United States Patent [191
`Bowes et al.
`
`lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll
`US005546547 A
`5,546,547
`1111 Patent Number:
`(45) Date of Patent:
`Aug. 13, 1996
`
`(54] MEMORY BUS ARBITER FOR A
`COMPUTER SYSTEM HAVING A DSP
`CO-PROCESSOR
`
`[7.5]
`
`IuvcnlCIIS: Michacl J. Dowc~, Cupc1tino; F;uid A.
`Yazdy, Belmont, both of Calif.
`
`[73] Assignee: Apple Computer, Inc., Cupertino,
`Calif.
`
`(21] Appl. No.: 189,138
`Jan. 28, 1994
`
`[22] Filed:
`lot. a.6
`........................... ...... ..................... G06F 13/00
`[5 1]
`[52] U.S. Cl .............................................. 395/294; 395/293
`[58] Field of Search ...................................... 3951325, 425
`
`(56]
`
`References Cited
`
`U.S. PATENT DOCUMENTS
`
`4,641,238
`4,928.234
`4,987.529
`5,263,163
`5,301,282
`
`211987 Kneib ...................................... 395/325
`5/1990 Kitamura et al ........................ 395/425
`111991 Craft et al . .............................. 3951325
`1111993 Holt et al. ............................... 3951325
`4/1994 Amini ct al . ............................ 395/325
`
`5/1994 Averill .................................... 3951325
`5,313,591
`5,353,417 1011994 Fuoco et al . ............................ 395/325
`Primary Examiner- Jack B. Harvey
`Assistalll Examiner-David A. WJ.Iey
`Attorney, Agent, or Firm- Blakely, Sokolotf, Taylor & zat.
`man
`[57]
`
`ABSTRACT
`
`An arbitration scheme for a computer system in which a
`digital signal processor resides on the computer system's
`memory bus without requiring a block of dedicated static
`random access memory. An arbitration cycle is divided into
`I 0 slices of which 5 slices are provided in each arbitration
`loop to the digital signal processor. Two slices are provided
`each to the system's 1/0 interface and to the peripheral bus
`controller. A final slice is provided to the system's CPU. A
`default state when no memory bus resource is requesting the
`system memory bus parks the memory bus on the CPU. The
`arbitration scheme provides sufficienl bandwidth for real(cid:173)
`time signal processing by the digital signal processor oper(cid:173)
`ating from the system's dynamic random access memory
`while also providing sufficient bandwidth for a local area
`network interface through the system's l/0 interface.
`
`8 Claims, 5 Drawing Sheets
`
`10
`
`200
`
`12
`
`14
`r'
`MAIN MEMORY
`SUBSYSTEM
`(DRAM)
`
`~ 110
`
`00
`N .-
`I
`
`NUBUS
`CONTROLLER -40
`
`Page 1 of 25
`
`ZTE EXHIBIT 1003
`
`

`
`U.S. Patent
`
`Aug. 13, 1996
`
`Sheet 1 of 5
`
`5,546,547
`
`10
`
`12
`
`CPU
`
`ROM
`
`14
`
`15
`
`16
`
`MAIN MEMORY
`SUBSYSTEM
`(DRAM)
`
`1/0
`INTERFACE
`
`MASS
`STORAGE
`
`MEMORY BUS
`
`20-
`
`DSP
`
`Rgure 1 (Prior Art)
`
`Page 2 of 25
`
`

`
`U.S. Patent
`
`Aug. 13, 1996
`
`Sheet 2 of 5
`
`5,546,547
`
`10
`
`200
`
`12
`
`14
`
`r
`MAIN MEMORY
`SUBSYSTEM
`(DRAM)
`
`-110
`
`co
`C'J
`,.-
`1
`
`z:
`:5
`
`Figure2
`
`Page 3 of 25
`
`

