Method and device

[0001] The present invention relates to preamble of the claims. Thus, the present invention is concerned with improvements to logic cell arrays.

[0002] Logic cell arrays, in particular also reconfigurable ones, are known. This type of chip includes, in particular, systolic arrays, neuronal networks, multi-processor systems, processors having a number of arithmetic-logic units and/or logic cells and/or communicative/peripheral cells (I/O), networking and/or network chips, such as crossbar switches, as well as known chips of the type FPGA, DPGA, Xputer, Chameleon, etc. In this connection, refer in particular to the following patent rights, which are attributable to the present applicant and for disclosure purposes are referenced in their entirety: German Patent 44 16 881, German Patent 197 81 412, German Patent 197 81 483, German Patent 196 54 846, German Patent 196 54 593, German Patent 197 04 044, German Patent 198 80 129, German Patent 198 61 088, German Patent 199 80 312, PCT/DE 00/01869, German Patent 100 36 627, German Patent 100 28 397, German Patent 101 10 530, German Patent 101 11 014, Patent Cooperation Treaty/European Patent 00/10516, European Patent 01 102 674 and family members of these. Logic cells are thus defined as any cells that contain logic and/or arithmetic and/or switches and/or memory and/or peripherals.

[0003] While for those types of systems in the related art, in particular those previously cited, a series of approaches are known in order to enable in principle an efficient processing of data, and these approaches may also be implemented well in hardware architectures, there nevertheless exists in practical implementation the desire to ensure the most favorable possible designs, which, in particular, can be situated in a space-saving manner on a wafer and/or can be operated in an energy-saving manner; beyond this it is desirable to find especially advantageous kinds of operation.

[0004] One of the serious difficulties with systems from the related art is seen in the fact that a large number of cells have to communicate with each other. The communication may be required in order to pass the data to be processed from one cell to another. This is the case, for example, if a cell is supposed to further process the results from another cell, e.g. by linking of the result obtained there to results obtained from one or more other cells. Furthermore, communication may be required to transmit status signals.

[0005] Busses for transmitting signals to one of a number of possible receivers have been known for a long time. Busses are bundles of wires, the number of wires typically being determined from the number of bits to be transmitted together, that is, typically in parallel, plus in some cases a series of status lines.

[0006] With known simple busses, as are used, for example, in PC's for the communication of plug-in boards with the CPU and/or with each other, the bus lines may be routed to all receivers, and then appropriate control signals transmitted along with them, that is, addressing, ensures that only those receivers respond that are supposed to receive the data. Such a system becomes problematic when a great many communicating units need access to the bus or busses. This is because the communication of data must wait, if necessary, until the bus has been released by other units and/or time-slice solutions must be implemented that grant a transmitting unit only a certain transmission time, which as a rule is independent of whether all data has been transmitted in this time, which might also make it necessary to use a number of time slices for the data transmission. For example, this approach is practiced in systems like the token ring network. In systems like logic cell arrays, in which very rapid communication is desired in order to ensure high data processing speeds, this is undesirable.

[0007] It has also been proposed already that the busses be segmented. If, for example, in a series of logic cells several units to be connected to each other are disposed close together in pairs, a bus line that passes along all units and consequently is long may be separated by means of switches in order to form several subbus systems. In this context each segment, like the entire bus, comprises the required number of parallel data lines and the additionally required protocol lines; the communication of a pair of logic cells that are disposed close together then does not disrupt the communication of another pair that are disposed close together. In this way, the data rate that is transmitted via the bus system may be substantially increased.

[0008] However, it is problematic that if bus systems of this type are integrated on semiconductor chips, such as in CMOS technology, the structure is typically complex and the operation is energy inefficient.

[0009] The object of the present invention is to provide something new for commercial application.

[0010] The achievement of this objective is claimed in an independent form. Preferred embodiments are found in the dependent claims.

[0011] A first essential aspect of the present invention thus proposes that, in a logic cell array having a number of logic cells and a segmented bus system for the logic cell communication, it is foreseen that the bus system comprises different segment lines having shorter and longer segments for connecting two points in order to be able to keep down the number of bus elements that have to be traversed when the communication start and end points are separated from one another.

