System and method for the management of a multiple-resource architecture

Abstract

In the field of programmable integrated circuits designed for applications of high-powered computation, a compact system of low complexity is proposed to implement a static schedule scheme for real-time resource management, capable of taking account of the completion of a task. This system, used for the management of a multiple-resource architecture in which several resources can simultaneously execute different tasks, comprises: at least one management means (Gij) for the execution of a part (j, 0≦j) of a set or subset (i, 0≦i) of one or more tasks; at least one means of control (Ci) of the means or plurality of means of management (Gij) of the execution of said set or subset of task(s); at least one means of verifying the completion of one or more given tasks, each completion verification means being associated with all or part of the management means (Gij) and/or all or part of the control means (Ci) and/or all or part of the resources; and/or an initialization means used to initialize all or part of the preceding means.

Description

BACKGROUND OF THE INVENTION

[0001] 1. Field of the Invention

[0002] The invention relates to the field of digital computation programmable integrated circuits designed for high computing power applications, for example in applications characterized by algorithms based on imbricated loops whose running depends little or not at all on the data to be processed. This type of application can be encountered especially in the first stages of digital signal processing.

[0003] The circuits considered contain several physical resources capable of executing a microprogram that implements a special function such as, for example, a computation (the performance of microprogrammed arithmetic operations on a set of data) or a data transfer (the conveyance of a set of data from one memorization zone to another zone). The physical resources work in parallel, each executing a microprogram sequentially, or as the case may be, in a loop. There are two modes of management of the execution of a program by these physical resources.

[0004] In the first mode, a single program is run commonly. A single instruction containing a field (a micro-instruction) for each of the physical resources is active at each elementary cycle. This is the case with the VLIW (Very Long Instruction Word) architectures.

[0005] Alternatively, the microprograms are specific to the physical resources, thus enabling the resources to execute and especially to repeat certain sequences independently and at their own pace.

[0006] The second mode is the one chosen by most digital signal processors (DSPs) because it allows the architecture to have high flexibility. The main resource is the operative part (the one carrying out arithmetic operations). It works at its own pace in parallel with the data transfer resources known as DMA (Dynamic Memory Access) controllers. Furthermore, it is assumed that the resources can be downloaded.

[0007] As can be seen in FIG. 1, each resource is capable of initially memorizing a set of microcommands or micro-instructions of its own, sufficient for the execution of its task. External intervention is not necessary except to provide these microcommands and initialize their execution. Each resource then works independently with respect to other resources; in return, it may report the end of execution of the task for which it has been configured. The micro-instruction memorization unit in each resource can be seen as a cache memory used to reduce the mean input bit rate of the microcommands coming from outside the circuit (this is a common difficulty with high-powered computation circuits and is related to the passband of the program).

[0008] In order to carry out the parallel execution of tasks by several resources, a device for managing the execution of tasks by resources according to a static scheduling scheme is implemented. The management of the execution by the resources of the microprograms is done by sequencing at one or two levels.

[0009] The standard approach used especially in digital signal microprocessors (DSPs) consists in including scheduling directives in the operative part of the program. The sequencing is then done only at one level. The microprograms are specific to each resource, which leaves them free to be executed at their own pace (especially for the repeating of certain sequences independently of their neighbors).

[0010] As can be seen in FIG. 1, the microsequencer proper to a given resource then runs the microprogram of this resource independently with respect to the other resources. Most of the instructions are designed for the computation resource. Some of them are used to parameterize the transfer resources. The main drawback of the sequencing at one level is the simultaneous management, by the microsequencers of the different resources, of numerous activities each producing events (end of an iteration, end of loop, etc.) desynchronized with the activity of the others. This gives rise to the need to be able to interrupt an activity in order to launch another one or take account of the completion of a task, giving rise to a relatively complex logic (management of interruptions, saving of context, etc.) inducing high design costs.

[0011] The applications envisaged by the invention are simple. They therefore provide the hope of simpler, high-performance circuit embodiments. This is why two distinct levels of sequencing are introduced:

[0012] a fast sequencing level for each of the resources that do not manage asynchronous events;

[0013] a total sequencing level controlling the downloading of the microprogram in each of the resources and their launching. The use of this second level provides for a reduction of the microprogram bit rate because the downloading can be done in parallel without slowing down the activity of the resources). This total sequence level may be more or less elaborated.

[0014] One elaboration consists in running a program by forcing the task launching times without taking account of the completion messages. This is direct launching at predetermined instants. This preparation is simpler and more reactive but it accepts no variation of the task execution time and becomes a problem in the case of multimode operations.

[0015] Another preparation consists of the use of a processing core type structure but this is a bulky and less reactive approach.

[0016] The present invention is used to overcome or at least reduce these drawbacks by proposing a compact system of low complexity to implement a static schedule scheme for the real-time resource management capable of taking account of the completion of a task.

