Patent Application
20070113046
The present invention relates to reconfigurable computing. In particular, the present invention relates to improvements in the architecture of reconfigurable devices.
Reconfigurable data procesing arrays are known in the art. Reference is being made to the previous applications and/or publications of the present applicant/assignee all of which are encorporated herein by way of reference. Accordingly, the devices described hereinafter may be multidimensional (n>1) arrays coaprinsing coarse grained computing and/or data operation elements allowing for runtime reconfiguration of the entire array or parts thereof, preferably in response to a signal indicating reconfigurability to a loading unit (CT, CM or the like).
Now, several of these data procesing arrays have been built (i.e. Xpp1, XPP128, XPP2, XPP64). It is however desirable to improve the known device further as well as to improve methods of its operation.
Accordingly, in order to achieve this object there will be described a number of improvements allowing separately or in commom to improve the performance and/or power consumption and/or cost of the device.
A first way to improve the known devices is to improve the functionability of each single processor element. It has been previously suggested to include a ring-memory (RINGSPEICHER) in the array, to store instructions in the ring-memory and to provide a pointer that points to one of the ring-memory adresses so as to select an instruction to be carried out next. Furthermore, it has been suggested to provide at least one “shadow configuration” and to switch over between several configurations/shadow configurations. Anotrher or additional suggestions has been designated as “wave reconfiguration”.
While these known methods improve the performance of a reconfigurable device, there seems to be both a need and a possibility for further improvements.
It is to be understood that while in the following description, a detailed example is given, for example with respect to thee number of registers given associated with eachj PAE, it is not deemed necessary to provide an ALU with exactly this number of registers. Rather, it will be understood by the average skilled person that deviations from the explicitly described embodiment are easily feasible and that the detailed level of description stems from an effort to provide an examplary PAE and not from the wish to restrict the scope of invention.
1 Overview of Changes vs. XPP XPP-II
1.1 ALU-PAE Architecture
In the suggested improved architecture, a PAE might e.g. comprise 4 input ports and 4 output ports. Embedded with each PAE is the FREG path newly named DF with its dataflow capabilities, like MERGE, SWAP, DEMUX as well as ELUT.
2 input ports Ri0 and Ri1 are directly connected to the ALU.
Two output ports receive the ALU results.
Ri2 and Ri3 are typically fed to the DF path which output is Ro2 and Ro3.
Alternatively Ri2 and Ri3 can serve as inputs for the ALU as well. This extension is needed to provide a suitable amount of ALU inputs if Function Folding (as described later) is used.
In this mode Ro2 and Ro3 serve as additional outputs.
Associated to each data register (Ri or Ro) is an event port (Ei or Eo).
It is possible, albeit not necessary to implement an additional data and event bypass BRi0-1, BEi0-. The decision depends on how often Function Folding will be used and how many inputs and outputs are required in average.
1.1.1 Other Extensions
SIMD operation is implemented in the ALUs to support 8 and 16 bit wide data words for i.e. graphics and imaging.
Saturation is supported for ADD/SUB/MUL instructions for i.e. voice, video and imaging algorithms.
1.2 Function Folding
1.2.1 Basics and Input/Output Paradigms
Within this chapter the basic operation paradigms of the XPP architecture are repeated for a better understanding based on Petri-Nets. In addition the Petri-Nets will be enhanced for a better understanding of the subsequently described changes of the current XPP architecture.
In most arrays each PAE operates as a data flow node as defined by Perti-Nets. (Some arrays might have parts that have other functions and should thus be not considered as a standard PAE). A Petri-Net supports a calculation of multiple inputs and produces one single output. Special for a Perti-Net is, that the operation is delayed until all input data is available.
For the XPP technology this means:
The quantity of data-and events is defined by the data and control flow, the availability is displayed at runtime by the handshake protocol RDY/ACK.
The thick arbor indicates the operation, the dot on the right side indicates that the operation is delayed until all inputs are available.
Enhancing the basic methodology function folding supports multiple operations—maybe even sequential—instead of one, defined as a Cycle. It is important that the basics of Petri-Nets remain unchanged.
Typical PAE-like Petri-Nets consume one input packet per one operation. For sequential operation multiple reads of the same input packet are supported. However, the interface model again keeps unchanged.
Data duplication occurs in the output path of the Petri-Net, which does not influence the operation basics again.
1.2.2 Method of Function Folding
One of the most important extensions is the capability to fold multiple PAE functions onto on PAE and execute them in a sequential manner. It is important to understand that the intention is not to support sequential processing or even microcontroller capabilities at all. The intention of Function Folding is just to take multiple dataflow operations and map them on a single PAE, using a register structure instead, of a network between each function.
One goal may be to save silicon area by rising to clock frequency locally in the PAE's. An additional expectation is to save power since the busses operate at a fraction of the clock frequencies of the PAEs. Data transfers over the busses, which consume much power, are reduced.
The internal registers can be implemented in different ways, e.g. in one of the following two:
1. Dataflow Model
Each register (r′) has a valid bit which is set as soon as data has been written into the register and reset after the data has been read. Data cannot be written if valid is set, data can not-be read if valid is not set. This approach implements a 100% compatible dataflow behaviour.
2. Sequencer Model
The registers have no associated valid bits. The PAE operates as a sequencer, whereas at the edges of the PAE (the bus connects) the paradigm is changed to the XPP-like dataflow behaviour.
Even if at first the dataflow model seems preferable, it has major down sides. One is that a high amount of register is needed to implement each data path and data duplication is quite complicated and not efficient. Another is that sometimes a limited sequential operation simplifies programming and hardware effort.
Therefore it is assumed consecutively that sequencer model is implemented. Since pure dataflow can be folded using automatic tools the programmer should stay within the dataflow paradigm and not be confused with the additional capabilities. Automatic tools must take care i.e. while register allocation that the paradigm is not violated.
The following figure shows that using sequencer model only 2 registers (instead of 4) are required:
For allowing complex function like i.e. address generation as well as algorithms like “IMEC”-like data stream operations the PAE has not only 4 instruction registers implemented but 8, whereas the maximum bus-clock vs. PAE-clock ration is limited to a factor of 4 for usual function folding.
It is expected that the size of the new PAE supporting Function Folding will increase by max. 25%. On the other hand 4 PAEs are reduced to 1.
Assuming that in average not the optimum but only about 3 functions can be folded onto a single PAE a XPP64 could be replaced by a XPP21. Taking the larger PAEs into account the functionality of a XPP64 XPP-II should be executable on a XPP XPP-III with an area of less than half.
The function folding method and apparatus as well as otrher further improvements will be described in even more detailed hereinafter.
