstream processing element indicates that storage in the downstream processing element is not available for an output of the processing element. The processor may include a flow control path network to carry the backpressure signal according to the dataflow graph. A dataflow token may cause an output from a dataflow operator receiving the dataflow token to be sent to an input buffer of a particular processing element of the plurality of processing elements. The second operation may include a memory access and the plurality of processing elements comprises a memory-accessing dataflow operator that is not to perform the memory access until receiving a memory dependency token from a logically previous dataflow operator. The plurality of processing elements may include a first type of processing element and a second, different type of processing element.
In another embodiment, a method includes decoding an instruction with a decoder of a core of a processor into a decoded instruction; executing the decoded instruction with an execution unit of the core of the processor to perform a first operation; receiving an input of a dataflow graph comprising a plurality of nodes; overlaying the dataflow graph into a plurality of processing elements of the processor and an interconnect network between the plurality of processing elements of the processor with each node represented as a dataflow operator in the plurality of processing elements; and performing a second operation of the dataflow graph with the interconnect network and the plurality of processing elements by a respective, incoming operand set arriving at each of the dataflow operators of the plurality of processing elements. The method may include stalling execution by a processing element of the plurality of processing elements when a backpressure signal from a downstream processing element indicates that storage in the downstream processing element is not available for an output of the processing element. The method may include sending the backpressure signal on a flow control path network according to the dataflow graph. A dataflow token may cause an output from a dataflow operator receiving the dataflow token to be sent to an input buffer of a particular processing element of the plurality of processing elements. The method may include not performing a memory access until receiving a memory dependency token from a logically previous dataflow operator, wherein the second operation comprises the memory access and the plurality of processing elements comprises a memory-accessing dataflow operator. The method may include providing a first type of processing element and a second, different type of processing element of the plurality of processing elements.
In yet another embodiment, an apparatus includes a data path network between a plurality of processing elements; and a flow control path network between the plurality of processing elements, wherein the data path network and the flow control path network are to receive an input of a dataflow graph comprising a plurality of nodes, the dataflow graph is to be overlaid into the data path network, the flow control path network, and the plurality of processing elements with each node represented as a dataflow operator in the plurality of processing elements, and the plurality of processing elements are to perform a second operation by a respective, incoming operand set arriving at each of the dataflow operators of the plurality of processing elements. The flow control path network may carry backpressure signals to a plurality of dataflow operators according to the dataflow graph. A dataflow token sent on the data path network to a dataflow operator may cause an output from the dataflow operator to be sent to an input buffer of a particular processing element of the plurality of processing elements on the data path network. The data path network may be a static, circuit switched network to carry the respective, input operand set to each of the dataflow operators according to the dataflow graph. The flow control path network may transmit a backpressure signal according to the dataflow graph from a downstream processing element to indicate that storage in the downstream processing element is not available for an output of the processing element. At least one data path of the data path network and at least one′flow control path of the flow control path network may form a channelized circuit with backpressure control. The flow control path network may pipeline at least two of the plurality of processing elements in series.
In another embodiment, a method includes receiving an input of a dataflow graph comprising a plurality of nodes; and overlaying the dataflow graph into a plurality of processing elements of a processor, a data path network between the plurality of processing elements, and a flow control path network between the plurality of processing elements with each node represented as a dataflow operator in the plurality of processing elements. The method may include carrying backpressure signals with the flow control path network to a plurality of dataflow operators according to the dataflow graph. The method may include sending a dataflow token on the data path network to a dataflow operator to cause an output from the dataflow operator to be sent to an input buffer of a particular processing element of the plurality of processing elements on the data path network. The method may include setting a plurality of switches of the data path network and/or a plurality of switches of the flow control path network to carry the respective, input operand set to each of the dataflow operators according to the dataflow graph, wherein the data path network is a static, circuit switched network. The method may include transmitting a backpressure signal with the flow control path network according to the dataflow graph from a downstream processing element to indicate that storage in the downstream processing element is not available for an output of the processing element. The method may include forming a channelized circuit with backpressure control with at least one data path of the data path network and at least one flow control path of the flow control path network.
In yet another embodiment, a processor includes a core with a decoder to decode an instruction into a decoded instruction and an execution unit to execute the decoded instruction to perform a first operation; a plurality of processing elements; and a network means between the plurality of processing elements to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the network means and the plurality of processing elements with each node represented as a dataflow operator in the plurality of processing elements, and the plurality of processing elements are to perform a second operation by a respective, incoming operand set arriving at each of the dataflow operators of the plurality of processing elements.
In another embodiment, an apparatus includes a data path means between a plurality of processing elements; and a flow control path means between the plurality of processing elements, wherein the data path means and the flow control path means are to receive an input of a dataflow graph comprising a plurality of nodes, the dataflow graph is to be overlaid into the data path means, the flow control path means, and the plurality of processing elements with each node represented as a dataflow operator in the plurality of processing elements, and the plurality of processing elements are to perform a second operation by a respective, incoming operand set arriving at each of the dataflow operators of the plurality of processing elements.
In one embodiment, a processor includes a core with a decoder to decode an instruction into a decoded instruction and an execution unit to execute the decoded instruction to perform a first operation; and an array of processing elements to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the array of processing elements with each node represented as a dataflow operator in the array of processing elements, and the array of processing elements is to perform a second operation when an incoming operand set arrives at the array of processing elements. The array of processing element may not perform the second operation until the incoming operand set arrives at the array of processing elements and storage in the array of processing elements is available for output of the second operation. The array of processing elements may include a network (or channel(s)) to carry dataflow tokens and control tokens to a plurality of dataflow operators. The second operation may include a memory access and the array of processing elements may include a memory-accessing dataflow operator that is not to perform the memory access until receiving a memory dependency token from a logically previous dataflow operator. Each processing element may perform only one or two operations of the dataflow graph.
In another embodiment, a method includes decoding an instruction with a decoder of a core of a processor into a decoded instruction; executing the decoded instruction with an execution unit of the core of the processor to perform a first operation; receiving an input of a dataflow graph comprising a plurality of nodes; overlaying the dataflow graph into an array of processing elements of the processor with each node represented as a dataflow operator in the array of processing elements; and performing a second operation of the dataflow graph with the array of processing elements when an incoming operand set arrives at the array of processing elements. The array of processing elements may not perform the second operation until the incoming operand set arrives at the array of processing elements and storage in the array of processing elements is available for output of the second operation. The array of processing elements may include a network carrying dataflow tokens and control tokens to a plurality of dataflow operators. The second operation may include a memory access and the array of processing elements comprises a memory-accessing dataflow operator that is not to perform the memory access until receiving a memory dependency token from a logically previous dataflow operator. Each processing element may performs only one or two operations of the dataflow graph.
In yet another embodiment, a non-transitory machine readable medium that stores code that when executed by a machine causes the machine to perform a method including decoding an instruction with a decoder of a core of a processor into a decoded instruction; executing the decoded instruction with an execution unit of the core of the processor to perform a first operation; receiving an input of a dataflow graph comprising a plurality of nodes; overlaying the dataflow graph into an array of processing elements of the processor with each node represented as a dataflow operator in the array of processing elements; and performing a second operation of the dataflow graph with the array of processing elements when an incoming operand set arrives at the array of processing elements. The array of processing element may not perform the second operation until the incoming operand set arrives at the array of processing elements and storage in the array of processing elements is available for output of the second operation. The array of processing elements may include a network carrying dataflow tokens and control tokens to a plurality of dataflow operators. The second operation may include a memory access and the array of processing elements comprises a memory-accessing dataflow operator that is not to perform the memory access until receiving a memory dependency token from a logically previous dataflow operator. Each processing element may performs only one or two operations of the dataflow graph.
In another embodiment, a processor includes a core with a decoder to decode an instruction into a decoded instruction and an execution unit to execute the decoded instruction to perform a first operation; and means to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the means with each node represented as a dataflow operator in the means, and the means is to perform a second operation when an incoming operand set arrives at the means.
In one embodiment, a processor includes a core with a decoder to decode an instruction into a decoded instruction and an execution unit to execute the decoded instruction to perform a first operation; a plurality of processing elements; and an interconnect network between the plurality of processing elements to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the interconnect network and the plurality of processing elements with each node represented as a dataflow operator in the plurality of processing elements, and the plurality of processing elements is to perform a second operation when an incoming operand set arrives at the plurality of processing elements. The processor may further comprise a plurality of configuration controllers, each configuration controller is coupled to a respective subset of the plurality of processing elements, and each configuration controller is to load configuration information from storage and cause coupling of the respective subset of the plurality of processing elements according to the configuration information. The processor may include a plurality of configuration caches, and each configuration controller is coupled to a respective configuration cache to fetch the configuration information for the respective subset of the plurality of processing elements. The first operation performed by the execution unit may prefetch configuration information into each of the plurality of configuration caches. Each of the plurality of configuration controllers may include a reconfiguration circuit to cause a reconfiguration for at least one processing element of the respective subset of the plurality of processing elements on receipt of a configuration error message from the at least one processing element. Each of the plurality of configuration controllers may a reconfiguration circuit to cause a reconfiguration for the respective subset of the plurality of processing elements on receipt of a reconfiguration request message, and disable communication with the respective subset of the plurality of processing elements until the reconfiguration is complete. The processor may include a plurality of exception aggregators, and each exception aggregator is coupled to a respective subset of the plurality of processing elements to collect exceptions from the respective subset of the plurality of processing elements and forward the exceptions to the core for servicing. The processor may include a plurality of extraction controllers, each extraction controller is coupled to a respective subset of the plurality of processing elements, and each extraction controller is to cause state data from the respective subset of the plurality of processing elements to be saved to memory.
In another embodiment, a method includes decoding an instruction with a decoder of a core of a processor into a decoded instruction; executing the decoded instruction with an execution unit of the core of the processor to perform a first operation; receiving an input of a dataflow graph comprising a plurality of nodes; overlaying the dataflow graph into a plurality of processing elements of the processor and an interconnect network between the plurality of processing elements of the processor with each node represented as a dataflow operator in the plurality of processing elements; and performing a second operation of the dataflow graph with the interconnect network and the plurality of processing elements when an incoming operand set arrives at the plurality of processing elements. The method may include loading configuration information from storage for respective subsets of the plurality of processing elements and causing coupling for each respective subset of the plurality of processing elements according to the configuration information. The method may include fetching the configuration information for the respective subset of the plurality of processing elements from a respective configuration cache of a plurality of configuration caches. The first operation performed by the execution unit may be prefetching configuration information into each of the plurality of configuration caches. The method may include causing a reconfiguration for at least one processing element of the respective subset of the plurality of processing elements on receipt of a configuration error message from the at least one processing element. The method may include causing a reconfiguration for the respective subset of the plurality of processing elements on receipt of a reconfiguration request message; and disabling communication with the respective subset of the plurality of processing elements until the reconfiguration is complete. The method may include collecting exceptions from a respective subset of the plurality of processing elements; and forwarding the exceptions to the core for servicing. The method may include causing state data from a respective subset of the plurality of processing elements to be saved to memory.
In yet another embodiment, a non-transitory machine readable medium that stores code that when executed by a machine causes the machine to perform a method including decoding an instruction with a decoder of a core of a processor into a decoded instruction; executing the decoded instruction with an execution unit of the core of the processor to perform a first operation; receiving an input of a dataflow graph comprising a plurality of nodes; overlaying the dataflow graph into a plurality of processing elements of the processor and an interconnect network between the plurality of processing elements of the processor with each node represented as a dataflow operator in the plurality of processing elements; and performing a second operation of the dataflow graph with the interconnect network and the plurality of processing elements when an incoming operand set arrives at the plurality of processing elements. The method may include loading configuration information from storage for respective subsets of the plurality of processing elements and causing coupling for each respective subset of the plurality of processing elements according to the configuration information. The method may include fetching the configuration information for the respective subset of the plurality of processing elements from a respective configuration cache of a plurality of configuration caches. The first operation performed by the execution unit may be prefetching configuration information into each of the plurality of configuration caches. The method may include causing a reconfiguration for at least one processing element of the respective subset of the plurality of processing elements on receipt of a configuration error message from the at least one processing element. The method may include causing a reconfiguration for the respective subset of the plurality of processing elements on receipt of a reconfiguration request message; and disabling communication with the respective subset of the plurality of processing elements until the reconfiguration is complete. The method may include collecting exceptions from a respective subset of the plurality of processing elements; and forwarding the exceptions to the core for servicing. The method may include causing state data from a respective subset of the plurality of processing elements to be saved to memory.
In another embodiment, a processor includes a core with a decoder to decode an instruction into a decoded instruction and an execution unit to execute the decoded instruction to perform a first operation; a plurality of processing elements; and means between the plurality of processing elements to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the m and the plurality of processing elements with each node represented as a dataflow operator in the plurality of processing elements, and the plurality of processing elements is to perform a second operation when an incoming operand set arrives at the plurality of processing elements.
In one embodiment, an apparatus (e.g., a processor) includes: a spatial array of processing elements comprising a communications network to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the spatial array of processing elements with each node represented as a dataflow operator in the spatial array of processing elements, and the spatial array of processing elements is to perform an operation by a respective, incoming operand set arriving at each of the dataflow operators; a plurality of request address file circuits coupled to the spatial array of processing elements and a cache memory, each request address file circuit of the plurality of request address file circuits to access data in the cache memory in response to a request for data access from the spatial array of processing elements; a plurality of translation lookaside buffers comprising a translation lookaside buffer in each of the plurality of request address file circuits to provide an output of a physical address for an input of a virtual address; and a translation lookaside buffer manager circuit comprising a higher level translation lookaside buffer than the plurality of translation lookaside buffers, the translation lookaside buffer manager circuit to perform a first page walk in the cache memory for a miss of an input of a virtual address into a first translation lookaside buffer and into the higher level translation lookaside buffer to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the first page walk in the higher level translation lookaside buffer to cause the higher level translation lookaside buffer to send the physical address to the first translation lookaside buffer in a first request address file circuit. The translation lookaside buffer manager circuit may simultaneously, with the first page walk, perform a second page walk in the cache memory, wherein the second page walk is for a miss of an input of a virtual address into a second translation lookaside buffer and into the higher level translation lookaside buffer to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the second page walk in the higher level translation lookaside buffer to cause the higher level translation lookaside buffer to send the physical address to the second translation lookaside buffer in a second request address file circuit. The receipt of the physical address in the first translation lookaside buffer may cause the first request address file circuit to perform a data access for the request for data access from the spatial array of processing elements on the physical address in the cache memory. The translation lookaside buffer manager circuit may insert an indicator in the higher level translation lookaside buffer for the miss of the input of the virtual address in the first translation lookaside buffer and the higher level translation lookaside buffer to prevent an additional page walk for the input of the virtual address during the first page walk. The translation lookaside buffer manager circuit may receive a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidate the mapping in the higher level translation lookaside buffer, and send shootdown messages to only those of the plurality of request address file circuits that include a copy of the mapping in a respective translation lookaside buffer, wherein each of those of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received. The translation lookaside buffer manager circuit may receive a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidate the mapping in the higher level translation lookaside buffer, and send shootdown messages to all of the plurality of request address file circuits, wherein each of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received.
In another embodiment, a method includes overlaying an input of a dataflow graph comprising a plurality of nodes into a spatial array of processing elements comprising a communications network with each node represented as a dataflow operator in the spatial array of processing elements; coupling a plurality of request address file circuits to the spatial array of processing elements and a cache memory with each request address file circuit of the plurality of request address file circuits accessing data in the cache memory in response to a request for data access from the spatial array of processing elements; providing an output of a physical address for an input of a virtual address into a translation lookaside buffer of a plurality of translation lookaside buffers comprising a translation lookaside buffer in each of the plurality of request address file circuits; coupling a translation lookaside buffer manager circuit comprising a higher level translation lookaside buffer than the plurality of translation lookaside buffers to the plurality of request address file circuits and the cache memory; and performing a first page walk in the cache memory for a miss of an input of a virtual address into a first translation lookaside buffer and into the higher level translation lookaside buffer with the translation lookaside buffer manager circuit to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the first page walk in the higher level translation lookaside buffer to cause the higher level translation lookaside buffer to send the physical address to the first translation lookaside buffer in a first request address file circuit. The method may include simultaneously, with the first page walk, performing a second page walk in the cache memory with the translation lookaside buffer manager circuit, wherein the second page walk is for a miss of an input of a virtual address into a second translation lookaside buffer and into the higher level translation lookaside buffer to determine a physical address mapped to the virtual address, and storing a mapping of the virtual address to the physical address from the second page walk in the higher level translation lookaside buffer to cause the higher level translation lookaside buffer to send the physical address to the second translation lookaside buffer in a second request address file circuit. The method may include causing the first request address file circuit to perform a data access for the request for data access from the spatial array of processing elements on the physical address in the cache memory in response to receipt of the physical address in the first translation lookaside buffer. The method may include inserting, with the translation lookaside buffer manager circuit, an indicator in the higher level translation lookaside buffer for the miss of the input of the virtual address in the first translation lookaside buffer and the higher level translation lookaside buffer to prevent an additional page walk for the input of the virtual address during the first page walk. The method may include receiving, with the translation lookaside buffer manager circuit, a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidating the mapping in the higher level translation lookaside buffer, and sending shootdown messages to only those of the plurality of request address file circuits that include a copy of the mapping in a respective translation lookaside buffer, wherein each of those of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received. The method may include receiving, with the translation lookaside buffer manager circuit, a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidate the mapping in the higher level translation lookaside buffer, and sending shootdown messages to all of the plurality of request address file circuits, wherein each of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received.
In another embodiment, an apparatus includes a spatial array of processing elements comprising a communications network to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the spatial array of processing elements with each node represented as a dataflow operator in the spatial array of processing elements, and the spatial array of processing elements is to perform an operation by a respective, incoming operand set arriving at each of the dataflow operators; a plurality of request address file circuits coupled to the spatial array of processing elements and a plurality of cache memory banks, each request address file circuit of the plurality of request address file circuits to access data in (e.g., each of) the plurality of cache memory banks in response to a request for data access from the spatial array of processing elements; a plurality of translation lookaside buffers comprising a translation lookaside buffer in each of the plurality of request address file circuits to provide an output of a physical address for an input of a virtual address; a plurality of higher level, than the plurality of translation lookaside buffers, translation lookaside buffers comprising a higher level translation lookaside buffer in each of the plurality of cache memory banks to provide an output of a physical address for an input of a virtual address; and a translation lookaside buffer manager circuit to perform a first page walk in the plurality of cache memory banks for a miss of an input of a virtual address into a first translation lookaside buffer and into a first higher level translation lookaside buffer to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the first page walk in the first higher level translation lookaside buffer to cause the first higher level translation lookaside buffer to send the physical address to the first translation lookaside buffer in a first request address file circuit. The translation lookaside buffer manager circuit may simultaneously, with the first page walk, perform a second page walk in the plurality of cache memory banks, wherein the second page walk is for a miss of an input of a virtual address into a second translation lookaside buffer and into a second higher level translation lookaside buffer to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the second page walk in the second higher level translation lookaside buffer to cause the second higher level translation lookaside buffer to send the physical address to the second translation lookaside buffer in a second request address file circuit. The receipt of the physical address in the first translation lookaside buffer may cause the first request address file circuit to perform a data access for the request for data access from the spatial array of processing elements on the physical address in the plurality of cache memory banks. The translation lookaside buffer manager circuit may insert an indicator in the first higher level translation lookaside buffer for the miss of the input of the virtual address in the first translation lookaside buffer and the first higher level translation lookaside buffer to prevent an additional page walk for the input of the virtual address during the first page walk. The translation lookaside buffer manager circuit may receive a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidate the mapping in a higher level translation lookaside buffer storing the mapping, and send shootdown messages to only those of the plurality of request address file circuits that include a copy of the mapping in a respective translation lookaside buffer, wherein each of those of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received. The translation lookaside buffer manager circuit may receive a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidate the mapping in a higher level translation lookaside buffer storing the mapping, and send shootdown messages to all of the plurality of request address file circuits, wherein each of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received.