SUMMARY OF THE INVENTION

[0017] To this end, an object of the invention is a system used for the management of a multiple-resource architecture in which several resources can simultaneously execute different tasks, wherein the method comprises:

[0018] at least one means for the management of the execution of a part of a set or subset of one or more tasks,

[0019] at least one means of control of the means or plurality of means of management of the execution of said set or subset of tasks,

[0020] at least one means of verifying the completion of one or more given tasks, each completion verification means being associated with all or part of the management means and/or all or part of the control means and/or all or part of the resources,

[0021] and/or an initialization means used to initialize all or part of the preceding means.

[0022] The invention relates to a programmable integrated circuit comprising a single program memory comprising:

[0023] an input receiving the address signals sent by said system, and

[0024] a output transmitting the tasks to be executed to said system and to the resources of said circuit.

[0025] The system of management of a multiple-resource architecture proposed by the present invention, hereinafter called an organizer, uses a method for the management of a multiple-resource architecture in which several resources can simultaneously execute different tasks, wherein the system comprises at least the following steps:

[0026] a the management of the execution of at least one part of a set or subset of one or more tasks,

[0027] the control of the management of each part or at least one set or subset of one or more tasks,

[0028] the verification of the completion of the execution of at least one task during the management and/or control and/or execution by the resources,

[0029] the initialization of all or part of the above steps.

BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The characteristics and advantages of the invention shall appear more clearly from the following description, given by way of an example and from the appended figures, of which:

[0031] FIG. 1 a basic drawing of a multiple-resource architecture according to the prior art,

[0032] FIG. 2 is an exemplary basic diagram of a multiple-resource architecture with two sequencing levels according to the invention,

[0033] FIG. 3 is an exemplary diagrammatic view of the organizer according to the invention,

[0034] FIG. 4 shows an exemplary architecture of the organizer according to the invention,

[0035] FIG. 5 shows an exemplary detailed embodiment of a part of the organizer of FIG. 4,

[0036] FIG. 6 exemplifies the principle of operation of an organizer according to the invention,

[0037] FIG. 7 exemplifies the detailed operation of the organizer during the performance of a scheduling directive,

[0038] FIG. 8 is an exemplary detailed operation of the organizer during the passage to the following iteration.

MORE DETAILED DESCRIPTION

[0039] FIG. 2 is an exemplary basic drawing of a multiple-resource architecture according to the invention. The initialization order 1 induces the sending, by the organizer 2, of a signal 3 containing an address to a single program memory 4. The code stored in the program memory 4 is formed out by:

[0040] code sequences specific to the resources for which thy are intended,

[0041] scheduling directives that reflect the scheduling pattern of the activities, intended for the organizer 2.

[0042] The task 5 that is at the address 3 is therefore executed either by one of the resources 6 if it is a processing instruction or by the organizer 2 if it is a scheduling directive. When a resource 6 has terminated the execution of the instruction, it reports the completion 7 of this instruction to the organizer 2.

[0043] FIG. 3 shows an exemplary schematic drawing of the organizer 2 according to the invention. The initialization command 1 is received through the input E1 by the initialization means 21 which together initialize the means of the organizer 2 (not shown in the figure) and send the first address 3 to the address output S3. The task to be executed 5 read at this address 3 is received by the resources and the organizer 2 at its input E5 and transmitted i to the control means Ci of the means Gij for the management of the execution of one or more of the parts j of a set or subset of task(s) i. The tasks 5 contain, for example, either the indication of their addressee or the indication of the type (computation, transfer, scheduling). During the execution of a part j of a set or subsets of task(s) i, the management means Gij sends an address 3 to the output S3. A completion verification means 8 ascertains that the completion awaited by the management means Gi′j′, namely the end of execution of one or more tasks expected by the management means Gi′j′; is the completion reported by one of the resources 6 or one of the management means Gij or one of the control means Ci.

[0044] Seen from a particular resource 6, this resource downloads the micro-instructions that reach it and activates their execution as soon as the launching order reaches it. These tasks 5 (micro-instructions and launching command) are read sequentially in the program memory 4 from the address transmitted by the organizer 2. The role of the organizer 2 is to manage the corresponding sequence of addresses for the right number of times if it is a loop, and to do so at the right times. The scheduling directives enable it to initialize and carry out the parallel management of as many program address pointers as there are parts j in the set or subset of tasks i.

[0045] Let us take the example of the management of a multiple-resource architecture by an organizer 2 for the performance of a set i of tasks forming a program loop i. A management means Gij of a part j of a set or subset i of tasks is then a means of management of the execution of the part or stage j of a main or imbricated program loop i, namely a means of management hereinafter called a LAYER Gij. The control means Ci of the LAYERS Gij is hereinafter called a KERNEL Ci. Each of the KERNELS Ci has a different loop level I corresponding to it. There are therefore several KERNELS Ci in hierarchical levels for the imbricated loops.

[0046] FIG. 4 proposes an exemplary architecture of the organizer 2 applied to the management of the execution of a program loop i. FIG. 5 gives a detailed exemplary embodiment of a KERNEL Ci, of the associated LAYERS Gij of the architecture proposed by FIG. 4.

