Registers for data transfers

[0001] Parallel processing is an efficient form of information processing of concurrent events in a computing process. Parallel processing demands concurrent execution of many programs, in contrast to sequential processing. In the context of parallel processing, parallelism involves doing more than one thing at the same time. Unlike a serial paradigm where all tasks are performed sequentially at a single station or a pipelined machine where tasks are performed at specialized stations, with parallel processing, many stations are provided, each capable of performing and carrying out various tasks and functions simultaneously. A number of stations work simultaneously and independently on the same or common elements of a computing task. Accordingly, parallel processing solves various types of computing tasks and certain problems are suitable for solution by applying several instruction processing units and several data streams.

[0002] FIG. 1 is a block diagram of a processing system.

[0003] FIG. 2 is a detailed block diagram of the processing system of FIG. 1 where one of the embodiments of the invention may be advantageously practiced.

[0004] FIG. 3 is a block diagram of a functional pipeline unit of the processing system of FIG. 1.

[0005] FIG. 4 is a block diagram illustrating details of the processing system of FIG. 1 where one of the embodiments of the invention may be advantageously practiced.

[0006] FIG. 5 is a simplified block diagram of a context pipeline process.

[0007] FIG. 6 is a flowchart illustrating the process of a context pipeline where one of the embodiments of the invention may be advantageously practiced.

[0008] FIG. 7 is a flowchart illustrating the process of determining the address of the Next Neighbor registers.

[0009] Architecture:

[0010] Referring to FIG. 1, a computer processing system 10 includes a parallel, hardware-based multithreaded network processor 12. The hardware-based multithreaded processor 12 is coupled to a memory system or memory resource 14. Memory system 14 includes dynamic random access memory (DRAM) 14a and static random access memory 14b (SRAM) . The processing system 10 is especially useful for tasks that can be broken into parallel subtasks or functions. Specifically, the hardware-based multithreaded processor 12 is useful for tasks that are bandwidth oriented rather than latency oriented. The hardware-based multithreaded processor 12 has multiple functional microengines or programming engines 16a-16h (collectively, programming engines 16) each with multiple hardware controlled threads that are simultaneously active and independently work on a specific task.

[0011] The programming engines 16 each maintain program counters in hardware and states associated with the program counters. Effectively, corresponding sets of context or threads can be simultaneously active on each of the programming engines 16 while only one is actually operating at any one time.

[0012] In this example, eight programming engines 16a-16h are illustrated in FIG. 1. Each programming engine 16a-16h processes eight hardware threads or contexts. The eight programming engines 16a-16h operate with shared resources including memory resource 14 and bus interfaces (not shown). The hardware-based multithreaded processor 12 includes a dynamic random access memory (DRAM) controller 18a and a static random access memory (SRAM) controller 18b. The DRAM memory 14a and DRAM controller 18a are typically used for processing large volumes of data, e.g., processing of network payloads from network packets. The SRAM memory 14b and SRAM controller 18b are used in a networking implementation for low latency, fast access tasks, e.g., accessing look-up tables, memory for the core processor 20, and the like.

[0013] The eight programming engines 16a-16h access either the DRAM memory 14a or SRAM memory 14b based on characteristics of the data. Thus, low latency, low bandwidth data are stored in and fetched from SRAM memory 14b, whereas higher bandwidth data for which latency is not as important, are stored in and fetched from DRAM memory 14a. The programming engines 16 can execute memory reference instructions to either the DRAM controller 18a or SRAM controller 18b.

[0014] The hardware-based multithreaded processor 12 also includes a processor core 20 for loading microcode control for the programming engines 16. In this example, although other types of processor cores may be used in embodiments of this invention, the processor core 20 is an XScale based architecture, designed by Intel Corporation, of Santa Clara, Calif.

[0015] The processor core 20 performs general-purpose computer type functions such as handling protocols, exceptions, and extra support for packet processing where the programming engines 16 pass the packets off for more detailed processing such as in boundary conditions.

[0016] The processor core 20 executes an operating system (not shown). Through the operating system (OS), the processor core 20 can call functions to operate on the programming engines 16a-16h. For the core processor 20 implemented as an XScale architecture, operating systems such as Microsoft NT real-time of Microsoft Corporation, of Seattle, Wash., VxWorks real-time operating system of WindRiver, of Alameda, Calif., or a freeware OS available over the Internet can be used.

[0017] Advantages of hardware multithreading can be explained by SRAM or DRAM memory accesses. As an example, an SRAM access requested by a context (e.g., Thread0), from one of the programming engines 16, e.g., programming engine 16a, will cause the SRAM controller 18b to initiate an access to the SRAM memory 14b. The SRAM controller 18b accesses the SRAM memory 14b, fetches the data from the SRAM memory 14b, and returns data to a requesting programming engine 16.

