Parallel task operation in processor and reconfigurable coprocessor configured based on information in link list including termination information for synchronization

Description

FIELD OF THE INVENTION

The present invention relates to reconfigurable processors. In particular, the present invention addresses connecting a reconfigurable processor to a standard processor in a particularly favorable manner.

BACKGROUND INFORMATION

A reconfigurable architecture is understood in the present case to be modules (VPUs) having a configurable function and/or interconnection, in particular integrated modules having a plurality of arithmetic and/or logic and/or analog and/or memory and/or internal/external interconnecting units that are configured in one or more dimensions and are interconnected directly or via a bus system.

Generic modules of this type include in particular systolic arrays, neural networks, multiprocessor systems, processors having a plurality of arithmetic units and/or logic cells and/or communicative/peripheral cells (IO), interconnecting and networking modules such as crossbar switches as well as conventional modules of the generic types FPGA, DPGA, Chameleon, XPUTER, etc. Reference is made in this context in particular to the following protective rights of the present applicant: P 44 16 881.0-53, DE 197 81 412.3, DE 197 81 483.2, DE 196 54 846.2-53, DE 196 54 593.5-53, DE 197 04 044.6-53, DE 198 80 129.7, DE 198 61 088.2-53, DE 199 80 312.9, PCT/DE 00/01869, DE 100 36 627.9-33, DE 100 28 397.7, DE 101 10 530.4, DE 101 11 014.6, PCT/EP 00/10516, EP 01 102 674.7, DE 196 51 075.9-53, DE 196 54 846.2-53, DE 196 54 593.5-53, DE 197 04 728.9, DE 198 07 872.2, DE 101 39 170.6, DE 199 26 538.0, DE 101 42 904.5, DE 101 10 530.4, DE 102 02 044.2, DE 102 06 857.7, DE 101 35 210.7-53, EP 02 001 331.4, 60/317,876. These are herewith incorporated to the full extent for disclosure purposes.

The aforementioned architecture is used as an example for illustration and is referred to below as VPU. This architecture is composed of any arithmetic cells, logic cells (including memories) and/or memory cells and/or networking cells and/or communicative/peripheral (IO) cells (PAEs) which may be arranged to form a one-dimensional or multidimensional matrix (PAC), the matrix optionally having different cells of any type; bus systems are also understood to be cells here. The matrix as a whole or parts thereof are assigned a configuration unit (CT) which influences the interconnection and function of the PA.

SUMMARY

A reconfigurable processor (VPU) is designed into a technical environment having a standard processor (CPU) such as a DSP, RISC or CISC processor or a (micro)controller. This design permits the simplest possible connection, which is nevertheless very efficient. Another feature is the simple programming of the resulting system. Continued use of existing programs of the CPU and code compatibility and simple integration of the VPU into existing programs with no problem are taken into account by the method described here.

Reconfigurable modules (VPUs) of different generic types (such as PACT XPP technology, Morphics, Morphosys, Chameleon) are generally incompatible with existing technical environments and programming methods.

The programs of the modules are also incompatible with pre-existing CPU programs. This necessitates enormous developing effort in programming, e.g., in particular for modules of the generic Morphics and Morphosys types. Chameleon already has a standard processor (ARC) integrated into the reconfigurable modules. Thus, the tools for programming are available. However, not all technical environments are suitable for use of ARC processors, and in particular existing programs, code libraries, etc. are often provided for any indeterminate other CPUs.

In accordance with the present invention, VPU (or a plurality of VPUs, without having to mention this specifically each time) is connected to a preferred CPU in such a way that it assumes the position and function of a coprocessor. The function as coprocessor permits the simple tie-in to existing program codes according to the pre-existing methods for handling coprocessors according to the related art.

This system may be designed in particular as a (standard) processor or unit and/or integrated into a semiconductor (system on chip, SoC).

In order to provide the coprocessor link between the CPU and the VPU, an exchange of data, i.e., information between the CPU and VPU, is necessary. In particular, the processor must typically relay data and instructions about what must be done to the data. The data exchange between the CPU and VPU may take place via memory linkage and/or IO linkage. The CPU and VPU may in principle share all the resources. In particular embodiments, however, it is also possible for the CPU and VPU to jointly use only some of the resources, while other resources are available explicitly and exclusively for the CPU or the VPU. The question of which variant is preferred will typically depend on, among other things, the overall layout of the system, the possible cost, available resources and the expected data load. It should be pointed out that whenever reference is made to a single CPU, this may also be understood to refer to a plurality of CPUs together.

To perform a data exchange, data records and/or configurations may be copied and/or written/read in memory areas provided specifically for this purpose and/or corresponding basic addresses may be set so that they point to the particular data ranges.