[0012] Consequently, the present invention, in an initial basic concept, utilizes the fact that, configuring the busses using segments of great length that are fashioned as a single line for bypassing long paths in a logic cell array produces an especially simple design and an especially efficient operation. By simultaneously providing short segment lines, it is ensured that all points are nevertheless addressable as needed.

[0013] Preferably, each of the segment lines is formed of a plurality of parallel lines whose number is determined by the bus width and/or the bus protocol. Each segment is therefore formed by a bundle of parallel lines, as is known per se from the related art. It is understood, that then all lines of a segment have essentially the same length; essentially, this means that the lines of a line bundle may all be routed to one and the same end point line, such as a data input of a logic cell, where they may be connected to the input members that are assigned to each of their bits.

[0014] It is preferable if data conversion logic cells like arithmetic units have three of the line bundles that are required for the processed bit width to be supplied to them as data supply lines. This permits carrying out a linking of operands A, B, C of the ABC type within the cell. This operation is of particular significance in the field of digital signal processing, and the bus structure is therefore especially suitable for logic cell arrays for the construction of real-time configurable DSP arrays.

[0015] It is especially preferred, in particular in a logic cell array with which an arithmetically multiplicative linking of operands may be carried out, if at least two of the line bundles for data that are required for the bit width being processed are led out from the cells, in particular with one line bundle for high-order bits and one line bundle for low-order bits; in this way, for a data word width of 32 bits, a 64-bit-wide output is created in order to output the result of a multiplication in full width.

[0016] In reconfigurable logic cells, it is preferable if control inputs are addressable in particular via the segmentable bus system and are disposed in such a manner that at least the logic cell's sequence control signals, such as the signals Reset, Step, Stop, Go and Reload, are transferable into the cell. These trigger and/or enable a reset, a reconfiguration and an activation. For this purpose a corresponding number of bit lines may be provided; preferably, at least two bit-wide data lines are therefore provided for control signal inputs. It is especially preferred if each of the control signals has a separate signal input assigned to it; alternatively, an input linking circuit on a several-bit-wide status line may provide for an orderly addressing of the cell. These signal inputs are control signal inputs in as much as they can carry the signals that control configuration and/or reconfiguration (flow control signals), especially trigger signals, such as are known in the aforementioned related art to the cell. The actual communication of the cell with the unit or a reconfiguring unit, such as a configuration manager (CT or CM) may thus be achieved in a way that is known per se, for example, via the so-called ready/ack/rej protocol, which permits a reconfiguration of the cell only under certain conditions. For the details of this protocol, refer to the applicant's patents that were comprehensively incorporated by reference in the introduction and any additional publicly accessible document about XPP/VPU architecture.

[0017] The bus system may include segments for connecting the logic cells to each other; this is typically the case and therefore of particular significance, because an especially high and efficient data throughput is needed inside the logic cell arrays without the processor being heated up due to, for example, high power dissipation. In this case, the present invention may be used particularly efficiently.

[0018] Alternatively and/or in addition, the bus system may be routed to internal and/or external units that configure and/or reconfigure the logic cells; the busses in particular may be routed past a plurality of data processing and/or storage logic cells to one or more such (re-)configuring units, so that either a communication of the cells with each other and/or with the or a configuring unit is possible. It should be mentioned in this connection that in principle it is possible to send information out to a plurality of receivers simultaneously in order to, for example, reset all cells simultaneously in the case of rebooting. The bus structure may be designed for this accordingly.

[0019] Furthermore, it is preferable and possible if the bus system is routed to I/O interfaces and/or to external data memories. In other words, the segmentation structure may be used to advantage both inside and outside the logic cell field. I/O interfaces transmit bus systems that are inside chips to the outside. In this context any bus systems, in some cases alternating in time, are connected for external output and/or input. In addition, the possibility may exist of combining several bus systems in such a manner that they are synchronized with each other. For example, any two output bus systems or an input bus and an output bus together may be synchronized.