[0047] For the management of the execution of the iteration number Ni of the loop i according to the initializing parameters of the loop i, the KERNEL Ci has an interaction counter iti that is initialized at the number of iterations of the loop i to be performed iti=Ni and decremented iti=iti−1 at each new iteration of the loop i. When the counter iti reaches the value 0 and as soon as the access arbitrator 27+ for access to the completion bus 27 permits access, the completion signal sender emi sends a completion signal to report the end of the processing of the loop i.

[0048] The authorization of access to the completion bus 27 delivered by the arbitrator 27+ gives the right to the use of the completion signal-conveying bus (hereinafter called a completion bus) 27 during one cycle for sending a completion signal 7. In the case of several simultaneous requests, due to the fact that each resource 6 and each KERNEL Ci that has ended its task wishes to send a completion signal 7, the authorization from the arbitrator 27+ depends on an order of priority that may be fixed.

[0049] The set of LAYERS Gij associated with a KERNEL Ci constitutes a shift register chain. The role of the KERNEL Ci is to manage a loop i, control the associated pipeline (the chain formed by the cascade-connected LAYERS Gij) and synchronize the LAYERS Gij with each other. For this purpose, each KERNEL Ci has a pipeline feed mechanism ci. The set formed by the associated KERNEL Ci and LAYERS Gij is called SEQPAR.

[0050] The information contained in each register of this chain is a state indicator e, a program address counter ad, an indicator a of the state of progress of the pipeline and a dimension indicator (not shown in the figures).

[0051] The organizer 2 then comprises a set of activity registers aij that indicate the state of progress of execution in the steps j of the loop i. Only the LAYERS Gij validated by the number of stages Ji of the loop i can be activated.

[0052] The organizer 2 also has a set of address counters adij that can be incremented. These counters drive each execution of a stage j of the loop i. Each LAYER Gij manages a sequence of addresses A[j]i thus enabling the parallel execution of the stages j of the loop i. Within a LAYER Gij, the running of the program is done by incrementation of the address counter adij=adij+1. The incrementation of the address counter adij=adij+1 is then done during the sending of the signal 3 containing the address A[j]i of the counter adij; to the program address bus 23. The basic address A0i initializes the address counter adij=A0i of the first LAYER Gi0, the address counters adij of the other LAYERS Gij are initialized by the address A[j−1]i contained in the address counter adij−1 of the previous LAYER Gij−1 at the beginning of the new iteration.

[0053] The state indicator ej indicates whether the LAYER Gij is idle, in sending mode, on standby or at end of processing. When a LAYER Gij is active and its state register eij is sending, the arbitrator 23+ of the address bus 23 notes a request for access to this bus 23 for this LAYER Gij.

[0054] The authorization of access to the completion bus 23 delivered by the arbitrator 23+ gives the right to the use of the address bus 23 so long as the requester maintains his request. In the case of several simultaneous requests, due to the fact that each LAYER Gij may wish to send a completion signal 3, the authorization from the arbitrator 23− depends on an order of priority that may be fixed.

[0055] (Observation mode) When a LAYER Gij is active and its state register eij is in the standby state, the completion detector dij of this LAYER Gij observes the completion signals 7 sent on the completion bus 27 by the completion signal senders emi of the KERNELS Ci of the organizer 2 and the resources. When the completion signal 7 received corresponds to the awaited completion code, the state register eij of the LAYER Gij comes out of the standby state.

[0056] (Memorization mode) When a completion signal 7 is received by the completion detector dij of the LAYER Gij while its state register eij is not in a standby state, the completion detector dij memorizes the completion signal 7 received to take it into account when the state register eij goes into a standby state.

[0057] When all the LAYERS Gij of a KERNEL Ci have their state registers eij at the “end” of processing, the pipeline feed mechanism ci activates the shifting of the registers of the chain associated with this KERNEL Ci according to the clock signal 29 given by this KERNEL Ci and decrements the iteration counter iti of this KERNEL Ci. The contents of a register are therefore passed on to the next LAYER Gij+1 when the pipeline i (the shift register string associated with the KERNEL Ci) moves forward.

[0058] The execution of the program loop i can thus be managed by making the data addressing information (contained in the dimension indicator register which is not shown) and the program adij+1=adij follow on throughout this chain. This information may thus also be specific to a given recurrence (multimode programming). Consequently, the codes linked to an iteration are necessarily sequential. Once the basic address A0i has been addressed, the organizer 2 controls all the operations of execution without external intervention.

[0059] Since it has only sequenced registers and counters, the organizer 2 proposes the use of a small volume of material. Furthermore, since it has only one task bus 25, the flow rate of the tasks 5 needed for the scheduling is low.

