• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    23 ABSTRACT In a digital signal processor (200) comprising at least one vector execution unit (203,205) and at least a first memory unit (230, 231) a third unit is arranged to provideaddressing data in the form of an address vector to be used for addressing the firstmemory unit (230, 231), said third unit being connectable to the first memory unit(230, 231) through the on-chip network (244), in such a Way that data provided fromthe third unit (250) can be used to control the reading from and/or the Writing to thefirst memory unit (230, 231). This enables fast reading from and Writing to a memory unit of data in any desired order.
    公开号:SE1151230A1
    申请号:SE1151230
    申请日:2011-12-20
    公开日:2013-06-21
    发明作者:Anders Nilsson；Eric Tell；Erik Alfredsson
    申请人:Mediatek Sweden Ab；
    IPC主号:
    专利说明:

    Digital signal processor and method for addressing a memory in a digital signal processor Technical FieldThe present invention relates to a digital signal processor according to the preamble of claim l. Such a processor is particularly suitable for OFDM systems.
    Background and Related ArtMany mobile communication devices use a radio transceiver that includes one or more digital signal processors (DSP).
    For increased performance and reliability many mobile terminals presently use a typeof DSP known as a baseband processor (BBP), for handling many of the signalprocessing functions associated with processing of the received the radio signal and preparing signals for transmission.
    Many of the functions frequently performed in such processors are performed on largenumbers of data samples. Therefore a type of processor known as Single InstructionMultiple Data (SIMD) processor is useful because it enables the same instruction to beperformed for a whole vector of data rather than on one integer at a time. This kind ofprocessor is able to process vector instructions, which means that a single instructionperforms the same function to a limited number of data units. Data are grouped into bytes or words and packed into a vector to be operated on.
    As a further development of SIMD architecture, Single Instruction stream MultipleTasks (SIMT) architecture has been developed. Traditionally in SIMT architecture oneor two vector execution units that use SIMD data-paths have been provided in association with an integer execution unit, which may be part of a core processor.
    International Patent Application WO 2007/018467 discloses a DSP according to the SIMT architecture, having a processor core including an integer execution unit and a program memory, and two vector execution units which are connected to, but notintegrated in the core. The vector execution units may be Complex Arithmetic LogicUnits (CALU) or Complex Multiply-Accumulate Units (CMAC). The data to beprocessed in the vector execution units are provided from data memory units connected to the vector execution units through a on-chip network.
    The memory units comprise address generation units which are arranged to control theread or write order at any given time. For increased flexibility, the address generationunit can enable different readout modes, or patterns, such as reading from every nthaddress in the memory. These modes have to provide a regular pattern, which limitsthe possible ways data can be read or written. Further, the available modes are preselected for a particular address generation unit, and cannot be changed.
    The article Nilsson, A and Tell, E: “An llmm2, 70mW fully programmable basebandprocessor for mobile WiMAX nad DVB-T/H in 0.12 um CMOS”, describes a SIMTtype DSP and briefly states that “as the memory banks can accept external addressingfrom the network, integer memories as well as accelerators can be used to provideaddress sequences for irregular vector addressing. This also provides the ability to doindirect vector addressing”. This article does not address any of the problems involvedin actually implementing such a solution, and also hence does not provide a workable solution.
    Summary of the InventíonIt is an objective of the present invention to enable a more flexible addressing of the data memories of a processor in SIMT architecture.
    This objective is achieved according to the present invention by a digital signalprocessor comprising at least one functional unit, which may be a vector executionunit, an integer execution unit or an accelerator, and at least a first memory unitarranged to provide data to be operated on by the functional unit, a third unit and an on-chip network connecting the functional unit, the first memory unit and the third unit. The digital signal processor is characterized in that the third unit is arranged toprovide addressing data in the form of an address vector to be used for addressing thefirst memory unit, said third unit being connectable to the first memory unit in such away that data provided from the third unit can be used to control the reading fromand/or the writing to the first memory unit and that the processor further comprises amemory address interface unit arranged to compensate for the latency between the first and the third unit.
    The invention also relates to a method of addressing a memory in a digital signalprocessor comprising at least one functional unit and at least a first memory unitarranged to provide data to be operated on by the functional unit, and a on-chipnetwork connecting the functional unit and the first memory unit, and a third unitarranged to provide addressing data for the first memory unit in the form of an addressvector, said method comprising the steps of 0 setting the first memory unit to receive addressing data from the third unit, providing addressing data from the third unit to the first memory unit,0 reading data from, or writing data to the first memory unit according to the addressing data.