In yet another embodiment, a method includes: overlaying an input of a dataflow graph comprising a plurality of nodes into a spatial array of processing elements comprising a communications network with each node represented as a dataflow operator in the spatial array of processing elements; coupling a plurality of request address file circuits to the spatial array of processing elements and a plurality of cache memory banks with each request address file circuit of the plurality of request address file circuits accessing data in the plurality of cache memory banks in response to a request for data access from the spatial array of processing elements;
providing an output of a physical address for an input of a virtual address into a translation lookaside buffer of a plurality of translation lookaside buffers comprising a translation lookaside buffer in each of the plurality of request address file circuits; providing an output of a physical address for an input of a virtual address into a higher level, than the plurality of translation lookaside buffers, translation lookaside buffer of a plurality of higher level translation lookaside buffers comprising a higher level translation lookaside buffer in each of the plurality of cache memory banks; coupling a translation lookaside buffer manager circuit to the plurality of request address file circuits and the plurality of cache memory banks; and performing a first page walk in the plurality of cache memory banks for a miss of an input of a virtual address into a first translation lookaside buffer and into a first higher level translation lookaside buffer with the translation lookaside buffer manager circuit to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the first page walk in the first higher level translation lookaside buffer to cause the first higher level translation lookaside buffer to send the physical address to the first translation lookaside buffer in a first request address file circuit. The method may include simultaneously, with the first page walk, performing a second page walk in the plurality of cache memory banks with the translation lookaside buffer manager circuit, wherein the second page walk is for a miss of an input of a virtual address into a second translation lookaside buffer and into a second higher level translation lookaside buffer to determine a physical address mapped to the virtual address, and storing a mapping of the virtual address to the physical address from the second page walk in the second higher level translation lookaside buffer to cause the second higher level translation lookaside buffer to send the physical address to the second translation lookaside buffer in a second request address file circuit. The method may include causing the first request address file circuit to perform a data access for the request for data access from the spatial array of processing elements on the physical address in the plurality of cache memory banks in response to receipt of the physical address in the first translation lookaside buffer. The method may include inserting, with the translation lookaside buffer manager circuit, an indicator in the first higher level translation lookaside buffer for the miss of the input of the virtual address in the first translation lookaside buffer and the first higher level translation lookaside buffer to prevent an additional page walk for the input of the virtual address during the first page walk. The method may include receiving, with the translation lookaside buffer manager circuit, a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidating the mapping in a higher level translation lookaside buffer storing the mapping, and sending shootdown messages to only those of the plurality of request address file circuits that include a copy of the mapping in a respective translation lookaside buffer, wherein each of those of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received. The method may include receiving, with the translation lookaside buffer manager circuit, a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidate the mapping in a higher level translation lookaside buffer storing the mapping, and sending shootdown messages to all of the plurality of request address file circuits, wherein each of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received.
In another embodiment, a system includes a core with a decoder to decode an instruction into a decoded instruction and an execution unit to execute the decoded instruction to perform a first operation; a spatial array of processing elements comprising a communications network to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the spatial array of processing elements with each node represented as a dataflow operator in the spatial array of processing elements, and the spatial array of processing elements is to perform a second operation by a respective, incoming operand set arriving at each of the dataflow operators; a plurality of request address file circuits coupled to the spatial array of processing elements and a cache memory, each request address file circuit of the plurality of request address file circuits to access data in the cache memory in response to a request for data access from the spatial array of processing elements; a plurality of translation lookaside buffers comprising a translation lookaside buffer in each of the plurality of request address file circuits to provide an output of a physical address for an input of a virtual address; and a translation lookaside buffer manager circuit comprising a higher level translation lookaside buffer than the plurality of translation lookaside buffers, the translation lookaside buffer manager circuit to perform a first page walk in the cache memory for a miss of an input of a virtual address into a first translation lookaside buffer and into the higher level translation lookaside buffer to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the first page walk in the higher level translation lookaside buffer to cause the higher level translation lookaside buffer to send the physical address to the first translation lookaside buffer in a first request address file circuit. The translation lookaside buffer manager circuit may simultaneously, with the first page walk, perform a second page walk in the cache memory, wherein the second page walk is for a miss of an input of a virtual address into a second translation lookaside buffer and into the higher level translation lookaside buffer to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the second page walk in the higher level translation lookaside buffer to cause the higher level translation lookaside buffer to send the physical address to the second translation lookaside buffer in a second request address file circuit. The receipt of the physical address in the first translation lookaside buffer may cause the first request address file circuit to perform a data access for the request for data access from the spatial array of processing elements on the physical address in the cache memory. The translation lookaside buffer manager circuit may insert an indicator in the higher level translation lookaside buffer for the miss of the input of the virtual address in the first translation lookaside buffer and the higher level translation lookaside buffer to prevent an additional page walk for the input of the virtual address during the first page walk. The translation lookaside buffer manager circuit may receive a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidate the mapping in the higher level translation lookaside buffer, and send shootdown messages to only those of the plurality of request address file circuits that include a copy of the mapping in a respective translation lookaside buffer, wherein each of those of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received. The translation lookaside buffer manager circuit may receive a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidate the mapping in the higher level translation lookaside buffer, and send shootdown messages to all of the plurality of request address file circuits, wherein each of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received.
In yet another embodiment, a system includes a core with a decoder to decode an instruction into a decoded instruction and an execution unit to execute the decoded instruction to perform a first operation; a spatial array of processing elements comprising a communications network to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the spatial array of processing elements with each node represented as a dataflow operator in the spatial array of processing elements, and the spatial array of processing elements is to perform a second operation by a respective, incoming operand set arriving at each of the dataflow operators; a plurality of request address file circuits coupled to the spatial array of processing elements and a plurality of cache memory banks, each request address file circuit of the plurality of request address file circuits to access data in (e.g., each of) the plurality of cache memory banks in response to a request for data access from the spatial array of processing elements; a plurality of translation lookaside buffers comprising a translation lookaside buffer in each of the plurality of request address file circuits to provide an output of a physical address for an input of a virtual address; a plurality of higher level, than the plurality of translation lookaside buffers, translation lookaside buffers comprising a higher level translation lookaside buffer in each of the plurality of cache memory banks to provide an output of a physical address for an input of a virtual address; and a translation lookaside buffer manager circuit to perform a first page walk in the plurality of cache memory banks for a miss of an input of a virtual address into a first translation lookaside buffer and into a first higher level translation lookaside buffer to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the first page walk in the first higher level translation lookaside buffer to cause the first higher level translation lookaside buffer to send the physical address to the first translation lookaside buffer in a first request address file circuit. The translation lookaside buffer manager circuit may simultaneously, with the first page walk, perform a second page walk in the plurality of cache memory banks, wherein the second page walk is for a miss of an input of a virtual address into a second translation lookaside buffer and into a second higher level translation lookaside buffer to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the second page walk in the second higher level translation lookaside buffer to cause the second higher level translation lookaside buffer to send the physical address to the second translation lookaside buffer in a second request address file circuit. The receipt of the physical address in the first translation lookaside buffer may cause the first request address file circuit to perform a data access for the request for data access from the spatial array of processing elements on the physical address in the plurality of cache memory banks. The translation lookaside buffer manager circuit may insert an indicator in the first higher level translation lookaside buffer for the miss of the input of the virtual address in the first translation lookaside buffer and the first higher level translation lookaside buffer to prevent an additional page walk for the input of the virtual address during the first page walk. The translation lookaside buffer manager circuit may receive a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidate the mapping in a higher level translation lookaside buffer storing the mapping, and send shootdown messages to only those of the plurality of request address file circuits that include a copy of the mapping in a respective translation lookaside buffer, wherein each of those of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received. The translation lookaside buffer manager circuit may receive a shootdown message from a requesting entity for a mapping of a physical address to a virtual address, invalidate the mapping in a higher level translation lookaside buffer storing the mapping, and send shootdown messages to all of the plurality of request address file circuits, wherein each of the plurality of request address file circuits are to send an acknowledgement message to the translation lookaside buffer manager circuit, and the translation lookaside buffer manager circuit is to send a shootdown completion acknowledgment message to the requesting entity when all acknowledgement messages are received.
In another embodiment, an apparatus (e.g., a processor) includes: a spatial array of processing elements comprising a communications network to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the spatial array of processing elements with each node represented as a dataflow operator in the spatial array of processing elements, and the spatial array of processing elements is to perform an operation by a respective, incoming operand set arriving at each of the dataflow operators; a plurality of request address file circuits coupled to the spatial array of processing elements and a cache memory, each request address file circuit of the plurality of request address file circuits to access data in the cache memory in response to a request for data access from the spatial array of processing elements; a plurality of translation lookaside buffers comprising a translation lookaside buffer in each of the plurality of request address file circuits to provide an output of a physical address for an input of a virtual address; and a means comprising a higher level translation lookaside buffer than the plurality of translation lookaside buffers, the means to perform a first page walk in the cache memory for a miss of an input of a virtual address into a first translation lookaside buffer and into the higher level translation lookaside buffer to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the first page walk in the higher level translation lookaside buffer to cause the higher level translation lookaside buffer to send the physical address to the first translation lookaside buffer in a first request address file circuit.
In yet another embodiment, an apparatus includes a spatial array of processing elements comprising a communications network to receive an input of a dataflow graph comprising a plurality of nodes, wherein the dataflow graph is to be overlaid into the spatial array of processing elements with each node represented as a dataflow operator in the spatial array of processing elements, and the spatial array of processing elements is to perform an operation by a respective, incoming operand set arriving at each of the dataflow operators; a plurality of request address file circuits coupled to the spatial array of processing elements and a plurality of cache memory banks, each request address file circuit of the plurality of request address file circuits to access data in (e.g., each of) the plurality of cache memory banks in response to a request for data access from the spatial array of processing elements; a plurality of translation lookaside buffers comprising a translation lookaside buffer in each of the plurality of request address file circuits to provide an output of a physical address for an input of a virtual address; a plurality of higher level, than the plurality of translation lookaside buffers, translation lookaside buffers comprising a higher level translation lookaside buffer in each of the plurality of cache memory banks to provide an output of a physical address for an input of a virtual address; and a means to perform a first page walk in the plurality of cache memory banks for a miss of an input of a virtual address into a first translation lookaside buffer and into a first higher level translation lookaside buffer to determine a physical address mapped to the virtual address, store a mapping of the virtual address to the physical address from the first page walk in the first higher level translation lookaside buffer to cause the first higher level translation lookaside buffer to send the physical address to the first translation lookaside buffer in a first request address file circuit.
In one embodiment, an apparatus (e.g., hardware accelerator) includes a data path having a first branch and a second branch, and the data path comprising at least one processing element; a switch circuit (for example, switch PE, e.g., PE 9) comprising a switch control input to receive a first switch control value to couple an input of the switch circuit to the first branch and a second switch control value to couple the input of the switch circuit to the second branch; a pick circuit (for example, a pick PE, e.g., another instance of PE 9) comprising a pick control input to receive a first pick control value to couple an output of the pick circuit to the first branch and a second pick control value to couple the output of the pick circuit to a third branch of the data path; a predicate propagation processing element to (e.g., simultaneously) output a first edge predicate value and a second edge predicate value based on (e.g., both of) a switch control value from the switch control input of the switch circuit and a first block predicate value (e.g., from another PE); and a predicate merge processing element to (e.g., simultaneously) output a pick control value to the pick control input of the pick circuit and a second block predicate value based on both of a third edge predicate value and one of the first edge predicate value or the second edge predicate value. The second branch and the third branch may be a same branch of the data path. A second predicate propagation processing element may be coupled to the predicate propagation processing element to send the first block predicate value to the predicate propagation processing element based on at least a switch control value from a switch control input of a second switch circuit of the data path. The second predicate propagation processing element may be coupled to the predicate merge processing element to send the third edge predicate value to the predicate merge processing element based on at least the switch control value from the switch control input of the second switch circuit of the data path. A second predicate merge processing element may be coupled to the predicate merge processing element to send the third edge predicate value to the predicate merge processing element based on at least a pick control value from a pick control input of a second pick circuit of the data path. The predicate propagation processing element may output: a false value as the first edge predicate value and a false value as the second edge predicate value when the first block predicate value is a false value; a true value as the first edge predicate value and a false value as the second edge predicate value when the first block predicate value is a true value and the switch control value is a false value; and a false value as the first edge predicate value and a true value as the second edge predicate value when the first block predicate value is a true value and the switch control value is a true value. The predicate merge processing element may output: a false value as the second block predicate value and no value for the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a false value; a true value as the second block predicate value and a false value as the pick control value to the pick control input when the third edge predicate value is a true value and the one of the first edge predicate value or the second edge predicate value is a false value; and a true value as the second block predicate value and a true value as the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a true value. The predicate merge processing element may output: a false value as the second block predicate value and no value for the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a false value; a true value as the second block predicate value and a false value as the pick control value to the pick control input when the third edge predicate value is a true value and the one of the first edge predicate value or the second edge predicate value is a false value; and a true value as the second block predicate value and a true value as the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a true value. The predicate propagation processing element may stall sending of the one of the first edge predicate value or the second edge predicate value to the predicate merge processing element when a backpressure signal from the predicate merge processing element indicates that storage in the predicate merge processing element is not available for the one of the first edge predicate value or the second edge predicate value.
In another embodiment, a method includes receiving, on a switch control input of a switch circuit, a first switch control value to couple an input of the switch circuit to a first branch of a data path or a second switch control value to couple the input of the switch circuit to a second branch of the data path, the data path comprising at least one processing element; receiving, on a pick control input of a pick circuit, a first pick control value to couple an output of the pick circuit to the first branch and a second pick control value to couple the output of the pick circuit to a third branch of the data path; simultaneously outputting, by a predicate propagation processing element, a first edge predicate value and a second edge predicate value based on both of a switch control value from the switch control input of the switch circuit and a first block predicate value; and simultaneously outputting, by a predicate merge processing element, a pick control value to the pick control input of the pick circuit and a second block predicate value based on both of a third edge predicate value and one of the first edge predicate value or the second edge predicate value. The method may include a second predicate propagation processing element sending the first block predicate value to the predicate propagation processing element based on at least a switch control value from a switch control input of a second switch circuit of the data path. The method may include the second predicate propagation processing element sending the third edge predicate value to the predicate merge processing element based on at least the switch control value from the switch control input of the second switch circuit of the data path. The method may include a second predicate merge processing element sending the third edge predicate value to the predicate merge processing element based on at least a pick control value from a pick control input of a second pick circuit of the data path. The method may include the predicate propagation processing element outputting: a false value as the first edge predicate value and a false value as the second edge predicate value when the first block predicate value is a false value; a true value as the first edge predicate value and a false value as the second edge predicate value when the first block predicate value is a true value and the switch control value is a false value; and a false value as the first edge predicate value and a true value as the second edge predicate value when the first block predicate value is a true value and the switch control value is a true value. The method may include the predicate merge processing element outputting: a false value as the second block predicate value and no value for the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a false value; a true value as the second block predicate value and a false value as the pick control value to the pick control input when the third edge predicate value is a true value and the one of the first edge predicate value or the second edge predicate value is a false value; and a true value as the second block predicate value and a true value as the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a true value. The method may include the predicate merge processing element outputting: a false value as the second block predicate value and no value for the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a false value; a true value as the second block predicate value and a false value as the pick control value to the pick control input when the third edge predicate value is a true value and the one of the first edge predicate value or the second edge predicate value is a false value; and a true value as the second block predicate value and a true value as the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a true value. The method may include the predicate propagation processing element stalling sending of the one of the first edge predicate value or the second edge predicate value to the predicate merge processing element when a backpressure signal from the predicate merge processing element indicates that storage in the predicate merge processing element is not available for the one of the first edge predicate value or the second edge predicate value.
In yet another embodiment, a non-transitory machine readable medium that stores code that when executed by a machine causes the machine to perform a method comprising: receiving, on a switch control input of a switch circuit, a first switch control value to couple an input of the switch circuit to a first branch of a data path or a second switch control value to couple the input of the switch circuit to a second branch of the data path, the data path comprising at least one processing element; receiving, on a pick control input of a pick circuit, a first pick control value to couple an output of the pick circuit to the first branch and a second pick control value to couple the output of the pick circuit to a third branch of the data path; simultaneously outputting, by a predicate propagation processing element, a first edge predicate value and a second edge predicate value based on both of a switch control value from the switch control input of the switch circuit and a first block predicate value; and simultaneously outputting, by a predicate merge processing element, a pick control value to the pick control input of the pick circuit and a second block predicate value based on both of a third edge predicate value and one of the first edge predicate value or the second edge predicate value. The method may include a second predicate propagation processing element sending the first block predicate value to the predicate propagation processing element based on at least a switch control value from a switch control input of a second switch circuit of the data path. The method may include the second predicate propagation processing element sending the third edge predicate value to the predicate merge processing element based on at least the switch control value from the switch control input of the second switch circuit of the data path. The method may include a second predicate merge processing element sending the third edge predicate value to the predicate merge processing element based on at least a pick control value from a pick control input of a second pick circuit of the data path. The method may include the predicate propagation processing element outputting: a false value as the first edge predicate value and a false value as the second edge predicate value when the first block predicate value is a false value; a true value as the first edge predicate value and a false value as the second edge predicate value when the first block predicate value is a true value and the switch control value is a false value; and a false value as the first edge predicate value and a true value as the second edge predicate value when the first block predicate value is a true value and the switch control value is a true value. The method may include the predicate merge processing element outputting: a false value as the second block predicate value and no value for the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a false value; a true value as the second block predicate value and a false value as the pick control value to the pick control input when the third edge predicate value is a true value and the one of the first edge predicate value or the second edge predicate value is a false value; and a true value as the second block predicate value and a true value as the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a true value. The method may include the predicate merge processing element outputting: a false value as the second block predicate value and no value for the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a false value; a true value as the second block predicate value and a false value as the pick control value to the pick control input when the third edge predicate value is a true value and the one of the first edge predicate value or the second edge predicate value is a false value; and a true value as the second block predicate value and a true value as the pick control value to the pick control input when the third edge predicate value is a false value and the one of the first edge predicate value or the second edge predicate value is a true value. The method may include the predicate propagation processing element stalling sending of the one of the first edge predicate value or the second edge predicate value to the predicate merge processing element when a backpressure signal from the predicate merge processing element indicates that storage in the predicate merge processing element is not available for the one of the first edge predicate value or the second edge predicate value.
In another embodiment, an apparatus (e.g., hardware accelerator) includes a data path having a first branch and a second branch, and the data path comprising at least one processing element; a switch circuit (for example, switch PE, e.g., PE 9) comprising a switch control input to receive a first switch control value to couple an input of the switch circuit to the first branch and a second switch control value to couple the input of the switch circuit to the second branch; a pick circuit (for example, a pick PE, e.g., another instance of PE 9) comprising a pick control input to receive a first pick control value to couple an output of the pick circuit to the first branch and a second pick control value to couple the output of the pick circuit to a third branch of the data path; a first means to (e.g., simultaneously) output a first edge predicate value and a second edge predicate value based on (e.g., both of) a switch control value from the switch control input of the switch circuit and a first block predicate value (e.g., from another PE); and a second means to (e.g., simultaneously) output a pick control value to the pick control input of the pick circuit and a second block predicate value based on both of a third edge predicate value and one of the first edge predicate value or the second edge predicate value.