In one preferred variant, for controlling the coprocessor, a data record containing the basic settings of a VPU such as, for example, certain basic addresses, is provided. In addition, status variables may also be provided in the data record for triggering and function control of a VPU by a CPU and/or for separate transmission and may be exchanged with or separately from data. In a particularly preferred variant, the addresses may be flexibly distributed and allocated. Thus preferably only one basic address in the I/O address space or the memory address space need be fixedly agreed upon to be used with its data record as a pointer to the flexibly defined addresses.

The data record may be exchanged via a common memory (RAM) and/or a common peripheral address base (IO). The addresses may be flexibly distributed and allocated.

For synchronization of the CPU and VPU, unidirectional or mutual interrupt methods (e.g., interrupt lines) may be provided and/or synchronization may be performed via polling methods. In addition, interrupts may also be used for synchronizing data transfers and/or DMA transfers. In one embodiment that is particularly preferred, a VPU is started by a CPU and then independently thereof it runs the application which has been started, i.e., instructed.

A preferred structure in which the VPU used provides its own mechanisms for loading and controlling configurations is particularly efficient. For example, PACT XPP and Chameleon belong to the generic type of these VPUs. The circuits according to the present invention permit a method for operation so that some or all configurations of the VPU together with the program of the CPU to be executed are loaded into a memory. During execution of the program, the CPU may refer to the memory locations (e.g., via addresses or pointers), each containing the particular configurations to be executed. The VPU may then automatically load the configurations without any further influence by the CPU. If and to the extent that the VPU, i.e., the reconfigurable field having particularly coarse-grained runtime-configurable elements, has a load logic for loading configurations, it may be sufficient if the processor issues instructions to the VPU to load a certain configuration. The call to the reconfigurable processor, which then functions as the coprocessor, may thus preferably be issued via a single instruction to the load logic. It should be pointed out that by prior agreement between the VPU and CPU, i.e., the calling host processor, it is possible to stipulate precisely which configuration is to be executed by which call. It should be pointed out here that suitable control means may be provided in the load logic unit, whether dedicated, implemented or formed by one or more reconfigurable cells of the reconfigurable processor. Execution begins immediately or, if necessary, is begun via additional information (e.g., interrupt and/or start instructions) by the CPU.

In a particularly preferred further embodiment, the VPU is able to independently read and write data within one or more memories, some of which may be shared with or independent of the CPU.

In a particularly preferred further embodiment, the VPU may also independently load new configurations out of the memory and reconfigure them as needed without requiring any additional influence by the CPU.

These embodiments permit operation of VPUs mostly independently of CPUs. Only synchronization exchange between the CPU and VPU, which is preferably bidirectional, should additionally be provided to coordinate the data processing and/or configuration execution sequences.

The sequence control of a VPU may be accomplished directly by a program executed on the CPU, which more or less constitutes the main program which swaps out certain subprograms to the VPU. This variant is particularly easy to implement.

However, mechanisms controlled via the operating system (in particular the scheduler) are preferably used for synchronization and sequence control. Whenever possible, a simple scheduler in particular may perform the following after transfer of the function to the VPU:

• • 1. allow the current main program to continue running on the CPU if it is able to run independently on a VPU and simultaneously with data processing; additionally and/or alternatively,
  • 2. if or as soon as the main program must wait for the end of data processing on the VPU, the task scheduler switches to another task (e.g., another main program). The VPU may continue working in the background regardless of the CPU task currently at hand.

Each newly activated task will typically (if it uses the VPU) check before use to determine whether it is available for data processing or whether it is currently still processing data in a manner which blocks the required VPU resources. It is then necessary to wait either for the end of data processing or, if preferable according to priority, for example, the task must be changed.

A simple and nevertheless efficient method may be created and/or implemented in particular on the basis of so-called descriptor tables which may be implemented as follows, for example:

For calling the VPU, each task generates one or more tables (VPUPROC) having a suitable defined data format in the memory area assigned to it. This table includes all the control information for a VPU, e.g., the program/configuration to be executed (or pointers to the corresponding memory locations) and/or memory location(s) (or pointers to each) and/or data sources (or pointers thereto) for the input data and/or the memory location(s) (or pointers thereto) for the operand or the result data.

For example, a table or a chained list (LINKLIST) may be found in the memory area of the operating system, pointing to all VPUPROC tables in the order in which they are created and/or called.

Data processing on the VPU is preferably performed so that a main program creates a VPUPROC and calls the VPU via the operating system. The operating system creates an entry in the LINKLIST. The VPU processes the LINKLIST and executes the particular VPUPROC referenced. The end of a particular data processing is preferably indicated by a corresponding entry in the LINKLIST and/or VPUCALL table which the VPU may query by regular polling, for example. As an alternative, interrupts may be used as an indicator from the VPU to the CPU and if necessary may also be used for exchanging the VPU status. It is not only possible here to indicate the fact that the end of the program has been reached but it is also possible to indicate the fact that a point in the subprogram has already been reached, and if so, which point.