    Hence, according to the invention, addressing can be achieved fast and efficiently inany order throughout the memory unit. Data can be addressed in a memory in anyorder, as efficiently as an ordered sequence of data since the addressing may behandled in parallel with the processing performed by the functional unit. The firstmemory unit may be arranged to receive addressing data from the third unit through a dedicated bus or through the on chip network.
    The memory address interface unit provides a solution to the problems caused bylatency between the first and the third unit. In particular, a read signal from the firstunit will take some clock cycles to reach the third unit. When starting read operations there will be a delay of several clock cycles before the first data item actually reaches the execution unit. By storing the first address samples in the memory address interface unit, before the first unit requests them, the delay at startup can be reduced.
    The third unit may be any unit in the processor, for example0 a memory unit, referred to as the second memory unit, preferably an integermemory unit, 0 a scalar execution unit, 0 a vector execution unit, or 0 an accelerator unit Different units may be used for the addressing of different memory banks.
    The first memory unit may be a complex memory or an integer memory.
    The memory address interface unit preferably comprises- Latency memory means arranged to store a number representative of the latencybetween the first and the third unit,- Sample memory means arranged to store the sample number, that is, the numberof address items to be transferred from the third unit to the first unit, - A register for forwarding the address items from the third unit to the first unit Accordingly, the memory address interface unit is preferably arranged to perform thefollowing functions:- reading at least a first address item from the third unit When the first unitconnects to the third unit for obtaining address information, without waiting fora read signal- when a read signal is received from the first unit, forwarding the first addressitem to the first unit- reading following address items from the third unit subsequently until alladdress items have been read and forwarding the following address items to the first unit subsequently when read signals are received from the first unit.
    The memory address interface unit preferably keeps track of the number of addressitems to read from the third unit by changing the sample counter each time an addressitem is read. This is advantageous because the memory address interface unit willcontinue to receive read requests from the first unit after it has read the last addresssample from the third unit. In this way the memory address interface unit knows whento stop retrieving address items from the third unit. When the retrieval stops, only the last address samples are sent from the memory address interface unit to the third unit.
    It is often important that all calculations are performed with the same timing,regardless of where the data are coming from. Since the latency may vary dependingon which units are involved, an advantage of the invention is that the timing can becontrolled by avoiding delay differences caused by different numbers of pipeline steps between different units.
    In a particularly advantageous embodiment, the second memory unit comprises anaddress generation unit arranged to control the reading of address data from the secondmemory unit according to a predefined pattern. In this embodiment, the method maycomprise the step of controlling reading of addresses from the second memory unit bymeans of an address generation unit arranged to control the reading of address datafrom the second memory unit according to a predefined pattern. This enables thehandling of subsets of the addresses kept in the second memory unit, which is particularly advantageous for OFDM-based digital TV applications.
    The processor according to the invention is particularly well suited for OFDM-basedtelecommunication standards, for example, according to the LTE and LTE advanced standards.
    Brief Description of the DrawíngsIn the following the invention will be described in more detail, by way of example,and with reference to the appended drawings.
    Fig. l illustrates an example of the SIMT architecture.
    Fig. 2 illustrates the function of address pointers Fig. 3 illustrates a first embodiment of the invention in a simplified example of theSIMT architecture.
    Fig. 4 illustrates a second embodiment of the invention in a simplified example of theSIMT architecture.
    Fig. 5 illustrates a memory address interface unit according to an embodiment of theinvention.
    Fig. 6 illustrates a third embodiment of the invention.
    Detailed description of Embodiments Figure 1 illustrates an example of a prior art baseband processor 200 according to theSIMT architecture. The processor 200 includes a controller core 201 and a first 203and a second 205 vector execution unit, which will be discussed in more detail below.A FEC unit 206 is connected to the on-chip network. In a concrete implementation, of course, the FEC unit 206 may comprise several different units.
    A host interface unit 207 provides connection to a host processor not shown in Fig. 1in a manner well known in the art. A digital front end unit 209 provides connection to a front end unit in a manner well known in the art.
    As is common in the art, the controller core 201 comprises a program memory 211 aswell as instruction issue logic and functions for multi-context support. For eachexecution context, or thread, supported this includes a program counter, stack pointerand register file (not shown explicitly in figure 1). Typically, 2-3 threads aresupported. The controller core 201 also comprises an integer execution unit 212 of a kind known in the art.
    The first vector execution unit 203 in this example is a CMAC vector execution unit,and the second vector execution unit is a CALU vector execution unit. Each vectorexecution unit 203, 205 comprises a vector controller 213, a vector load/store unit 215 and a number of data paths 217. The vector controller of each vector execution unit is connected to the program memory 211 of the controller core 201 via the issue logic, to receive issue signals related to instructions from the program memory.