In one embodiment, an apparatus (e.g., an accelerator circuit) includes a plurality of processing elements; an interconnect network between the plurality of processing elements to transfer values between the plurality of processing elements; and a first processing element of the plurality of processing elements comprising: a configuration register within the first processing element to store a configuration value that causes the first processing element to perform an operation according to the configuration value, a plurality of input queues, an input controller to control enqueue and dequeue of values into the plurality of input queues according to the configuration value, a plurality of output queues, and an output controller to control enqueue and dequeue of values into the plurality of output queues according to the configuration value. Wherein, when at least one of the plurality of input queues stores a value, the input controller may send a not empty value to operation circuitry of the first processing element to indicate the first processing element may begin the operation on the value stored in the at least one of the plurality of input queues. When at least one of the plurality of output queues is not full, the output controller may send a not full value to operation circuitry of the first processing element to indicate the first processing element may begin the operation on a value stored in the at least one of the plurality of input queues. When at least one of the plurality of input queues stores a value, the input controller may send a not empty value to operation circuitry of the first processing element and when at least one of the plurality of output queues is not full, the output controller may send a not full value to operation circuitry of the first processing element, and the operation circuitry of the first processing element may begin the operation on the value stored in the at least one of the plurality of input queues after both the not empty value and the not full value are received. When at least one of the plurality of input queues is not full, the input controller may send a ready value to an upstream processing element of the plurality of processing elements. When at least one of a plurality of output queues of the upstream processing element stores a value, an output controller of the upstream processing element may send a valid value to the input controller of the first processing element, and the input controller of the first processing element may enqueue the value into the at least one of the plurality of input queues from the at least one of the plurality of output queues of the upstream processing element after both the ready value is asserted by the first processing element and the valid value is received from the upstream processing element. When at least one of the plurality of output queues stores a value, the output controller may send a valid value to a downstream processing element of the plurality of processing elements. When at least one of a plurality of input queues of the downstream processing element is not full, an input controller of the downstream processing element may send a ready value to the output controller of the first processing element, and the output controller of the first processing element may dequeue the value from the at least one of the plurality of output queues after both the valid value is asserted by the first processing element and the ready value is received from the downstream processing element.
In another embodiment, a method includes coupling a plurality of processing elements together by an interconnect network between the plurality of processing elements to transfer values between the plurality of processing elements; storing a configuration value in a configuration register within a first processing element of the plurality of processing elements that causes the first processing element to perform an operation according to the configuration value; controlling enqueue and dequeue of values into a plurality of input queues of the first processing element according to the configuration value with an input controller in the first processing element; and controlling enqueue and dequeue of values into a plurality of output queues of the first processing element according to the configuration value with an output controller in the first processing element. When at least one of the plurality of input queues stores a value, the input controller may send a not empty value to operation circuitry of the first processing element to indicate the first processing element may begin the operation on the value stored in the at least one of the plurality of input queues. When at least one of the plurality of output queues is not full, the output controller may send a not full value to operation circuitry of the first processing element to indicate the first processing element may begin the operation on a value stored in the at least one of the plurality of input queues. When at least one of the plurality of input queues stores a value, the input controller may send a not empty value to operation circuitry of the first processing element and when at least one of the plurality of output queues is not full, the output controller may send a not full value to operation circuitry of the first processing element, and the operation circuitry of the first processing element may begin the operation on the value stored in the at least one of the plurality of input queues after both the not empty value and the not full value are received. When at least one of the plurality of input queues is not full, the input controller may send a ready value to an upstream processing element of the plurality of processing elements. When at least one of a plurality of output queues of the upstream processing element stores a value, an output controller of the upstream processing element may send a valid value to the input controller of the first processing element, and the input controller of the first processing element may enqueue the value into the at least one of the plurality of input queues from the at least one of the plurality of output queues of the upstream processing element after both the ready value is asserted by the first processing element and the valid value is received from the upstream processing element. When at least one of the plurality of output queues stores a value, the output controller may send a valid value to a downstream processing element of the plurality of processing elements. When at least one of a plurality of input queues of the downstream processing element is not full, an input controller of the downstream processing element may send a ready value to the output controller of the first processing element, and the output controller of the first processing element may dequeue the value from the at least one of the plurality of output queues after both the valid value is asserted by the first processing element and the ready value is received from the downstream processing element.
In yet another embodiment, a processor includes a core with a decoder to decode an instruction into a decoded instruction and an execution unit to execute the decoded instruction to perform a first operation; a plurality of processing elements; an interconnect network between the plurality of processing elements to transfer values between the plurality of processing elements; and a first processing element of the plurality of processing elements comprising: a configuration register within the first processing element to store a configuration value that causes the first processing element to perform a second operation according to the configuration value, a plurality of input queues, an input controller to control enqueue and dequeue of values into the plurality of input queues according to the configuration value, a plurality of output queues, and an output controller to control enqueue and dequeue of values into the plurality of output queues according to the configuration value. Wherein, when at least one of the plurality of input queues stores a value, the input controller may send a not empty value to operation circuitry of the first processing element to indicate the first processing element may begin the second operation on the value stored in the at least one of the plurality of input queues. When at least one of the plurality of output queues is not full, the output controller may send a not full value to operation circuitry of the first processing element to indicate the first processing element may begin the second operation on a value stored in the at least one of the plurality of input queues. When at least one of the plurality of input queues stores a value, the input controller may send a not empty value to operation circuitry of the first processing element and when at least one of the plurality of output queues is not full, the output controller may send a not full value to operation circuitry of the first processing element, and the operation circuitry of the first processing element may begin the second operation on the value stored in the at least one of the plurality of input queues after both the not empty value and the not full value are received. When at least one of the plurality of input queues is not full, the input controller may send a ready value to an upstream processing element of the plurality of processing elements. When at least one of a plurality of output queues of the upstream processing element stores a value, an output controller of the upstream processing element may send a valid value to the input controller of the first processing element, and the input controller of the first processing element may enqueue the value into the at least one of the plurality of input queues from the at least one of the plurality of output queues of the upstream processing element after both the ready value is asserted by the first processing element and the valid value is received from the upstream processing element. When at least one of the plurality of output queues stores a value, the output controller may send a valid value to a downstream processing element of the plurality of processing elements. When at least one of a plurality of input queues of the downstream processing element is not full, an input controller of the downstream processing element may send a ready value to the output controller of the first processing element, and the output controller of the first processing element may dequeue the value from the at least one of the plurality of output queues after both the valid value is asserted by the first processing element and the ready value is received from the downstream processing element.
In another embodiment, an apparatus (e.g., an accelerator circuit) includes a plurality of processing elements; an interconnect network between the plurality of processing elements to transfer values between the plurality of processing elements; and a first processing element of the plurality of processing elements comprising: a configuration register within the first processing element to store a configuration value that causes the first processing element to perform an operation according to the configuration value, a plurality of input queues, a first means to control enqueue and dequeue of values into the plurality of input queues according to the configuration value, a plurality of output queues, and a second means to control enqueue and dequeue of values into the plurality of output queues according to the configuration value.
In yet another embodiment, an apparatus (e.g., an accelerator circuit) includes a plurality of processing elements; an interconnect network between the plurality of processing elements to transfer values between the plurality of processing elements; and a first processing element of the plurality of processing elements comprising: a configuration register within the first processing element to store a configuration value that causes the first processing element to perform an operation according to the configuration value, a plurality of input queues, a plurality of output queues, and means to control enqueue and dequeue of values into the plurality of input queues according to the configuration value and control enqueue and dequeue of values into the plurality of output queues according to the configuration value.
In another embodiment, an apparatus comprises a data storage device that stores code that when executed by a hardware processor causes the hardware processor to perform any method disclosed herein. An apparatus may be as described in the detailed description. A method may be as described in the detailed description.
In yet another embodiment, a non-transitory machine readable medium that stores code that when executed by a machine causes the machine to perform a method comprising any method disclosed herein.
An instruction set (e.g., for execution by a core) may include one or more instruction formats. A given instruction format may define various fields (e.g., number of bits, location of bits) to specify, among other things, the operation to be performed (e.g., opcode) and the operand(s) on which that operation is to be performed and/or other data field(s) (e.g., mask).
Some instruction formats are further broken down though the definition of instruction templates (or subformats). For example, the instruction templates of a given instruction format may be defined to have different subsets of the instruction format's fields (the included fields are typically in the same order, but at least some have different bit positions because there are less fields included) and/or defined to have a given field interpreted differently. Thus, each instruction of an ISA is expressed using a given instruction format (and, if defined, in a given one of the instruction templates of that instruction format) and includes fields for specifying the operation and the operands. For example, an exemplary ADD instruction has a specific opcode and an instruction format that includes an opcode field to specify that opcode and operand fields to select operands (source1/destination and source2); and an occurrence of this ADD instruction in an instruction stream will have specific contents in the operand fields that select specific operands. A set of SIMD extensions referred to as the Advanced Vector Extensions (AVX) (AVX1 and AVX2) and using the Vector Extensions (VEX) coding scheme has been released and/or published (e.g., see Intel® 64 and IA-32 Architectures Software Developer's Manual, June 2016; and see Intel® Architecture Instruction Set Extensions Programming Reference, February 2016).
Exemplary Instruction Formats
Embodiments of the instruction(s) described herein may be embodied in different formats. Additionally, exemplary systems, architectures, and pipelines are detailed below. Embodiments of the instruction(s) may be executed on such systems, architectures, and pipelines, but are not limited to those detailed.
Generic Vector Friendly Instruction Format
A vector friendly instruction format is an instruction format that is suited for vector instructions (e.g., there are certain fields specific to vector operations). While embodiments are described in which both vector and scalar operations are supported through the vector friendly instruction format, alternative embodiments use only vector operations the vector friendly instruction format.
While embodiments of the disclosure will be described in which the vector friendly instruction format supports the following: a 64 byte vector operand length (or size) with 32 bit (4 byte) or 64 bit (8 byte) data element widths (or sizes) (and thus, a 64 byte vector consists of either 16 doubleword-size elements or alternatively, 8 quadword-size elements); a 64 byte vector operand length (or size) with 16 bit (2 byte) or 8 bit (1 byte) data element widths (or sizes); a 32 byte vector operand length (or size) with 32 bit (4 byte), 64 bit (8 byte), 16 bit (2 byte), or 8 bit (1 byte) data element widths (or sizes); and a 16 byte vector operand length (or size) with 32 bit (4 byte), 64 bit (8 byte), 16 bit (2 byte), or 8 bit (1 byte) data element widths (or sizes); alternative embodiments may support more, less and/or different vector operand sizes (e.g., 256 byte vector operands) with more, less, or different data element widths (e.g., 128 bit (16 byte) data element widths).
The class A instruction templates in
The generic vector friendly instruction format 16300 includes the following fields listed below in the order illustrated in
Format field 16340— a specific value (an instruction format identifier value) in this field uniquely identifies the vector friendly instruction format, and thus occurrences of instructions in the vector friendly instruction format in instruction streams. As such, this field is optional in the sense that it is not needed for an instruction set that has only the generic vector friendly instruction format.
Base operation field 16342— its content distinguishes different base operations.
Register index field 16344— its content, directly or through address generation, specifies the locations of the source and destination operands, be they in registers or in memory. These include a sufficient number of bits to select N registers from a PxQ (e.g. 32×512, 16×128, 32×1024, 64×1024) register file. While in one embodiment N may be up to three sources and one destination register, alternative embodiments may support more or less sources and destination registers (e.g., may support up to two sources where one of these sources also acts as the destination, may support up to three sources where one of these sources also acts as the destination, may support up to two sources and one destination).
Modifier field 16346— its content distinguishes occurrences of instructions in the generic vector instruction format that specify memory access from those that do not; that is, between no memory access 16305 instruction templates and memory access 16320 instruction templates. Memory access operations read and/or write to the memory hierarchy (in some cases specifying the source and/or destination addresses using values in registers), while non-memory access operations do not (e.g., the source and destinations are registers). While in one embodiment this field also selects between three different ways to perform memory address calculations, alternative embodiments may support more, less, or different ways to perform memory address calculations.
Augmentation operation field 16350— its content distinguishes which one of a variety of different operations to be performed in addition to the base operation. This field is context specific. In one embodiment of the disclosure, this field is divided into a class field 16368, an alpha field 16352, and a beta field 16354. The augmentation operation field 16350 allows common groups of operations to be performed in a single instruction rather than 2, 3, or 4 instructions.
Scale field 16360— its content allows for the scaling of the index field's content for memory address generation (e.g., for address generation that uses 2scale*index+base).
Displacement Field 16362A— its content is used as part of memory address generation (e.g., for address generation that uses 2scale*index+base+displacement).
Displacement Factor Field 16362B (note that the juxtaposition of displacement field 16362A directly over displacement factor field 16362B indicates one or the other is used)—its content is used as part of address generation; it specifies a displacement factor that is to be scaled by the size of a memory access (N)— where N is the number of bytes in the memory access (e.g., for address generation that uses 2scale*index+base+scaled displacement). Redundant low-order bits are ignored and hence, the displacement factor field's content is multiplied by the memory operands total size (N) in order to generate the final displacement to be used in calculating an effective address. The value of N is determined by the processor hardware at runtime based on the full opcode field 16374 (described later herein) and the data manipulation field 16354C. The displacement field 16362A and the displacement factor field 16362B are optional in the sense that they are not used for the no memory access 16305 instruction templates and/or different embodiments may implement only one or none of the two.
Data element width field 16364— its content distinguishes which one of a number of data element widths is to be used (in some embodiments for all instructions; in other embodiments for only some of the instructions). This field is optional in the sense that it is not needed if only one data element width is supported and/or data element widths are supported using some aspect of the opcodes.
Write mask field 16370— its content controls, on a per data element position basis, whether that data element position in the destination vector operand reflects the result of the base operation and augmentation operation. Class A instruction templates support merging-writemasking, while class B instruction templates support both merging- and zeroing-writemasking. When merging, vector masks allow any set of elements in the destination to be protected from updates during the execution of any operation (specified by the base operation and the augmentation operation); in other one embodiment, preserving the old value of each element of the destination where the corresponding mask bit has a 0. In contrast, when zeroing vector masks allow any set of elements in the destination to be zeroed during the execution of any operation (specified by the base operation and the augmentation operation); in one embodiment, an element of the destination is set to 0 when the corresponding mask bit has a 0 value. A subset of this functionality is the ability to control the vector length of the operation being performed (that is, the span of elements being modified, from the first to the last one); however, it is not necessary that the elements that are modified be consecutive. Thus, the write mask field 16370 allows for partial vector operations, including loads, stores, arithmetic, logical, etc. While embodiments of the disclosure are described in which the write mask field's 16370 content selects one of a number of write mask registers that contains the write mask to be used (and thus the write mask field's 16370 content indirectly identifies that masking to be performed), alternative embodiments instead or additional allow the mask write field's 16370 content to directly specify the masking to be performed.
Immediate field 16372— its content allows for the specification of an immediate. This field is optional in the sense that is it not present in an implementation of the generic vector friendly format that does not support immediate and it is not present in instructions that do not use an immediate.
Class field 16368— its content distinguishes between different classes of instructions. With reference to
Instruction Templates of Class A
In the case of the non-memory access 16305 instruction templates of class A, the alpha field 16352 is interpreted as an RS field 16352A, whose content distinguishes which one of the different augmentation operation types are to be performed (e.g., round 16352A.1 and data transform 16352A.2 are respectively specified for the no memory access, round type operation 16310 and the no memory access, data transform type operation 16315 instruction templates), while the beta field 16354 distinguishes which of the operations of the specified type is to be performed. In the no memory access 16305 instruction templates, the scale field 16360, the displacement field 16362A, and the displacement scale filed 16362B are not present. No-Memory Access Instruction Templates—Full Round Control Type Operation
In the no memory access full round control type operation 16310 instruction template, the beta field 16354 is interpreted as a round control field 16354A, whose content(s) provide static rounding. While in the described embodiments of the disclosure the round control field 16354A includes a suppress all floating point exceptions (SAE) field 16356 and a round operation control field 16358, alternative embodiments may support may encode both these concepts into the same field or only have one or the other of these concepts/fields (e.g., may have only the round operation control field 16358).
SAE field 16356— its content distinguishes whether or not to disable the exception event reporting; when the SAE field's 16356 content indicates suppression is enabled, a given instruction does not report any kind of floating-point exception flag and does not raise any floating point exception handler.
Round operation control field 16358— its content distinguishes which one of a group of rounding operations to perform (e.g., Round-up, Round-down, Round-towards-zero and Round-to-nearest). Thus, the round operation control field 16358 allows for the changing of the rounding mode on a per instruction basis. In one embodiment of the disclosure where a processor includes a control register for specifying rounding modes, the round operation control field's 16350 content overrides that register value.
No Memory Access Instruction Templates—Data Transform Type Operation
In the no memory access data transform type operation 16315 instruction template, the beta field 16354 is interpreted as a data transform field 16354B, whose content distinguishes which one of a number of data transforms is to be performed (e.g., no data transform, swizzle, broadcast).
In the case of a memory access 16320 instruction template of class A, the alpha field 16352 is interpreted as an eviction hint field 16352B, whose content distinguishes which one of the eviction hints is to be used (in
Vector memory instructions perform vector loads from and vector stores to memory, with conversion support. As with regular vector instructions, vector memory instructions transfer data from/to memory in a data element-wise fashion, with the elements that are actually transferred is dictated by the contents of the vector mask that is selected as the write mask.
Memory Access Instruction Templates—Temporal
Temporal data is data likely to be reused soon enough to benefit from caching. This is, however, a hint, and different processors may implement it in different ways, including ignoring the hint entirely.
Memory Access Instruction Templates—Non-Temporal
Non-temporal data is data unlikely to be reused soon enough to benefit from caching in the 1st-level cache and should be given priority for eviction. This is, however, a hint, and different processors may implement it in different ways, including ignoring the hint entirely.
Instruction Templates of Class B
In the case of the instruction templates of class B, the alpha field 16352 is interpreted as a write mask control (Z) field 16352C, whose content distinguishes whether the write masking controlled by the write mask field 16370 should be a merging or a zeroing.
In the case of the non-memory access 16305 instruction templates of class B, part of the beta field 16354 is interpreted as an RL field 16357A, whose content distinguishes which one of the different augmentation operation types are to be performed (e.g., round 16357A.1 and vector length (VSIZE) 16357A.2 are respectively specified for the no memory access, write mask control, partial round control type operation 16312 instruction template and the no memory access, write mask control, VSIZE type operation 16317 instruction template), while the rest of the beta field 16354 distinguishes which of the operations of the specified type is to be performed. In the no memory access 16305 instruction templates, the scale field 16360, the displacement field 16362A, and the displacement scale filed 16362B are not present.
In the no memory access, write mask control, partial round control type operation 16310 instruction template, the rest of the beta field 16354 is interpreted as a round operation field 16359A and exception event reporting is disabled (a given instruction does not report any kind of floating-point exception flag and does not raise any floating point exception handler).
Round operation control field 16359A— just as round operation control field 16358, its content distinguishes which one of a group of rounding operations to perform (e.g., Round-up, Round-down, Round-towards-zero and Round-to-nearest). Thus, the round operation control field 16359A allows for the changing of the rounding mode on a per instruction basis. In one embodiment of the disclosure where a processor includes a control register for specifying rounding modes, the round operation control field's 16350 content overrides that register value.
In the no memory access, write mask control, VSIZE type operation 16317 instruction template, the rest of the beta field 16354 is interpreted as a vector length field 16359B, whose content distinguishes which one of a number of data vector lengths is to be performed on (e.g., 128, 256, or 512 byte).