[0060] In order to initialize the execution of the program loop i, the initializing means 21 receive the command 1 to initialize the means of the organizer 2 as described in the step (S1) of the principle of operation of FIG. 6, given by way of an example. Then, this initialization means 21 sends a signal 3 on the bus 23 containing the first address received from a register loaded beforehand (step S2). The organizer 2 receives the first task 5 read at this first address which comprises a scheduling directive, namely an instruction designed for the organizer (steps S3 and S4). The organizer 2 then executes this directive (S5).

[0061] In the present exemplary embodiment, there are four scheduling directives that reflect the activity schedule pattern:

[0062] loop initializing instruction: it configures one of the KERNELS Ci because it initializes the activity register ai0 in active mode and comprises the initializing parameters of the loop i (basic address A0i and/or number of iterations Ni and/or number of stages Ji of the loop and/or the level of imbrication i of the loop and/or the completion code or codes of the loop i and/or of the stages j of the loop i etc.)

[0063] standby instruction: it identifies the standby at the end of execution of a task and comprises for example the awaited completion code. It therefore makes it possible to wait for a task to end and for the resource 6 performing this task to be free before the next task of a same stage j is executed. It can also be used to program a standby between tasks of two different stages j and j′ of one and the same loop i or of two imbricated loops i and i′. A standby of this type is justified for example by a dependence of data between the tasks or again by the use of a common resource 6,

[0064] boundary instruction between two stages j and j+1: it comprises the completion code of a stage j, for example, and

[0065] end of loop instruction: it indicates the end of execution of a last stage Ji−1 of a loop i.

[0066] The directives described here above by way of an example comprise associated parameters and may give rise to several consecutive words of the program memory 4. One variant would consist of the use, for these three last types of directives, of a field inserted in the initialization instructions and the processing instructions.

[0067] The processing instructions comprise the parameters needed for the task (configuration parameters or microprogram micro-instructions). Indeed, information has to be given to the resource 6 before execution. This information may consist of parameters (number of iterations of a loop, address parameters, etc.) or a microprogram that the resources 6 uses, if necessary, to control the execution. The programming in this case is hierarchical because the program makes use of tasks that themselves use a microprogram.

[0068] FIG. 7 shows an exemplary principle of operation of the organizer 2 during the execution of these scheduling directives.

[0069] (START) The directive read at the first address is an initialization instruction of the loop 0. The execution of this instruction induces the selection of a KERNEL C0 of the organizer 2 (step S53d). The register of activity a00 then goes into active mode and all the other activity registers a0j(j≠0) go into inactive mode (step S54d). And the pipeline feed mechanism c0 of the KERNEL C0 receives the initializing parameters of the loop 9 (step S55d). This pipeline feed mechanism c0 initializes the iteration counter at the number of the iteration to be performed by the loop 0 (it0=N0) and places the state registers e0j of the LAYERS G0j associated with KERNEL C0 in the idle state. Then, it transmits the basic address to the address counter of the LAYER G00 (ad00=A00) (step S56d). This makes the state register e00 of this LAYER G00 go into sending mode and the completion detector d00 of this LAYER G00 go into memorization mode (step S57d).

[0070] (SENDING) The LAYER G00 is then ready to send the address A00, contained in its address counter ad00 (step S8e), on the address bus 23 as soon as it is allowed to do so by the arbitrator 23+ of access to this bus 23 (step S7e). The instruction contained in the address sent A00 is read and executed by the resource 6 or the organizer 2 for which it is intended. The address counter ado gets incremented ad00=ad00+1 (step S9e). So long as the instruction read is not a scheduling directive (step S3-S4), the state register e00 remains in sending mode (step S6), the LAYER G00 keeps its access to the bus 23, a new address A[0]0 contained in the address counter ad00 is sent (step S8e) and the address counter ad00 is incremented (step S9e).

[0071] (STANDBY) If the instruction read is the scheduling directive “standby instruction” (steps S3 to S5), the state register e00 of the LAYER G00 goes into standby mode and its completion detector d00 into observation mode. Upon reception of the completion code awaited by the LAYER G00 from the KERNEL C0, the LAYER G00 goes into a state of sending. This indicates that it is again requesting access to the program address bus 23. The process then takes place in the same way as in the first request for access to the address bus 23 by the LAYER G00.

[0072] (BOUNDARY) If the instruction read is the scheduling directive “boundary between two stages instruction” (steps S3 to S5), the state register e00 of the LAYER G00 goes to “end” of processing.

[0073] (END OF PROCESSING) Since the LAYER G00 is the only active one (either because the loop has only one stage 0 or because the other LAYERS G0j (j≠0) are not activated), the pipeline feed mechanism c0 manages the passage to the next iteration (step S8f). For this purpose, it decrements the iteration counter it0=it0−1 (step S81f and places the state registers e0j of the LAYERS G0j associated with the KERNEL C0 in the idle state. Then it shifts the activity registers a0j=a0j−1(j≧1) and, for the active LAYERS G0j−1, the address counters ad0j=ad0j−1(j≧1) (step S82f). Finally, if the iteration counter it0 has reached the value 0, the activity register a00 of the first LAYER G00 of the KERNEL C0 is inactivated, meaning that this LAYER G00 is in a partial end phase of the loop (step S83f-S84f). Otherwise, the activity register a00 of the first LAYER G00 of the KERNEL C0 is activated (step S83f-S84f) and the pipeline feed mechanism c0 sends the base address A00 to the address counter ad00 of the first LAYER G00 (step S85f). The state registers e0j of the active LAYER G0j go into sending mode and the completion detectors d0j of these LAYERS G0j go into memorization mode (step S86f).