[0020] In a practical design of the bus system, a series of interline elements is provided; here line means in particular as a first-class conductor, like a continuous metal line within a semiconductor wafer. Interline elements means those elements that are disposed between any two lines that are assigned to each other. First, interline elements may be connecting switches, in particular the control means that respond to the requirements of logic cells and/or the communication thereof and/or other units; thus, the switch may, for example, be or become open or closed at the request of a configuration manager, that is, a unit that configures the logic cell field. This makes it possible in particular, as is preferable, to use a compiler to establish how the bus structure is to be designed; in particular the compiler or another configuring or controlling unit is able to determine whether the communication between two cells that are separated is to be accomplished via segment lines having shorter or longer segments, and moreover, if there are a number of longer segments, a preference may also be specified through which one of a multiplicity of segment lines the communication is to occur. In this way the power loss through the switch may be minimized and/or an optimal adaptation of the data conversion and/or processing to signal propagation times may be provided along the bus lines.

[0021] The interline elements may additionally or alternatively include multiplexers in order to feed a signal from a line to one of a series of destinations, such as logic cells, and/or lines that continue further and/or, in order to feed it to a multiplicity of destinations, in particular selectable destinations, simultaneously.

[0022] Furthermore, registers may be provided as interline elements. These perform different tasks. In this manner data may at first be kept on the bus until a destination is ready to retrieve them. Furthermore, when busses are very long, one can prevent the data processing from being properly completed only if the long periods of time until the data arrive at the destination when paths are long (and in some cases the return of the reception confirmation) are taken into account. Buffering the data in registers does increase the total amount of time (delay) until a data packet traverses the bus, but the interim time until arrival of the data (latency) in which the array or individual cells may not be used meaningfully is reduced. It should be pointed out that here again a tradeoff between latency and delay may be required, the register being switched in, for example, only under predetermined and/or established conditions, such as very long bus paths. It is then desirable if the switching on of the register is controlled, e.g. by a control unit or similar item, like a configuration manager, which in turn is able to operate in response to compiler-generated commands. Furthermore, it makes sense to provide a register in order to ensure a synchronous arrival of bits of different significance at a destination point. If busses are fed back, that is, a connection is provided from a cell output to the input of the same cell, it is especially preferable to use registers because such high-frequency feedback loops, which are capable of destroying components, may be prevented by the register-imposed delay time.

[0023] In an especially preferred variant, registers may be designed all or in part as additive and/or multiplicative registers; specific links may be implemented without problem and with little circuitry effort, and, thus, the logic cell structure may be relieved on a net basis via shifting of simple linking tasks. In particular, a register may be designed precisely for the purpose of linking two operands algebraically; in busses for transmitting status signals that indicate the status of a logic cell or a series of logic cells and/or trigger a change thereof, that is, in busses for so-called trigger vectors, the registers may preferably be designed to implement Boolean links and/or lookup tables, especially modifiable lookup tables.

[0024] Other interline elements that can be provided and that preferably can also optionally be connected in addition are line drivers.

[0025] This may be necessary in particular for multiplex systems that provide a very strong signal fanout.

[0026] In an especially preferred variant of the present invention, it is possible to provide a change from one segment line having longer segments to a segment line having shorter segments and/or vice-versa along a data transmission path. This permits the provision of roughly equal signal propagation times for the communication of a larger number of cells in the same direction, that is, along the same bus structure, via an appropriate combination of short and long segments, even if buffer registers are provided. The bus structure in this context is comparable to a street that has fast lanes and crawler lanes and enables a lane change at predetermined intermediate positions.

[0027] It is generally preferable if the bus system includes a multiplicity of parallel segment lines, several parallel segment lines being provided with longer segments. The longer segments of the segment lines that have longer segments do not all need to be of the same length; a staggered arrangement may be provided.