In the case of a memory access 16320 instruction template of class B, part of the beta field 16354 is interpreted as a broadcast field 16357B, whose content distinguishes whether or not the broadcast type data manipulation operation is to be performed, while the rest of the beta field 16354 is interpreted the vector length field 16359B. The memory access 16320 instruction templates include the scale field 16360, and optionally the displacement field 16362A or the displacement scale field 16362B.
With regard to the generic vector friendly instruction format 16300, a full opcode field 16374 is shown including the format field 16340, the base operation field 16342, and the data element width field 16364. While one embodiment is shown where the full opcode field 16374 includes all of these fields, the full opcode field 16374 includes less than all of these fields in embodiments that do not support all of them. The full opcode field 16374 provides the operation code (opcode).
The augmentation operation field 16350, the data element width field 16364, and the write mask field 16370 allow these features to be specified on a per instruction basis in the generic vector friendly instruction format.
The combination of write mask field and data element width field create typed instructions in that they allow the mask to be applied based on different data element widths.
The various instruction templates found within class A and class B are beneficial in different situations. In some embodiments of the disclosure, different processors or different cores within a processor may support only class A, only class B, or both classes. For instance, a high performance general purpose out-of-order core intended for general-purpose computing may support only class B, a core intended primarily for graphics and/or scientific (throughput) computing may support only class A, and a core intended for both may support both (of course, a core that has some mix of templates and instructions from both classes but not all templates and instructions from both classes is within the purview of the disclosure). Also, a single processor may include multiple cores, all of which support the same class or in which different cores support different class. For instance, in a processor with separate graphics and general purpose cores, one of the graphics cores intended primarily for graphics and/or scientific computing may support only class A, while one or more of the general purpose cores may be high performance general purpose cores with out of order execution and register renaming intended for general-purpose computing that support only class B. Another processor that does not have a separate graphics core, may include one more general purpose in-order or out-of-order cores that support both class A and class B. Of course, features from one class may also be implement in the other class in different embodiments of the disclosure. Programs written in a high level language would be put (e.g., just in time compiled or statically compiled) into an variety of different executable forms, including: 1) a form having only instructions of the class(es) supported by the target processor for execution; or 2) a form having alternative routines written using different combinations of the instructions of all classes and having control flow code that selects the routines to execute based on the instructions supported by the processor which is currently executing the code.
Exemplary Specific Vector Friendly Instruction Format
It should be understood that, although embodiments of the disclosure are described with reference to the specific vector friendly instruction format 16400 in the context of the generic vector friendly instruction format 16300 for illustrative purposes, the disclosure is not limited to the specific vector friendly instruction format 16400 except where claimed. For example, the generic vector friendly instruction format 16300 contemplates a variety of possible sizes for the various fields, while the specific vector friendly instruction format 16400 is shown as having fields of specific sizes. By way of specific example, while the data element width field 16364 is illustrated as a one bit field in the specific vector friendly instruction format 16400, the disclosure is not so limited (that is, the generic vector friendly instruction format 16300 contemplates other sizes of the data element width field 16364).
The generic vector friendly instruction format 16300 includes the following fields listed below in the order illustrated in
EVEX Prefix (Bytes 0-3) 16402—is encoded in a four-byte form.
Format Field 16340 (EVEX Byte 0, bits [7:0])—the first byte (EVEX Byte 0) is the format field 16340 and it contains 0×62 (the unique value used for distinguishing the vector friendly instruction format in one embodiment of the disclosure).
The second-fourth bytes (EVEX Bytes 1-3) include a number of bit fields providing specific capability.
REX field 16405 (EVEX Byte 1, bits [7-5])— consists of a EVEX.R bit field (EVEX Byte 1, bit [7]—R), EVEX.X bit field (EVEX byte 1, bit [6]—X), and 16357BEX byte 1, bit[5]—B). The EVEX.R, EVEX.X, and EVEX.B bit fields provide the same functionality as the corresponding VEX bit fields, and are encoded using is complement form, e.g., ZMM0 is encoded as 1111B, ZMM15 is encoded as 0000B. Other fields of the instructions encode the lower three bits of the register indexes as is known in the art (rrr, xxx, and bbb), so that Rrrr, Xxxx, and Bbbb may be formed by adding EVEX.R, EVEX.X, and EVEX.B.
REX′ field 16310— this is the first part of the REX′ field 16310 and is the EVEX.R′ bit field (EVEX Byte 1, bit [4]—R′) that is used to encode either the upper 16 or lower 16 of the extended 32 register set. In one embodiment of the disclosure, this bit, along with others as indicated below, is stored in bit inverted format to distinguish (in the well-known x86 32-bit mode) from the BOUND instruction, whose real opcode byte is 62, but does not accept in the MOD R/M field (described below) the value of 11 in the MOD field; alternative embodiments of the disclosure do not store this and the other indicated bits below in the inverted format. A value of 1 is used to encode the lower 16 registers. In other words, R′Rrrr is formed by combining EVEX.R′, EVEX.R, and the other RRR from other fields.
Opcode map field 16415 (EVEX byte 1, bits [3:0]—mmmm)—its content encodes an implied leading opcode byte (0F, 0F 38, or 0F 3).
Data element width field 16364 (EVEX byte 2, bit [7]—W)—is represented by the notation EVEX.W. EVEX.W is used to define the granularity (size) of the datatype (either 32-bit data elements or 64-bit data elements).
EVEX.vvvv 16420 (EVEX Byte 2, bits [6:3]-vvvv)—the role of EVEX.vvvv may include the following: 1) EVEX.vvvv encodes the first source register operand, specified in inverted (1s complement) form and is valid for instructions with 2 or more source operands; 2) EVEX.vvvv encodes the destination register operand, specified in 1s complement form for certain vector shifts; or 3) EVEX.vvvv does not encode any operand, the field is reserved and should contain 1111b. Thus, EVEX.vvvv field 16420 encodes the 4 low-order bits of the first source register specifier stored in inverted (1 s complement) form. Depending on the instruction, an extra different EVEX bit field is used to extend the specifier size to 32 registers.
EVEX.U 16368 Class field (EVEX byte 2, bit [2]—U)—If EVEX.U=0, it indicates class A or EVEX.U0; if EVEX.U=1, it indicates class B or EVEX.U1.
Prefix encoding field 16425 (EVEX byte 2, bits [1:0]-pp)—provides additional bits for the base operation field. In addition to providing support for the legacy SSE instructions in the EVEX prefix format, this also has the benefit of compacting the SIMD prefix (rather than requiring a byte to express the SIMD prefix, the EVEX prefix requires only 2 bits). In one embodiment, to support legacy SSE instructions that use a SIMD prefix (66H, F2H, F3H) in both the legacy format and in the EVEX prefix format, these legacy SIMD prefixes are encoded into the SIMD prefix encoding field; and at runtime are expanded into the legacy SIMD prefix prior to being provided to the decoder's PLA (so the PLA can execute both the legacy and EVEX format of these legacy instructions without modification). Although newer instructions could use the EVEX prefix encoding field's content directly as an opcode extension, certain embodiments expand in a similar fashion for consistency but allow for different meanings to be specified by these legacy SIMD prefixes. An alternative embodiment may redesign the PLA to support the 2 bit SIMD prefix encodings, and thus not require the expansion.
Alpha field 16352 (EVEX byte 3, bit [7]—EH; also known as EVEX.EH, EVEX.rs, EVEX.RL, EVEX.write mask control, and EVEX.N; also illustrated with α)—as previously described, this field is context specific.
Beta field 16354 (EVEX byte 3, bits [6:4]—SSS, also known as EVEX.s2-0, EVEX.r2-0, EVEX.rr1, EVEX.LL0, EVEX.LLB; also illustrated with βββ)—as previously described, this field is context specific.
REX′ field 16310— this is the remainder of the REX′ field and is the EVEX.V′ bit field (EVEX Byte 3, bit [3]—V′) that may be used to encode either the upper 16 or lower 16 of the extended 32 register set. This bit is stored in bit inverted format. A value of 1 is used to encode the lower 16 registers. In other words, V′VVVV is formed by combining EVEX.V′, EVEX.vvvv.
Write mask field 16370 (EVEX byte 3, bits [2:0]—kkk)—its content specifies the index of a register in the write mask registers as previously described. In one embodiment of the disclosure, the specific value EVEX.kkk=000 has a special behavior implying no write mask is used for the particular instruction (this may be implemented in a variety of ways including the use of a write mask hardwired to all ones or hardware that bypasses the masking hardware).
Real Opcode Field 16430 (Byte 4) is also known as the opcode byte. Part of the opcode is specified in this field.
MOD R/M Field 16440 (Byte 5) includes MOD field 16442, Reg field 16444, and R/M field 16446. As previously described, the MOD field's 16442 content distinguishes between memory access and non-memory access operations. The role of Reg field 16444 can be summarized to two situations: encoding either the destination register operand or a source register operand, or be treated as an opcode extension and not used to encode any instruction operand. The role of R/M field 16446 may include the following: encoding the instruction operand that references a memory address, or encoding either the destination register operand or a source register operand.
Scale, Index, Base (SIB) Byte (Byte 6)—As previously described, the scale field's 5450 content is used for memory address generation. SIB.xxx 16454 and SIB.bbb 16456— the contents of these fields have been previously referred to with regard to the register indexes Xxxx and Bbbb.
Displacement field 16362A (Bytes 7-10)— when MOD field 16442 contains 10, bytes 7-10 are the displacement field 16362A, and it works the same as the legacy 32-bit displacement (disp32) and works at byte granularity.
Displacement factor field 16362B (Byte 7)— when MOD field 16442 contains 01, byte 7 is the displacement factor field 16362B. The location of this field is that same as that of the legacy x86 instruction set 8-bit displacement (disp8), which works at byte granularity. Since disp8 is sign extended, it can only address between −128 and 127 bytes offsets; in terms of 64 byte cache lines, disp8 uses 8 bits that can be set to only four really useful values −128, −64, 0, and 64; since a greater range is often needed, disp32 is used; however, disp32 requires 4 bytes. In contrast to disp8 and disp32, the displacement factor field 16362B is a reinterpretation of disp8; when using displacement factor field 16362B, the actual displacement is determined by the content of the displacement factor field multiplied by the size of the memory operand access (N). This type of displacement is referred to as disp8*N. This reduces the average instruction length (a single byte of used for the displacement but with a much greater range). Such compressed displacement is based on the assumption that the effective displacement is multiple of the granularity of the memory access, and hence, the redundant low-order bits of the address offset do not need to be encoded. In other words, the displacement factor field 16362B substitutes the legacy x86 instruction set 8-bit displacement. Thus, the displacement factor field 16362B is encoded the same way as an x86 instruction set 8-bit displacement (so no changes in the ModRM/SIB encoding rules) with the only exception that disp8 is overloaded to disp8*N. In other words, there are no changes in the encoding rules or encoding lengths but only in the interpretation of the displacement value by hardware (which needs to scale the displacement by the size of the memory operand to obtain a byte-wise address offset). Immediate field 16372 operates as previously described.
Full Opcode Field
Register Index Field
Augmentation Operation Field
When U=1, the alpha field 16352 (EVEX byte 3, bit [7]—EH) is interpreted as the write mask control (Z) field 16352C. When U=1 and the MOD field 16442 contains 11 (signifying a no memory access operation), part of the beta field 16354 (EVEX byte 3, bit [4]-S0) is interpreted as the RL field 16357A; when it contains a 1 (round 16357A.1) the rest of the beta field 16354 (EVEX byte 3, bit [6-5]—S2-1) is interpreted as the round operation field 16359A, while when the RL field 16357A contains a 0 (VSIZE 16357.A2) the rest of the beta field 16354 (EVEX byte 3, bit [6-5]—S21) is interpreted as the vector length field 16359B (EVEX byte 3, bit [6-5]—L1-0). When U=1 and the MOD field 16442 contains 00, 01, or 10 (signifying a memory access operation), the beta field 16354 (EVEX byte 3, bits [6:4]—SSS) is interpreted as the vector length field 16359B (EVEX byte 3, bit [6-5]—L1-0) and the broadcast field 16357B (EVEX byte 3, bit [4]—B).
Exemplary Register Architecture
In other words, the vector length field 16359B selects between a maximum length and one or more other shorter lengths, where each such shorter length is half the length of the preceding length; and instructions templates without the vector length field 16359B operate on the maximum vector length. Further, in one embodiment, the class B instruction templates of the specific vector friendly instruction format 16400 operate on packed or scalar single/double-precision floating point data and packed or scalar integer data. Scalar operations are operations performed on the lowest order data element position in an zmm/ymm/xmm register; the higher order data element positions are either left the same as they were prior to the instruction or zeroed depending on the embodiment.
Write mask registers 16515—in the embodiment illustrated, there are 8 write mask registers (k0 through k7), each 64 bits in size. In an alternate embodiment, the write mask registers 16515 are 16 bits in size. As previously described, in one embodiment of the disclosure, the vector mask register k0 cannot be used as a write mask; when the encoding that would normally indicate k0 is used for a write mask, it selects a hardwired write mask of 0xFFFF, effectively disabling write masking for that instruction.
General-purpose registers 16525—in the embodiment illustrated, there are sixteen 64-bit general-purpose registers that are used along with the existing x86 addressing modes to address memory operands. These registers are referenced by the names RAX, RBX, RCX, RDX, RBP, RSI, RDI, RSP, and R8 through R15.
Scalar floating point stack register file (x87 stack) 16545, on which is aliased the MMX packed integer flat register file 16550—in the embodiment illustrated, the x87 stack is an eight-element stack used to perform scalar floating-point operations on 32/64/80-bit floating point data using the x87 instruction set extension; while the MMX registers are used to perform operations on 64-bit packed integer data, as well as to hold operands for some operations performed between the MMX and XMM registers.
Alternative embodiments of the disclosure may use wider or narrower registers. Additionally, alternative embodiments of the disclosure may use more, less, or different register files and registers.
Exemplary Core Architectures, Processors, and Computer Architectures
Processor cores may be implemented in different ways, for different purposes, and in different processors. For instance, implementations of such cores may include: 1) a general purpose in-order core intended for general-purpose computing; 2) a high performance general purpose out-of-order core intended for general-purpose computing; 3) a special purpose core intended primarily for graphics and/or scientific (throughput) computing. Implementations of different processors may include: 1) a CPU including one or more general purpose in-order cores intended for general-purpose computing and/or one or more general purpose out-of-order cores intended for general-purpose computing; and 2) a coprocessor including one or more special purpose cores intended primarily for graphics and/or scientific (throughput). Such different processors lead to different computer system architectures, which may include: 1) the coprocessor on a separate chip from the CPU; 2) the coprocessor on a separate die in the same package as a CPU; 3) the coprocessor on the same die as a CPU (in which case, such a coprocessor is sometimes referred to as special purpose logic, such as integrated graphics and/or scientific (throughput) logic, or as special purpose cores); and 4) a system on a chip that may include on the same die the described CPU (sometimes referred to as the application core(s) or application processor(s)), the above described coprocessor, and additional functionality. Exemplary core architectures are described next, followed by descriptions of exemplary processors and computer architectures.
Exemplary Core Architectures
In-Order and Out-of-Order Core Block Diagram
In
The front end unit 16630 includes a branch prediction unit 16632 coupled to an instruction cache unit 16634, which is coupled to an instruction translation lookaside buffer (TLB) 16636, which is coupled to an instruction fetch unit 16638, which is coupled to a decode unit 16640. The decode unit 16640 (or decoder or decoder unit) may decode instructions (e.g., macro-instructions), and generate as an output one or more micro-operations, micro-code entry points, micro-instructions, other instructions, or other control signals, which are decoded from, or which otherwise reflect, or are derived from, the original instructions. The decode unit 16640 may be implemented using various different mechanisms. Examples of suitable mechanisms include, but are not limited to, look-up tables, hardware implementations, programmable logic arrays (PLAs), microcode read only memories (ROMs), etc. In one embodiment, the core 16690 includes a microcode ROM or other medium that stores microcode for certain macro-instructions (e.g., in decode unit 16640 or otherwise within the front end unit 16630). The decode unit 16640 is coupled to a rename/allocator unit 16652 in the execution engine unit 16650.
The execution engine unit 16650 includes the rename/allocator unit 16652 coupled to a retirement unit 16654 and a set of one or more scheduler unit(s) 16656. The scheduler unit(s) 16656 represents any number of different schedulers, including reservations stations, central instruction window, etc. The scheduler unit(s) 16656 is coupled to the physical register file(s) unit(s) 16658. Each of the physical register file(s) units 16658 represents one or more physical register files, different ones of which store one or more different data types, such as scalar integer, scalar floating point, packed integer, packed floating point, vector integer, vector floating point, status (e.g., an instruction pointer that is the address of the next instruction to be executed), etc. In one embodiment, the physical register file(s) unit 16658 comprises a vector registers unit, a write mask registers unit, and a scalar registers unit. These register units may provide architectural vector registers, vector mask registers, and general purpose registers. The physical register file(s) unit(s) 16658 is overlapped by the retirement unit 16654 to illustrate various ways in which register renaming and out-of-order execution may be implemented (e.g., using a reorder buffer(s) and a retirement register file(s); using a future file(s), a history buffer(s), and a retirement register file(s); using a register maps and a pool of registers; etc.). The retirement unit 16654 and the physical register file(s) unit(s) 16658 are coupled to the execution cluster(s) 16660. The execution cluster(s) 16660 includes a set of one or more execution units 16662 and a set of one or more memory access units 16664. The execution units 16662 may perform various operations (e.g., shifts, addition, subtraction, multiplication) and on various types of data (e.g., scalar floating point, packed integer, packed floating point, vector integer, vector floating point). While 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 multiple execution units that all perform all functions. The scheduler unit(s) 16656, physical register file(s) unit(s) 16658, and execution cluster(s) 16660 are shown as being possibly plural because certain embodiments create separate pipelines for certain types of data/operations (e.g., a scalar integer pipeline, a scalar floating point/packed integer/packed floating point/vector integer/vector floating point pipeline, and/or a memory access pipeline that each have their own scheduler unit, physical register file(s) unit, and/or execution cluster—and in the case of a separate memory access pipeline, certain embodiments are implemented in which only the execution cluster of this pipeline has the memory access unit(s) 16664). It should also be understood that where separate pipelines are used, one or more of these pipelines may be out-of-order issue/execution and the rest in-order.
The set of memory access units 16664 is coupled to the memory unit 16670, which includes a data TLB unit 16672 coupled to a data cache unit 16674 coupled to a level 2 (L2) cache unit 16676. In one exemplary embodiment, the memory access units 16664 may include a load unit, a store address unit, and a store data unit, each of which is coupled to the data TLB unit 16672 in the memory unit 16670. The instruction cache unit 16634 is further coupled to a level 2 (L2) cache unit 16676 in the memory unit 16670. The L2 cache unit 16676 is coupled to one or more other levels of cache and eventually to a main memory.
By way of example, the exemplary register renaming, out-of-order issue/execution core architecture may implement the pipeline 16600 as follows: 1) the instruction fetch 16638 performs the fetch and length decoding stages 16602 and 16604; 2) the decode unit 16640 performs the decode stage 16606; 3) the rename/allocator unit 16652 performs the allocation stage 16608 and renaming stage 16610; 4) the scheduler unit(s) 16656 performs the schedule stage 16612; 5) the physical register file(s) unit(s) 16658 and the memory unit 16670 perform the register read/memory read stage 16614; the execution cluster 16660 perform the execute stage 16616; 6) the memory unit 16670 and the physical register file(s) unit(s) 16658 perform the write back/memory write stage 16618; 7) various units may be involved in the exception handling stage 16622; and 8) the retirement unit 16654 and the physical register file(s) unit(s) 16658 perform the commit stage 16624.