    The function of the data paths 217, 227 and the vector load/store units 215, 225 is well known in the art and will not be discussed in any detail in this document.
    There could be an arbitrary number of vector execution units, including only CMACunits, only CALU units or a suitable number of each type. There may also be othertypes of vector execution unit than CMAC and CALU. As explained above, a vectorexecution unit is a processor that is able to process vector instructions, which meansthat a single instruction performs the same function to a number of data units. Datamay be complex or real, and are grouped into bytes or words and packed into a vectorto be operated on by a vector execution unit. In this document, CALU and CMACunits are used as examples, but it should be noted that vector execution units may be used to perform any suitable function on vectors of data.
    As is known in the art, a number of accelerators 242 are typically used, since theyenable efficient implementation of certain baseband functions such as channel codingand interleaving. Such accelerators are well known in the art and will not be discussedin any detail here. The accelerators may be configurable to be reused by many different standards.
    An on-chip network 244 connects the controller core 201, the digital front end unit209, the host interface unit 207, the vector execution units 203, 205, the memory banks230, 232, the integer bank 238 and the accelerators 242. Vector execution units, scalarexecution units, integer execution units and accelerators are collectively referred to inthis document as functional units. A scalar execution unit is only able to process one sample at a time, but this sample may have a real or complex value.
    To enable several concurrent vector operations, the processor preferably has a distributed memory system where the memory is divided into several memory banks, represented in Figure 1 by Memory bank 0 230 to Memory bank N 231. Each memorybank 230, 231 has its own memory 232, 233 and address generation unit AGU 234,235 respectively. The memories 232, 233 are typically, but not necessarily, complexmemories. This arrangement in conjunction With the on-chip network improves thepower efficiency of the memory system and the throughput of the processor asmultiple address calculations can be performed in parallel. The PBBP of Fig. 2preferably also includes integer memory banks 238, each including a memory 239 and an address generation unit 240.
    Each memory has address pointers indicating the position in the memory that should be read or Written next.
    For example, the commandsOut rO, CDMO_ADDROut rl, CDMI_ADDRset the positions in complex data memory 0 and complex data memory 1, respectively, that should be read from or Written to.
    Each address generation unit 234, 235 performs an address calculation to control theorder in Which data are to be read from or Written to the corresponding memory 232,233. For increased flexibility, the address generation unit can also be arranged toenable tWo or more different modes. Several such modes are known in the art. Theaddress generation logic can for example perform: linear, bit reversed, modulo and2D-addressing With different increments, including negative increments. For example,the address generator can be arranged to read every Kzth data item according to theaddressing function, K being an integer. Alternatively, the address generator might bearranged to address the memory backvvards. Hence, as an example, if the start addressis 0, the address generator can be arranged to read from the corresponding memoryaccording to three different modes:0 Read each address consecutively, that is, 0, 1, 2, 3, 4, 5, 6, 70 Read every Kzth address, that is, if K=2; 0, 2, 4, 6 If the start address is 10 and K=-20 Read backwards, that is 10, 8, 6, 4, 2, 0 In order, for example, to make a vector execution unit multiply data items from twodifferent memories, the command might look as follows: cmac.n CDMO, CDMln being the length of the vector to be operated on. This Would then be performed on avector of n data items from each of the memories CDMO and CDMl, starting With the data item indicated by the pointer of the respective memory.
    As long as the data of a particular memory is to be read consecutively, or in an ordersupported by its address generation unit, there is no problem. But the address generatordoes not support a situation Where the vector to be operated on by the vector executionunit Was held in the memory in an irregular order. For example, multiplying together the sequences CDMO[l, 3, 7, 5,11]and CDMl [0,l,2,3,4] Would be a very complicated operation because there is no regular pattern in thesequence to be read from CDMO. According to the prior art, therefore the addressWould have to be set in the AGU manually, before each reading of data from thememory. This Would cause a significant delay in reading the data Which Would reduce the processor” s overall performance.
    Figure 2 illustrates the function of the address pointer in a first and a second memoryunit. Each memory unit comprises a number of data items, including a sequence of data that is to be provided as input data to a vector execution unit.
    The first memory unit 230 is CDMO and the relevant sequence of data is shown as ablock 230a in this memory unit. An address pointer 230b points to the start of this block, to indicate where reading of data should start.
    The second memory unit 231 is CDM1 which also has a sequence of data to be used asinput data by a functional unit, such as a vector execution unit. This sequence of datais shown as a block 231a, with an address pointer pointing 231b to the start of thisblock. As can be seen, the location of the sequence of data 231a in the second memory231 may be different from the location of the sequence of data 230a in the first memory 230.