[0074] If the loop 0 has only one stage 0 and the LAYER G00 is in a partial end phase of the loop, i.e. if it is inactive and at the end of the processing phase in the present example, the KERNEL C0 enters an end-of-loop phase comprising the sending of the signal 7 of completion of the loop 0 by the completion signal sender em0 on the completion bus 27 after agreement by the arbitrator 27+ (step S9f-S10f).

[0075] If the loop 0 has several stages j, then the second LAYER G01 is active and the address counter ad01 of this LAYER G01 contains the last address A[0]0 computed by the address counter ad00 of the first LAYER G00. If the loop has several iterations, the first LAYER G00 is still active and the address counter ad00 of this LAYER G00 contains the basic address A00.

[0076] (SENDING) The two LAYERS G00 and G01 are in a sending state, namely they request access to the address bus 23. The arbitrator 23+ gives priority permission for example to the second LAYER G01 (with the order of priority ranging in this case from the last to the first LAYER of a KERNEL Ci). The address counter ad01 of this LAYER G01 then sends its address A[1]0 and gets incremented ad01=ad01+1 (steps S7e to S9e). The instruction read at the address sent A[1]0 is then the first instruction of the second stage of the loop 0 designed for one of the resources 6 or the organizer 2.

[0077] So long as the instruction read is not a scheduling directive (step S3-S4), the state register e01 remains in sending mode (step S6), the LAYER G01 therefore keeps its access to the bus 23, a new address A[1]0 is sent (step S8e) and the address counter ad01 incremented (step S9e).

[0078] (STANDBY) If the instruction read is the “standby instruction” scheduling directive (steps S3 to S5), the state register e01 of the LAYER G01 goes into standby mode and its completion detector d01 into observation mode (step S51a). The arbitrator 22 then matches the access authorization to the address bus 23 with the first LAYER G00 which is then still a requesting layer. Immediately upon reception of the completion code awaited by the second LAYER G01 of the KERNEL C0, the LAYER G01 goes into the sending state (step S7a-S8a). It then again asks for access to the address bus 23. The process then runs in the same way as for the first iteration.

[0079] (BOUNDARY) If the instruction read is the “boundary between two stages instruction” scheduling directive (steps S3 to S5), then the state register e0j(j=0,1) of the LAYER G0j(j=0,1) in the sending phase, namely in the sending state and having permission to access the address bus 23, takes the end-of-processing value (step S51f).

[0080] (END OF PROCESSING) When all the state registers e0j(j=0.1) of the active LAYERS G0j(j=0.1) are at the end of processing, the pipeline feed mechanism c0 can pass to the next iteration (step S7f-S8f). For this purpose, it decrements the iteration counter it0=it0−1 (step S81f), shifts the activity registers a0j=a0j−1(j≧1) and for the active LAYER G0j−1, the address counters ad0j=ad0j−1(j≧1) (step S82f).

[0081] If all the LAYERS G0j(j=0.1) are in the partial end phase, namely that the iteration counter it0 has reached the value 0 and all the LAYERS G0j(j=0,1) are inactive, the execution of the loop 0 is terminated. The KERNEL C0 is then in an end-of-loop phase that includes the sending of the completion signal of the loop 0 on the completion bus 27 by the completion signal sender em0 after authorization by the arbitrator 27+.

[0082] We have deliberately limited the above example to the control by the KERNEL Ci of two LAYERS G1j(j=0,1) but it is clear that, using the same scheme as the one given in this example, the KERNEL Ci can control only one LAYER or more than two LAYERS Gij(jε).

[0083] (START) When a program contains imbricated loops, the instruction read at the address sent A[j]i by a LAYER Gij may be a loop initialization instruction (step S4-S5). In this case, said LAYER Gij in sending phase, namely in a sending state and having access to the program address bus, goes into the standby state and its completion detector d1j goes into the mode of observation of the completion signal of the new KERNEL Ci′ (i′≠i) (step 51d-53d).

[0084] This new KERNEL Ci′ may be selected from among those available, for example as a function of the data elements contained in the field NUMSEQ of the initialization instruction of the loop i′. The loop initializing instruction will work in the same way for the imbricated loop i′ as for the main loop 0. The new KERNEL C1′ could also implement several LAYERS Gi′j′ and work in the same way as the KERNEL C0 controlling the management of the execution of the main loop 0.