The core 16690 may support one or more instructions sets (e.g., the x86 instruction set (with some extensions that have been added with newer versions); the MIPS instruction set of MIPS Technologies of Sunnyvale, Calif.; the ARM instruction set (with optional additional extensions such as NEON) of ARM Holdings of Sunnyvale, Calif.), including the instruction(s) described herein. In one embodiment, the core 16690 includes logic to support a packed data instruction set extension (e.g., AVX1, AVX2), thereby allowing the operations used by many multimedia applications to be performed using packed data.
It should be understood that the core may support multithreading (executing two or more parallel sets of operations or threads), and may do so in a variety of ways including time sliced multithreading, simultaneous multithreading (where a single physical core provides a logical core for each of the threads that physical core is simultaneously multithreading), or a combination thereof (e.g., time sliced fetching and decoding and simultaneous multithreading thereafter such as in the Intel® Hyperthreading technology).
While register renaming is described in the context of out-of-order execution, it should be understood that register renaming may be used in an in-order architecture. While the illustrated embodiment of the processor also includes separate instruction and data cache units 16634/16674 and a shared L2 cache unit 16676, alternative embodiments may have a single internal cache for both instructions and data, such as, for example, a Level 1 (L1) internal cache, or multiple levels of internal cache. In some embodiments, the system may include a combination of an internal cache and an external cache that is external to the core and/or the processor. Alternatively, all of the cache may be external to the core and/or the processor.
Specific Exemplary In-Order Core Architecture
The local subset of the L2 cache 16704 is part of a global L2 cache that is divided into separate local subsets, one per processor core. Each processor core has a direct access path to its own local subset of the L2 cache 16704. Data read by a processor core is stored in its L2 cache subset 16704 and can be accessed quickly, in parallel with other processor cores accessing their own local L2 cache subsets. Data written by a processor core is stored in its own L2 cache subset 16704 and is flushed from other subsets, if necessary. The ring network ensures coherency for shared data. The ring network is bi-directional to allow agents such as processor cores, hf caches and other logic blocks to communicate with each other within the chip. Each ring data-path is 1012-bits wide per direction.
Thus, different implementations of the processor 16800 may include: 1) a CPU with the special purpose logic 16808 being integrated graphics and/or scientific (throughput) logic (which may include one or more cores), and the cores 16802A-N being one or more general purpose cores (e.g., general purpose in-order cores, general purpose out-of-order cores, a combination of the two); 2) a coprocessor with the cores 16802A-N being a large number of special purpose cores intended primarily for graphics and/or scientific (throughput); and 3) a coprocessor with the cores 16802A-N being a large number of general purpose in-order cores. Thus, the processor 16800 may be a general-purpose processor, coprocessor or special-purpose processor, such as, for example, a network or communication processor, compression engine, graphics processor, GPGPU (general purpose graphics processing unit), a high-throughput many integrated core (MIC) coprocessor (including 30 or more cores), embedded processor, or the like. The processor may be implemented on one or more chips. The processor 16800 may be a part of and/or may be implemented on one or more substrates using any of a number of process technologies, such as, for example, BiCMOS, CMOS, or NMOS.
The memory hierarchy includes one or more levels of cache within the cores, a set or one or more shared cache units 16806, and external memory (not shown) coupled to the set of integrated memory controller units 16814. The set of shared cache units 16806 may include one or more mid-level caches, such as level 2 (L2), level 3 (L3), level 4 (L4), or other levels of cache, a last level cache (LLC), and/or combinations thereof. While in one embodiment a ring based interconnect unit 16812 interconnects the integrated graphics logic 16808, the set of shared cache units 16806, and the system agent unit 16810/integrated memory controller unit(s) 16814, alternative embodiments may use any number of well-known techniques for interconnecting such units. In one embodiment, coherency is maintained between one or more cache units 16806 and cores 16802-A-N.
In some embodiments, one or more of the cores 16802A-N are capable of multi-threading. The system agent 16810 includes those components coordinating and operating cores 16802A-N. The system agent unit 16810 may include for example a power control unit (PCU) and a display unit. The PCU may be or include logic and components needed for regulating the power state of the cores 16802A-N and the integrated graphics logic 16808. The display unit is for driving one or more externally connected displays.
The cores 16802A-N may be homogenous or heterogeneous in terms of architecture instruction set; that is, two or more of the cores 16802A-N may be capable of execution the same instruction set, while others may be capable of executing only a subset of that instruction set or a different instruction set.
Exemplary Computer Architectures
Referring now to
The optional nature of additional processors 16915 is denoted in
The memory 16940 may be, for example, dynamic random access memory (DRAM), phase change memory (PCM), or a combination of the two. For at least one embodiment, the controller hub 16920 communicates with the processor(s) 16910, 16915 via a multi-drop bus, such as a frontside bus (FSB), point-to-point interface such as QuickPath Interconnect (QPI), or similar connection 16995.
In one embodiment, the coprocessor 16945 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, or the like. In one embodiment, controller hub 16920 may include an integrated graphics accelerator.
There can be a variety of differences between the physical resources 16910, 16915 in terms of a spectrum of metrics of merit including architectural, microarchitectural, thermal, power consumption characteristics, and the like.
In one embodiment, the processor 16910 executes instructions that control data processing operations of a general type. Embedded within the instructions may be coprocessor instructions. The processor 16910 recognizes these coprocessor instructions as being of a type that should be executed by the attached coprocessor 16945. Accordingly, the processor 16910 issues these coprocessor instructions (or control signals representing coprocessor instructions) on a coprocessor bus or other interconnect, to coprocessor 16945. Coprocessor(s) 16945 accept and execute the received coprocessor instructions.
Referring now to
Processors 17070 and 17080 are shown including integrated memory controller (IMC) units 17072 and 17082, respectively. Processor 17070 also includes as part of its bus controller units point-to-point (P-P) interfaces 17076 and 17078; similarly, second processor 17080 includes P-P interfaces 17086 and 17088. Processors 17070, 17080 may exchange information via a point-to-point (P-P) interface 17050 using P-P interface circuits 17078, 17088. As shown in
Processors 17070, 17080 may each exchange information with a chipset 17090 via individual P-P interfaces 17052, 17054 using point to point interface circuits 17076, 17094, 17086, 17098. Chipset 17090 may optionally exchange information with the coprocessor 17038 via a high-performance interface 17039. In one embodiment, the coprocessor 17038 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, 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 17090 may be coupled to a first bus 17016 via an interface 17096. In one embodiment, first bus 17016 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 the present disclosure is not so limited.
As shown in
Referring now to
Referring now to
Embodiments (e.g., of the mechanisms) disclosed herein may be implemented in hardware, software, firmware, or a combination of such implementation approaches. Embodiments of the disclosure 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 17030 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 the 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.
Emulation (Including Binary Translation, Code Morphing, Etc.)
In some cases, an instruction converter may be used to convert an instruction from a source instruction set to a target instruction set. For example, the instruction converter may translate (e.g., using static binary translation, dynamic binary translation including dynamic compilation), morph, emulate, or otherwise convert an instruction to one or more other instructions to be processed by the core. The instruction converter may be implemented in software, hardware, firmware, or a combination thereof. The instruction converter may be on processor, off processor, or part on and part off processor.
The present application is a continuation application of U.S. patent application Ser. No. 16/024,854 filed Jun. 30, 2018, now U.S. Pat. No. 11,200,186, which is incorporated herein by reference.
This invention was made with Government support under contract number H98230-13-D-0124 awarded by the Department of Defense. The Government has certain rights in this invention.
| Number | Name | Date | Kind |
|---|---|---|---|
| 672177 | Metcalf | Apr 1901 | A |
| 4314349 | Batcher | Feb 1982 | A |
| 4965716 | Sweeney | Oct 1990 | A |
| 5093920 | Agrawal et al. | Mar 1992 | A |
| 5141747 | Scholz | Aug 1992 | A |
| 5239634 | Buch | Aug 1993 | A |
| 5241635 | Papadopoulos et al. | Aug 1993 | A |
| 5314503 | Bobrove et al. | May 1994 | A |
| 5393454 | Mondin et al. | Feb 1995 | A |
| 5393536 | Brandt et al. | Feb 1995 | A |
| 5410722 | Cornaby | Apr 1995 | A |
| 5465368 | Davidson et al. | Nov 1995 | A |
| 5502661 | Glunz | Mar 1996 | A |
| 5560032 | Nguyen et al. | Sep 1996 | A |
| 5574944 | Stager | Nov 1996 | A |
| 5581767 | Katsuki et al. | Dec 1996 | A |
| 5604120 | Yano et al. | Feb 1997 | A |
| 5625630 | Abramovici et al. | Apr 1997 | A |
| 5655096 | Branigin | Aug 1997 | A |
| 5689719 | Miura et al. | Nov 1997 | A |
| 5725364 | Mirazita | Mar 1998 | A |
| 5787029 | de Angel | Jul 1998 | A |
| 5790821 | Pflum | Aug 1998 | A |
| 5805827 | Chau et al. | Sep 1998 | A |
| 5925099 | Futral et al. | Jul 1999 | A |
| 5930484 | Tran et al. | Jul 1999 | A |
| 5933429 | Bubenik et al. | Aug 1999 | A |
| 5948081 | Foster | Sep 1999 | A |
| 6020139 | Schwartz et al. | Feb 2000 | A |
| 6088780 | Yamada et al. | Jul 2000 | A |
| 6141747 | Witt | Oct 2000 | A |
| 6205533 | Margolus | Mar 2001 | B1 |
| 6247064 | Alferness et al. | Jun 2001 | B1 |
| 6314503 | D'Errico et al. | Nov 2001 | B1 |
| 6393454 | Chu | May 2002 | B1 |
| 6393536 | Hughes et al. | May 2002 | B1 |
| 6460131 | Trimberger | Oct 2002 | B1 |
| 6507947 | Schreiber et al. | Jan 2003 | B1 |
| 6515333 | Riccobene | Feb 2003 | B1 |
| 6553448 | Mannion | Apr 2003 | B1 |
| 6553482 | Witt | Apr 2003 | B1 |
| 6604120 | De Angel | Aug 2003 | B1 |
| 6615333 | Hoogerbrugge et al. | Sep 2003 | B1 |
| 6640267 | Raza | Oct 2003 | B1 |
| 6721774 | Lee et al. | Apr 2004 | B1 |
| 6725364 | Crabill | Apr 2004 | B1 |
| 6728945 | Wang | Apr 2004 | B1 |
| 6886085 | Shuf et al. | Apr 2005 | B1 |
| 6947416 | Nee et al. | Sep 2005 | B1 |
| 7000072 | Aisaka et al. | Feb 2006 | B1 |
| 7058785 | Ochotta | Jun 2006 | B1 |
| 7181578 | Guha et al. | Feb 2007 | B1 |
| 7200137 | Dorr et al. | Apr 2007 | B2 |
| 7203936 | Gillies et al. | Apr 2007 | B2 |
| 7257665 | Niell et al. | Aug 2007 | B2 |
| 7290096 | Jeter et al. | Oct 2007 | B2 |
| 7308514 | Hewitt | Dec 2007 | B1 |
| 7379067 | Deering et al. | May 2008 | B2 |
| 7380108 | Uht et al. | May 2008 | B1 |
| 7486678 | Devanagondi et al. | Feb 2009 | B1 |
| 7509484 | Golla et al. | Mar 2009 | B1 |
| 7546331 | Islam | Jun 2009 | B2 |
| 7630324 | Li et al. | Dec 2009 | B2 |
| 7647435 | Check et al. | Jan 2010 | B2 |
| 7660911 | McDaniel | Feb 2010 | B2 |
| 7724740 | Wang et al. | May 2010 | B1 |
| 7817652 | Macadam et al. | Oct 2010 | B1 |
| 7911960 | Aydemir et al. | Mar 2011 | B1 |
| 7936753 | Colloff et al. | May 2011 | B1 |
| 7987479 | Day | Jul 2011 | B1 |
| 8001510 | Miller et al. | Aug 2011 | B1 |
| 8010766 | Bhattacharjee et al. | Aug 2011 | B2 |
| 8051227 | Gewirtz et al. | Nov 2011 | B1 |
| 8055880 | Fujisawa et al. | Nov 2011 | B2 |
| 8115659 | Hurley et al. | Feb 2012 | B2 |
| 8156284 | Vorbach et al. | Apr 2012 | B2 |
| 8160975 | Tang et al. | Apr 2012 | B2 |
| 8225073 | Master et al. | Jul 2012 | B2 |
| 8332597 | Bennett | Dec 2012 | B1 |
| 8356162 | Muff et al. | Jan 2013 | B2 |
| 8495341 | Busaba et al. | Jul 2013 | B2 |
| 8561194 | Lee | Oct 2013 | B2 |
| 8578117 | Burda et al. | Nov 2013 | B2 |
| 8619800 | Finney et al. | Dec 2013 | B1 |
| 8806248 | Allarey et al. | Aug 2014 | B2 |
| 8812820 | Vorbach et al. | Aug 2014 | B2 |
| 8924596 | Beeson et al. | Dec 2014 | B1 |
| 8935515 | Colavin et al. | Jan 2015 | B2 |
| 8966457 | Ebcioglu et al. | Feb 2015 | B2 |
| 8990452 | Branson et al. | Mar 2015 | B2 |
| 8996906 | Townley et al. | Mar 2015 | B1 |
| 9009712 | Schmit et al. | Apr 2015 | B2 |
| 9026705 | Feehrer et al. | May 2015 | B2 |
| 9026769 | Jamil et al. | May 2015 | B1 |
| 9052890 | Phillips et al. | Jun 2015 | B2 |
| 9104474 | Kaul et al. | Aug 2015 | B2 |
| 9110846 | Buchheit et al. | Aug 2015 | B2 |
| 9135051 | Chaplik et al. | Sep 2015 | B2 |
| 9135057 | Branson et al. | Sep 2015 | B2 |
| 9170846 | Delling et al. | Oct 2015 | B2 |
| 9213571 | Ristovski et al. | Dec 2015 | B2 |
| 9268528 | Tannenbaum et al. | Feb 2016 | B2 |
| 9268704 | Fleischer et al. | Feb 2016 | B2 |
| 9285860 | Hofmann | Mar 2016 | B2 |
| 9429983 | Chall et al. | Aug 2016 | B1 |
| 9473144 | Thiagarajan et al. | Oct 2016 | B1 |
| 9520876 | Chadwick et al. | Dec 2016 | B1 |
| 9594521 | Blagodurov et al. | Mar 2017 | B2 |
| 9658676 | Witek et al. | May 2017 | B1 |
| 9696928 | Cain et al. | Jul 2017 | B2 |
| 9753658 | Beeson et al. | Sep 2017 | B2 |
| 9760291 | Beale et al. | Sep 2017 | B2 |
| 9762563 | Davis et al. | Sep 2017 | B2 |
| 9847783 | Teh et al. | Dec 2017 | B1 |
| 9886072 | Venkataraman | Feb 2018 | B1 |
| 9923905 | Amiri et al. | Mar 2018 | B2 |
| 9946718 | Bowman et al. | Apr 2018 | B2 |
| 10108417 | Krishna et al. | Oct 2018 | B2 |
| 10187467 | Nagai | Jan 2019 | B2 |
| 10331583 | Ahsan et al. | Jun 2019 | B2 |
| 10467183 | Fleming et al. | Nov 2019 | B2 |
| 20010001325 | Fujimoto et al. | May 2001 | A1 |
| 20020026493 | Scardamalia et al. | Feb 2002 | A1 |
| 20020090751 | Srigg et al. | Jul 2002 | A1 |
| 20020103943 | Lo et al. | Aug 2002 | A1 |
| 20020178285 | Donaldson et al. | Nov 2002 | A1 |
| 20020184291 | Hogenauer | Dec 2002 | A1 |
| 20030023830 | Hogenauer | Jan 2003 | A1 |
| 20030028750 | Hogenauer | Feb 2003 | A1 |
| 20030120802 | Kohno | Jun 2003 | A1 |
| 20030126233 | Bryers et al. | Jul 2003 | A1 |
| 20030163649 | Kapur et al. | Aug 2003 | A1 |
| 20030177320 | Sah et al. | Sep 2003 | A1 |
| 20030212726 | Luick | Nov 2003 | A1 |
| 20030225814 | Saito et al. | Dec 2003 | A1 |
| 20030233643 | Thompson et al. | Dec 2003 | A1 |
| 20040001458 | Dorenbosch et al. | Jan 2004 | A1 |
| 20040017807 | Dorr | Jan 2004 | A1 |
| 20040017820 | Garinger et al. | Jan 2004 | A1 |
| 20040022094 | Radhakrishnan et al. | Feb 2004 | A1 |
| 20040022107 | Zaidi et al. | Feb 2004 | A1 |
| 20040030859 | Doerr et al. | Feb 2004 | A1 |
| 20040044877 | Myers | Mar 2004 | A1 |
| 20040054818 | Kirsch | Mar 2004 | A1 |
| 20040103230 | Emerson | May 2004 | A1 |
| 20040124877 | Parkes | Jul 2004 | A1 |
| 20040128401 | Fallon et al. | Jul 2004 | A1 |
| 20040230849 | Dhong et al. | Nov 2004 | A1 |
| 20040263524 | Lippincott | Dec 2004 | A1 |
| 20050025120 | O'Toole et al. | Feb 2005 | A1 |
| 20050076187 | Claydon | Apr 2005 | A1 |
| 20050108776 | Carver et al. | May 2005 | A1 |
| 20050134308 | Okada et al. | Jun 2005 | A1 |