[0018] During an SRAM access, if one of the programming engines 16a-16h has a single thread that could operate, that programming engine would be dormant until data was returned from the SRAM memory 14b.

[0019] By employing hardware context swapping within each of the programming engines 16a-16h, the hardware context swapping enables other contexts with unique program counters to execute in that same programming engine. Thus, another thread e.g., Thread1 can function while the first thread, Thread0, is awaiting the read data to return. During execution, Thread1 may access the DRAM memory 14a. While Thread1 operates on the DRAM unit, and Thread0 is operating on the SRAM unit, a new thread, e.g., Thread2 can now operate in the programming engine 16. Thread2 can operate for a certain amount of time until it needs to access memory or perform some other long latency operation, such as making an access to a bus interface. Therefore, simultaneously, the multi-threaded processor 12 can have a bus operation, an SRAM operation, and a DRAM operation all being completed or operated upon by one of the programming engines 16 and have one more threads or contexts available to process more work.

[0020] The hardware context swapping also synchronizes the completion of tasks. For example, two threads can access the shared memory resource, e.g., the SRAM memory 14b. Each one of the separate functional units, e.g., the SRAM controller 18b, and the DRAM controller 18a, when they complete a requested task from one of the programming engine threads or contexts reports back a flag signaling completion of an operation. When the programming engines 16a-16h receive the flag, the programming engines 16a-16h can determine which thread to turn on.

[0021] One example of an application for the hardware-based multithreaded processor 12 is as a network processor. As a network processor, the hardware-based multithreaded processor 12 interfaces to network devices such as a Media Access Controller (MAC) device, e.g., a 10/100 BaseT Octal MAC or a Gigabit Ethernet device compliant with IEEE 802.3. In general, as a network processor, the hardware-based multithreaded processor 12 can interface to any type of communication device or interface that receives or sends large amount of data. The computer processing system 10 functioning in a networking application can receive network packets and process those packets in a parallel manner.

[0022] Registers in Programming Engines:

[0023] Referring to FIG. 2, one exemplary programming engine 16a from the programming engines 16, is shown. The programming engine 16a includes a control store 30, which in one example includes a RAM of 4096 instructions, each of which is 40-bits wide. The RAM stores a microprogram that the programming engine 16a executes. The microprogram in the control store 30 is loadable by the processor core 20 (FIG. 1).

[0024] In addition to event signals that are local to an executing thread, the programming engine 16a employs signaling states that are global. With signaling states, an executing thread can broadcast a signal state to all programming engines 16a-16h. Any and all threads in the programming engines can branch on these signaling states.

[0025] As described above, the programming engine 16a supports multi-threaded execution of eight contexts. This allows one thread to start executing just after another thread issues a memory reference and must wait until that reference completes before doing more work. Multi-threaded execution is critical to maintaining efficient hardware execution of the programming engine 16a because memory latency is significant. Multi-threaded execution allows the programming engines 16 to hide memory latency by performing useful independent work across several threads.

[0026] The programming engine 16a, to allow for efficient context swapping, has its own register set, program counter, and context specific local registers. Having a copy per context eliminates the need to move context specific information to and from shared memory and programming engine registers for each context swap. Fast context swapping allows a context to perform computations while other contexts wait for input-output (I/O), typically, external memory accesses to complete or for a signal from another context or hardware unit.

[0027] General Purpose Registers

[0028] The programming engine 16a executes the eight contexts by maintaining eight program counters and eight context relative sets of registers. A number of different types of context relative registers, such as general purpose registers (GPRs) 32, inter-programming agent registers (not shown), Static Random Access Memory (SRAM) input transfer registers 34, Dynamic Random Access Memory (DRAM) input transfer registers 36, SRAM output transfer registers 38, DRAM output transfer registers 40.

[0029] The GPRs 32 are used for general programming purposes. The GPRs 32 are read and written exclusively under program control. The GPRs 32, when used as a source in an instruction, supply operands to an execution datapath 44.

[0030] The execution datapath 44 can take one or two operands, perform an operation, and optionally write back a result. The execution datapath 44 includes a content addressable memory (CAM) 45. Each entry of the CAM 45 stores a 32-bit value, which can be compared against a source operand. All entries are compared in parallel and the result of the lookup is a 6-bit value.

[0031] When used as a destination in an instruction, the GPRs 32 are written with the result of the execution datapath 44. The programming engine 16a also includes I/O transfer registers 34, 36, 38 and 40 which are used for transferring data to and from the programming engine 16a and locations external to the programming engines 16a, e.g., the DRAM memory 14a, the SRAM memory 14b, and the like.

[0032] Transfer Registers

[0033] The programming engine 16a also includes transfer registers 34, 36, 38 and 40. Transfer registers 34, 36, 38 and 40 are used for transferring data to and from the programming engine 16a and locations external to the programming engine, e.g., DRAMs, SRAMs etc. There are four types of transfer registers as illustrated in FIG. 2, namely, input transfer registers and output transfer registers.