    As the reading proceeds, the pointer will move to point to the next address to be readfrom at any given time. Traditionally, the pointer information is taken from the address generation unit 234, 235, of the corresponding memory 230, 231.
    Figure 3 is a simplified drawing showing only the parts of the SIMT architecture thatare particularly relevant for the present invention. The parts shown, using the samereference numerals as in Figure 1, are: one of the vector execution units, in thisexample, the CALU vector execution unit 205, the first 230 and second 231 memorybank, and the on-chip network 244 connecting these three units together. As before,each of the memory banks 230 and 231 comprises a memory 232, 233, and an addressgeneration unit 234, 235. There is also a third memory bank 250 comprising a memory252 and an address generation unit 254. The third memory bank is also connected to the other units through the network 244.
    The third memory bank 250 is typically an integer memory bank, which makes itsuitable for holding address information in the form of a data vector which may bereferred to as an address vector. This memory bank is sometimes referred to as theaddress memory in this document. Otherwise, it may be of exactly the same type as thefirst and second memory banks 230, 231, which are referred to as data memories.
    According to the invention the memory 252 of the third memory bank 250 holds an ll address vector to be used for the addressing of the second memory bank 231. Asshown symbolically by the arrow connecting the second and the third memory bankthe connection bypasses the address generation unit 235 of the second memory bank 231 to address the memory 233 directly.
    Since the data in the memory 252 of the third memory bank 250 can be changed quite easily, this provides a very flexible way of addressing the second memory bank 231.
    Reading address data from a separate memory unit introduces additional latency in thesystem. When the execution unit is ready to start receiving data from the secondmemory bank, it will send a read signal to the second memory bank, which is the bankthat is to provide the data for the calculations performed by the execution unit. Thesecond memory bank will then send a read signal to the address memory bank. Theaddress memory bank will respond by sending its first address item to the secondmemory bank. Only then can the second memory bank send the data item to theexecution unit. Hence there will be a latency which will cause a delay at startup of a vector execution.
    In order to overcome this delay, in this embodiment a memory address interface unit256 is arranged between the address bank 250 and the second memory bank 231. Thememory address interface unit 256 serves as a memory address interface unit forintermediate storage of the address items retrieved from the third memory bank 250.The design and functions of the memory address interface unit 25 6 will be discussed in more detail in connection with Figure 5.
    As a complement, the address generation unit 254 of the third memory bank 250 canalso be used to set a readout mode, as discussed above in connection with Figure 1, forexample to read every other data item from the third memory bank. This means that insome situations the same contents of the third memory bank can be used for different applications. For example, a repeat function could be achieved. 12 As will be understood, a similar arrangement might be used also for addressing thefirst memory bank 230, or the address generation unit 234 of the first memory bankcould be used in a conventional way. Further, any number of memory banks andfunctional units, of any kind, might be provided. There might be a number of memorybanks that could be used as address memories. A memory bank and a memory bankcould easily change the order in which its entries were read or written, by connectingto the appropriate address memory, since all units are interconnected through the network 244.
    Figure 4 illustrates another embodiment of the invention. The parts shown, using thesame reference numerals as in Figure l, are: the CMAC vector execution unit 203, theCALU vector execution unit 205, the first 230 and second 231 memory bank, and thenetwork 244 connecting these three units together. As before, each of the memorybanks 230 and 231 comprises a memory 232, 233, and an address generation unit 234,235. A third memory bank 250 is also shown, comprising a memory 252 and anaddress generation unit 254. The third memory bank is also connected to the otherunits through the network 244 and may be used as an address memory as discussed inconnection with Figure 3. In the embodiment shown in Figure 4, the second memorybank 235 is addressed from the CMAC vector execution unit 203. This means that theaddresses to be read from the second memory bank 231 are being calculated in theCMAC vector execution unit 203. Of course, this is only an example. As the skilledperson would realize the addressing could be performed from a CALU vectorexecution unit or from any other type of functional unit such as a vector execution unit, another execution unit or accelerator.
    As will be understood, the embodiments shown in Figures 4 and 5 could be combinedso that some memory banks would be addressed by means of their internal addressgeneration units 231, 233, others would be addressed from separate memory banks 250, and yet others from functional units 203, 205. 13 By using another vector execution unit to calculate the addresses from Which to read in a particular memory, memory addressing can be achieved in a very flexible Way. This is particularly useful in applications such as Pilot extraction and user separation in OFDM systems Rake finger processing in CDMA systems A method according to the invention of enabling the addressing of one memory unit from another unit of the network, for example, integer data memory IDM: 1) 2) 3) 4) Set the first memory unit, for example, CDMO to use IDM as an addresssource Set the second memory unit, for example, CDMl to use linear addressing bymeans of its address generation unit Process the data in the vector execution unit, for example calculations, on thedata provided from the first and second memory unit in the order they areprovided.