| 20050138323 | Snyder | Jun 2005 | A1 |
| 20050166038 | Wang et al. | Jul 2005 | A1 |
| 20050172103 | Inuo et al. | Aug 2005 | A1 |
| 20050223131 | Goekjian et al. | Oct 2005 | A1 |
| 20060026359 | Ross et al. | Feb 2006 | A1 |
| 20060041872 | Poznanovic et al. | Feb 2006 | A1 |
| 20060101237 | Mohl et al. | May 2006 | A1 |
| 20060130030 | Kwiat et al. | Jun 2006 | A1 |
| 20060179255 | Yamazaki | Aug 2006 | A1 |
| 20060179429 | Eggers et al. | Aug 2006 | A1 |
| 20060200647 | Cohen | Sep 2006 | A1 |
| 20060221931 | Apostol et al. | Oct 2006 | A1 |
| 20060236008 | Asano et al. | Oct 2006 | A1 |
| 20070011436 | Bittner, Jr. | Jan 2007 | A1 |
| 20070033369 | Kasama et al. | Feb 2007 | A1 |
| 20070074014 | Musoll et al. | Mar 2007 | A1 |
| 20070079036 | Mukherjee | Apr 2007 | A1 |
| 20070118332 | Meyers et al. | May 2007 | A1 |
| 20070143546 | Narad | Jun 2007 | A1 |
| 20070180315 | Aizawa | Aug 2007 | A1 |
| 20070203967 | Dockser | Aug 2007 | A1 |
| 20070204137 | Tran | Aug 2007 | A1 |
| 20070226458 | Stuttard et al. | Sep 2007 | A1 |
| 20070256038 | Manohar | Nov 2007 | A1 |
| 20070266223 | Nguyen | Nov 2007 | A1 |
| 20070276976 | Gower et al. | Nov 2007 | A1 |
| 20070299980 | Amini et al. | Dec 2007 | A1 |
| 20080005392 | Amini et al. | Jan 2008 | A1 |
| 20080072113 | Tsang et al. | Mar 2008 | A1 |
| 20080082794 | Yu et al. | Apr 2008 | A1 |
| 20080133889 | Glew | Jun 2008 | A1 |
| 20080133895 | Sivtsov et al. | Jun 2008 | A1 |
| 20080155135 | Garg et al. | Jun 2008 | A1 |
| 20080184255 | Watanabe et al. | Jul 2008 | A1 |
| 20080218203 | Arriens et al. | Sep 2008 | A1 |
| 20080263330 | May et al. | Oct 2008 | A1 |
| 20080270689 | Gotoh | Oct 2008 | A1 |
| 20080307258 | Challenger et al. | Dec 2008 | A1 |
| 20090013329 | May et al. | Jan 2009 | A1 |
| 20090037697 | Ramani et al. | Feb 2009 | A1 |
| 20090055624 | Kirsch | Feb 2009 | A1 |
| 20090063665 | Bagepalli et al. | Mar 2009 | A1 |
| 20090113169 | Yang et al. | Apr 2009 | A1 |
| 20090119456 | Park et al. | May 2009 | A1 |
| 20090133023 | Li et al. | May 2009 | A1 |
| 20090144522 | Vorbach et al. | Jun 2009 | A1 |
| 20090175444 | Douglis et al. | Jul 2009 | A1 |
| 20090182993 | Fant | Jul 2009 | A1 |
| 20090222252 | Chen et al. | Sep 2009 | A1 |
| 20090228682 | Mejdrich et al. | Sep 2009 | A1 |
| 20090300324 | Inuo | Dec 2009 | A1 |
| 20090300325 | Paver et al. | Dec 2009 | A1 |
| 20090300388 | Mantor et al. | Dec 2009 | A1 |
| 20090309884 | Lippincott et al. | Dec 2009 | A1 |
| 20090328048 | Khan et al. | Dec 2009 | A1 |
| 20100017761 | Higuchi et al. | Jan 2010 | A1 |
| 20100115168 | Bekooij | May 2010 | A1 |
| 20100180105 | Asnaashari | Jul 2010 | A1 |
| 20100191911 | Heddes | Jul 2010 | A1 |
| 20100217915 | O'Connor et al. | Aug 2010 | A1 |
| 20100228885 | McDaniel | Sep 2010 | A1 |
| 20100228925 | Doerr et al. | Sep 2010 | A1 |
| 20100235416 | Oliver | Sep 2010 | A1 |
| 20100254262 | Kantawala et al. | Oct 2010 | A1 |
| 20100262721 | Asnaashari et al. | Oct 2010 | A1 |
| 20100302946 | Yang et al. | Dec 2010 | A1 |
| 20110004742 | Hassan | Jan 2011 | A1 |
| 20110008300 | Wouters et al. | Jan 2011 | A1 |
| 20110040822 | Eichenberger et al. | Feb 2011 | A1 |
| 20110083000 | Rhoades et al. | Apr 2011 | A1 |
| 20110099295 | Wegener | Apr 2011 | A1 |
| 20110107337 | Cambonie | May 2011 | A1 |
| 20110133825 | Jones et al. | Jun 2011 | A1 |
| 20110202747 | Busaba et al. | Aug 2011 | A1 |
| 20110216861 | Melzer et al. | Sep 2011 | A1 |
| 20110292708 | Kang et al. | Dec 2011 | A1 |
| 20110302358 | Yu et al. | Dec 2011 | A1 |
| 20110314238 | Finkler et al. | Dec 2011 | A1 |
| 20110320724 | Mejdrich et al. | Dec 2011 | A1 |
| 20120017066 | Vorbach et al. | Jan 2012 | A1 |
| 20120054255 | Buxbaum et al. | Mar 2012 | A1 |
| 20120066483 | Boury et al. | Mar 2012 | A1 |
| 20120079168 | Chou et al. | Mar 2012 | A1 |
| 20120089812 | Smith | Apr 2012 | A1 |
| 20120124117 | Yu et al. | May 2012 | A1 |
| 20120126850 | Wasson et al. | May 2012 | A1 |
| 20120126851 | Kelem et al. | May 2012 | A1 |
| 20120128107 | Oren | May 2012 | A1 |
| 20120131288 | Box et al. | May 2012 | A1 |
| 20120144126 | Nimmala et al. | Jun 2012 | A1 |
| 20120174118 | Watanabe et al. | Jul 2012 | A1 |
| 20120239853 | Moshayedi | Sep 2012 | A1 |
| 20120260239 | Martinez et al. | Oct 2012 | A1 |
| 20120278543 | Yu et al. | Nov 2012 | A1 |
| 20120278587 | Caufield et al. | Nov 2012 | A1 |
| 20120303932 | Farabet et al. | Nov 2012 | A1 |
| 20120303933 | Manet et al. | Nov 2012 | A1 |
| 20120317388 | Driever et al. | Dec 2012 | A1 |
| 20120324180 | Asnaashari et al. | Dec 2012 | A1 |
| 20120330701 | Hyder et al. | Dec 2012 | A1 |
| 20130021058 | Huang et al. | Jan 2013 | A1 |
| 20130024875 | Wang et al. | Jan 2013 | A1 |
| 20130036287 | Chu et al. | Feb 2013 | A1 |
| 20130067138 | Schuette et al. | Mar 2013 | A1 |
| 20130067203 | Chung et al. | Mar 2013 | A1 |
| 20130070588 | Steele et al. | Mar 2013 | A1 |
| 20130080652 | Cradick et al. | Mar 2013 | A1 |
| 20130080993 | Stravers et al. | Mar 2013 | A1 |
| 20130081042 | Branson et al. | Mar 2013 | A1 |
| 20130125127 | Mital et al. | May 2013 | A1 |
| 20130145203 | Fawcett | Jun 2013 | A1 |
| 20130147515 | Wasson et al. | Jun 2013 | A1 |
| 20130151919 | Huynh | Jun 2013 | A1 |
| 20130160028 | Black | Jun 2013 | A1 |
| 20130166879 | Sun et al. | Jun 2013 | A1 |
| 20130315211 | Balan et al. | Nov 2013 | A1 |
| 20140013083 | Jha | Jan 2014 | A1 |
| 20140032860 | Yamada et al. | Jan 2014 | A1 |
| 20140052955 | Moll et al. | Feb 2014 | A1 |
| 20140095891 | Wolrich et al. | Apr 2014 | A1 |
| 20140098890 | Sermadevi et al. | Apr 2014 | A1 |
| 20140115300 | Bodine | Apr 2014 | A1 |
| 20140143564 | Tannenbaum et al. | May 2014 | A1 |
| 20140188968 | Kaul et al. | Jul 2014 | A1 |
| 20140215189 | Airaud et al. | Jul 2014 | A1 |
| 20140281409 | Abdallah et al. | Sep 2014 | A1 |
| 20140351551 | Doerr et al. | Nov 2014 | A1 |
| 20140380024 | Spadini et al. | Dec 2014 | A1 |
| 20150007182 | Rossbach et al. | Jan 2015 | A1 |
| 20150026434 | Basant et al. | Jan 2015 | A1 |
| 20150033001 | Ivanov | Jan 2015 | A1 |
| 20150067305 | Olson et al. | Mar 2015 | A1 |
| 20150067368 | Henry et al. | Mar 2015 | A1 |
| 20150082011 | Mellinger et al. | Mar 2015 | A1 |
| 20150082075 | Denman et al. | Mar 2015 | A1 |
| 20150089162 | Ahsan et al. | Mar 2015 | A1 |
| 20150089186 | Kim et al. | Mar 2015 | A1 |
| 20150100757 | Burger et al. | Apr 2015 | A1 |
| 20150106596 | Morbach et al. | Apr 2015 | A1 |
| 20150113184 | Stanford-Jason et al. | Apr 2015 | A1 |
| 20150188847 | Chopra et al. | Jul 2015 | A1 |
| 20150199173 | Lutz et al. | Jul 2015 | A1 |
| 20150220345 | Corbal et al. | Aug 2015 | A1 |
| 20150261528 | Ho et al. | Sep 2015 | A1 |
| 20150268963 | Etsion et al. | Sep 2015 | A1 |
| 20150278097 | Kelm et al. | Oct 2015 | A1 |
| 20150317134 | Kim et al. | Nov 2015 | A1 |
| 20150378731 | Lai et al. | Dec 2015 | A1 |
| 20150379670 | Koker et al. | Dec 2015 | A1 |
| 20160077568 | Kandula et al. | Mar 2016 | A1 |
| 20160098279 | Glew | Apr 2016 | A1 |
| 20160098420 | Dickie et al. | Apr 2016 | A1 |
| 20160239265 | Duong et al. | Aug 2016 | A1 |
| 20160364364 | Fleischer | Dec 2016 | A1 |
| 20160381431 | Patterson et al. | Dec 2016 | A1 |
| 20170013495 | Chae et al. | Jan 2017 | A1 |
| 20170031866 | Nowatzki et al. | Feb 2017 | A1 |
| 20170062075 | Barber et al. | Mar 2017 | A1 |
| 20170083313 | Sankaralingam et al. | Mar 2017 | A1 |
| 20170092371 | Harari | Mar 2017 | A1 |
| 20170116154 | Palmer et al. | Apr 2017 | A1 |
| 20170163543 | Wang et al. | Jun 2017 | A1 |
| 20170185655 | Jahankhani et al. | Jun 2017 | A1 |
| 20170242700 | Hughes et al. | Aug 2017 | A1 |
| 20170255414 | Gerhart et al. | Sep 2017 | A1 |
| 20170262383 | Lee et al. | Sep 2017 | A1 |
| 20170286169 | Ravindran et al. | Oct 2017 | A1 |
| 20170293766 | Schnjakin et al. | Oct 2017 | A1 |
| 20170295149 | Lu et al. | Oct 2017 | A1 |
| 20170315815 | Smith et al. | Nov 2017 | A1 |
| 20170315978 | Boucher et al. | Nov 2017 | A1 |
| 20170325124 | Mitra et al. | Nov 2017 | A1 |
| 20170371836 | Langhammer | Dec 2017 | A1 |
| 20180081806 | Kothinti et al. | Mar 2018 | A1 |
| 20180081834 | Wang et al. | Mar 2018 | A1 |
| 20180088647 | Suryanarayanan et al. | Mar 2018 | A1 |
| 20180095728 | Hasenplaugh et al. | Apr 2018 | A1 |
| 20180101502 | Nassif et al. | Apr 2018 | A1 |
| 20180113797 | Breslow et al. | Apr 2018 | A1 |
| 20180188983 | Fleming et al. | Jul 2018 | A1 |
| 20180188997 | Fleming et al. | Jul 2018 | A1 |
| 20180189063 | Fleming et al. | Jul 2018 | A1 |
| 20180189231 | Fleming et al. | Jul 2018 | A1 |
| 20180189239 | Nurvitadhi et al. | Jul 2018 | A1 |
| 20180189675 | Nurvitadhi et al. | Jul 2018 | A1 |
| 20180210730 | Sankaralingam et al. | Jul 2018 | A1 |
| 20180218767 | Wolff | Aug 2018 | A1 |
| 20180225403 | Nicol et al. | Aug 2018 | A1 |
| 20180248994 | Lee et al. | Aug 2018 | A1 |
| 20180285385 | West et al. | Oct 2018 | A1 |
| 20180293162 | Tsai et al. | Oct 2018 | A1 |
| 20180300181 | Hetzel et al. | Oct 2018 | A1 |
| 20180316760 | Chernin et al. | Nov 2018 | A1 |
| 20180332342 | Wu et al. | Nov 2018 | A1 |
| 20180365181 | Cottam et al. | Dec 2018 | A1 |
| 20180373509 | Zhang et al. | Dec 2018 | A1 |
| 20190004878 | Adler et al. | Jan 2019 | A1 |
| 20190004945 | Fleming et al. | Jan 2019 | A1 |
| 20190004955 | Adler et al. | Jan 2019 | A1 |
| 20190004994 | Fleming et al. | Jan 2019 | A1 |
| 20190005161 | Fleming et al. | Jan 2019 | A1 |
| 20190007332 | Fleming et al. | Jan 2019 | A1 |
| 20190018815 | Fleming et al. | Jan 2019 | A1 |
| 20190026237 | Bannon et al. | Jan 2019 | A1 |
| 20190042217 | Glossop et al. | Feb 2019 | A1 |
| 20190042218 | Zhang | Feb 2019 | A1 |
| 20190042513 | Fleming et al. | Feb 2019 | A1 |
| 20190089354 | Venugopal et al. | Mar 2019 | A1 |
| 20190095369 | Fleming et al. | Mar 2019 | A1 |
| 20190095383 | Fleming et al. | Mar 2019 | A1 |
| 20190101952 | Diamond et al. | Apr 2019 | A1 |
| 20190102179 | Fleming et al. | Apr 2019 | A1 |
| 20190102338 | Tang et al. | Apr 2019 | A1 |
| 20190129720 | Ivanov | May 2019 | A1 |
| 20190205263 | Fleming et al. | Jul 2019 | A1 |
| 20190205269 | Fleming et al. | Jul 2019 | A1 |
| 20190205284 | Fleming et al. | Jul 2019 | A1 |
| 20190258919 | Lie et al. | Aug 2019 | A1 |
| 20190286987 | Lie et al. | Sep 2019 | A1 |
| 20190303153 | Halpern et al. | Oct 2019 | A1 |
| 20190303168 | Fleming et al. | Oct 2019 | A1 |
| 20190303263 | Fleming et al. | Oct 2019 | A1 |
| 20190303297 | Fleming et al. | Oct 2019 | A1 |
| 20190303312 | Ahsan et al. | Oct 2019 | A1 |
| 20190377688 | Basu et al. | Dec 2019 | A1 |
| Number | Date | Country |
|---|---|---|
| 101174200 | May 2008 | CN |
| 102103479 | Jun 2011 | CN |
| 102203752 | Sep 2011 | CN |
| 102349049 | Feb 2012 | CN |
| 103765400 | Apr 2014 | CN |
| 104468302 | Mar 2015 | CN |
| 104539430 | Apr 2015 | CN |
| 104641346 | May 2015 | CN |
| 104778028 | Jul 2015 | CN |
| 104937539 | Sep 2015 | CN |
| 102018005181 | Jan 2019 | DE |
| 0410777 | Jan 1991 | EP |
| 2317446 | May 2011 | EP |
| 2660716 | Nov 2013 | EP |
| 2854026 | Apr 2015 | EP |
| 3005139 | Apr 2016 | EP |
| 2374684 | Nov 2009 | RU |
| 8901203 | Feb 1989 | WO |
| 2003100602 | Dec 2003 | WO |
| 2007031696 | Mar 2007 | WO |
| 2012138957 | Oct 2012 | WO |
| 2014035449 | Mar 2014 | WO |
| 2014193878 | Dec 2014 | WO |
| 2015044696 | Apr 2015 | WO |
| 2015196343 | Dec 2015 | WO |
| Entry |
|---|
| Abandonment from U.S. Appl. No. 15/640,544, dated Mar. 20, 2018, 2 pages. |
| Advisory Action from U.S. Appl. No. 14/037,468, dated Aug. 11, 2017, 3 pages. |
| Amdahl, Gene M., “Validity of the Single Processor Approach to Achieving Large Scale Computing Capabilities”, AFIPS Spring Joint Computer Conference, 1967, pp. 1-4. |
| Applicant Initiated Interview Summary, U.S. Appl. No. 15/721,816, dated Mar. 10, 2020, 3 pages. |
| Arvind., et al., “Executing a Program on the MIT Tagged-Token Dataflow Architecture”, Mar. 1990, IEEE Transactions on Computers, vol. 39 (3), pp. 300-318. |
| Asanovic K, et al., “The Landscape of Parallel Computing Research: A View from Berkeley,” Dec. 18, 2006, Electrical Engineering and Computer Sciences University of California at Berkeley, Technical Report No. UCB/EECS-2006-183, http://www.eecs.berkeley.edu/Pubs/TechRpts/2006/EECS-2006-183.html, 56 pages. |
| Ball T., “What's in a Region? or Computing Control Dependence Regions in Near-Linear Time for Reducible Control Flow,” Dec. 1993, ACM Letters on Programming Languages and Systems, 2(1-4):1-16, 24 pages. |
| Benchmarking DNN Processors, Nov. 2016, 2 pages. |
| Bluespec, “Bluespec System Verilog Reference Guide,” Jun. 16, 2010, Bluespec, Inc, 453 pages. |
| Bohm I., “Configurable Flow Accelerators,” Mar. 3, 2016, XP055475839. retrieved from http://groups.inf.ed.ac.uk/pasta/rareas_cfa.html on Oct. 25, 2018, 3 pages. |
| Brodman et al., “Writing Scalable SIMD Programs with ISPC”, Proceedings of the 2014 Workshop on Programming models for SIMD/Vector Processing, Feb. 2014, pp. 25-32. |
| Budiu et al., “Spatial Computation”, ASPLOS'04, Oct. 7-13, 2004, pp. 1-13. |
| Burger D., et al., “Scaling to the End of Silicon with EDGE Architectures,” Jul. 12, 2004, vol. 37 (7), pp. 44-55. |
| Burkowski, F. J., “Instruction Set Design Issues Relating to a Static Dataflow Computer”, ACM Sigarch Computer Architecture News, vol. 10, No. 3, Apr. 1982, pp. 101-111. |
| Canis A., et al., “LegUp: An Open-Source High-Level Synthesis Tool for FPGA-Based Processor/Accelerator Systems,” ACM Transactions on Embedded Computing Systems, vol. 1(1), Article 1, Jul. 2012, 25 pages. |
| Carloni L. P., et al., “The Theory of Latency Insensitive Design,” Sep. 2001, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 20 (9), 18 pages. |
| Chandy K. M., et al., “Parallel Program Design: A Foundation,” Addison-Wesley Publishing Company, Aug. 1988, 552 pages. |
| Chen Y., et al., “Eyeriss v2: A Flexible and High-Performance Accelerator for Emerging Deep Neural Networks,” Jul. 2018, 14 pages. |
| Chen Y., et al., “Eyeriss: A Spacial Architecture for Energy-Efficient Dataflow for Convolutional Neural Networks,” Jun. 2016, 53 pages. |
| Chen Y., et al., “Eyeriss: An Energy-Efficient Reconfigurable Accelerator For Deep Convolutional Neural Networks,” IEEE International Solid-State Circuits Conference, ISSCC 2016, Digest of Technical Papers, retrieved from eyeriss-isscc2016, spreadsheet, http://eyeriss.mit.edu/benchmarking.html, 2016, 7 pages. |
| Chen Y., et al., “Eyeriss: An Energy-Efficient Reconfigurable Accelerator for Deep Convolutional Neural Networks,” IEEE Journal of Solid State Circuits (JSSC), ISSCC Special Issue, Jan. 2017, vol. 52 (1), pp. 127-138. |
| Chen Y., et al., “Understanding the Limitations of Existing Energy-Efficient Design Approaches for Deep Neural Networks”, Feb. 2018, 3 pages. |
| Chen Y., et al., “Using Dataflow to Optimize Energy Efficiency of Deep Neural Network Accelerators,” IEEE Micro's Top Picks from the Computer Architecture Conferences, May/Jun. 2017, pp. 12-21. |
| Clark et al., “Application-Specific Processing on a General-Purpose Core Via Transparent Instruction Set Customization”, Proceedings of the37th International Symposium on Microarchitecture (Micro-37 2004), IEEE, 2004, 11 pages. |
| Clark et al., “VEAL: Virtualized Execution Accelerator for Loops”, International Symposium on Computer Architecture, 2008, 12 pages. |
| Compton K., et al., “Reconfigurable Computing: A Survey of Systems and Software,” ACM Computing Surveys, Jun. 2002, vol. 34 (2), pp. 171-210. |
| Cong J., et al., “Supporting Address Translation for Accelerator-Centric Architectures,” Feb. 2017, IEEE International Symposium on High Performance Computer Architecture (HPCA), 12 pages. |
| Coral Collaboration: Oak Ridge, Argonne, Livermore, Benchmark codes, downloaded from https://asc.llnl.gov/CORAL-benchmarks/ on Nov. 16, 2018, 6 pages. |
| Corrected Notice of Allowability, U.S. Appl. No. 16/024,802, dated Nov. 5, 2020, 7 pages. |
| Corrected Notice of Allowability, U.S. Appl. No. 16/024,802, dated Oct. 1, 2020, 2 pages. |
| Corrected Notice of Allowability, U.S. Appl. No. 16/024,849, dated Oct. 19, 2020, 5 pages. |
| Corrected Notice of Allowability, U.S. Appl. No. 16/236,423, dated May 7, 2020, 2 pages. |
| Corrected Notice of Allowance from U.S. Appl. No. 14/037,468, dated Apr. 1, 2019, 10 pages. |
| Corrected Notice of Allowance from U.S. Appl. No. 14/037,468, dated May 29, 2019, 12 pages. |