[0034] The input transfer registers, when used as a source in an instruction, supply operands to the execution datapath 44, whereas output transfer registers are written with the result from the execution datapath 44 when utilized as a destination in an instruction.

[0035] Local Control and Status Registers (CSRs)

[0036] Local control and status registers (CSRs) 37 are external to the execution datapath 44 and hold specific purpose information. They can be read and written by special instructions (local_csr_rd and local_csr_wr) and are typically accessed less frequently than datapath registers.

[0037] Next Neighbor Registers

[0038] The programming engine 16a also includes one hundred and twenty eight (128) Next Neighbor (NN) registers, collectively referred to as NN registers 35. Each NN Register 35, when used as a source in an instruction, also supplies operands to the execution datapath 44. Each NN register 35 is written either by an external entity, not limited to, an adjacent programming engine, or by the same programming engine 16a where each NN register 35 resides. The specific register is selected by a context-relative operation where the register number is encoded in the instruction, or as a ring operation, selected via, e.g., NN_Put (NN write address) and NN_Get (NN read address) in the CSR Registers.

[0039] NN_Put registers are used when the previous neighboring programming engine executes an instruction with NN_Put as a destination. The NN register selected by the value in this register is written, and the value in NN_Put is then incremented (a value of 127 wraps back to 0). The value in this register is compared to the value in NN_Get register to determine when to assert NN_Full and NN_Empty status signals.

[0040] NN_Get registers are used when the NN register 35 is accessed as a source, which is specified in the source field of the instruction. The NN register selected by the value in this register is read, and the value in NN_Put is then decremented (a value of 127 wraps back to 0). The value in this register is compared to the value in the NN_Put register to determine when to assert NN_Full and NN_Empty status signals.

[0041] Specifically, when each NN register 35 is used as a destination in an instruction, the instruction result data are sent out of the programming engine 16a, typically to another, adjacent programming engine. On the other hand, when the NN register 35 is used as a destination in an instruction, the instruction result data are written to the selected NN Register 35 in the programming engine 16a. The data are not sent out of the programming engine 16a as it would be when each NN register 35 is used as a destination. Each NN register 35 is used in a context pipelining method, as described below.

[0042] A local memory 42 is also used. The local memory 42 includes addressable storage located in the programming engine 16a. The local memory 42 is read and written exclusively under program control. The local memory 42 also includes variables shared by all the programming engines 16. Shared variables are modified in various assigned tasks during functional pipeline stages by the programming engines 16a-16h, which are described next. The shared variables include a critical section, defining the read-modify-write times. The implementation and use of the critical section in the computing processing system 10 is also described below.

[0043] Functional Pipelining and Pipeline Stages

[0044] Referring to FIG. 3, the programming engine 16a is shown in a functional pipeline unit 50. The functional pipeline unit 50 includes the programming engine 16a and a data unit 52 that includes data, operated on by the programming engine, e.g., network packets 54. The programming engine 16a is shown having a local register unit 56. The local register unit 56 stores information from the data packets 54.

[0045] In the functional pipeline unit 50, the contexts 58 of the programming engines 16a, namely, Programming Engine0.1 (PE0.1) through Programming Engine0.n (PEO.n), remain with the programming engine 16a while different functions are performed on the data packets 54 as time 66 progresses from time0 to timet. A programming execution time is divided into m functional pipeline stages or pipe-stages 60a-60m. Each pipeline stage of the pipeline stages 60a-60m performs different pipeline functions 62a, 64, or 62p on data in the pipeline.

[0046] The pipeline stage 60a is, for example, a regular time interval within which a particular processing function, e.g., the function 62a is applied to one of the data packets 54. A processing function 62 can last one or more pipelines stages 60. The function 64, for example, lasts two pipeline stages, namely pipeline stages 60b and 60c.

[0047] A single programming engine such as the programming engine 16a can constitute a functional pipeline unit 50. In the functional pipeline unit 50, the functions 62a, 64, and 62p move through the functional pipeline unit 50 from one programming engine (e.g., programming engine 16a), to another programming engine (e.g., programming engine 16b), as will be described next.

[0048] Referring to FIG. 4, the data packets 54 are assigned to programming engine contexts 58 in order. Thus, if n threads or contexts 58 execute in the programming engine 16a, the first context 58, PE0.1 completes processing of the data packet 54 before the data packets 54 from the PE0.n context arrives. With this approach the programming engine 16b can begin processing the n1 packet.

[0049] Dividing the execution time of the programming engine 16a, for example, into functional pipeline stages 60a-60c results in more than one of the programming engines 16 executing an equivalent functional pipeline unit 70 in parallel. The functional pipeline stage 60a is distributed across two programming engines 16a and 16b, with each of the programming engines 16a and 16b executing eight contexts each.