Equality of Internal Data Registers and Bus Transfers
The function fold concept realises two different models of data processing:
Due to the unpredictability of the arrival of data at the input registers (IR) a deadlock or at a least significant reduction in performance could occur if the commands in RC0 . . . RCn were to be performed in a linear manner. In particular, if feed-backs of the PAE outputs to the inputs of the PAE are present, deadlocks might occur. This can be avoided if the instructions are not to be processed in a given order but rather according to the possibility of their processing, that is, one instruction can be carried out as soon as all conditions of the VPU-mbdel are fulfilled. Therefore, for example, once all RDY-handshakes of incoming data, ACK-handshakes of outgoing data and, if necessary, triggers, (including their handshakes) are valid, then the instruction can be carried out. As the FF PAE has data additionally stored in internal registers, their validity and status has to be checkable as well in a preferred embodiment. Therefor, every internal data register (RD0 . . . RDn) is separately assigned a valid bit indicating whether or not valid data are present in the register. When writing data into the register, valid is set, when reading, valid is reset. Data can be read only if “valid” is set and can be written only if “valid” is not set. Accordingly, the valid flag corresponds most closely to the status that is produced in the state machines of bus systems by the transmittal of RDY/ACK-handshakes. It is a preferred embodiment and considered to be inventive to provide a register with a status bit in that way.
It is therefore possible to carry out instructions at the time when all conditions for the execution—again very similar to PETRI-nets are fulfilled.
Basically, there are two methods available for selection of instruction and control of their execution described herein after.
Method A: FF PAE Program Pointer
(Finite State Machine & Program Pointer-Approach)
According to the control principle of sequential processors, a program counter is used to select a certain instruction within the instruction memory. A finite state machine controls the program counter. This finite state machine now checks whether or not all conditions for the instruction in RC (PC), that is the instruction, onto which the PC. (Program Counter) points, are fulfilled. To do so, the respective RDY- and/or ACK-handshakes of the in- and/or outputs needed for the execution of the instructions are checked. Furthermore, the valid-flags of the internal registers to be read (RD0 . . . RDn) are checked so as to control whether or not they are set, and the valid-flags of those internal registers (RD0 . . . RDn) into which is to be written, are checked whether they are not set. If one of the conditions is not fulfilled, the instructions will not be carried out. PC is controlled to count further, the instruction is skipped and the next instruction is selected and checked as described.
The advantage of this method is the compatibility with sequential processor models. The disadvantage resides in the necessity to test and to skip instructions. Both of which might result in significant losses of performance under certain circumstances.
Method B: FF PAE Program Pointer
(Enabler & Arbiter-Approach)
This method is based upon the possibility to test all instructions in Rc0 . . . Rcn in parallel. In order to save the expense of the complete decoding of array instructions, each RC is assigned an entry in an evaluation mask field, the length of which corresponds to the maximum number of states to be tested; therefore, for every possible RDY- or ACK-trigger-signal (as well the RDY/ACKs of the triggers) as well as for every valid bit in RD0 . . . RDn two bits are available indicating whether or not the respective signal is to be set or not set; or, whether the state of the signal is unimportant for the execution of the instruction.
The mask shows only some entries. At In-Trigger, both the state of the trigger (set, not set) as well as the value of the trigger (trigger value) can be tested via RDY-value.
A test logic testing via for example the Line Control described herein after all instructions in parallel. Using an arbiter, an instruction of the set of all executables is selected. The arbiter controls the instruction multiplexer via ISel according to the transferral of the selected instructions to the PAE.
The Line Control has one single line of Boolean test logic for every single instruction. By means of an ExOR-gate (e) the value of the signal to be tested against the setting in em of the line is checked. By means of an OR-gate (+) respectively, a selection is carried out, whether the checked signal is relevant (don't care). The results of all checked signals are ANDed. A logic 1 at the -output of the AND-gates (&) shows an executable instruction. For every RC, a different test-line exists. All test-lines are evaluated in parallel. An arbiter having one of a number of possible implementations such as a priority arbiter, Round-Robin-Arbiter and so forth, selects one instruction for execution out of all executable instructions. There are further implementations possible obvious to the average skilled person. Those variants might be widely equivalent in the way of operation and function. In particular, the possibility of using “negative logic” is to be mentioned.
The following figure gives an overview of the entire circuitry:
Advantages of the Method are:
Disadvantages:
All busses assigned to a certain PAE are connected to the input registers (IR) or the output registers of the PAE are connected to all busses respectively (compare for example DE 100 50 442.6 or the XPP/VPU-handbooks of the applicant).
It has been realised that PAEs, in particular FF PAEs, allow for a depopulation of bus interconnects, in particular, if more IR/OR will be available compared to the State of the Art of the XPP as previously known. The depopulation, that is the reductions of the possibilities to connect the IR or ER onto the busses can be symmetrically or asymmetrically. The depopulation will typically amount to 20 to 70%. It is significant that the depopulation will not or not significantly effect the interconnectability and/or the routability of an algorithm in a negative way.
The method of depopulation is particularly relevant in view of the fact that several results can be achieved. The hardware-expense and thus the costs of the bus systems can be reduced significantly; the speed of the busses is increased since the gate delay is reduced by the minimisation of connecting points; simultaneously, the power consumption of the busses is reduced.
A preferred depopulation according to the VPU-architecture according to the State of the Art, however, with more IR/OR is shown in the following figure.
In particular, reference is being made to an optional extension of the bus architecture allowing for a direct next neighbour data transfer of two adjacent PAEs, in particular two PAEs placed one onto the other. Here, the outputs (OR) of one PAE are directly connected to a dedicated bus which is then directly connected to the inputs (IR) of a neighbouring PAE (compare next figure). The figure only shows an horizontal next neighbour bus, however, in general, vertical busses are possible as well.
In the figure, the shaded circles stand for possible bus connects: MUX. Double circuits stand for a connection from the bus: DeMUX.
Changes of the PAE IO
The following figure shows the State of the Art of a PAE implementation as known from XPU128, XPP64A and described in DE 100 50 442:6
The known PAE has a main data flow in the direction from top to bottom to the main ALU in the PAE-core. At the left and right side, data channels are placed additionally transmitting data along the main data flow direction, once the same direction as the main data flow (FREG) and once in the reverse direction (BREG). On both sides of the PAE, data busses are provided that run in the reverse direction of the main data flow of the PAE and onto which the PAE as well as FREG and BREG are connected. The architecture of the State of the Art requires eight data busses for each PAE side as well as four transfer channels for FREG/BREG for typical applications.
The bus system of the State of the Art has switching elements, register elementsa (R), each at the side of the PAEs. The switching elements allow for the disruption of a bus segment or disconnection to a neighbouring bus, the register elements allow the construction of an efficient pipelining by transferring data through the register, so as to allow for higher transferral band-width. The typical latency in vertical direction for next-neighbour-transmitting is 0 per segment, however is 0.5-1 in horizontal direction per segment and higher frequencies.