[0085] The functions of the organizer 2 according to the invention are numerous. It can for example manage the execution of a sequence of tasks:

[0086] in multimode. Indeed, the organizer 2 may work in multimode because the initializing means 21 can be used to insert an iteration of a loop i′ of a program different from that of the loop i which it was in the process of managing because the multimode operation, as its name indicates, has several modes. For this purpose, it is enough for example that the initialization means should give a different address A00 corresponding to a loop i′ instead of the loop i (see the step S85f). A mode corresponds to a processing chain fixed in advance that is applied to a set of data elements during a time interval called a recurrence. The successive recurrences may bear different modes, for example:

Recurrence N: Recurrence N + 1:
{             {
Loop 0        Loop 2
Loop 1        }
}

[0087] by anticipation. The anticipation function of the organizer 2 is used to anticipate the programming of the resources 6, i.e. to avoid waiting for the end of the execution 7 of the task 5 in a resource 6′ to program the next task for the next resource 6″. Indeed, this makes it possible, when a task 5 has just been launched, to program the resources 6 for the next task. For this purpose, the resource 6″ must accept two memory zones for its program. To program the first task for each LAYER Gij during the previous iteration, the activity register chain may for example be duplicated and the task may be defined.

[0088] in taking account of the iteration to access data. A multidimensional iteration counter, may be used to associate, with each iteration, accesses to multidimensional data tables. This is an example of use of an iteration counter iti to compute a memory address to access different data as a function of the iterations of the loop. The value of the multidimensional counter iti[j] may then be passed from LAYER Gij to LAYER Gij+1 through a shift register just as the activity registers aij and the address counters ad are shifted so that the value truly corresponds to the progress of each LAYER Gij of the chain in the loop i.

[0089] in associating a data buffer with a KERNEL Ci The iteration counter iti of each KERNEL Ci provides for a data buffer management associated with the pipeline (with the chain formed by the KERNEL Ci and the associated LAYERS Gij). This means that the reservation of a memory zones for each iteration of the loop I managed by the organizer 2 is possible. When the iteration is ended, the memory zone can be released for a new iteration either by computing the value of the iteration counter iti modulo the number of LAYERS Gij used or by using a counter working at the same pace as the iteration counter iti of the KERNEL Ci, whose value has passed from LAYER Gij to LAYER Gij+1 through a shift register.

[0090] The list of functions presented here above is not restricted.

[0091] The first variant of the organizer 2 lies in the fact that the LAYERS G are not physically linked to a special KERNEL Ci. Any kernel KERNEL Ci may use any LAYER G by multiplexing the information exchanged between a KERNEL Ci and the LAYERS G.

[0092] The second variant of the organizer 2 groups together the registers contained in the LAYERS Gij in a generalized bank of registers accessible by all the LAYERS Gij. This solution requires that all the activity registers aij and all the state registers adij of the LAYERS Gij associated with a kernel KERNEL Ci should be capable of being read at each cycle.

[0093] In a third variant, the scheduling directives can either be covered by an instruction in the same way as all the instructions that program the resources 6 or else be covered only by instruction fields associated with the other instructions.

Claims

1. A system used for the management of a multiple-resource architecture in which several resources can simultaneously execute different tasks, wherein the method comprises: at least one management means (Gij) for the execution of a part (j, 0≦j) of a set or subset (i, 0≦i) of one or more tasks, at least one means of control (Ci) of the means or plurality of means of management (Gij) of the execution of said set or subset (i) of task(s), at least one means of verifying the completion of one or more given tasks, each completion verification means being associated with all or part of the management means (Gij) and/or all or part of the control means (Ci) and/or all or part of the resources, and/or an initialization means used to initialize all or part of the preceding means.

2. A system according to the preceding claim, wherein each completion verification means comprises: at least one completion signal sender (em) associated with all or part of the management means (Gij) and/or all or part of the control means (Ci), each sender (em) sending a signal (7) indicating the end of execution, the completion, of one or more given tasks, at least one completion detector (d) associated directly or indirectly with all or part of the management means (Gij), each detector (d) receiving completion signals from all the completion signal senders (em) and/or from all the resources.

3. A system according to claim 2 wherein each completion verification means comprises: at least one completion signal sender (emi) associated with the control means (Ci), each sender (si) sending a signal indicating the end of execution of one or more given tasks for which the management of execution is controlled by said control means (Ci), at least one completion detector (dij) directly associated with the management means (Gij), each detector (dij) receiving completion signals from all the completion signal senders (emi) and/or from all the resources (6).

4. A system according to one of the claims 2 or 3 wherein each completion verification means comprises: a single completion bus transmitting all the completion signals sent by each completion signal sender (em) or (emi) and/or each resource to all the completion detectors (d) or (dij), and a means of arbitration of access to said bus.

5. A system according to one of the above claims, comprising: one or more outputs (S3) used to send the address of the task to be executed comprising a link with the initializing means and/or the management means (Gij) and/or the control means (Ci), one or more inputs (E5) to receive the task to be executed comprising a link with all the control means (Ci), and/or one or more inputs (E1) to receive the initializing command and comprising a link with and initializing means and/or one or more inputs to transmit the completion signals from the resources to the completion verification means associated with the control means (Ci) and/or the management means (Gij).