[0050] In operation, each of the data packets 54 remains with one of the contexts 58 for a longer period of time as more programming engines 16 are added to the functional pipeline units 50 and 70. In this example, the data packet 54 remains with a context sixteen data packet arrival times (8 contexts2 programming engines) because context PE0.1 is not required to accept another data packet 58 until the other contexts 58 have received their data packets.

[0051] In this example, function 62a of the functional pipeline stage 60a can be passed from the programming engine 16a to the programming engine 16b. Passing of the function 62a is accomplished by using Next Neighbor registers, as illustrated by dotted lines 80a-80c in FIG. 4.

[0052] The number of functional pipeline stages 60a-60m is equal to the number of the programming engines 16a and 16b in the functional pipeline units 50 and 70. This ensures that a particular pipeline stage executes in only one programming engine 16 at any one time.

[0053] Context Pipelining:

[0054] Each of the programming engine 16 supports multi-threaded execution of eight contexts. One reason for this is to allow one thread to start executing just after another thread issues a memory reference and must wait until that reference completes before doing more work. This behavior is critical to maintaining efficient hardware execution of the programming engines 16a-16f because memory latency is significant. Stated differently, if only a single thread execution was supported, the programming engine would sit idle for a significant number of cycles waiting for references to complete and thereby reduce overall computational throughput. Multi-threaded execution allows a programming engine to hide memory latency by performing useful independent work across several threads.

[0055] The programming engines 16a-16h (FIG. 1) each have eight available contexts. To allow for efficient context swapping, each of the eight contexts in the programming engine has its own register set, program counter, and context specific local registers. Having a copy per context eliminates the need to move context specific information to/from shared memory and programming engine registers for each context swap. Fast context swapping allows a context to do computation while other contexts wait for I/O, typically external memory accesses, to complete or for a signal from another context or hardware unit.

[0056] Referring now to FIG. 5, the context for a specific assigned task is maintained on the programming engines 16a-16c using CAM 45a-45c. The packets are processed in a pipelined fashion similar to an assembly line using NN registers 35a-35c to pass data from one programming engine to a subsequent, adjacent programming engine. Data are passed from one stage 90a to a subsequent stage 90b and then from stage 90b to stage 90c of the pipeline, and so forth. In other words, data are passed to the next stage of the pipeline allowing the steps in the processor cycle to overlap. In particular, while one instruction is being executed, the next instruction can be fetched, which means that more than one instruction can be in the pipe at any one time, each at a different stage of being processed.

[0057] For example, data can be passed forward from one programming engine 16 to the next programming engine 16 in the pipeline using the NN registers 35a-35c, as illustrated by example in FIG. 5. This method of implementing pipelined processing has the advantage that the information included in CAM 45a-45c for each stage 90a-c is consistently valid for all eight contexts of the pipeline stage. The context pipeline method may be utilized when minimal data from the packet being processed must advance through the context pipeline.

[0058] Referring to FIG. 6, as described above, context pipelining requires that the data resulting from a pipe stage, such as pipe stage P, be sent to the next pipe stage, e.g., pipe stage P1 (100). Then, Next Neighbor registers can be written from the ALU output of the processing engine 16a in pipe stage P (102), and the Next Neighbor registers can be read as a source operand by the next programming engine 16b at the pipe stage P1 (104).

[0059] Referring to FIG. 7, two processes may be used to determine the address of the Next Neighbor registers to be written in the programming engine 16b. In one process, each context of the programming engine 16a may write to the same Next Neighbor registers for the same context in programming engine 16b (200). In another method, a write pointer register in the programming engine 16a and a read pointer register in the programming engine 16b may be used (300) to implement an inter processing engine FIFO (302). The values of write pointer register in the programming engine 16a and the read pointer register in the programming engine 16b are used to produce a full indication checked by the programming engine 16 before inserting data onto the FIFO (304), and an empty indication may be used the programming engine 16b before removing data from the FIFO (306). The FIFO Next Neighbor configuration may provide the elasticity between contexts in the pipe stages P and P1. When a context in the pipe stage P1 finds the Next Neighbor FIFO is empty, that context can perform a No-op function, allowing the pipe stage to maintain a predetermined execution rate or beat even if the previous pipe stage may not be supplying an input at this same rate.

[0060] Other Embodiments:

[0061] In the examples described above in conjunction with FIGS. 1-7, the computer processing system 10 may implement programming engines 16 using a variety of network processors.

[0062] It is to be understood that while the invention has been described in conjunction with the detailed description thereof, the foregoing description is intended to illustrate and not limit the scope of the invention, which is defined by the scope of the appended claims. Other aspects, advantages, and modifications are within the scope of the following claims.