Now, a modified PAE structure is suggested, wherein two ALUs, each having a different main data flow direction are provided in each PAE, allowing for significantly improved routability. On one hand, the tools used for routing are better and simpler, on the other hand, a significant reduction in hardware resources is achieved. First tests shows that the number of busses necessary in horizontal direction is reduced by about 25% over the State of the Art. The vertical connects in FREG/BREG (=BYPASS) can even be reduced by about 50%. Also, it is no more necessary to distinguish between FREG and BREG as was necessary in DE 100 50 442.6.
The double-ALU structure has been further developed to an ALU-PAE having inputs and outputs in both directions. Using automatic routers as well as hand-routed applications, further additional significant improvements of the network topology can be shown. The number of busses necessary seems to be reduced to about 50% over the State of the Art, the number of vertical connects in the FREG/BREG (=BYPASS) can be reduced by about 75%.
For this preferred embodiment which can be used for conventional as well as for function fold ALUs, it is possible to place register and switching elements in the busses in the middle of the PAE instead of at the sides thereof (see FIG. below).
In this way, it is possible even for high frequencies to transmit data in horizontal direction to the respective neighbouring PAE without having to go through a register element. Accordingly, it is possible to set, up next neighbour connections in vertical and horizontal directions which are latency free (compare State of the Art and drawings referring to depopulated busses). The example of the interconnections shown in the respective figure allows transferral having zero latency in vertical direction and horizontally from left to right. Using an optimisation of PAE interface structure a latency free next neighbouring transmission in both horizontal directions can be achieved. If in every corner of the PAE input register (IR, arrow of bus into PAE) from bus and output register (OR, arrow from PAE to bus) to the bus are implemented, each neighbouring PAE can exchange data without latency (see FIG.).
It is possible to further optimise the above disclosed PAE arrangement. This can be done by using no separate bypass at all in all or some of the PAEs. The perferred embodiment comprises two ALUs, one of these being “complete” and having all necessary functions, for example multiplication and BarrelShift while the second has a reduced instruction set eliminating functions that require larger arrays such as multiplication and BarrelShift. The second ALU is in a way replacing BYPASS (as drawn). There are several possible positions for the register in switching elements per bus system, and two of the preferred positions per bus are shown in
Both ALUs comprise additional circuits to transfer data between the busses so as to implement the function of the bypass. A number of possible ways of implementations exist and two of these shall be explained as an example.
a) Multiplexer
Configurable multiplexers within the ALU are connected so that ALU inputs are bypassing the ALU and are directly connected to their outputs.
b) MOVE Instruction
A MOVE instruction, stored in Rc0 . . . Rcn is transferring within the respective processing clock of the function fold the data according to the input specified within the instruction to the specified output.
Superscalarity/Pipelining
It is possible and suggested as first way of improving performance to provide roughly superscalar FF ALU-PAEs which calculate for example 2, 4, 8 operations per bus clock @ FF=2, 4, 8, even while using the MUL opcode.
The basic concept is to make use of the VALID-flags of each internal register. MUL is implemented as one single opcode which is pipelined over two stages.
MUL takes its operands from the input registers Ri and stores the results into internal data registers Rd. VALID is set if data is stored into Rd. ADD (or any other Opcode, such as BSFT) uses the result in Rd if VALID is set; if not the execution is skipped according to the specified VALID behaviour. In addition the timing changes for all OpCodes, if the MUL instruction is used inside a PAE configuration. In this case all usually single cycle OpCodes will change to pipelined 2 cycle, OpCodes. The change is achieved by inserting a bypass able multiplexer into the data stream as well as into control.
The following program will be explained in detail:
In the first bus-cycle after configuration (t0) MUL is executed (assuming the availability of data at Ri0/1). The register pair Rd0/1 is invalid during the whole bus-cycle, which means during both FF-PAE internal clock cycles. Therefore ADD is not executed in the 2nd clock cycle. After t0 the result of MUL is written into the register pair, which VALID flags are set at the same time.
In t1 new data is multiplied. Since the VALID is set for Rd0/1 now the ADD command is executed in the 2nd clock cycle, but takes 2 clock cycles for over all execution. Therefore operand read and result write is inline for both operations, MUL as well as ADD.
The result of a MUL-ADD combination is available with 2 clocks latency in a FF=2 ALU-PAE. For FF>=6 no latency is inserted.
However since multiplication and all other commands are processed in parallel the machine streams afterwards without any additional delays.
If there are OpCodes besides MUL which require 2 clock cycles for execution (e.g. BSTF) the architecture must be modified to allow at least 3 data writes to registers after the second internal clock cycle.
The data path output multiplexer gets 2 times larger as well as the bus system to the output registers (OR) and the feedback path to the internal data registers (Rd).
If accordingly defined for the OpCodes, more than 4 internal registers can be used without increasing the complexity by using enables (en) to select the specific register to write in the data. Multiple registers are connected to the same bus, e.g. Rd0, Rd4, Rd8, Rd12. However not all combinations of register transfers are possible with this structure. If e.g. MUL uses Rd0 and Rd1 the following registers are blocked for the OpCode executed in parallel: Rd4, 5, 8, 9, 12, 13.
Register Map:
Datapath Architecture:
The Sequencer PAEs
Since there is a need to be able to run control flow dominated applications on the XPP III as well, Sequencer PAEs will be introduced. Such a PAE can be thought of as a very simple kind of processor which is capable to run sequential code within the XPP. This allows the efficient implementation of control flow oriented applications like the H.264 Codec on the array whereas with SEQ-PAEs missing the realization would be more difficult and resource consuming.
The SEQ-PAEs are not build from scratch. Instead such a tile will be build up by a closely coupling of a ALU-PAE and neighboring RAM-PAE, which can be seen in
Therefore the functionality of the ALU- as well as RAM-PAE has to be enhanced to be able to fulfill the requirements of such a SQE-PAE. This information will be given next.
ALU-PAE Enhancements
The extended version of the ALU-PAE is given in
Further more, since the ALU-PAE marks the execution part of the tiny processor, there is a need to transfer values to and from the internal register directly to the RAM. Therefore a additional multiplexer AM1 is inserted in the multiplexer hierarch of section 2. In the normal mode this multiplexer feeds the word from its predecessor to the next stage whereas in the SEQ mode an immediate value provided by the Imm. Register will be delivered. In addition in SEQ mode a value of one of the internal registers can be delivered to the RAM-PAE via the output of the multiplexer. However, it has also to be considered to provide a “LOAD reg, imm” since this is not much slower than “ADD reg, reg, imm”
To enable the RAM-PAE to write data to the internal register of the ALU-PAE another multiplexer is inserted in the multiplexer chain of section 4. Similar to the scenario given above this multiplexer will only be activated in SEQ mode whereas in normal mode this multiplexer will just forward the data of its predecessor. In one preferred embodiment, it is suggested to place RS2 behind BSFT-Mux in view of the delay. Data could be written into the internal registers via this. (LOAD reg, imm)]
As it has already been discussed, data can be processed during one or two cycles by the ALU-PAE depending on the selected arithmetic function. Due to the auto synchronization feature of the XPP and due to the fact that in normal mode a successive operation will not start before the previous one is finished, it does not really care if an operation lasts one or two clock cycles. Whereas the tile is working in SEQ mode there is a difference since we assume to have a pipeline character. This means that a one cycle operation could run in parallel with a two cycle module where the operation would be executed in stage two at this time. Due to the limited multiplexing capacities of a word—16 Bit—only one result could be written to the connected registers whereas the other one would be lost. In general there are three possibilities to solve this problem.