[0028] It is also generally preferable if, when there is a greater number of parallel segment lines, the segment line ends and/or interline elements are disposed within the segment lines at an offset in relation to each other in the bus direction. Typically, interline elements, such as, switches, registers, multiplexers and drivers are provided at the segment line ends or leads. The hardware technology implementation of these elements then requires substantial space, which may be considerable compared to that of the lines to be disposed in the intermediate layers. The preferred offset disposition of these elements then ensures that space only has to be provided for interline element arrays of, for example, two or three segment lines, but not for interline element dispositions of all available segment lines. Also it is possible not to provide drivers or registers for all interline element dispositions, but only to provide them every nth segment end. However, then it is preferable that at least three segment lines that have longer segments for at least two segment lines have segment line switching circuits, in particular multiplexers provided at predetermined positions as interline elements. In this way, the desired segment change may be configured as required. Preferably, the bus structure is comparable to a pigtail made out of a multiplicity of strands that repeatedly cross each other, a segment change being at the strand crossings and changes being possible among segment pairs or segment groups that vary along the bus. It is then possible and preferable that the segment line switching circuits for the change from a first to a second segment line are provided at a first position and the segment line switching circuits for the switch from a second to a third segment line at a second position.

[0029] In addition to the switching only by pairs, it is preferred to provide also the possibility of selecting, at a position or a multiplicity of positions, among several segment lines to switch to and/or to which data are simultaneously output.

[0030] In an additional aspect of the present invention, two-way communication of the cells is possible for the logic cell array. In bus systems having interline elements, such as drivers and/or registers, directions of travel are defined; in order to enable the communication of the cells in two directions, separate bus systems are therefore preferably provided for opposite running directions. At least in one direction it is once again possible to provide at least number different, segment lines with shorter and longer segments, in particular ones that are once again generally parallel.

[0031] If the segment lines are separate for the two directions of travel, a register is preferably provided for at least one direction of travel; as explained previously, the register is preferably provided in the bus system that is routed in the reverse direction, i.e. to that bus system, with which signals may be routed back from an element output to an element input. It should be mentioned that according to another aspect of the present invention, a first bus system may be provided for the transmission of data to be processed and a second bus system for the transmission of status and/or monitoring or control information may be provided; once again one or both bus systems may be formed with segment lines having short and long segments and preferably the respective bus systems or bus arrays may be configurable separately from each other, that is, definable in their circuitry and/or regarding the operations in linking registers or the output of lookup tables.

[0032] Normally, the bus system will preferably be used in a logic cell array in which a plurality of logic cells are arranged adjacent to each other in a row; the longer segments then bypass at at least one logic cell; if the logic cell array includes even more logic cells in a row, the longest segments preferably bypass more than one logic cell.

[0033] It should be pointed out that, in at least two-dimensional logic fields, that is, a disposition of logic cells in at least rows and columns, a segmented bus system may be provided in each row and each column and have the previously described structure having long and short segments in parallel segment lines.

[0034] The described bus structure is advantageous in all arrays in which data processing units, thus logic units, are to be linked to each other. However, special advantages are offered if the logic cells are selected from, include and/or form arithmetic-logic units, DPGA's, Kress arrays, systolic processors and RAW machines; digital signal processors (DSP's), that have more than one arithmetic-logic unit, configurable (programmable) logic cells or other cores and/or PAE's, ALU, input/output (I/O) functions and/or memory management units (I/O) and memories. A successful implementation of a bus structure having segment lines that have short and long segments is the VPU/XPP processor unit of the applicant.

[0035] Such a processor will preferably have a multiplicity of different communicating logic units disposed in an array, the logic units having at least memory storage and data conversion units and the memory units being disposed closed to the edge of the array. This makes it possible to have data run through the array and be buffered, if necessary, at the edge in order to bring about reconfiguration as required. The flow may also occur via parallel rows or columns and/or in a meandering way, in order to thereby provide increased computing power.

[0036] Also, if need be, input/output units may be disposed closer to the edge than the storage units; this allows the buffering of data before processing by passing through the array. However, in any case it is preferred if for communication of the logic units, at least from edge to edge, at least one bus system is provided that is, in particular, segmentable as previously described. Then, it is possible if for the operation of a processor, data are read from a first memory that is close to the edge and be changed as required in a first data conversion unit that is close to this edge, transferred from there into at least one other data conversion unit in order to carry out further data modifications as required, the data, after flowing through a plurality of data conversion units into a second memory close to the edge are stored at a position distant from the first memory, a reconfiguration of the data conversion units is carried out for the re-determination of the data conversion, and the data is routed through at least one part of the data conversion units, possibly in the opposite direction of travel. It may also be provided for operation that the outgoing run is made in a row and the return run in that row or a row situated beneath it in order to take into consideration pipeline effects.