    For each data item read from the memory unit by the vector execution unit, thememory is programmed to retrieve a neW address from the network, that is,either from the third memory unit, in the embodiment of Fig. 3, or from thesecond vector execution unit, in the embodiment of Fig. 4. The neW addressWill indicate the position in the first memory unit from Which data is to be read next.
    Alternatively, for Writing the results of the processing performed by a vector execution unit to a data memory: 1) 2) 3) Set the first memory unit, for example, CDMO to use IDM as an addresssource Set the second memory unit, for example, CDMl to use linear addressing bymeans of its address generation unit Process data in the vector execution unit, for example calculations, and Write the result to a data memory. 14 4) For each data item written to the data memory unit by the vector executionunit, the memory is programmed to retrieve a new address from the network,that is, either from the third memory unit, in the embodiment of Fig. 3, or fromthe second vector execution unit, in the embodiment of Fig. 4. The newaddress will indicate the position in the data memory to which data should be written next.
    In the example methods above, of course the addressing data could be obtained from avector execution unit or from some other unit in the processor, instead of the address memory.
    In both the examples shown in Figures 3 and 4, the address information to be providedfrom the address memory, or from the appropriate vector execution unit, respectively,must be timed in such a way that the next address to be read from, or written to,reaches the second memory bank one clock cycle before it should actually be readfrom or written to. That is: there should be a read ahead of data elements because ofpipelining. The amount of data that is read ahead may be controlled by control signalsthat are propagated from the memory block that is addressed through the network tothe address source. Alternatively, it may be controlled by a fixed value programmedinto the memory address interface unit. The read ahead can also be implemented byaddress sources pushing a pre-defined amount of address data over the network, where the number of pipeline stages is encoded in the hardware.
    In order to overcome the problems caused by latency between the vector executionunit 203 providing the address data and the memory bank 231 that is to use the addressdata, a memory address interface unit 256 is arranged between the vector executionunit 203 and the second memory bank 231. The memory address interface unit 256 issimilar to the memory address interface unit 256 of Figure 3 and serves as a memoryaddress interface unit for intermediate storage of the address items retrieved from the third memory bank 250.
    Figure 5 shows a memory address interface unit 256 according to a preferredembodiment of the invention. The memory interface has a memory 258 and a controlunit 260. The control unit 260 comprises two memories: - a latency memory 262 holding the number of pipeline steps required to read from the third unit to the first memory.- a sample counter 264 arranged to keep track of the number of address samples to be read from the third unit 203, 250 The latency memory 262 is typically, but not necessarily, hardcoded. The sample count memory is arranged to be set for each operation as needed.
    As discussed above the third unit is the one providing the address information. Thismay be a memory unit 250 as shown in Figure 3 or a vector execution unit 203 as shown in Figure 4.
    When the data memory bank (not shown in Figure 5) is setup to receive addressingdata from the third, address providing, unit, the memory address interface unit 256reads the first address data items from the third unit to its memory 258. In this way,when the execution unit that is to receive the data from the data memory bank sends aread signal to signal that it is ready to receive the first address item, this first addressitem is already stored in the memory address interface unit and can be sent to thesecond unit without any delay. Without the memory address interface unit, theprocedure would be - the execution unit sends a read signal to the memory unit - the memory unit sends a read signal to the unit that is to provide the address - the unit that is to provide the address responds by sending the first address.
    - The memory unit, upon receiving the first address, sends the data item to the execution unit.Therefore, it would take several clock cycles before the execution unit could startworking. Once started, however, addresses could be delivered at the appropriate pace. 16 For configuring the system the following steps are performed: l. The core orders external addressing of the data memory by sending a signal to thememory interface to fill the queue to the data memory or by writing its controlregisters. 2. The memory address interface unit performs a sufficient number of readoperations from the unit providing the address data to have in its memory anumber of address items corresponding to the network latency. This means that thenumber of address items should correspond to the number of pipeline steps that have to be performed to retrieve the address.
    During execution, the memory address interface unit continues to send address dataitems from its register to the data memory unit and to retrieve new address data items consecutively.