6. A system according to the preceding claim comprising: a single input (E1) to receive the initialization command comprising a link with the initialization means, a single output to send the address of the task to be executed comprising a link with the initialization means and the management means (Gij), a single input (E5) to receive the task to be executed comprising a link with the control means (Ci), and a single input (E7) to receive the completion signals from the resources comprising a link with the completion detectors (dij).

7. A system according to the preceding claim comprising: A single address bus connecting the initialization means and the management means (Gij) to the address output (S3), and A means of arbitration of access to said bus.

8. A system according to one of the above claims wherein, when the initialization means receives an initialization command, said initialization means: sends the first address signal, of the address at which an initialization instruction is located, activates the first management means (Gi0) of each control means (Ci) and deactivates the other management means (Gij, j≠0).

9. A system according to one of the above claims wherein: the address (A[j]i) contained in a management means (Gij) is sent if said management means (Gij) is in sending phase, namely in a state of sending with permission to send and, in that this sending induces the incrementation of said address (A[j]i=A[j]i+1) contained in said management means (Gij).

10. A system according to the preceding claim wherein a management means (Gij) in sending state goes into the sending phase of this state when said arbitration means (23+) for the access to the address bus permits said management means (Gij) to send on this bus.

11. A system according to one of the above claims wherein, when the task to be executed received by the control means (Ci) comprises an instruction for the initialization of a set or subset (i) of tasks: one of the control means (Ci) among those available receives the initialization parameters contained in this initialization instruction comprising the number of parts (Ji) of said set or subset (i) of tasks, the first management means (Gi0) associated with said control means (Ci) receives the basic address (A0i), said first management means (Gi0) goes into a sending state, and the completion verification means (8) associated with said management means (Gi0) goes into completion memorization mode.

12. A system according to the preceding claim wherein, when the task to be executed received by the control means (Ci) comprises an instruction for the initialization of a set or subset (i) of tasks, and when a means for the management (Gij) of the execution of a part (i) of a set or subset (i, i≠i′) of task(s) is in the sending phase, the management means Gij in the sending state goes into a standby state.

13. A system according to one of the above claims wherein, when the task to be executed, received by the control means (Ci), comprises an instruction for placing on standby, the management means Gij in the sending state goes into a standby state.

14. A system according to one of the claims 12 or 13 wherein, when a management means (Gij) goes into a standby state, the completion verification means (8) associated with said management means Gij goes into mode of observation of the awaited completion, namely the completion of the sub-set (i′) of task(s) indicated by the instruction of initialization or of completion of the task or tasks indicated by the standby instruction, it being known that a completion verification means (8) in observation mode performs the following operations: observation of the instances of completion and detection of the awaited completion among those memorized during the memorization mode and those observed, and wherein: when the completion verification means associated with said management means Gij detects the awaited completion, said management means (Gij) goes into a sending state, the completion verification means (8) associated with said management means (Gij) goes into the completion memorization mode.

15. A system according to one of the above claims wherein, when the task to be executed received by the control means (Ci) comprises an instruction indicating the boundary between two parts (j) and (j+1), the management means (Gij) in sending state go into an end-of-processing state.

16. A system according to the preceding claim wherein, if a set or a subset (i) of a task is a program loop (i) and if all the management means or active LAYERS (Gij) of a control means or KERNEL (Ci) are in an end-of-processing state: said KERNEL (Ci) goes from the management of an iteration (iti) to the management of the next iteration (iti+1) of the loop (i), each LAYER (Gij) of said KERNEL (Ci) takes at least a part of the parameters of the previous LAYER (Gij−1) comprising, if said LAYER (Gij) is not in an end-of-processing state of the iteration (Ni) of said loop (i), the address (A[j]i=A[j−1]i). the first LAYER (Gi0) receives the basic address (A0i) if said first LAYER (Gi0) is not in a state of end of processing of the iteration (Ni). of said loop (i), each LAYER (Gi0) having received an address A[j]i goes into a sending state, and the completion verification means associated with the LAYERS (Gij) in a sending state go into a state of memorizing completions.

17. A system according to one of the claims 15 or 16 wherein, when the set of the management means (Gij) are at the end of processing and, in the case of a loop (i), in the state of the end of processing of the iteration (Ni) of said loop (i), the control means (Ci) indicates the completion of the execution of the set or subset (i).

18. A programmable integrated circuit comprising a single program memory comprising: an input receiving the address signals sent by said system, and an output transmitting the tasks to be executed to said system and to the resources of said circuit.

19. A method for the management of a multiple-resource architecture in which several resources can simultaneously execute different tasks, wherein the system comprises at least the following steps: the management of the execution of at least one part j, 0≦j of a set or subset i 0≦I of one or more tasks, the control of the management of each part or at least one set or subset (i) of one or more tasks, the verification of the completion of the execution of at least one task during the management and/or control and/or execution by the resources, the initialization of all or part of the above steps.