The first one could be that the compiler is capable to handle this problem. This would mean that it has to know about the pipeline structure of the whole SEQ-PAE as well as of a tile in detail. To prohibit a parallel execution the compile would have to add a NOP to every two cycle instruction for the structure given above. However this idea seems not to be convenient due to the strong relation between the hardware structure and the compiler. The drawback would be that every time changes are made to the hardware the compile would most likely have to be trimmed to the new structure.
The second idea could be to recognize such a situation in the decode stage of the pipeline. If a two cycle instruction is directly followed by an instruction accessing a one stage arithmetic unit it has to be delayed by one clock cycle as well.
The last possibility is to make the complete ALU-PAE look like a two stage execution unit. Therefore only one register has to be included in the multiplexer chain of section four right after the crossover from the multiplexer separating the one stage of the two stage modules. Obviously, this is preferred.
Comparing the last to ideas the third one seems to be the best one since only one register has to be inserted If we a closer look to the second solution special logic would be needed for analyzing the disallowed combination of instructions as well as logic for stopping the program counter (PC) and the instruction retardation. It has to be assumed that this logic would require much more area than the registers as well as the fact that the delay of the logic would possibly increase the critical path.
Since it has to be distinguished between the SEQ and the normal mode where a one cycle execution should still be available. This possibility is given by a multiplexer which allows to bypass the RS2 Register as shown in the corresponding figure. (
The RAM-PAE
A Short Description of the Stages
To get the SEQ-PAE working there still has to be provided more functionality. Right now the RAM-PAE will take care of it. As a first approach for realizing the sequencer a four stage pipeline has been chosen. The stages are, as it can be seen in
In the fetch stage the program counter for the next clock cycle will be calculated. This. means that it will be either incremented by 1 via a local adder or one of the program counters from the decode or execution stage 2 will be selected. The program counter of the execution stage thereby provides the address if a call instruction occurred whereas the program counter of the execution stage provides the PC if there has been a conditional jump. Right now the branch address can either be calculated out of the current PC and a value which either be an immediate value or a value from a internal registers of the ALU-RAM—indirect addressing mode—or an absolute value. This e.g. is necessary if there is return from a subroutine to the previous context whereas the according absolute PC will be provided by the stack bank.
In the decode stage the instruction coming from the code bank will be decoded. Necessary control signals and, if needed, the immediate value, for the internal execution stage 1 as well as for the execution stage 1 of the ALU-PAE will be generated. The signals include the control information for the multiplexers and gating stages of section two of the ALU-PAE, the operation selection of the ALU's tiles, e.g. signed or unsigned multiplication, and the information whether the stack pointer (SP) should be in/decremented or kept unchanged in the next stage depending on the fact if the instruction is either a call or jump. In case a call instruction occurred a new PC will be calculated in parallel and delivered to the fetch stage.
Furthermore the read address and read enable signal to the data bank will be generated in case of a load instruction. In the execution stage 1, which by the way is the first stage available on the ALU as well as on the RAM-PAE, the control signals for execution stage 2 of the ALU-PAE are generated. Those signal will take care that the correct output of one of the arithmetical tiles will be selected and written to the enabled registers. If the instruction should be a conditional jump or return the stack pointer will be modified in this stage. In parallel the actual PC will be saved to the stack bank at the address give by the Rsp EX1 register in case of a branch. Otherwise, in case of a return, the read address as well as the read enable signal will be applied to the stack bank.
In execution stage 2 the value of the PC will be calculated and provided to the multiplexer in the fetch stage in case of a jump. At the time write address and write enable signal to the data bank are generated if data from the ALU have to be saved.
Instead of two adders, it is possible to provide only one in the rpp path.
Pipeline Actions
In the following section a short overview of the actions that are taking place in the four stages will be given for some basic instructions. It should help to understand the behaviour of the pipeline. Since the instruction which is going to be discussed will be available at the instruction register the actins of the fetch stage will be omitted in this representation.
Instruction: Load value from data bank to R[n]
Instruction: Sore value from R[n] to data bank
1.3 Array Structure
First advantages over the prior art are obtained by using function folding PAEs. These as well as other PAEs can be improved.
The XPP-II structure of the PAEs consumes much area for FREG and BREG and their associated bus interfaces. In addition feed backs through the FREGs require the insertion of registers into the feedback path, which result not only in an increased latency but also in a negative impact onto the throughput and performance of the XPP.
A new PAE structure and arrangement is proposed with the expectation to minimize latency and optimize the bus interconnect structure to achieve an optimized area.
The XPP-III PAE structure does not include BREGs any more. As a replacement the ALUs are alternating flipped horizontally which leads to improved placement and routing capabilities especially for feedback paths i.e. of loops.
Each PAE contains now two ALUs and two BP paths, one from top to bottom and one flipped from bottom to top.
1.4 Bus Modifications
Within this chapter optimizations are described which might reduce the required area and the amount of busses. However, those modifications comprise several proposals, since they have to be evaluated based on real algorithms. It is possible to e.g. compose a questionnaire to collect the necessary input from the application programmes.
1.4.1 Next Neighbour
In XPP-II architecture a direct horizontal data path between two PAEs block a vertical data bus. This effect increases the required vertical busses within a XPP and drives cost unnecessarily. Therefore in XPP-III a direct feed path between horizontal PAEs is proposed.
In addition horizontal busses-of different length are proposed, i.e. next neighbour, crossing 2 PAES, crossing 4 PAEs.
1.4.2 Removal of Registers in Busses
In XPP-II registers are implemented in the vertical busses which can be switched on by configuration for longer paths. This registers can furthermore be preloaded by configuration which requires a significant amount of silicon area. It is proposed to not implement registers in the busses any more, but to use an enhanced DF or Bypass (PB) part within the PAEs which is able to reroute a path to the same bus using the DF or BP internal registers instead:
Here, it might be to decide how many resources are saved for the busses and how many are needed for the PAEs and/or how often must registers be inserted, are 1 or max. 2 paths enough per PAE (limit is two since DF/BP offers max. 2 inputs
1.4.3 Shifting n:1, 1:n Capabilities from Busses to PAEs
In XPP-II n:1 and 1:n transitions are supported by the busses which requires a significant amount of resources i.e. for the sample-and-hold stage of the handshake signals.
Depending on the size of n two different capabilities are provided with the new PAE structure:
This method saves a significant amount of static resources in silicon but requires dedicated PAE resources at runtime.