`
`"'0
`0)
`<0
`CD
`~
`
`0 -1\.)
`
`01
`
`301
`
`320-
`
`RESET
`-
`
`309- I -~:;
`
`I
`
`I
`
`\ -dsp & "~-- ~ "'-'-~~"'
`
`I " \
`
`/'\
`
`?
`
`J-303
`
`iab
`
`308- I ·-; l·
`
`~ \
`
`I
`
`} ................ _.,..-
`
`I
`
`'·--
`
`J ~
`I
`
`'
`
`~y t..
`, ;
`
`" · ijl
`
`"i
`
`1
`
`)-304
`
`Figure 3
`
`d •
`00
`•
`~
`~
`~
`
`~ = ~
`
`> = ~
`
`!-"
`.. w
`
`~
`\C
`\C
`0\
`
`fJ). ::r
`~ .....
`w
`0 .....
`tA
`
`01
`....
`01
`~
`...
`0\
`til
`~
`-...l
`
`

`
`U.S. Patent
`
`Aug. 13, 1996
`
`Sheet 4 of S
`
`5,546,547
`
`DSF CLCK
`
`_....
`
`DSP
`STATE
`MACHINE
`
`-210
`
`BCLK
`
`r
`
`r--+
`
`NON-DSP
`STATE
`MACHINES
`
`-230
`
`MCA
`
`;::uu
`
`212-
`
`DSP BUS
`INTERfACE UNIT
`
`L
`........
`
`DRAM
`..
`220- INTERFACE UNIT
`
`./
`·"'-
`
`.. BUS INTERFACE
`235-
`..
`
`NON-DSP
`
`UNIT
`
`./
`
`........
`
`'
`
`/
`
`>
`
`'
`
`/
`
`>
`
`'
`
`/
`
`>
`
`... BB-
`.....
`DSP REO
`NUE REO
`lOB REO
`
`....
`
`LOGIC
`
`- BUS ARBITRATION UNIT
`..
`..
`...
`..
`
`DSP
`241- WATCHDOG
`TIMER
`
`CPUBGN, DSPBGN, NUBB ~N, IOBBGN
`
`-
`
`DSPLOCK
`
`CPU LOCK
`
`-
`
`...
`
`210
`
`Figure 4
`
`Page 5 of 25
`
`

`
`-u
`ru
`(Q
`CD
`0'>
`
`0 -N
`
`01
`
`CLOCK
`(BCLK)
`DSP BR_
`
`DSPBG_
`
`NUBUS BR_
`
`NVBUS BG~
`
`lOB BA_
`
`IOBBG_
`
`CPUBG_
`
`BB_
`
`1
`
`I
`
`I
`
`·Figure 5
`
`NUBUS CONTROLLER OWNS THE BUS
`DRIVEN BY THE NUBUS CONTROLLER
`
`DSP OWNS THE BUS DRIVEN BY THE ARBITER
`WHEN DSP IS THE BUS OWNER
`
`DSP OWNS THE BUS AGAIN
`DRIVEN BY THE ARBITER
`
`CPU OWNS THE BUS
`DRIVEN BY THE CPU
`
`TIME
`
`~ •
`00 •
`~
`~ .....
`~ = .....
`
`~
`
`Jo-'
`JM
`~
`\Q
`\Q
`C'l
`
`!;I} =(cid:173)~
`~ ....
`
`ll1
`0
`~
`ll1
`
`~
`
`Ot
`
`Ot ... 0\
`~ 01 ... -....l
`
`