In this method, which is preferred according to the present invention, the VPU works largely independently of the CPU. In particular, the CPU and the VPU may perform independent and different tasks per unit of time. The operating system and/or the particular tasks need only monitor the tables (LINKLIST and/or VPUPROC).

As an alternative, LINKLIST may also be omitted by chaining the VPUPROCs to one another using pointers, as is known from lists, for example. VPUPROCs that have been processed are removed from the list and new ones are inserted. Programmers are familiar with this method which therefore need not be explained in greater detail here.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows an example VPU.

FIG. 2 shows an example CPU system.

FIG. 3 shows an exemplary system.

FIG. 4 shows an example interface structure.

DETAILED DESCRIPTION OF EXAMPLE EMBODIMENTS

FIG. 1 shows a particularly preferred VPU design. Configuration managers (CTs) (0101), preferably hierarchical, control and administer a system of reconfigurable elements (PACs) (0102). The CTs are assigned a local memory for configurations (0103). The memory also has an interface (0104) to a global memory which supplies the configuration data. The configuration sequences are controllable via an interface (0105). There is an interface of reconfigurable elements (0102) to sequence control and event management (0106); there is also an interface to data exchange (0107).

FIG. 2 shows a detail of an exemplary CPU system, e.g., a DSP of the C6000 type from Texas Instruments or a microcontroller from ARM (0201). Program memories (0202), data memories (0203), any peripherals (0204) and an EMIF (0205) are shown. A VPU is integrated (0208) as a coprocessor via a memory bus (0206) and a peripheral bus (0207). A DMA controller (EDMA) (0209) may perform any DMA transfers, e.g., between memory (0203) and VPU (0208) or memory (0203) and peripherals (0204). The VPU and/or the CPU may also access the memory independently without the assistance of a DMA. The shared memory may be also be designed as a dual port memory or multiport memory in particular. Additional units, in particular reconfigurable FPGAs, may be assigned to the system to permit fine-grained processing of individual signals or data bits and/or to be able to establish flexible adaptable interfaces (e.g., various serial interfaces (V24, USB, etc.)), various parallel interfaces (hard drive interfaces, Ethernet, telecommunications interfaces (a/b, TO, ISDN, DSL, etc.)).

FIG. 3 shows a more abstract system definition. A CPU (0301) is assigned a memory (0302) to which it has read and/or write access. A VPU (0303) is connected to the memory. The VPU is divided into a CT part (0309) and the reconfigurable elements for data processing (0310).

To increase memory accesses, the memory may have a plurality of independent access buses (multiport) which under some circumstances may be used simultaneously. In a particularly preferred embodiment, the memory is segmented into multiple independent segments (memory banks), each bank optionally being accessed independently. All segments are preferably within a uniform address space.

One segment is preferably available mainly for CPU (0304), another segment is available mainly for data processing by VPU (0305) and yet another segment is mainly available for the configuration data of VPU (0306).

Typically and preferably a fully embodied VPU has its own address generators and/or DMAs to perform data transfers. Alternatively and/or additionally, it is possible for a DMA (0307) to be provided inside the system (FIG. 3) for data transfers with the VPU.

The system contains IO means (0308) to which the CPU and VPU may have access.

Both the CPU and VPU may have dedicated memory areas and IO areas to which the other does not have access.

A data record (0311) which may be in the memory area and/or in the IO area and/or partially in one of the two, as shown graphically, is used for communication between the CPU and VPU, e.g., for exchange of basic parameters and control information. The data record may include the following information, for example, and thus constitutes a basic settings data record:

• • 1. Basic address(es) of the CT memory area in 0306 for localizing the configurations,
  • 2. Basic address(es) of data transfers with 0305,
  • 3. I/O addresses of data transfers with 0308,
  • 4. Synchronization information, e.g., reset, stop, starting the VPU,
  • 5. Status information on the VPU, e.g., error or status of data processing.

The CPU and VPU are synchronized by polling status data and/or status information and/or preferably by interrupt control (0312).

The basic setting data record may contain a LINKLIST and/or VPUCALLs or alternatively may point to the LINKLIST and/or VPUCALLs or to the first entry thereof by pointers.

FIG. 4 shows an example embodiment of the interface structure of a VPU for tying into a system like that in FIG. 3. The VPU here is assigned a memory/DMA interface and/or an IO interface for data transfer (0401). Another system interface (0402) takes over sequence control such as the management of interrupts, starting and stopping processing, exchange of error states, etc.

The memory/DMA interface and/or an IO interface is connected to a memory bus and/or an IO bus.

The system interface is preferably connected to an IO bus, but alternatively or additionally according to 0311 it may also be connected to a memory. Interfaces (0401, 0402) may be designed for adaptation of different working frequencies of the CPU and/or VPU and/or system and may have a clock matching circuit; for example, the system, i.e., the CPU, may operate at 400 MHz and the VPU at 200 MHz.