Here, it. might be worthwhile to evaluate how much silicon area is saved per bus how often occurs n=2, 2≦n≦4, n>4 the ratio between saved silicon area and required PAE resource and to decide on the exact bus structure in repsonse to one or all of said criteria.
1.5 FSM in RAM-PAEs
In the XPP-II architecture implementing control structures is very costly, a lot of resources are required and programming is quite difficult.
However memories can be used for a simple FSMs implementation. The following enhancement of the RAM-PAEs offers a cheap and easy to program solution for many of the known control issues, including HDTV.
Basically the RAM-PAE is enhanced by an feedback from the data output to the address input through a register (FF) to supply subsequent address within each stage. Furthermore additional address inputs from the PAE array can cause conditional jumps, data output will generate event signals for the PAE array. Associated counters which can be reloaded and stepped by the memory output generate address input for conditional jumps (i,e, end of line, end of frame of a video picture).
A typical RAM PAE implementation has about 16-32 data bits but only 8-12 address bits. To optimize the range of input vectors it is therefore suggested to insert some multiplexers at the address inputs to select between multiple vectors, whereas the multiplexers are controlled by some of the output data bits.
One implementation for an XPP having 24 bit wide data busses is sketched in the next figure. 4 event inputs are used as input, as well as the lower for bits of input port Ri0. 3 counters are implemented, 4 events are generated as well as the lower 10 bits of the Ro0 port.
The memory organisation suggested here may be as follows:
The width of the counters is according to the bus width of the data busses.
For a 16 bit implementation it is suggested to use the carry signal of the the counters as their own reload signal (auto reload), also some of the multiplexers are not driven by the memory but “hard wired” by the configuration.
The proposed memory organisation is as follows:
1.6 IOAG Interface
1.6.1 Address Generators and Bit Reversal Addressing
Implemented within the IO interfaces are address generators to support e.g. 1 to 3 dimensional addressing directly without any ALU-PAE resources. The address generation is then done by 3 counters, each of them has e.g. configurable base address, length and step width.
The first counter (CNT1) has a step input to be controlled by the array of ALU-PAEs. Its carry is connected to the step input of CNT2, which carry again is connected to the step input of CNT3.
Each counter generates carry if the value is equal to the configured length. Immediately with carry the counter is reset to its configured base address.
One input is dedicated for addresses from the array of ALU-PAEs which can be added to the values of the counters. If one or more counters are not used they are configured to be zero.
In addition CNT1 supports generation of bit reversal addressing by supplying multiple carry modes.
1.6.2 Support for Different Word Width
In general it is necessary to support multiple word width within the PAE array. 8 and 16 bit wide data words are preferred for a lot of algorithms i.e. graphics. In addition to the already described SIMD operation, the IOAG allows the split and merge of such smaller data words.
Since the new PAE structure allows 4 input and 4 output ports, the IOAG can support word splitting and merging as follows:
Input ports are merged within the IOAG for word writes to the IO.
For output ports the read word is split according to the configured word width.
1.7 Multi-Voltage Power Supply and Frequency Stepping
PAEs and busses are build to perform depending on the workload. Therefore the clock frequency is configurable according to the data bandwidth, in addition clock gating for registers is supported, busses are decoupled using row of AND gates. Dynamically clock pulses are gated, whenever no data can be processed.
Depending on the clock frequency in the PAEs and the required bandwidth for the busses the voltage is scaled in an advanced architecture. Within the 4S project such methods are evaluated and commercially usable technologies are researched.
1.8 XPP/μP Coupling
For a closed coupling of a μP and a XPP a cache and register interface would be the preferable structure for high level tools like C-compilers. However such a close coupling is expected not to be doable in a very first step.
Yet, two different kind of couplings may be possible for a tight coupling:
2.1 Overview
In a preferred embodiment, the ALU-PAE comprises 3 paths:
Then, each of the paths contains 2 data busses and 1 event bus. The busses of the DF path can be rerouted to the ALU path by configuration.
2.2 ALU Path Registers
The ALU path comprises 12 data registers:
Eight instruction registers are implemented, each of them is 24 bit wide according to the opcode format.
Three special purpose registers are implemented:
2.3 Data Duplication and Multiple Input Reads
Since Function Folding can operate in a purely data stream mode as well as in a sequential mode (see 1.2) it is useful to support Ri reads in dataflow mode (single read only) and sequential mode (multiple read). The according protocols are described below:
Each input register Ri can be configured to work in one of two different modes:
Dataflow Mode
This is the standard protocol of the XPP-II implementation: A data packet is taken read from the bus if the register is empty, an ACK handshake is generated. If the register is not empty ACK the data is not latched and ACK is not generated.
If the register contains data, it can be read once. Immediately with the read access the register is marked as empty. An empty register cannot be read.
Simplified the protocol is defined as follows:
Sequencer Mode
The input interface is according to the bus protocol definition: A data packet is taken read from the bus if the register is empty, an ACK handshake is generated. If the register is not empty ACK the data is not latched and ACK is not generated.
If the register contains data it can be read multiple times during a sequence. A sequence is defined from Rpp=0 to Rpp=Rjb. During this time no new data can be written into the register. Simultaneously with the reset of Rpp to 0 the register content is cleared an new data is accepted from the bus.
Simplified the protocol is defined as follows:
2.4 Data Register and Event Handling
Data registers are directly addressed, each data register can be, individually selected. Three address opcode form is used, rt←rs1, rs0. An virtual output register is selected by adding ‘o’ behind the register. The result will be stored in rt and copied to the virtual output register rv as well according to the rule op out (rv, rt)←rs1, rs0.
Please note, accessing input and (virtual) output registers follow the rules defined in chapter 2.3.
Events are used equal to data registers. All input and internal events can be addressed directly, output events are used whenever an ‘o’ is added behind the event.
2.4.1. ACCumulator Mode
To achieve low power consumption and for better supporting DSPlike algorithms an accumulator register is available which can be addressed by just one set bit for the result register (ao) and operand register (ai).
For commutative operations always operand register 1 is replaced by ai. For non commutative operations as SUBtract operand register 1 selects, whether ai is the first or second operand. register 2 defines the accordingly other operand.
It is to be noted that it has to be clarified whether a real ACCumulator mode makes sense or just a MAC-command should be implemented to handle the multiply accumulate in a single command consuming two clock cycles with an implicit hidden accumulator access.
2.4.2. Parameter Stack Mode (PSTACK)
Unused entries in the Opcode Registers Rc can operate as stack for constants and parameters. At Rpp==0000 the Rps PStack registers points to Rjb+1, which means the PStack area starts immeadiately behind the last entry in the Opcode register file.
To access the PStack, the FF-PAE must be in the Fast-Parameter Mode. Each read access to Ri3 is redirected to read from the PStack, whereas after each read access the pointer incremented with one. There is no check for an overflow of the PStack pointer implemented, an overflow is regarded as a program bug.