    The functional unit receiving the data items will continue to send read signals to thememory unit until it has received the appropriate number of data items, and thememory unit will in turn send read signals to the memory address interface unit. Sincesome data items were already read from the address providing unit before thefunctional unit started sending read signals, this means that some read signals will be sent after all the address items have been read from the address providing unit.
    Hence, the latency means that the memory address interface unit would continue toread address items from the third unit after it should stop. More precisely it would readas many address items too many as the number it stored in its register when it wasbeing configured in step 2 above. To keep this from happening, the sample counterkeeps track of the number of address samples that have been retrieved. When thedesired number of address samples have been read from the third unit to the memoryaddress interface unit, the memory address interface unit stops retrieving new address samples although it will continue to receive read signals from the data memory. 17 Instead, for the last address items, the memory address interface unit will empty its memory to provide these items to the data memory.
    In a preferred embodiment, the address information is timed in such a way that thenext address to be read from, or written to, reaches the second memory bank one clock cycle before it should actually be read from or written to.
    Figure 6 shows an advantageous embodiment of a processor according to theinvention, in which a number of memory units can share a smaller number of memoryaddress interface units. The same reference numbers as above are used for the sameunits as shown in previous Figures. As can be seen the processor according to thisembodiment has the same units as the processors shown in Figures 3 and 4, allconnected through the on-chip network 244. In addition, the embodiment of Figure 6has an address crossbar 270 arranged to provide address information to the desiredmemory unit 230, 231. In Figure 6, two memory address interface units 256 are shown, both of which are connected to the address crossbar 270. The address crossbarthen functions as a concentrator, selecting for each memory unit 230, 231 which one of the memory address interface units 256 it should receive addressing data from.
    The embodiments of the invention are particularly useful in applications in whichcomplex address patterns are used, which cannot be hard-coded at design time or areinfeasible to store pre-defined in system memory. Such patterns could be based on run time parameters and must be computed dynamically.
    For example, the OFDM-based telecommunications standard known as LTE (LongTerm Evolution) uses dynamic allocation of frequencies to users. It is necessary toselect the frequencies allocated to one user. In other situations it is desired to select allpilot tones, which may be done based on a table. Traditionally this is achieved bylooking in a table to obtain the address information of the pilot tones, then load the desired data item from memory, shuffle the data in the frequencies to place the 18 relevant data points adj acent each other and then store the data points back in the memory.
    This type of address pattern cannot be programmed into a traditional addressgeneration unit. This means that, for example, to perform an FFT of the pilot tones, theaddress pointer will have to be set several times for each operation, meaning that justthe administration of the data will be so complicated that it will lower the performance (utilization) of the DSP processor significantly.
    If, instead, the memory unit is programmed to retrieve a new address from the on-chipnetwork, as discussed above, each data point to be used can be addressed directly,reducing the capacity needed for the administration of data, and thereby increasing the utilization and performance of the DSP.
    Another application in Which the present invention is particularly useful is in digitalTV applications. Each OFDM symbol in DVB-T2 consists of up to 32768 subcarriers,resulting in a set of 32768 data points. This signal comprises pilot tones to be used asreference data, which are unevenly distributed throughout the frequency spectrum.Such an address pattern cannot be handled easily in traditional address generationunits. According to the invention, the addresses could simply be stored in the thirdmemory unit and picked from there by the memory unit providing data to the vector execution unit.
    For digital TV it is also possible to use only a subset of the 32768 (32k) points. Thesubset may be 16k, 8k, 4k, 2k or lk points, that is, half, quarter, etc., down to l/32 ofthe points. According to the invention, it would only be necessary to store one table ofthe addresses, as a subset of this table could be selected by setting the addresses accordingly in the address memory.
    The invention is also useful when addressing data is to be received from units that have unpredictable timing. Examples of such units are programmable co-processors or 19 error correction units such as turbo decoders. Typically programmable co-processorscan deliver an address stream With an average throughput matching the requirement,but With data delivered in small bursts. In the same Way, error correction blocks WillWork iteratively on a set of data until it is correct, and it is impossible to predictexactly hoW many cycles this Will take. Therefore the output from such units Will beunpredictable bursts of data. A memory address interface unit according to theinvention may be used to even out the bursts of data. If the register 258 shown inFigure 5 is replaced With a FIFO queue the memory address interface unit can store the number of data contained in a burst and send them consecutively to the next unit.