| Corrected Notice of Allowance from U.S. Appl. No. 15/396,395, dated Jun. 7, 2019, 8 pages. |
| Corrected Notice of Allowance from U.S. Appl. No. 15/640,534, dated Apr. 26, 2019, 21 pages. |
| Corrected Notice of Allowance from U.S. Appl. No. 15/640,534, dated Jul. 2, 2019, 12 pages. |
| Corrected Notice of Allowance from U.S. Appl. No. 15/640,535, dated Feb. 13, 2019, 7 pages. |
| Corrected Notice of Allowance from U.S. Appl. No. 15/721,802, dated Dec. 18, 2018, 8 pages. |
| Corrected Notice of Allowance from U.S. Appl. No. 15/721,802, dated Jun. 12, 2019, 11 pages. |
| Corrected Notice of Allowance from U.S. Appl. No. 15/855,964, dated Jun. 25, 2019, 7 pages. |
| Dagum et al., “OpenMP: An Industry Standard API for Shared-Memory Programming”, IEEE Computational Science and Engineering, Jan.-Mar. 1998, vol. 5, No. 1, pp. 46-55. |
| Dally W.J., et al., “Principles and Practices of Interconnection Networks,” Morgan Kaufmann, 2003, 584 pages. |
| Davidson A., “The Intel(Registered) HyperFlex(Trademark) FPGA Architecture meets the Performance Requirements of Next-Generation Systems,” White Paper, A New FPGA Architecture and Leading-Edge FinFET Process Technology Promise to Meet Next-Generation System Requirements, 6 pages. |
| Dennis et al., “A Preliminary Architecture for a Basic Data-Flow Processor”, Proceedings of the 2nd annual symposium on Computer architecture, Jan. 1975, pp. 126-132. |
| Dijkstra E.W., “Guarded Commands, Nondeterminacy and Formal Derivation of Programs,” Aug. 1975, Communications of the ACM, vol. 18 (8), pp. 453-457. |
| Eisenhardt S., et al., “Optimizing Partial Reconfiguration of Multi-Context Architectures,” Dec. 2008, 2008 International Conference on Reconfigurable Computing and FPGAs, 6 pages. |
| Emer J., et al., “Asim: A Performance Model Framework”, Feb. 2002, Computer, vol. 35 (2), pp. 68-76. |
| Emer J., et al., “Hardware Architectures for Deep Neural Networks at CICS/MTL Tutorial”, Mar. 27, 2017, 258 pages. |
| Emer J., et al., “Hardware Architectures for Deep Neural Networks at ISCA Tutorial”, Jun. 24, 2017, 290 pages. |
| Intel, “Intel(Registered) 64 and IA-32 Architectures Software Developer Manuals,” Oct. 12, 2016, Updated—May 18, 2018, 19 pages. |
| Intel, “Intel(Registered) oneAPI Programming Guide”, Jun. 28, 2021, 86 pages. |
| Intention to Grant, EP App. No. 20157809.3, dated Jun. 21, 2021, 7 pages. |
| International Preliminary Report on Patentability for Application No. PCT/RU2011/001049, dated Jul. 10, 2014, 6 pages. |
| International Preliminary Reporton Patentability, PCT App. No. PCT/US2017/050663, dated Apr. 11, 2019, 11 pages. |
| International Preliminary Reporton Patentability, PCT App. No. PCT/US2017/055849, dated Apr. 25, 2019, 6 pages. |
| International Preliminary Reporton Patentability, PCT App. No. PCT/US2019/019965, dated Oct. 15, 2020, 6 pages. |
| International Preliminary Reporton Patentability, PCT App. No. PCT/US2019/020243, dated Oct. 15, 2020, 7 pages. |
| International Preliminary Reporton Patentability, PCT App. No. PCT/US2019/020270, dated Oct. 15, 2020, 7 pages. |
| International Preliminary Reporton Patentability, PCT App. No. PCT/US2019/020287, dated Oct. 15, 2020, 6 pages. |
| International Preliminary Reporton Patentability, PCT App. No. PCT/US2019/034358, dated Jan. 14, 2021, 6 pages. |
| International Preliminary Report on Patentability, PCT App. No. PCT/US2019/034400, dated Jan. 14, 2021, 8 pages. |
| International Preliminary Report on Patentability, PCT App. No. PCT/US2019/034433, dated Jan. 14, 2021, 7 pages. |
| International Preliminary Report on Patentability, PCT App. No. PCT/US2019/034441, dated Jan. 14, 2021, 7 pages. |
| International Search Report and Written Opinion for Application No. PCT/RU2011/001049, dated Sep. 20, 2012, 6 pages. |
| International Search Report and Written Opinion for Application No. PCT/US2019/019965, dated Jun. 13, 2019, 9 pages. |
| International Search Report and Written Opinion for Application No. PCT/US2019/020243, dated Jun. 19, 2019, 11 pages. |
| International Search Report and Written Opinion for Application No. PCT/US2019/020270, dated Jun. 14, 2019, 11 pages. |
| International Search Report and Written Opinion for Application No. PCT/US2019/020287, dated Jun. 12, 2019, 9 pages. |
| International Search Report and Written Opinion, PCT App. No. PCT/US2017/050663, dated Dec. 28, 2017, 12 pages. |
| International Search Report and Written Opinion, PCT App. No. PCT/US2017/055849, dated Dec. 26, 2017, 8 pages. |
| International Search Report and Written Opinion, PCT App. No. PCT/US2019/034358, dated Sep. 18, 2019, 10 pages. |
| International Search Report and Written Opinion, PCT App. No. PCT/US2019/034400, dated Sep. 19, 2019, 11 pages. |
| International Search Report and Written Opinion, PCT App. No. PCT/US2019/034433, dated Sep. 20, 2019, 10 pages. |
| International Search Report and Written Opinion, PCT App. No. PCT/US2019/034441, dated Sep. 23, 2019, 10 pages. |
| Kalte H., et al., “Context Saving and Restoring for Multitasking in Reconfigurable Systems”, International Conference on Field Programmable Logic and Applications, Aug. 2005, pp. 223-228. |
| Kim et al., “Energy-Efficient and High Performance CGRA-based Multi-Core Architecture,” Journal of Semiconductor Technology and Science, vol. 14 (3), Jun. 2014, 16 pages. |
| King M., et al., “Automatic Generation of Hardware/Software Interfaces”, Proceedings of the 17th International Conference on Architectural Support for Programming Languages and Operating Systems, ASPLOS'12, Mar. 2012, 12 pages. |
| Knuth D. E., et al., “Fast Pattern Matching In Strings,” Jun. 1977, SIAM Journal of Computing, vol. 6(2), pp. 323-350. |
| Kumar et al., “Needle: Leveraging Program Analysis to Analyze and Extract Accelerators from Whole Programs”, IEEE International Symposium on High Performance Computer Architecture (HPCA), IEEE, 2017, 12 pages. |
| Lee T., et al., “Hardware Context-Switch Methodology for Dynamically Partially Reconfigurable Systems,” Journal of Information Science and Engineering, vol. 26, Jul. 2010, pp. 1289-1305. |
| Lengauer et al., “A Fast Algorithm for Finding Dominators in a Flowgraph”, ACM Transactions on Programming Languages and Systems, vol. 1, No. 1, Jul. 1979, pp. 121-141. |
| Lewis D., et al., “The Stratix™ 10 Highly Pipelined FPGA Architecture,” FPGA'16, ACM, Feb. 21-23, 2016, pp. 159-168. |
| Lis., etal., “Case Study: Computing Black-Scholes with Intel® Advanced Vector Extensions,” Sep. 6, 2012, 20 pages. |
| Marquardt A., et al., “Speed and Area Trade-OFFS in Cluster-Based FPGA Architectures,” Feb. 2000, IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 8 (1), 10 pages. |
| Matsen F.A., et al., “The CMU warp processor,” In Supercomputers: Algorithms, Architectures, and Scientific Computation, 1986, pp. 235-247. |
| Mccalpin J.D., “Memory Bandwidth and Machine Balance in Current High Performance Computers,” IEEE Computer Society Technical Committee on Computer Architecture (TCCA) Newsletter, Dec. 1995, 7 pages. |
| Mccalpin J.D., “Stream: Sustainable memory bandwidth in high performance computers,” 2016, 4 pages. |
| Mei B. et al., “ADRES: An Architecture with Tightly Coupled VLIW Processor and Coarse-Grained Reconfigurable Matrix”, 2003, In Proceedings of International Conference on Field-Programmable Logic and Applications, 10 pages. |
| Merrill D., et al., “Revisiting sorting for GPGPU stream architectures,” In Proceedings of the 19th International Conference on Parallel Architectures and Compilation Techniques (PACT'10), Feb. 2010, 17 pages. |
| Mirsky E., at al., “Matrix: A Reconfigurable Computing Architecture with Configurable Instruction Distribution and Deployable Resources,” 1996, In Proceedings of the IEEE Symposium on FPGAs for Custom Computing Machines, pp. 157-166. |
| Natalie E. J., et al., “On-Chip Networks,” Synthesis Lectures on Computer Architecture, Morgan and Claypool Publishers, 2009, 148 pages. |
| Non-Final Office Action from U.S. Appl. No. 13/994,582, dated Feb. 7, 2018, 12 pages. |
| Non-Final Office Action from U.S. Appl. No. 13/994,582, dated Mar. 23, 2017, 9 pages. |
| Non-Final Office Action from U.S. Appl. No. 14/037,468, dated Aug. 27, 2015, 10 pages. |
| Non-Final Office Action from U.S. Appl. No. 14/037,468, dated Dec. 2, 2016, 16 pages. |
| Non-Final Office Action from U.S. Appl. No. 14/037,468, dated Oct. 19, 2017, 19 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/283,295, dated Apr. 30, 2018, 18 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/396,038, dated Oct. 5, 2018, 38 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/396,049, dated Jun. 15, 2018, 33 pages. |
| Examination Report, DE App. No. 102018005172.3, dated Feb. 2, 2022, 20 pages (Original Document Only). |
| Notice of Allowance, U.S. Appl. No. 15/944,546, dated Mar. 28, 2022, 4 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,540, dated Jul. 1, 2019, 36 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,540, dated Mar. 14, 2019, 39 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,543, dated Jul. 29, 2019, 43 pages. |
| Notice of Allowance from U.S. Appl. No. 15/719,281, dated Jan. 24, 2019, 36 pages. |
| Notice of Allowance from U.S. Appl. No. 15/719,285, dated Jul. 23, 2019, 26 pages. |
| Notice of Allowance from U.S. Appl. No. 15/721,802, dated Jul. 31, 2018, 6 pages. |
| Notice of Allowance from U.S. Appl. No. 15/721,802, dated Mar. 18, 2019, 23 pages. |
| Notice of Allowance from U.S. Appl. No. 15/721,802, dated Nov. 30, 2018, 30 pages. |
| Notice of Allowance from U.S. Appl. No. 15/721,809, dated Jun. 6, 2019, 32 pages. |
| Notice of Allowance from U.S. Appl. No. 15/855,964, dated Apr. 24, 2019, 15 pages. |
| Notice of Allowance from U.S. Appl. No. 15/859,454, dated Jun. 7, 2019, 55 pages. |
| Notice of Allowance from U.S. Appl. No. 15/859,454, dated Sep. 12, 2019, 8 pages. |
| Notice of Allowance from U.S. Appl. No. 15/859,466, dated May 17, 2019, 56 pages. |
| Notice of Allowance from U.S. Appl. No. 15/944,761, dated Sep. 12, 2019, 75 pages. |
| Notice of Allowance from U.S. Appl. No. 16/024,801, dated Jun. 5, 2019, 64 pages. |
| Notice of Allowance, U.S. Appl. No. 16/458,032, dated Nov. 25, 2020, 9 pages. |
| Notice of Allowance, U.S. Appl. No. 15/396,038, dated Dec. 10, 2019, 2 pages. |
| Notice of Allowance, U.S. Appl. No. 15/396,038, dated Jan. 23, 2020, 2 pages. |
| Notice of Allowance, U.S. Appl. No. 15/396,038, dated Oct. 2, 2019, 9 pages. |
| Notice of Allowance, U.S. Appl. No. 15/396,049, dated Oct. 16, 2019, 2 pages. |
| Notice of Allowance, U.S. Appl. No. 15/396,395, dated Aug. 7, 2019, 12 pages. |
| Notice of Allowance, U.S. Appl. No. 15/396,402, dated Jan. 8, 2020, 2 pages. |
| Notice of Allowance, U.S. Appl. No. 15/396,402, dated Sep. 16, 2019, 15 pages. |
| Notice of Allowance, U.S. Appl. No. 15/640,535, dated Aug. 21, 2019, 13 pages. |
| Notice of Allowance, U.S. Appl. No. 15/640,538, dated Sep. 20, 2019 2 pages. |
| Notice of Allowance, U.S. Appl. No. 15/640,538, dated Sep. 5, 2019, 8 pages. |
| Notice of Allowance, U.S. Appl. No. 15/640,541, dated Aug. 13, 2019, 19 pages. |
| Notice of Allowance, U.S. Appl. No. 15/719,281, dated Apr. 21, 2020, 5 pages. |
| Notice of Allowance, U.S. Appl. No. 15/719,281, dated Apr. 26, 2021, 5 pages. |
| Notice of Allowance, U.S. Appl. No. 15/721,809, dated Sep. 5, 2019, 8 pages. |
| Notice of Allowance, U.S. Appl. No. 15/721,816, dated Nov. 19, 2020, 9 pages. |
| Notice of Allowance, U.S. Appl. No. 15/859,473, dated Sep. 24, 2019, 6 pages. |
| Notice of Allowance, U.S. Appl. No. 15/941,888, dated Dec. 18, 2020. |
| Notice of Allowance, U.S. Appl. No. 15/941,888, dated Sep. 8, 2020, 9 pages. |
| Notice of Allowance, U.S. Appl. No. 15/943,608, dated Oct. 16, 2020, 8 pages. |
| Notice of Allowance, U.S. Appl. No. 15/944,546, dated Dec. 21, 2021, 8 pages. |
| Notice of Allowance, U.S. Appl. No. 15/944,546, dated Mar. 2, 2022, 4 pages. |
| Notice of Allowance, U.S. Appl. No. 15/944,761, dated Jan. 8, 2020, 2 pages. |
| Notice of Allowance, U.S. Appl. No. 15/944,761, dated Nov. 1, 2019, 2 pages. |
| Notice of Allowance, U.S. Appl. No. 16/024,801, dated Sep. 12, 2019, 10 pages. |
| Notice of Allowance, U.S. Appl. No. 16/024,802, dated Aug. 28, 2020, 11 pages. |
| Notice of Allowance, U.S. Appl. No. 16/024,802, dated Dec. 9, 2020, 6 pages. |
| Notice of Allowance, U.S. Appl. No. 16/024,849, dated Jul. 7, 2020, 8 pages. |
| Notice of Allowance, U.S. Appl. No. 16/024,854, dated Aug. 6, 2021, 11 pages. |
| Notice of Allowance, U.S. Appl. No. 16/221,379, dated Oct. 28, 2020, 7 pages. |
| Notice of Allowance, U.S. Appl. No. 16/236,423, dated Feb. 6, 2020, 8 pages. |
| Notice of Allowance, U.S. Appl. No. 16/370,915, dated Jun. 11, 2020, 12 pages. |
| Notice of Allowance, U.S. Appl. No. 16/370,928, dated Apr. 24, 2020, 9 pages. |
| Notice of Allowance, U.S. Appl. No. 16/370,928, dated Aug. 11, 2020, 9 pages. |
| Notice of Allowance, U.S. Appl. No. 16/370,928, dated Jan. 11, 2021, 2 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/396,395, dated Jul. 20, 2018, 18 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/396,402, dated Nov. 1, 2018, 22 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/640,533, dated Apr. 19, 2018, 8 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/640,534, dated Apr. 26, 2018, 8 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/640,535, dated May 15, 2018, 13 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/640,540, dated Oct. 26, 2018, 8 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/640,541, dated Apr. 12, 2019, 61 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/640,542, dated Apr. 2, 2019, 59 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/640,543, dated Jan. 2, 2019, 46 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/719,285, dated Feb. 25, 2019, 47 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/721,809, dated Jun. 14, 2018, 12 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/855,964, dated Dec. 13, 2018, 13 pages. |
| Non-Final Office Action from U.S. Appl. No. 15/859,473, dated Oct. 15, 2018, 10 pages. |
| Non-Final Office Action, U.S. Appl. No. 15/640,542, dated Mar. 4, 2020, 26 pages. |
| Non-Final Office Action, U.S. Appl. No. 15/721,802, dated Mar. 8, 2018, 8 pages. |
| Non-Final Office Action, U.S. Appl. No. 15/721,816, dated Dec. 5, 2019, 20 pages. |
| Non-Final Office Action, U.S. Appl. No. 15/943,608, dated Mar. 12, 2020, 8 pages. |
| Non-Final Office Action, U.S. Appl. No. 15/944,546, dated Nov. 25, 2019, 22 pages. |
| Non-Final Office Action, U.S. Appl. No. 16/024,849, dated Nov. 29, 2019, 10 pages. |
| Non-Final Office Action, U.S. Appl. No. 16/024,854, dated Feb. 19, 2021, 11 pages. |
| Non-Final Office Action, U.S. Appl. No. 16/024,854, dated Jan. 13, 2020, 10 pages. |
| Non-Final Office Action, U.S. Appl. No. 16/221,379, dated Mar. 31, 2020, 9 pages. |
| Non-Final Office Action, U.S. Appl. No. 16/236,423, dated Aug. 21, 2019, 75 pages. |
| Non-Final Office Action, U.S. Appl. No. 16/443,717, dated Sep. 30, 2019, 8 pages. |
| Non-Final Office Action, U.S. Appl. No. 16/458,032, dated Jul. 24, 2020, 9 pages. |
| Notice of Allowability, U.S. Appl. No. 15/943,608, dated Nov. 24, 2020, 6 pages. |
| Notice of Allowability, U.S. Appl. No. 16/443,717, dated Nov. 5, 2020, 2 pages. |
| Notice of Allowability, U.S. Appl. No. 16/458,020, dated Nov. 12, 2020, 10 pages. |
| Notice of Allowance from U.S. Appl. No. 13/994,582, dated Aug. 7, 2018, 8 pages. |
| Notice of Allowance from U.S. Appl. No. 14/037,468, dated Aug. 28, 2018, 9 pages. |
| Notice of Allowance from U.S. Appl. No. 14/037,468, dated Mar. 7, 2019, 51 pages. |
| Notice of Allowance from U.S. Appl. No. 15/283,295, dated Apr. 10, 2019, 49 pages. |
| Notice of Allowance from U.S. Appl. No. 15/283,295, dated Jan. 3, 2019, 7 pages. |
| Notice of Allowance from U.S. Appl. No. 15/396,049, dated Jul. 2, 2019, 70 pages. |
| Notice of Allowance from U.S. Appl. No. 15/396,395, dated Dec. 28, 2018, 36 pages. |