2.4.3 n:1 Transitions
n:1 transitions are not supported within the busses any more. Alternatively simple writes to multiple output registers Ro and event outputs Eo are supported. The Virtual Output registers (Rv) and Virtual Event (Ev) are translated to real Output registers (Ro) and real Events (Eo), whereas a virtual register can be mapped to multiple output registers.
To achieve this a configurable translation table is implemented for both data registers and event registers:
Rv3 unused
2.4.4 Accessing Input and Output Registers (Ri/Rv) and Events (Ei/Ev)
Independently from the opcode accessing input or output registers or events is defined as follows:
Reading an Input Register:
Writing to an Output Register:
2.4.5.Multi-Config Mode
The Multi-Config Mode allows for selecting 1 out of maximum 4 stored configurations. Incoming events on Fui0,1 and Fvi0,1 select one of the 4 configurations. Only one Event shall be active at a clock cycle.
The selection is done by a simple translation, each event points to a specific memory address.
Long configurations may use more than 3 opcode by using the next code segments as well. In this case, the according events can not be used.
2.5 Opcode Format
24 bit wide 3 address opcodes are used in a preferred embodiment:
op rt←ra, rb
Source registers can be Ri and Rd, target registers are Rv and Rd. A typical operation targets only Rd registers. If the source register for ra is Ri[x] the target register will be Rd[x].
The translation is shown is the following table:
Each operation can target a Virtual Output Register Rv by adding an out tag ‘o’ as a target identifier to the opcode:
op (rt, rot)←ra, rb
Data is transferred to the virtual output register and to the according internal register as well:
2.5.1 Conditional Execution
The SKIPE command supports conditional execution. Either an event or ALU flag is tested for a specific value. Depending on the check either the next two addresses are executed (Rpp+1) or skipped (Rpp+3). If an incoming event is checked, the program execution stops until the event is arrived at the event port (RDY handshake set).
SKIPE supports conditional execution of any OpCode which is not larger than two memory entries.
In SEQ-PAEs, which support CALL and RET OpCodes, also stack based subroutine calls are supported.
2.6 Clock
The PAE can operate at a configurable clock frequency of
2.7 The DF Path
The DataFlow path comprises the data registers Bri0 . . . 3 and Bro0 . . . 3 as well as the event register Bui/Bvi0 . . . 3 and Buo/Bvo0 . . . 3.
The main purpose of the DF path is to establish bus connections in the vertical direction. In addition the path includes a 4 stage FIFO for each of the data and event paths.
The DF path supports numerous instructions, whereas the instruction is selected by configuration and only one of them can be performed during a configuration, function folding is not available.
The following instructions are implemented in the DF path:
2.9 Parameter Broadcast and Update
Parameters and constants can be updated fast and synchronous using input register Ri3 and event input Ei7.
Depending on the update mode, data packets at the input register Ri3 are copied subsequently into Rd3, Rd2 and Rd1 at each access of the according register by the PAE, if the event Ei7 is set. Afterwards all input data at Ri3 is propagated to the output register Ro3, also the Eo7 event output is set, to indicate following PAEs the occurrence of a fast parameter update, which allows to chain PAEs, together (i.e. in a multi-TAP FIR filter) and updating all parameters in the chain.
Also the OpCode UPDATE updates all registers subsequently if Ei7 is set, depending on the Update Parameter Mode (upmcfg=nn10).
Also the register update can be configured to occur whenever Rpp==0 and Ei7 is set by upmcfg=nn01.
In both cases nn indicates the number of registers to be updated (1-3).
Ei7 must be 0 for at least one clock cycle to indicate the end of a running parameter update and the start of a new update.
3 Input Output Address Generators (IOAG)
The IOAGs are located in the RAM-PAEs and share the same registers to the busses. An IOAG comprises 3 counters with forwarded carries. The values of the counters and an immediate address input from the array are added to generate the address.
One counter offers reverse carry capabilities.
3.1 Adressing Modes
Several addressing modes are supported by the IOAG to support typical DSP-like addressing:
3.1.1 Immediate Addressing
The address is generated in the array and directly fed through the adder to the address output. All counters are disabled and set to 0.
3.1.2 xD Counting
Counters are enabled depending on the required dimension (x-dimensions require x counters). For each counter a base address and the step width as well as the maximum address are configured. Each carry is forwarded to the next higher and enabled counter; after carry the counter is reloaded with the start adddress.
A carry at the highest enabled counter generates an event, counting stops.
3.1.3 xD Circular
The operation is exactly the same as for xD counting, with the difference that a carry at the highest enabled counter generates an event, all counters are reloaded to their base address and continue counting.
3.1.4 Stack
One counter (CNT1) is used to decrement after data writes and increment after data reads. The base value of the counter can either be configured (base address) or loaded by the PAE array.
3.1.5 Reverse Carry
Typically carry is forwarded from LSB to MSB. Forwarding the carry to the opposite direction (reverse carry) allows generating address patterns which are very well suited for applications like FFT and the like. The carry is discarded at MSB.
For using reverse carry a value larger than LSB must be added to the actual value to count, wherefore the STEP register is used.
Step=1000 b
The counter is implemented to allow reverse carry at least for STEP values of −2, −1, +1, +2.
4. ALU/RAM Sequencers˜SEQ-PAEs
Each ALU-PAE at the left or right edge of the array can be closely coupled to the neighbouring RAM-PAEs as an IP option, thus allowing for configure a sequencer. For compatibility reasons, the data and opcode width of the sequencer is 16 bits.
The ALU-PAEs can operate exactly as array internal ALU-PAEs but have several extensions. Operation is Sequencer mode the register file is 8 data registers wide, Fu and Fv flags are, used as carry, sign, null, overflow and parity ALU flag word.
The address width is accordingly 16 bit. However since the RAM-PAE size is limited it is segmented into 16 segments. Those segments are used for code, data and stack and must be individually preloaded by the compiler.
4 segment registers point to the specific segments:
AuxiliaryBank Points to any Segment (but Code), Allowing Copy Operations Between Segments
The compiler has to take care that necessary data segments are preloaded and available. For cost reasons there is no automatic TLB installed.
Also segments have to be physically direct addressed due to the absence of TLBs. This means that the compiler has to implement range checking functions for according addresses.
Code segments behave accordingly to data segments. The compiler has to preload them before execution jumps into them. Also jumps are physically direct addressed, due to the absence of TLBs again.
A relocation of any segments is not possible, the mapping is fixed by the compiler.
The memory layout is shown below. A simple check mechanism is implemented to validate or invalidate memory segments.
At least the CodeBank (CB) and StackBank (SB) must be set. The first CodeBank must start at location 0000 h. For all other banks 0000 h is an illegal entry. Loading segments to the memory validates them, accordingly flushing invalidates them.
Memory banks are updates in terms of loaded or flushed in the background by a DMA engine controlled by the following opcodes
The address generators in the IOAG interfaces can be reused as DMA engine.