[0037] The processor may have at least one programmable gate array (PGA) and several data modification units having ALU's, it being possibly preferable if at least one or more PGA's are surrounded by other data conversion units, in particular ALU's, in order to be able to obtain data easily for linking purposes and/or output them.

[0038] It is possible that, in an array made up of reconfigurable units, several (sub-) arrays, especially identical ones, are provided that may be brought into communication with each other. For this purpose, each may be provided a reconfiguration unit that may be designed for reconfiguration of the reconfigurable element of the array. Running inside the array are the bus structures appropriately described above. A configuration manager, which may in particular be disposed on an array side on which no memory units and/or I/O units are provided is thus assigned to each array, the reconfiguration unit preferably extending over the breadth of the array, which offers the advantage that an optimal chip form may be produced.

[0039] A supervising configuration manager which is common to the arrays and which extends in particular over the breadth of more than one array and which is designed for communication with at least several, preferably all of the supervising configuration managers, is then provided, if desired.

[0040] Alternatively or additionally, given such a processor having arrays that may be brought into communication with each other, in particular identical ones that have reconfigurable data conversion units, it is possible for the arrays to have switching units (VPAE's) between them that are disposed among the fields for communication switching.

[0041] Using such switching units (VPAE's), data and/or control signals (monitoring signals, trigger vectors, etc.) may be transmitted.

[0042] It is possible that the configuration units of at least two of the arrays are in direct connection with each other. As a result, a supervising configuration unit may be omitted and/or a fault tolerance may be increased. Preferably, for this purpose the communication of the configuration units of the at least two arrays that are in direct connection with each other is determined via dedicated fixed lines and/or data lines that are pre-configured, that is established during bootup.

[0043] It should be mentioned that a possible and preferred design of the switching units (VPAE's) is comprised in the fact that they have two configuration registers, the first configuration register being addressable by the first array and the second configuration register being addressable by the second array. Only if access to the switching units has been permitted for both arrays, are data transmitted accordingly. For this purpose, the first array may permit the insertion of data in the switching unit by appropriate configuration of the first configuration register and signal the insertion of the data to the other configuration units via the dedicated line. Afterwards, the data are retrievable, which is ascertainable by configuration of the second configuration register. All in all, a sluice-like transmission of data is enabled by the VPAE's. The separation of a larger array into sub-arrays in other respects generally allows the reduction of the configuration effort to be handled by a configuration unit and as a result may contribute to an efficient array operation that is not impaired by a bottleneck imposed by a configuration unit.

[0044] It is plausibly. preferable if the sub-arrays may be brought into communication with each other via multiple data conversion and/or transfer units (VPAE's) so that also in this case no data rate limitation is produced in the data processing.

[0045] It is preferable in the geometric hardware implementation of the described system, if the communication-enabling data conversion units are connectable on both sides via busses running along the array, the busses being designed in particular to run past the array edge and there in particular may be routed further to I/O interface units.

[0046] As explained above, it is especially preferred if in the processor one or more data relaying rule implementation means, in particular a lookup table, is/are provided that are designed to enable relaying of data and in particular data conversion in response to the joint fulfillment of a multiplicity of criteria, in particular event-related criteria. The data to be relayed may in particular be data or trigger vectors that are involved in the reconfiguration and/or reconfigurability of logic cells. The function or the conversion of the relevant data relaying rule implementation means is configured accordingly.

[0047] It is possible also to design the processor directly specifically using segmented bus structures so that data conversion units and data storage means are provided, at least a few of the data conversion units being locally assigned data storage means and in particular means being provided to locally read out data that is locally written into the local data storage means for the further local recent data modification. In other words, interim results may be stored in a logic cell and be further processed along with other data and/or using other links.