    权利要求:
    Claims (15)
    [1] 1. l. A digital signal processor (200) comprising at least one functional unit, Which may be a vector execution unit (203, 205), a scalar execution unit or an accelerator, andat least a first memory unit (230, 231) arranged to provide data to be operated onby the functional unit, a third unit (250) and an on-chip network (244) connectingthe functional unit, the first memory unit and the third unit, said digital signalprocessor being characterized in that the third unit (250) is arranged to provideaddressing data in the form of an address vector to be used for addressing the firstmemory unit (230, 231), said third unit being connectable to the first memory unit(230, 231) in such a Way that data provided from the third unit (250) can be usedto control the reading from and/or the Writing to the first memory unit (230, 231)and that the processor further comprises a memory address interface unit (256) arranged to compensate for the latency between the first and the third unit.
    [2] 2. A processor according to claim l, Wherein the third unit is an address memoryunit, preferably an integer memory unit, holding address data for addressing the first memory unit.
    [3] 3. A processor according to claim l, Wherein the third unit is a second functional unit.
    [4] 4. A processor according to claim l or 2, Wherein the first memory unit (230, 231) is a complex memory.
    [5] 5. A processor according to any one of the preceding claims 2 or 4 When dependenton claim 2, Wherein the second memory unit comprises an address generation unitarranged to control the reading of address data from the second memory unit according to a predefined pattern. 21
    [6] 6. A processor according to any one of the preceding claims, further comprising an address crossbar (270) interconnecting at least one memory address interface unit(256) and at least a first and a second memory unit (230, 231), to enable addressdata to be provided to a selected one of the first and second memory unit through the memory address interface unit (256).
    [7] 7. A processor according to any one of the preceding claims, adapted for telecommunications, for example, according to the LTE and/or LTE advanced standard.
    [8] 8. A processor according to any one of the claims l - 7, adapted for digital television signals.
    [9] 9. A method of addressing a memory in a digital signal processor (200) comprising atleast one functional unit, which may be a vector execution unit (203, 205), a scalarexecution unit or an accelerator, and at least a first memory unit (230, 231)arranged to provide data to be operated on by the vector execution unit, and an on-chip network (244) connecting the vector execution unit and the first memory unit,and a third unit (250) arranged to provide addressing data for the first memory unitin the form of an address vector, said method comprising the steps of 0 setting the first memory unit to receive addressing data from the third unit, 0 providing addressing data from the third unit to the first memory unit, 0 reading data from, or writing data to the first memory unit according to the addressing data,0 compensating for the latency between the first and the third unit by means of amemory address interface unit (256) arranged between the first and the third unit.
    [10] 10. l0. A method according to claim 9, wherein the third unit is an address memory unit, preferably an integer memory unit, holding address data for addressing the first memory unit.
    [11] 11.
    [12] 12.
    [13] 13.
    [14] 14.
    [15] 15. 22 A method according to claim 9, Wherein the third unit is a functional unit. A method according to any one of the claims 9 - 11, Wherein the first memory unit (230, 231) is a complex memory. A method according to any one of the claims 10 or 12 When dependent on claim10, comprising the step of controlling reading of addresses from the addressmemory unit by means of an address generation unit arranged to control thereading of address data from the second memory unit according to a predefined pattern. A method to any one of the claims 9 - 13, comprising the step of timing theaddress information in such a Way that the next address to be read from, or Writtento, reaches the first memory unit one clock cycle before it should actually be read from or Written to. A method according to claim 14, Wherein the amount of data that is read ahead iscontrolled by address sources pushing a pre-defined amount of address data over the network, the number of pipeline stages being encoded in the hardware.
    类似技术:
    公开号 | 公开日 | 专利标题
    US20090287859A1|2009-11-19|DMA Engine
    US10657050B2|2020-05-19|Tile based interleaving and de-interleaving for digital signal processing
    US20140324890A1|2014-10-30|Data Driven Parallel Sorting System and Method
    KR20200138761A|2020-12-10|Multi-threaded, self-scheduled reconfigurable computing fabric with reentrant queues to control loop execution
    WO2019191741A1|2019-10-03|Execution control of a multi-threaded, self-scheduling reconfigurable computing fabric
    US20210255864A1|2021-08-19|Multiple Types of Thread Identifiers for a Multi-Threaded, Self-Scheduling Reconfigurable Computing Fabric
    US20210224067A1|2021-07-22|Backpressure Control Using a Stop Signal for a Multi-Threaded, Self-Scheduling Reconfigurable Computing Fabric
    EP3776242A1|2021-02-17|Multi-threaded self-scheduling reconfigurable computing fabric
    EP3776188A1|2021-02-17|Conditional branching control for a multi-threaded, self- scheduling reconfigurable computing fabric
    WO2019191739A1|2019-10-03|Efficient loop execution for a multi-threaded, self-scheduling reconfigurable computing fabric
    CN104111817B|2017-05-10|Arithmetic processing device
    SE537423C2|2015-04-21|Digital signal processor and method for addressing a memory in a digital signal processor
    SE536462C2|2013-11-26|Digital signal processor and baseband communication device
    US8843807B1|2014-09-23|Circular pipeline processing system
    EP2592562B1|2016-05-18|Memory access for digital signal processing
    US10303630B2|2019-05-28|Configurable hardware accelerators
    US20140281368A1|2014-09-18|Cycle sliced vectors and slot execution on a shared datapath
    KR101927255B1|2018-12-12|Processing device of register window overflow/underflow
    US10387118B2|2019-08-20|Arithmetic operation unit and method of controlling arithmetic operation unit
    US20210055964A1|2021-02-25|Computational Partition for a Multi-Threaded, Self-Scheduling Reconfigurable Computing Fabric
    CN110096307B|2022-02-01|Communication processor
    US11275710B2|2022-03-15|Loop thread order execution control of a multi-threaded, self-scheduling reconfigurable computing fabric
    EP3776241A1|2021-02-17|Loop thread order execution control of a multi-threaded, self-scheduling reconfigurable computing fabric
    Wang et al.2006|On the Efficient Implementation of Pipelined Heaps for Network Processing.