| Notice of Allowance from U.S. Appl. No. 15/396,395, dated May 15, 2019, 23 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,533, dated Feb. 14, 2019, 43 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,533, dated May 22, 2019, 19 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,533, dated Oct. 10, 2018, 8 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,533, dated Sep. 12, 2019, 16 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,534, dated Apr. 2, 2019, 9 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,534, dated Jan. 4, 2019, 37 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,534, dated May 31, 2019, 9 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,534, dated Sep. 12, 2018, 7 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,535, dated Feb. 6, 2019, 38 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,535, dated Jun. 21, 2019, 8 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,535, dated May 24, 2019, 19 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,535, dated Oct. 9, 2018, 7 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,538, dated Jul. 3, 2019, 76 pages. |
| Notice of Allowance from U.S. Appl. No. 15/640,538, dated Oct. 17, 2018, 10 pages. |
| Notice of Allowance, U.S. Appl. No. 16/443,717, dated Jul. 8, 2020, 7 pages. |
| Nowatzki et al., “Exploring the Potential of Heterogeneous Von Neumann/Dataflow Execution Models”, 42nd Annual International Symposium on Computer Architecture (ISCA), Jun. 13-17, 2015, 13 pages. |
| Nvidia, “CUDA C++ Programming Guide”, Design Guide, PG-02829-001_v11.2, Jun. 2021, 375 pages. |
| Office Action, DE App. No. 102018005181.2, dated Jan. 31, 2022, 19 pages (Original Document Only). |
| Office Action, EP App. No. 17857177.4, dated May 11, 2021, 07 pages. |
| Panesar G., et al., “Deterministic Parallel Processing,” International Journal of Parallel Programming, Aug. 2006, vol. 34 (4), pp. 323-341. |
| Parashar A., et al., “Efficient Spatial Processing Element Control via Triggered Instructions,” IEEE Micro, vol. 34 (3), Mar. 10, 2014, pp. 120-137. |
| Parashar et al., “Leap: A Virtual Platform Architecture for FPGAs,” Intel Corporation, 6 pages, 2010. |
| Pellauer M., et al., “Efficient Control and Communication Paradigms for Coarse-Grained Spatial Architectures,” Sep. 2015, ACM Transactions on Computer Systems, vol. 33 (3), Article 10, 32 pages. |
| Pellauer M., et al., “Soft Connections: Addressing the Hardware-Design Modularity Problem,” 2009, In Proceedings of the 46th ACM/IEEE Design Automation Conference (DAC'09), pp. 276-281. |
| Raaijmakers S., “Run-Time Partial Reconfiguration on the Virtex-11 Pro,” 2007, 69 pages. |
| Requirement for Restriction/Election, U.S. Appl. No. 16/024,802, dated Mar. 19, 2020, 8 pages. |
| Schauser et al., “Compiler-Controlled Multithreading for Lenient Parallel Languages”, Aug. 1991, pp. 1-21. |
| Schmit H., et al., “PipeRench: A Virtualized Programmable Datapath in 0.18 Micron Technology,” 2002, IEEE 2002 Custom Integrated Circuits Conference, pp. 63-66. |
| Seznec et al., “A Case for (partially) TAgged GEometric History Length Branch Prediction”, Computer Science, J. Instr. Level Parallelism, 2006, pp. 1-23. |
| Sharifian et al., “Chainsaw: Von-Neumann Accelerators to Leverage Fused Instruction Chains”, 49th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), 2016, 14 pages. |
| Shin T., et al., “Minimizing Buffer Requirements for Throughput Constrained Parallel Execution of Synchronous Dataflow Graph,” ASPDAC '11 Proceedings of the 16th Asia and South Pacific Design Automation Conference , Jan. 2011, 6 pages. |
| Smith A., et al., “Compiling for EDGE Architectures,” Appears in the Proceedings of the 4th International Symposium Jn code Generation and Optimization, 2006, 11 pages. |
| Smith A., et al., “Dataflow Predication,” 2006, In Proceedings of the 39th Annual IEEE/ACM International Symposium on Microarchitecture, 12 pages. |
| Srinath et al., “Architectural Specialization for Inter-Iteration Loop Dependence Patterns”, 47th Annual IEEE/ACM International Symposium on Microarchitecture, (MICRO-47), Dec. 2014, 13 pages. |
| Suleiman A., et al., “Towards Closing the Energy Gap Between HOG and CNN Features for Embedded Vision,” IEEE International Symposium of Circuits and Systems (ISCAS), May 2017, 4 pages. |
| Supplemental Notice of Allowability, U.S. Appl. No. 15/721,816, dated Feb. 17, 2021, 2 pages. |
| Supplemental Notice of Allowability, U.S. Appl. No. 16/370,915, dated Sep. 23, 2020, 4 pages. |
| Supplementary European Search Report and Search Opinion, EP App. No. 19776406.1, dated Nov. 26, 2021, 10 pages. |
| Swanson et al., “WaveScalar”, Appears in the 36th International Symposium on Microarchitecture, (MICRO-36 2003), Dec. 3, 2003, pp. 1-12. |
| Swanson S., et al., “The WaveScalar Architecture,” May 2007, ACM Transactions on Computer Systems, vol. 25 (2), Article No. 4, 35 pages. |
| Sze V. et al., “Efficient Processing of Deep Neural Networks: A Tutorial and Survey,” arxiv.org, Cornell University Library, Mar. 27, 2017, pp. 1-31. |
| Sze V., “Designing Efficient Deep Learning Systems,” in Mountain View, CA, Mar. 27-28, 2019, retrieved from https://professional.mit.edu/programs/short-programs/designing-efficienl-deep-learning-systems-OC on Feb. 14, 2019, 2 pages. |
| Sze V., et al., “Efficient Processing of Deep Neural Networks: A Tutorial and Survey,” Proceedings of the IEEE, Dec. 2017, vol. 105 (12), pp. 2295-2329. |
| Sze V., et al., “Hardware for Machine Learning: Challenges and Opportunities,”IEEE Custom Integrated Circuits Conference (CICC), Oct. 2017, 9 pages. |
| Taylor M.B., et al., “The Raw Microprocessor: A Computational Fabric for Sollware Circuits and General-Purpose Programs,” 2002, IEEE Micro, vol. 22 (2), pp. 25-35. |
| The LLVM Compiler Infrastructure, retrieved from http://www.llvm.org/, on May 1, 2018, maintained by the llvm-admin team, 4 pages. |
| Truong D.N., et al., “A 167-Processor Computational Platform in 65 nm CMOS,” IEEE Journal of Solid-State Circuits, Apr. 2009, vol. 44 (4), pp. 1130-1144. |
| Tutorial at Micro-50, The 50th Annual IEEE/ACM International Symposium on Microarchitecture, Oct. 14-18, 2017, retrieved from https://www.microarch.org/micro50/ on Feb. 14, 2019, 3 pages. |
| Tutorial on Hardware Architectures for Deep Neural Networks at ISCA 2017, The 44th International Symposium on Computer Architecture, Jun. 24-28, 2017, retrieved from http://isca17.ece.utoronto.ca/doku.php on Feb. 14, 2019, 2 pages. |
| Van De Geijn R.A., et al., “SUMMA: Scalable Universal Matrix Multiplication Algorithm,” 1997, 19 pages. |
| Veen, Arthur H., “Dataflow Machine Architecture”, ACM Computing Surveys, vol. 18, No. 4, Dec. 1986, pp. 365-396. |
| Vijayaraghavan M., et al., “Bounded Dataflow Networks and Latency-Insensitive Circuits,” In Proceedings of the 7th IEEE/ACM International Conference on Formal Methods and Models for Codesign (Memocode'09), Jul. 13-15, 2009, pp. 171-180. |
| Wall, David W., “Limits of Instruction-Level Parallelism”, ACM SIGOPS Operating Systems Review, vol. 25, Apr. 1991 pp. 176-188. |
| Wikipedia, “TRIPS Architecture,” retrieved from https://en.wikipedia.org/wiki/TRIPS_architecture on Oct. 14, 2018, 4 pages. |
| Wikipedia, The Free Encyclopedia, “Priority encoder,” https://en.wikipedia.org/w/index.php?Title=Priority_encoder&oldid=746908667, revised Oct. 30, 2016, 2 pages. |
| Wikipedia, The Free Encyclopedia, “Truth table,” Logical Implication Table, https://enwikipedia.org/wiki/Truth_table#Logical_implication, revised Nov. 18, 2016, 9 pages. |
| Williston, Roving Reporter, Intel® Embedded Alliance, “Roving Reporter: FPGA + Intel® Atom TM = Configurable Processor,” Dec. 2010, 5 pages. |
| Yang T., et al., “A Method to Estimate the Energy Consumption of Deep Neural Networks,” Asilomar Conference on Signals, Systems and Computers, Oct. 2017, 5 pages. |
| Yang T., et al., “Deep Neural Network Energy Estimation Tool”, IEEE Conference on Computer Vision and Pattern Recognition CVPR 2017, Jul. 21-26, 2017, retrieved from https://energyestimation.mit.edu/ on Feb. 21, 2019, 4 pages. |
| Yang T., et al., “Designing Energy-Efficient Convolutional Neural Networks using Energy-Aware Pruning,” IEEE CVPR, Mar. 2017, 6 pages. |
| Yang T., et al., “NetAdapt: Platform-Aware Neural Network Adaptation for Mobile Applications,” European Conference on Computer Vision (ECCV), Version 1, Apr. 9, 2018, 16 pages. |
| Yang T., et al., “NetAdapt: Platform-Aware Neural Network Adaptation for Mobile Applications,” European Conference on Computer Vision (ECCV), Version 2, Sep. 28, 2018, 16 pages. |
| Ye Z.A., et al., “Chimaera: A High-Performance Architecture with a Tightly-Coupled Reconfigurable Functional Unit,” Proceedings of the 27th International Symposium on Computer Architecture (ISCA'00), 2000, 11 pages. |
| Yu Z., et al., “An Asynchronous Array of Simple Processors for DSP Applications,” IEEE International Solid-State Circuits Conference, ISSCC'06, Feb. 8, 2006, 10 pages. |
| Emer J., et al., “Hardware Architectures for Deep Neural Networks at MICR0-49 Tutorial,” Oct. 16, 2016, 300 pages. |
| Emer J., et al., “Tutorial on Hardware Architectures for Deep Neural Networks”, Nov. 2016, 8 pages. |
| Emer J.S., et al., “A Characterization of Processor Performance in the VAX-11/780,” In Proceedings of the 11th Annual International Symposium on Computer Architecture, Jun. 1984, vol. 12 (3), pp. 274-283. |
| European Search Report and Search Opinion, EP App. No 19781074.0, dated Dec. 13, 2021, 8 pages. |
| European Search Report and Search Opinion, EP App. No 19781298.5, dated Dec. 13, 2021, 11 pages. |
| European Search Report and Search Opinion, EP App. No 20157807.7, dated Aug. 25, 2020, 6 pages. |
| European Search Report and Search Opinion, EP App. No. 20157809.3, dated Sep. 17, 2020, 7 pages. |
| European Search Report and Search Opinion, EP App. No. 20163890.5, dated Nov. 25, 2020, 10 pages. |
| European Search Report and Search Opinion, EP App. No. 20164075.2, dated Sep. 14, 2020, 6 pages. |
| Ex Parte Quayle Action, U.S. Appl. No. 15/944,546, dated Jul. 12, 2021, 12 pages. |
| Examiner initialed interview summary, U.S. Appl. No. 16/024,802, dated Mar. 19, 2020, 1 page. |
| Extended European Search Report and Search Opinion, EP App. No. 17857177.4, dated Mar. 30, 2020, 7 pages. |
| Extended European Search Report for Application No. 17207172.2, dated Oct. 1, 2018, 14 pages. |
| Extended European Search Report for Application No. 17210484.6, dated May 29, 2018, 8 pages. |
| Final Office Action from U.S. Appl. No. 13/994,582, dated Oct. 3, 2017, 11 pages. |
| Final Office Action from U.S. Appl. No. 14/037,468, dated Jun. 1, 2017, 18 pages. |
| Final Office Action from U.S. Appl. No. 14/037,468, dated Jun. 15, 2018, 7 pages. |
| Final Office Action from U.S. Appl. No. 14/037,468, dated May 16, 2016, 24 pages. |
| Final Office Action from U.S. Appl. No. 14/037,468, dated Oct. 5, 2016, 17 pages. |
| Final Office Action from U.S. Appl. No. 15/396,038, dated Mar. 11, 2019, 36 pages. |
| Final Office Action from U.S. Appl. No. 15/396,049, dated Dec. 27, 2018, 38 pages. |
| Final Office Action from U.S. Appl. No. 15/396,402, dated May 17, 2019, 85 pages. |
| Final Office Action from U.S. Appl. No. 15/721,809, dated Dec. 26, 2018, 46 pages. |
| Final Office Action from U.S. Appl. No. 15/859,473, dated Feb. 26, 2019, 13 pages. |
| Final Office Action, U.S. Appl. No. 15/640,542, dated Aug. 7, 2019, 46 pages. |
| Final Office Action, U.S. Appl. No. 15/640,542, dated Aug. 24, 2020, 31 pages. |
| Final Office Action, U.S. Appl. No. 15/721,816, dated Jun. 12, 2020, 15 pages. |
| Final Office Action, U.S. Appl. No. 15/944,546, dated May 1, 2020, 17 pages. |
| Final Office Action, U.S. Appl. No. 16/024,849, dated Apr. 8, 2020, 9 pages. |
| Final Office Action, U.S. Appl. No. 16/024,854, dated Jul. 22, 2020, 11 pages. |
| Final Office Action, U.S. Appl. No. 16/443,717, dated Apr. 15, 2020, 4 pages. |
| Fleming et al., U.S. Patent App. titled “Memory Ordering in Acceleration Hardware,” 81 pages, filed Dec. 30, 2016., U.S. Appl. No. 15/396,038. |
| Fleming et al., U.S. Patent App. titled “Runtime Address Disambiguation in Acceleration Hardware,” filed Dec. 30, 2016, 97 pages., U.S. Appl. No. 15/396,049. |
| Fleming K., et al., “Configurable Spatial Accelerator: An Accelerator For High-performance Computing,” Dec. 2, 2016, 19 pages. |
| Fleming K., et al., “Leveraging Latency-Insensitivity to Ease Multiple FPGA Design”, In Proceedings of the ACM/SIGDA International Symposium on Field Programmable Gate Arrays, FPGA'12, Feb. 22-24, 2012, pp. 175-184. |
| Fleming K.E., et al., “Leveraging Latency-Insensitive Channels to Achieve Scalable Reconfigurable Computation,” Feb. 2013, 197 pages. |
| Full Day Tutorial Held at MICRO-49, Oct. 15-19, 2016, retrieved from https://www.microarch.org/micro49/ on Feb. 14, 2019, 2 pages. |
| Gao, Guang R., “Maximum Pipelining Linear Recurrence on Static Data Flow Computers”, International Journal of Parallel Programming, vol. 15, No. 2, 1986, pp. 127-149. |
| Govindaraju et al., “Dynamically Specialized Datapaths for Energy Efficient Computing”, 17th IEEE International Symposium on High Performance Computer Architecture, 2011, 12 pages. |
| Govindaraju et al.,“DySER: Unifying Functionality and Parallelism Specialization for Energy-Efficient Computing,” Published by the IEEE Computer Society, Sep./Oct. 2012, pp. 38-51. |
| Gupta et al., “Bundled Execution of Recurring Traces for Energy-Efficient General Purpose Processing”, Micro 44, Dec. 3-7, 2011, 12 pages. |
| Gurd et al., “The Manchester Prototype Dataflow Computer”, Communications of the ACM, vol. 28 No. 1, Jan. 1985, pp. 34-52. |
| Han S., et al., “Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding,” ICLR, Feb. 2016, 14 pages. |
| Hans., et al., “EIE: Efficient Inference Engine on Compressed Deep Neural Network,” ISCA, May 2016, 12 pages. |
| Hauser J. R. et al., “Garp: a MIPS processor with a Reconfigurable Coprocessor”, Proceedings of the 5th Annual IEEE Symposium on Field-Programmable Custom Computing Machines, 1997, 10 pages. |
| Hoogerbrugge J., et al., “Transport-Triggering vs. Operation-Triggering”, 1994, In Compiler Construction, Lecture Notes in Computer Science, vol. 786, Springer, pp. 435-449. |
| Hsin Y., “Building Energy-Efficient Accelerators for Deep Learning,” at Deep Learning Summit Boston, May 2016, retrieved from https://www.re-work.co/events/deep-learning-boston-2016 on Feb. 14, 2019, 10 pages. |
| Hsin Y., “Deep Learning & Artificial Intelligence,” at GPU Technology Conference, Mar. 26-29, 2018, retrieved from http://www.gputechconf.com/resources/poster-gallery/2016/deep-learning-artificial-intelligence on Feb. 14, 2019, 4 pages. |
| Ibrahim Eng., Walaa Abd El Aziz, “Binary Floating Point Fused Multiply Add Unit”, Faculty Of Engineering, Cairo University Giza, Egypt, 2012, 100 Pages. |
| Intel(Registered) Architecture, “Instruction Set Extensions and Future Features Programming Reference,” 319433-034, May 2018, 145 pages. |
| Intention to grant, EP App. No. 20164075.2, dated May 23, 2022, 7 pages. |
| Office Action, EP App. No. 20157807.7, dated Apr. 22, 2022, 3 pages. |
| Office Action, EP App. No. 20163890.5, dated May 11, 2022, 7 pages. |
| Zaidi, Ali Mustafa, “Accelerating control-flow intensive code in spatial hardware”, University of Cambridge, Technical Report, No. 870, May 31, 2015, pp. 1-170. |
| Chun-Jiang et al., “Formal Description of Design Space of SIMD Instruction Sets”, Computer Science, vol. 40, No. 6, Jun. 2013, pp. 32-36 (English Abstract Submitted). |
| Examination Report, DE App. No 102018005181.2, dated Nov. 22, 2022, 5 pages of Original Document Only. |
| Office Action, EP App. No. 20164075.2, dated Sep. 30, 2022, 5 pages. |
| Office Action, EP App. No. 20164075.2, dated Dec. 5, 2022, 5 pages. |
| First Office Action, CN App. No. 201780053811.X, dated Oct. 8, 2022, 29 pages (18 pages of English Translation and 11 pages of Original Document). |
| Office Action, CN App. No. 201810001530.6, dated Oct. 14, 2022, 10 pages of Original Document Only. |
| Office Action, CN App. No. 201810002672.4, dated Sep. 27, 2022, 15 pages of Original Document Only. |
| European Communication pursuant to Article 94(3) EPC, EP App. No. 19781074.0, dated Jul. 18, 2022, 5 pages. |
| Office Action, DE App. No. 102018005181.2, dated Aug. 14, 2022, 4 pages of Original Document Only. |
| Number | Date | Country | |
|---|---|---|---|
| 20220107911 A1 | Apr 2022 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 16024854 | Jun 2018 | US |
| Child | 17550875 | US |