[0048] If there is a reconfiguration of a cell array made up of logic cells that are reconfigurable while in operation, which allows in particular a reconfiguration in response to the propagation of specific pre-determined data, as is done in the publications that describe the wave reconfiguration and are referenced, especially patent applications of the present applicant, then it is especially preferable if one may select whether the reconfiguration occurs in response to the data propagation for all logic cells into which the data that was pre-specified and/or derived therefrom flow and/or only for a part, in particular a processing branch (e.g., an if/then branch), while at least one other branch remains unchanged, which reduces the reconfiguration effort and/or only a subchain of a longer row of logic cells, in particular, sequential data-processing ones, is reconfigured, in particular up to a predetermined logic cell, and/or in order to permit the reconfiguration immediately or after that data processing or further data processing. This substantially increases the practical use of the wave reconfiguration. It should be mentioned that such a selective wave reconfiguration is possible using only a few commands, in particular, for example, the commands Reset, Step, Stop, Go and Reload. Consequently, the number of trigger inputs or control inputs that are routed to a reconfigurable logic cell, such as a PAE, is small.

[0049] It should be mentioned to be especially advantageous and preferable that the manner in which a logic cell reacts to a specific characteristic data packet in wave reconfiguration is capable of being and/or is pre-loadable. This may happen in particular by pre-loading the lookup tables that are provided in the busses. The busses in this context may in part be routed via logic cells and/or formed by and/or with them.

[0050] It is possible without any other measures to use the processor array having the bus system of the present invention in order to approximate the behavior of non-linear systems quickly and simply in calculations. This preferably happens, by approximations that are particularly appropriate being determined for one row of working ranges that are adjacent to each other at one temporary working range (n) and at the working ranges that border it below and above ((n1), (n1)), the parameters that enable the approximations are pre-loaded, the behavior with the central (n) working range is approximated until that working range is exceeded inside of which the approximation is suitable, then that approximation (n1) is used for further work which belongs to the working range into which the exceeding proceeded, and preferably the approximation at that working range (n1) into which the exceeding did not proceed is overwritten using an approximation (n2) that, together with the previously used approximation (n), encompasses the currently used approximation (n1) or the accompanying working range. By overwriting that working range in which the operation value is not run, only three variants need to be pre-loaded in order to ensure a rapid further processing after the range is exceeded. Because the overwriting of the preloaded working range happens during the data processing, a constant data computation is possible during a steady transition of the operation values without greater leaps or abrupt changes. It is said to be plausible that this manner of approximation may not transferred to multidimensional cases due to operation value encompassing in more than one direction and corresponding overwriting of the intervals that no longer adjoin the current operation value interval after the overwriting of an n-dimensional operation value interval and is seen as included within patent protection.

[0051] The present invention is described below only by examples with respect to the drawing. The figures show:

[0052] FIG. 1 the overall structure of a logic cell array configured in accordance with the present invention,

[0053] FIG. 2 a detailed view of this having sub-arrays called clusters, VPAE's in between and connection busses,

[0054] FIG. 3 an example for a forward register of a configuration bus,

[0055] FIG. 4 the hierarchy of configuration managers of a logic cell array that is configured according to the present invention,

[0056] FIG. 5 details of a PAE having data, the details of which are partially to be inferred in different registrations. Recognizable are the three data inputs A B C in the logic cell, the five trigger inputs for the trigger vectors Reset, Stop, Step, GO and Reload, the two data outputs H(igh) and L(ow) as well as the status signal outputs UVW and the bus coupling,

[0057] FIG. 6a the structure of the logic cell in the block diagram,

[0058] FIG. 6b the structure of a register,

[0059] FIG. 6c the register structure of a reverse register for event signals having the optional lookup tables that are new in this context,

[0060] FIG. 6d a forward register having pre-loading capability,

[0061] FIG. 7 the detailed structure of a memory cell that has an input that is distinguishable from data conversion logic cells and has a correspondingly different required bus width,

[0062] FIG. 8 another detailed view of a reverse register in the configuration bus having a lookup table for the implementation of Boolean links from trigger vector bits,

[0063] FIG. 9 additional processor details, in particular of the busses,

[0064] FIG. 10 unsynchronized I/O circuit having two busses,

[0065] FIG. 11 synchronized I/O circuit having two busses,

[0066] According to FIG. 1, a processor 1 generally marked with a 1, which may form a unit that may be characterized as an XPP (extreme processing platform), includes two fields or clusters