    同族专利:
    公开号 | 公开日
    CN103999074A|2014-08-20|
    WO2013095257A1|2013-06-27|
    CN103999074B|2017-04-12|
    US9557996B2|2017-01-31|
    SE537423C2|2015-04-21|
    ES2653951T3|2018-02-09|
    EP2751705B1|2017-10-18|
    KR20140103343A|2014-08-26|
    EP2751705A1|2014-07-09|
    US20140351555A1|2014-11-27|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    US5809552A|1992-01-29|1998-09-15|Fujitsu Limited|Data processing system, memory access device and method including selecting the number of pipeline stages based on pipeline conditions|
    US6594710B1|1999-05-26|2003-07-15|Nec Electronics, Inc.|Apparatus and method for a random access peripheral unit|
    US6665790B1|2000-02-29|2003-12-16|International Business Machines Corporation|Vector register file with arbitrary vector addressing|
    US6976147B1|2003-01-21|2005-12-13|Advanced Micro Devices, Inc.|Stride-based prefetch mechanism using a prediction confidence value|
    US7017028B2|2003-03-14|2006-03-21|International Business Machines Corporation|Apparatus and method for updating pointers for indirect and parallel register access|
    US7334110B1|2003-08-18|2008-02-19|Cray Inc.|Decoupled scalar/vector computer architecture system and method|
    US7620047B2|2004-11-23|2009-11-17|Emerson Network Power - Embedded Computing, Inc.|Method of transporting a RapidIO packet over an IP packet network|
    US20070198815A1|2005-08-11|2007-08-23|Coresonic Ab|Programmable digital signal processor having a clustered SIMD microarchitecture including a complex short multiplier and an independent vector load unit|
    CN101282477A|2008-05-06|2008-10-08|艾诺通信系统（苏州）有限责任公司|Method and system for processing multicore DSP array medium based on RapidIO interconnection|
    US20130185538A1|2011-07-14|2013-07-18|Texas Instruments Incorporated|Processor with inter-processing path communication|US20160226544A1|2015-02-04|2016-08-04|GM Global Technology Operations LLC|Adaptive wireless baseband interface|
    法律状态:
    2020-07-28| NUG| Patent has lapsed|
    优先权:
    申请号 | 申请日 | 专利标题
    SE1151230A|SE537423C2|2011-12-20|2011-12-20|Digital signal processor and method for addressing a memory in a digital signal processor|SE1151230A| SE537423C2|2011-12-20|2011-12-20|Digital signal processor and method for addressing a memory in a digital signal processor|
    EP12816532.1A| EP2751705B1|2011-12-20|2012-11-28|Digital signal processor and method for addressing a memory in a digital signal processor|
    KR1020147019866A| KR20140103343A|2011-12-20|2012-11-28|Digital signal processor and method for addressing a memory in a digital signal processor|
    ES12816532.1T| ES2653951T3|2011-12-20|2012-11-28|Digital signal processor and method to address a memory in a digital signal processor|
    CN201280063320.0A| CN103999074B|2011-12-20|2012-11-28|Digital signal processor and method for addressing a memory in a digital signal processor|
    PCT/SE2012/051320| WO2013095257A1|2011-12-20|2012-11-28|Digital signal processor and method for addressing a memory in a digital signal processor|
    US14/364,619| US9557996B2|2011-12-20|2012-11-28|Digital signal processor and method for addressing a memory in a digital signal processor|
    [返回顶部]