The interfaces may translate the bus protocols using a protocol matching circuit, e.g., the VPU-internal protocol may be converted to an external AMBA bus protocol or vice versa.

The memory/DMA interface and/or IO interface supports the memory access of the CT to an external memory which is preferably direct (memory mapped). The data transfer of the CT(s) and/or PAC(s) may be buffered, e.g., via FIFO stages. The external memory (e.g., 0308, 0203) may be addressed directly, and DMA-internal and/or external DMA transfers may be performed.

Data processing, e.g., initializing and/or startup of configurations, is controlled via the system interface. In addition, status and/or error states are exchanged. Interrupts for the control and synchronization between the CTs and a CPU may be supported.

The system interface may convert VPU-internal protocols to be implemented on external (standard) protocols (e.g., AMBA).

It should be pointed out that bus interfaces, RAM cells, I/O cells and the like may be provided as parts (PAEs) of a VPU. This is also true when these units are to be used for processor-coprocessor linkage.

A preferred method for generating code for the system described here is described in the PACT20 patent application (i.e., U.S. patent application Ser. No. 09/967,498), the full content of which is herewith incorporated for disclosure purposes. This method includes a compiler which splits the program code into a CPU code and a VPU code. The split between the different processors is performed by various methods. In one particularly preferred embodiment, the particular split codes are expanded by adding interface routines for communication between CPU and VPU. This expansion may also be performed automatically by the compiler.

The advantage according to the present invention is that management complexity and/or interface complexity as well as programming of the system according to the present invention are simple and inexpensive.

The following tables show examples of communications between a CPU and a VPU. The particular active function units are assigned to the columns (CPU, system DMA and DMA interface (EDMA), i.e., memory interface (memory IF), system interface (system IF, 0402), CTs and the PAC). The rows show the individual cycles in order of execution. K1 references configuration 1, which is to be executed.

The first table shows a sequence using the DMA (EDMA) system for data transfer as an example. Each row indicates a control process taking place sequentially. The columns show the particular activity in the corresponding module:

It should be pointed out that the EDMA and VPU are automatically synchronized via interface 0401, i.e., DMA transfers take place only when the VPU is ready for it.

A second table shows a preferred optimized sequence as an example. The VPU itself has direct access to configuration memory (0306). In addition, the data transfers are executed by a DMA circuit within the VPU, which may be fixedly implemented, for example (PACT03, i.e., U.S. Pat. No. 6,513,077) and/or result from the configuration of configurable parts of the PAC.

The operating and synchronization complexity for the CPU is minimal, so that maximum performance is achieved.

In addition, according to this method a plurality of configurations may be executed in different areas of the VPU, i.e., in different PAEs or on the same resources by using a time multiplexing method.

In particular, a type of double buffering may be used for particularly simple and rapid reconfiguration in which a plurality of VPUs are provided, some optionally being reconfigured at a point in time of the VPUs, while others perform computations and possibly yet others may be inactive. The data connections, trigger connections, status connections, etc. are exchanged in a suitable way among the plurality of VPUs and optionally interconnected through addressed buses and/or multiplexers/demultiplexers according to the VPUs currently active and/or to be reconfigured.

The full content of all the PACT patent applications identified above as well as their family members is herewith incorporated for disclosure purposes.

Other further embodiments and combinations of the present inventions mentioned above are, of course, possible. In this regard, it should be pointed out in particular that instead of connecting a VPU to a CPU using the VPU as the coprocessor, such a connection is also possible using the CPU as the coprocessor. Such a case is preferred in particular to have instruction structures recognized as having only minor parallelism and/or minor vector components processed sequentially as program parts in compiling. It is then possible in particular for the VPU to call the CPU via linklists or tables. The linklists or tables may contain information indicating where data is to be retrieved, at which address the CPU is able to access program information to be processed by it, etc. The inquiry as to whether the CPU is then finished with processing the program parts to be executed by it may in turn be handled via polling or the like. Here again, the operating system may be used to assign tasks to the CPU and/or to monitor the tasks to be executed by it. In principle, all the methods described here may thus be used for both linking a CPU to a VPU as a coprocessor as well as the converse. The only thing that may be important here is which type of linkage the operating system is designed for. It should be pointed out that it is possible in particular to provide an operating system that permits mutual linkage, i.e., in particular, optionally the CPU to the VPU and/or parts thereof and the converse. The latter is particularly advantageous when entire program blocks having mainly sequential portions are to be delivered by the VPU as host to the CPU as coprocessor and these program blocks still have strongly vectorial or parallel code in some cases which may be more or less transmitted back by the CPU, in particular in response to a current or predicted VPU load that has been determined.