[0067] of a number of arithmetic logic cells (ALU-PAE'sarithmetic logic unit processing array element),

[0068] memory cells (RAM-PAE'sRandom Access Memory processing array element)

[0069] several illustrated as four input/output cells (I/O elements, input/output processing array element),

[0070] a bus system running through each field, and

[0071] a configuration manager that is assigned to the field,

[0072] these elements together as a cluster forming a unit, which by itself is capable of processing data and doing so as a reconfigurable unit

[0073] and a higher-order configuration manager being assigned to the two configuration managers (CM) of the cluster in order to form thereby a master cluster or the XPP, it being said that in particular it is possible using tree-like hierarchical structures to have more than two fields of processor elements cooperate; in this manner a higher-order configuration may be placed over more than two configuration managers (CM), which on their part are each assigned to only one individual PAE cluster and/or it may be provided that a multistage configuration manager hierarchy is built up, in which a configuration manager central level is actually assigned to first configuration managers (CM) and in turn is placed under one or more hierarchical levels.

[0074] FIG. 1 thus correctly depicts the arrangement of the cells to each other in a two-dimensional matrix and its relative size by way of approximation insofar as the individual cells are disposed very regularly and in particular the PAE's, that is, the memory storage and arithmetic logic cells have approximately the same size, which makes it possible to carry out a page addressing in the generally rectangular and/or regular cell structure having a roughly equal linking to the bus system; the shown page addressing in this context, however, is arbitrary regarding the disposition of memories and arithmetic units in relation to each other, insofar as other dispositions, also involving the I/O cells and/or using still other elements without problems, are possible and are adaptable to a desired special computing task without any further measures. The illustration that is shown is also exemplary insofar as, instead of a two-dimensional matrix, a three-dimensional disposition of individual cells within the space is possible, provided that the manufacturing technology of the semiconductor technology that is used permits this.

[0075] An individual PAE then includes a multiplicity of inputs and outputs. In so doing, the individual PAE is provided, on the one hand, with data inputs and outputs (A,B,C) and, on the other hand, with control or monitoring inputs and outputs (U,V,W). As is evident from FIG. 1, the PAE's are disposed in horizontal rows, the cells being connected among themselves via a bus system; these horizontal bus systems are connected in the center of the field to the PAE's that are situated above and below them, and to be precise with them only above and below in the shown preferred example. Thus, the transfer of data and status signals (i.e. control signals) upward and downward is accomplished via the PAE's and, to be precise, preferably as illustrated in FIG. 1, regardless of the previous example involves a RAM-PAE or an ALU-PAE; consequently all types of PAE's are preferably to be connected in the same manner to the bus system and internally wired accordingly. Different from what is illustrated, it would even be possible in principle also to configure the I/O cells for connection to bus lines that are situated one on top of the other, which, however, does not occur in the present example.

[0076] Consequently, the PAE's are used for the relaying of data, that is, for data to be processed as well as control and monitoring data in the vertical direction. The PAE's in this context are configured in such a manner that this relaying may be carried out with or without data modification.

[0077] The data modification may at first be accomplished in the actual functional unit of the PAE, e.g. if it is equipped as an arithmetic-logic unit and the data are linked in the arithmetic logic unit; in order to ensure that data may be supplied sufficiently rapidly to be linked in the cell and bring about an optimal utilization of the cell cycling, three lines A,B,C having the bit width that is able to be processed in the cell are routed to the cell (see FIG. 5). In this context a preferred additive input length is also possible that allows determining operations of the type (AB)C in a time-efficient and space-efficient manner using the cell results. If no data conversion is required, the data may be passed on to the functional unit. In this way, for example, data that are needed in a row further below are passed through into it without being subjected to a modification in the PAE. It is said that in the simultaneous routing of, for example, three data words corresponding to three inputs A, B, C they may be relayed in different ways. Therefore, it is possible, for example, to relay only the data word to line C unconverted, while A and B are linked to each other in the ALU, for example, by A/B division. Alternatively, two data words may also be rerouted, while a data word in the ALU, for example, is linked to a constant or another value temporarily stored there. The convertibility to other data word amounts that are simultaneously routed to the PAE is obvious, that is, it may also if desired be routed to the PAE as three data words; the individual word in this context has the processor-specific bit width, in the present case, for example, 32 bits.