20. A method according to the preceding claim, wherein the verification of completion comprises: the sending of a signal indicating the end of execution (completion) of one or more given tasks during the management and/or control, and the detection of the completion signal awaited during the management comprising the reception of the completion signals sent during the management and/or control and/or execution.

21. A method according to one of the claims 19 or 20 wherein the initialization comprises at least the sending of the first address at which an initialization instruction is found.

22. A method according to one of the claims 19 to 21, wherein the management of the execution of a part (j) of a set or subset (i) in the sending phase, namely in the sending state with permission to send, comprises: the sending of the address A[j]i at which the task to be executed is located, the simultaneous passage from the address A[j]i to the address A[j]i+1.

23. A method according to one of the claims 19 to 22 wherein: the control of the management of the execution of the parts (j) of a set or subset (i) of task(s) comprises: the execution of an instruction of initialization of said set or subset (i) of task(s), the reception, during said execution, of the parameters of initialization of said set or subset (i) of task(s) contained in this instruction of initialization, said parameters comprising the number of parts (Ji) of said set or subset (i) of task(s) and the basic address (A0i), the transmission of said basic address A0i received, and wherein the management of the execution of the first part of said set or subset (i) of task(s) comprises: the reception of said basic address A0i transmitted during the control of the management of said set or subset (i) of task(s), the passage to the sending state activated by said reception, the verification of completion comprises the passage into mode of memorization of the instances of completion activated by said passage into sending state.

24. A method according to the preceding claim wherein, when the instruction of initialization of a subset (i′) is executed during the control of the management of the execution of a subset (i′) and when the management of the execution of a part (j) of a set or subset (i, i≠i′) is in a sending phase: the control of management of the execution of said part (j) of said set or subset (i) comprises the execution of said instruction of initialization of the subset (i′), and the management of said part (j) of the set or subset (i) comprises the passage to the standby state activated by the execution, during said control of the management of the execution of the set or subset (i), of said initialization instruction of the subset (i′) whose completion is awaited.

25. A method according to one of the claims 19 to 24 wherein: the control of the management of the execution of a set or subset (i) comprises the execution of an instruction for placing on standby if the management of execution of a part (j) of said set or subset (i) is in sending phase, and wherein: the management of execution of said part (j) of said set or subset (i) in sending phase comprises the passage to the standby state activated by said execution indicating the task or tasks whose completion is awaited.

26. A method according to one of the claims 24 or 25 wherein: the verification of completion comprises the passage into observation mode activated by the passage into standby state of the management of the execution of a part (j) of a set or subset (i) of task(s), the passage into observation mode comprising at least: the observation of the instances of completion and the detection of the awaited completion among those memorized during the memorization mode and those observed, wherein the management of execution of a part (j) of a set or subset (i) of task(s) in sending phase comprises the passage to the standby state activated by said detection of the awaited completion, and wherein the verification of completion comprises the passage into mode of memorization of the instances of completion activated by said passage into sending state.

27. A method according to one of the claims 19 to 26 wherein: if the management of execution of a part (j) of said set or subset (i) is in sending phase, the control of the management of the execution of a set or subset (i) of task(s) comprises the execution of an instruction indicating the boundary between two parts (j) and (j+1), and wherein the management of the execution of said part (j) of said set or subset (i) comprises the passage into the end-of-processing state activated by said execution of the instruction indicating the boundary between said part (j) and the following part (j+1).

28. A method according to the preceding claim wherein, if the set or subset (i) of task(s) is a program loop and if the management of execution of all the parts or stages (j) of the loop (i) is in the end-of-processing state: the control of said management of the execution of the loop (i) comprises: the passage to the next iteration (iti)=(iti)+1, the transmission of the address (A[j]i) obtained during the management of the execution of a state (j) of said loop (i), if the management of the execution of the following stage (j+1) of said loop (i) is not in a state of end of processing of the iteration (Ni). Of said loop (i) the transmission of the basic address (A0i), if the management of the execution of the first stage of said loop (i) in the state of the end of processing of the iteration (Ni). of said loop (i), and wherein: the management of a stage (j+1) of said loop (i) comprises: the reception, as the case may be, of said transmitted address: basic address (A0i) for the first state and address (A[j+1]i)=(A[j]i) coming from the management of the execution of the previous stage (j) of said loop (i), and the passage into the sending state activated by said reception, the verification of completion comprises the passage into mode of memorization of instances of completion activated by said passage into sending state

29. A method according to one of the claims 27 or 28 wherein, the control of the management of the execution of a set or subset (i) of task(s) comprises the reporting of the completion of said set or subset (i) of task(s) as soon as the management of each part (j) of said set or subset (i) of task(s) is in an end-of-processing state and, in the case of a loop (i), in a state of end of processing of the iteration (Ni) of said loop (i).