Memory banks can be specifically validated or invalidated as follows:
The bank pointers are added to the address of any memory access. Since the address pointer can be larger than the 6 bits addressing a 64 line range, segment boarders are not “sharp”, which means, can be crossed without any limitation. However the programmer or compiler has to take care that no damage occurs while crossing them. If an invalid segment is reached a flag or trap is generated indicating the fault, eventually just wait states are inserted if a segment preload is running already in the background.
Alternatively a more advanced valid checking scheme can be implemented as shown below:
In difference to PAEs which require 24-bit instructions sequencers use 16-bit instructions only. To use the same instruction set and to keep the decoders simple, just the last 8 bits are discarded in sequencer mode.
4.1 IOAGs
IOAGs may comprise a 4-8 stage data output buffer to balance external latency and allow reading the same data address directly after the data has been written, regardless of external bus or memory latencies (up to the number of buffer stages).
In the follwoing, a number of OpCodes and their meanings is suggested:
Appendix B
In the following, an exaple for the use of function folding is given:
Function Folding and Fast Parameter Update Example FIR
upmcfg=1100
Alternative using MAC opcode, parameter pop and looping read Rd0, Ri1;
upmcfg=1110
upmcfg=1101
In the above, an improved data processor array has been described. Although only in some instances, it has been pointed out that reference to a certain number of registers, bit width etc. is for explanation only, it is to be understood that this also holds where such reference is not found.
If the array is to be very large or in case a real time process is run where two different fragments of an array unknown at compile time have to communicate with each other so as to enable data processing, it is advantageous to improve the performance by ensuring that a communication path can be set up. Several suggestions have been made already, e.g. Lee-Routing and/or the method described in PACT 7. It is to be understood that the following part of an improved array design might result in an improved circuitry for certain applications but that it is not deemed absolutely and inevitably necessary to implement it with e.g. a function fold PAE. Rather, the other suggestions for improvement will result in significant improvements on their own as will be understood by the average skilled person.
Routing Improvement
The suggested improvement described hereinafter concerns the static routing network for reconfigurable array architectures. Hereby this static network is enhanced by implementing additional logic to adaptive runtime routing.
The routing technique described in this document is instruction based which means that each routing process must be started by an instruction. If the user wants to establish a routing between two cells, he has to bring a specific instruction into the source cell. The hardware within the array calculates based on the instruction fields values the desired routing direction and establishes the logic stream. The routing process happens stepwise from one functional unit to another whereby each cell decides which direction should be taken next. On the way to an established route we defined three valuable states of the routing resources. The first state is the physical route or link. This means that the resources of this route are not used and available to routing processes. The second state is named temporal route or link. This state describes the temporarily not available link, which means that this link is in use for routing purposes but the mentioned routing is not confirmed yet. The problem here is that this route can be confirmed in the future or released if the successor cells are able to realise the desired routing. The last state is the logical route or link. This state represents an established route on the array which is able to transport calculation data.
This routing technique uses coordinates on the array to calculation routings. Each FU possesses unique coordinate's und on the basis of this information it is able to determine the routing direction to each desired cell Within the array. This concept is the basis for the adaptive runtime routing described in this document. The needed control logic for adaptive routing is implemented within the routing unit, especially within the routing controller which controls the interconnect matrix at runtime. Therefore the routing controller is able to analyze the incoming data of all input ports of the concerned FU and come to a decision what to do next.
Routing Establishment
For the purpose of incoming data analyzing and data buffering each input port owns so called in-registers (InReg). Additional to those standard registers there are InReg-controllers implemented (InRegCtrl). Those finite state machines (FSMs) have the job to store the actual state of the input links and in dependency of the actual state to trigger routing requests or release not required routings. To fulfil its job each InRegCtrl is connected to an in-controller (InCtrl) which is implemented exactly once per FU. Important requirement for requesting of new routings is that the mentioned input resource (InReg, InRegCtrl) are not used and so in the state of physical link.
InCtrl gets requests of all InRegCtrls all over the time and forwards one request after another to the routing controller (RoutCtrl). The selection which InRegCtrl should be served first is dependant on the routing priority of the input link and/or which input link was served last. Based on the coordinate information of the target cell and the coordinates of the actual FU the RoutCtrl calculates the forward direction for the requested input link. Thereby the RoutCtrl takes into account additional parameters like optimum bit (will be described later), the network utilisation towards the desired direction, etc.
If the direction calculation within the RoutCtrl was successful the RoutCtrl forwards the request with additional information about the output port to the interconnect matrix, which connects the input port with calculated output port. If this is done the RoutCtrl signals the successful routing operation to InCtrl. Because the actual reached routing state is not final it is necessary to store the actual state. This happens within the queue-request-registerfile (QueueRRF). Therefore the InCtrl is directly connected to the QueueRRF and is able to store the desired information. At this point the related input and output links reach the temporal link state and are temporarily not available for other routing processes.
Due the fact that the QueueRRF is able to store more than one routing entry, the InCtrl is able to hold multiple routing processes at the same time. But for the purpose of high hardware area consumption the direction calculation is realized once within the RoutCtrl.
The established temporal routing stays stored within the QueueRRF till the point the successor cell acknowledges the routing. In this case the InCtrl clear the according entry in the QueueRRF and signals the,successful routing to the InCtrl. The InRegCtrl changes into the state logical route and signal the predecessor cell the successfully finished routing process.
The other case can happen if the successor cell is not able to establish the desired route. In this case the InCtrl forwards a new request to the RoutCtrl based on the QueueRRF-entry. This request leads to new routing suggestion which will be stored within the QueueRRF.
If all available and expedient directions are checked and routing trials failed the InCtrl signals to InRegCtrl the failed routing. The InCtrl signals the same routing miss to the predecessor cell and finishes the routing process in the current cell.
Within the routing process there are two exceptions how the routing unit establishes a desired routing. Those exceptions affect the source and the target cell. The exception in both cases is that as well the source cell as the target cell do not need to route the started/ending routing through the interconnect matrix. To connect the FMs to the output links of cells simple multiplexers are used. Those multiplexers are implemented after the interconnect matrix and have to be switched explicitly. This happens after the routing process is finished. The exception lies in the finishing state. Here the InRegCtrl doesn't have to acknowledge the successful routing the predecessor it just has to consume the actual routing instruction in the InReg instead. This happens after the InCtrl signals the successful routing. Additionally the InReg switches the output multiplexer associated to the output port of the FM and finishes the routing establishment. The information needed the switch the right output multiplexer gets the InCtrl from the RoutCtrl.
Otherwise if the routing fails the InCtrl asserts cell specific interrupt line and signals the failure to the system.
The second exception concerns the target routing cell. Here it is important to connect the new route with the input ports of the local FM. Therefore simple multiplexers are used which are implemented before the interconnect matrix. If an ongoing routing process reaches the target cell the InCtrl identifies the target achievement and switches. the associated input multiplexer to forward the incoming data to the input port of the FM. This is the point where, the successful route establishment signal is generated by the InRegCtrl after InCtrl signals the success. Here the InRegctrl has the last job to finish the routing process by deleting the routing instruction and going to logical state.