[0078] Any inputs and/or outputs of PAE's may be configured using constants of the configuration unit (CT/CM). In this context constants may remain unchanged during the entire operation of a configuration, or be overwritten during the operation, that is the execution of a configuration of new data of the configuration unit (CT/CM) and/or other sources, such as PAE'S.

[0079] It is also understandable that, different from the case involving PAE's having ALU's specifically in the RAM-PAE's, instead of a linking of all or a part of the data, a complete or partial memorization and/or a complete or partial relaying may occur.

[0080] The relaying of the data, furthermore, may occur while bypassing the actual data conversion PAE core unit, like the ALU, in such a manner that a conversion of the data occurs before or during the relaying. This may occur by modifying the data while consulting the lookup tables and other linking units, for example, units that sort the relevant data words according to value. These units are preferably disposed between the bus coupling and the input into the PAE core unit and configurable in their function, whereupon whether the data are relayed unchanged or not is configurable; if they are not relayed unchanged, but rather changed, how the change occurs within the given possibilities is configured. What is especially important for the status signal relaying in this context is the provision of lookup tables, because predetermined output signals may be generated with them in response to predetermined input signals without any additional measures.

[0081] Furthermore, the data relaying also of unchanged data may be accomplished in a configurable manner immediately or delayed by register. The illustrated individual PAE for this purpose has a configurable, connectable forward register and a configurable, connectable reverse register for the data to be processed and a configurable, connectable forward register and a configurable, connectable reverse register for control data, that is, trigger vectors or status signal and/or control flow data; it is indicated that registers having increased (temporary memory) depth, in particular as FIFO registers, would be foreseeable, precisely in order to temporarily store data on a multi-cycle basis. Assigned to the register in question is a configurable multiplexer with which the register in question may be connected into the data path or removed from it, as desired. The multiplexers are also configurable such that results from the cell may be connected on the bus as desired and required. Therefore, for example, the top and bottom bit word that is obtained in a multiplication are output (H,L) and/or only a result data word (H) is output and a data word relayed and output on the bus, even a desired exchange (A<->H) being configurable. Consequently, a PAE is characterized in the present processor modeland this is shown in its significanceby a structure that has a central functionality lie an ALU, a RAM and/or, for example, also an FPGA and the additional data relaying functionality, it being able to possess, preferably, a forward and/or reverse register data relaying functionality, thus, comprised of three basic modules, such as FPGA, ALU, RAM and the two (forward/reverse) register data relaying functionality modules, the latter of these being able to implement additional functions. Central functionality in this context does not mean spatial central functionality, but rather functionality in the sense of assigning the central meaning and effect of the particular PAE. The additional functions, which are implemented via the register data relaying functionality modules, include in particular, preferably in the forward register for the data flow, control operations like multiplexing or demultiplexing and copying and/or other manipulations of data; in the reverse register an ALU functionality is preferably implemented for the data flow and/or a Boolean linking of event signals, i.e. trigger signals. Thus, the PAE is comprised of modules for the manipulation of data streams, for the operative linking and/or generation of data packets and the processing for the most part via logic linking and generation of event packets.

[0082] The input/output modules are likewise adapted in their structure in a particular way to the requirements of the data processing task. For this purpose, each of the input/output modules is provided with several channels, that are able to work independently of each other in the so-called PORT mode (FIG. 10) or in synchronization with each other in the so-called RAM mode (FIG. 11). Each I/O module may be used simultaneously by a plurality of algorithms while using a plurality of algorithms, for which sorting algorithms may also be brought in.

[0083] In other words, the input/output module (I/O interface) works in memory (RAM) mode (FIG. 11) as an interface to the external memory. In this context writing data and address data of any channel are synchronized, that is, only if both data packets are present, is an external writing operation started. In reading mode, the I/O unit sorts from externally transmitted packets back to the internal channels. The switching between reading and writing mode may occur, for example, via configuration or, for example, via signals.