Releasing Established Routing
For releasing of the logically established routings we introduced special instructions, so called end packets. The only purpose of those instructions is the route-dissolving by inject the necessary end packet into the logic established routing. There are two ways how the routings can be released. The first possibility is the global releasing. This means that all routes which are following the route where the end packet is injected will be released. This function is useful to delete whole configurations with one single instruction. For this purpose it is important that the FMs are able to forward the end packet unaltered through the internal datapaths.
The second way for route releasing is the local route releasing. Here it is possible to release single established routes between output and input ports of FMs. The end packets are not propagated through the FMs. In this case the end packet will be consumed by the last InRegCtrl.
The internal RU communication is similar to the routing process. If the InRegCtrl determines incoming end packet and the InRegctrl is in the logic route state, the InRegCtrl forwards the route release request to the InCtrl. The InCtrl clears the entries either within the interconnect matrix or within the input multiplexers registers or within the output multiplexer registers. Meanwhile the InRegCtrl consumes (in case of the local end packet and last cell in the chain) the instruction and goes to the idle state. If the end packet was a global instruction the InRegCtrl forwards alway the end packet to the successor.
Additional Features
For the purpose of priority control, we introduced a priority system to influence the order in which the RU serves the incoming routing requests. Therefore the instructions contain priority fields which describe the priority level. Higher values in this field result in higher priority und will be preferred by the RU during the runtime-routing. The priority field has direct influence on the selection of the incoming routing requests from the InRegCtrls to InCtrl.
Some inner configuration communication streams require strictly defined latency to reach the desired performance. Therefore it is very important to keep the maximum register chain length. To decrease the latency of the routed streams its is necessary to ensure that the array chose always the best routing between source and target, but this requirement may lead to not routable streams if this feature will be always required. To ease this problem we introduced a special bit within the routing instruction, so called optimum bit (OptBit). This bit has to be activated if the optimum routing is definitely required. In this case the array tries to reach this requirement und delivers an interrupt if fails.
The alternative to reach the required latency is the speed path counter. This counter gives the possibility to bypass a specific number of registers before buffering again. Therefore we defined a reference value and the counter value. Both numbers are stored within the instruction field. Each passed cell respective the RU compares the counter value and the reference value. If both values are equal then the actual cell buffers the stream and resets the counter. If the counter is smaller than the reference value the current buffer will be bypassed and the counter incremented by one. In this way it is possible to bypass a number of buffers which equals exactly to reference value.
Multi-Grained Communication Links
In addition to the coarse-grained point-to-point links we introduced more flexible multi-grained point-to-point links. Hereby one single point-to-point link connects two neighbor cells respective the RUs within those cells. One coarse-grained link consists of a set of wires, e.g. 32 wires for one 32 link, and additionally protocol signals. The whole vector is handled by a single set of control signals which makes this communication resource not usable for multi-grained communication.
To reach this requirement we divided the whole 32 bit vector into single strips, e.g. with groups of 8 times 1 bit strips and 3 times 8 bit strips. Each strip obtained separate control signals and is able to operate independently from other strips.
The idea behind this division is to combine those strips to logical multi-grained sub-links. If you have one multi-grained link you can use the whole vector as one interrelated 32 bit vector or split the whole vector into sub-channels. In this configuration each strip can be one single sub-channel or a group of strips can be gathered to a single sub-channel of desired bit-width. You just have—in respect of hardware costs—to consider that one sub-channel has to fit into one multi-grained link.
Multi-Grained Routing
In order to route multi-grained channels it's necessary to use the coarse grained links to support the routing process. The idea is to route two links in parallel, one coarse-grained link to support multi-grained routing and one multi-grained link, which will contain the final multi-grained stream. Therefore we defined a two packet routing instruction with needed data fields. The first instruction packet contains—compared to coarse-grained routing instruction—additional bit mask to specify used multi-grained sub-links and multi-grained link ID to identify the associated multi-grained link. The other features like described above—optimum bit, speed path, priority routing—are support in this routing mode as well. The routing process within the RU is performed similar to the coarse-grained routing.
The first packet which arrives in a cell is analyzed by the InRegCtrl and a request is generated and forwarded to the InCtrl. InCtrl forwards the request to the RoutCtrl and wait for the acknowledgement. If RoutCtrl finds one possible routing direction, the InCtrl gets the successful acknowledgement and the temporal routing will be established by the RoutCtrl. Next, the actual job will be stored within the QueueRRP and the InCtrl waits for the acknowledgement from the successor cell. If RoutCtrl is not able to find a possible routing, the InCtrl gets negative acknowledgement and which will be forwarded to the associated InRegCtrl, which generates the route unable signal to the predecessor cell and quits the routing process within this cell.
If the successor cell signals successful routing, the InRegCtrl clears the related entry in the QueueRRP and finishes the routing. If the successor cell is not able to establish a rout to the destination cell, it generates negative acknowledgement signal. Hereupon, the InCtlr starts new request to the RoutCtrl and handle the responses as described above.
The difference between the cbarse-grained routing and multi-grained routing lies in the handling of the multi-grained interconnect matrix. Each strip of a multi-grained link is handled separately. The RoutCtrl forwards the switch request to the strip matcher. Strip matcher has the job to analyze the input strips and to match them to the output link according to already used strips. What strip matcher is doing is to map the problem of strip matching into the time domain and switches the needed switchboxes for each strip separately one after another.
Routing Packet for Coarse-Grained Streams:
Routing Instruction for Multi-Grained Streams (First Packet):
Second Packet of the Routing Instruction for Multi-Grained Streams:
End Packet Instruction:
Data Packet:
Figures Relating to Improved Way of Routing:
| Number | Date | Country | Kind |
|---|---|---|---|
| 03019428.6 | Aug 2003 | EP | regional |
| 03025911.3 | Nov 2003 | EP | regional |
| 103 57 284.8 | Dec 2003 | DE | national |
| 03028953.2 | Dec 2003 | EP | regional |
| 03079015.8 | Dec 2003 | EP | regional |
| 04002604.9 | Feb 2004 | EP | regional |
| 04002719.5 | Feb 2004 | EP | regional |
| 04003258.3 | Feb 2004 | EP | regional |
| 04004885.2 | Mar 2004 | EP | regional |
| 04075654.6 | Mar 2004 | EP | regional |
| 04005403.3 | Mar 2004 | EP | regional |
| 04013557.6 | Jun 2004 | EP | regional |
| 04018267.7 | Aug 2004 | EP | regional |
| 04077206.3 | Aug 2004 | EP | regional |
| Filing Document | Filing Date | Country | Kind | 371c Date |
|---|---|---|---|---|
| PCT/EP04/09640 | 8/30/2004 | WO | 11/10/2006 |