`
`5,546,547
`
`1
`MEMORY BUS ARBITER FOR A
`COMPUTER SYSTEM HAVING A DSP
`CO-PROCESSOR
`
`BACKGROUND OF THE INVENTION
`
`I. Related Applications
`This application is related to U.S. patent application Scr.
`No. 08/189,139, entitled "Dual Bus Concurre111 Multi-Chan(cid:173)
`nel Direct Memory Access Corn roller and Method", Ser. No.
`081189,132, entitled "Multiple Register Set Direct Memory
`Access Channel Architecture", and Ser. No. 081189,131,
`entitled "Direct Memory Access Channel Architecture and
`Method for Reception of Network lnfonnation", each of
`which is assigned to the assignee of the present invention 15
`and filed concurrently herewith.
`2. Field of the Invention
`The present invention relates to digital computer system
`architecture. More particularly, the present invention relates
`to a computer architecture in which multiple processing 20
`units share a common memory bus and support real-lime
`applications.
`3. Description of Related Art
`Until recently, telecommunications and computing were 25
`considered to be entirely separate disciplines. Telecommu(cid:173)
`nications was analog and done in real-time whereas com(cid:173)
`puting was digital and performed at a rate determined by the
`processing speed of a computer. Today, such technologies as
`speech processing, sound processing, electronic facsimile .
`30
`and image processing have blurred these lines. In the corning
`years, computing and telecommunications will become
`almost indistinguishable in a race to support a broad range
`of new multimedia (i.e., voice, video and traditional data)
`applications. These applications are made possible by
`emerging digital-processing technologies, which include:
`compressed audio (both high fidelity audio and speed), high
`resolution still images, video, and high speed signal trans(cid:173)
`mission such as by means of modem or facsimile exchange.
`The emerging technologies will allow for collaboration at a
`distance such as by video conferencing.
`Each of these aspects of real-time information processing
`may require dedicated processors designed for their imple(cid:173)
`mentation. However, it is becoming more and more c.ommon
`to usc programmable digital signal processors (DSP) avail- 45
`able on the market today, such as the AT&T® DSP3210.
`DSPs are autonomous processors having their own real-time
`operating systems. As such, they arc ideally suited to real(cid:173)
`Lime audio and image signal processing.
`In handling real-time information such as speech recog- so
`nition and modem functionality, a DSP requires a large
`amount of bandwidth to memory for processing the sheer
`volume of data required to effectuate real-time compULing.
`FIG. 1 illustrates a typical computer architecture in which a
`CPU 10 is coupled to a memory bus 100. The memory bus ss
`100 may also be referred to as the system bus or CPU bus.
`In any event, it is this bus which couples the system's CPU
`10 to the IJO interface 15 and the various components of the
`memory subsystem. In FIG. 1, the CPU is in communication
`with a ROM 12 and the main memory subsystem 14 through 60
`the memory bus 100. The main memory subsystem 14
`usually comprises a memory controller and a large array of
`dynamic random access memory (DRAM) for supporting
`operating applications and data for access by the CPU over
`the memory bus 100. The main memory subsystem is 65
`distinguished from mass storage 16 which may comprise
`hard magnetic disk drives or a CD-ROM which provide for
`
`40
`
`35
`
`2
`relatively slow access, high volume storage of information.
`Access times to mass storage are slower in part because of
`the need to process requests through the 110 interface and the
`inherently slower nature of mass storage devices. The
`memory subsystem's DRAM, on the other hand, is semi(cid:173)
`conductor memory which provides for fairly quick storage
`and retrieval for operating applications and which may be
`fed from the slower mass storage 16 through the 110
`interface 15 over the memory bus 100 to meet the require-
`tO ments of the CPU 10.
`ln many computer systems, the l/0 interface 15 is also a
`direct memory access (DMA) controller which manages the
`transfer of data between 110 devices and the main memory
`subsystem without requiring the CPU to perform that task.
`The J/0 interface 15 may also be used for coupling the
`computer system to other computer systems over a network
`such as an Ethernet local area network.
`Also shown coupled to the memory bus 100 in FIG. 1 is
`a digital signal processor (DSP) 20. The logic or controller
`needed to couple the DSP to tile memory bus is not shown.
`The DSP 20 may be an off-the-shelf DSP such as the
`AT&T® DSP321 0. Most DSPs include an on-board cache of
`static random access memory (SRAM) which in the case of
`the AT&T DSP3210 is an 8-Kbyte SRAM cache. In prior art
`computer systems, because of the high bandwidth required
`for real-time processing by a DSP, it has not been possible
`for the DSP to run off of the computer system's DRAM in
`the way 1he CPU 10 utilizes it without adversely affecting
`the rest of the computer system. Thus, there has been
`provided a large block of SRAM 24 for use by the DSP 20.
`This has allowed the memory bus to be relatively free of
`DSP requests yielding the freedom to process CPU requests
`and requests from 110 devices or networks through the 110
`interface without having to contend for the bandwidth
`required by the DSP 20 .
`A significant disadvantage to the prior art computer
`architecture of FIG. 1 is the requirement of a substantial
`block of static random access memory 24. SRAMs are
`significantly more expensive than DRAM which greatly
`increases the cost of computer systems which incorporate
`SRAM. One object of the emerging multimedia technolo(cid:173)
`gies is to bring these technologies to the mass market in
`configurations as inexpensive as possible. It is therefore one
`object of the present invention to provide a computer
`architecture which incorporates DSP technology for real(cid:173)
`time data processing without requiring the inclusion of
`expensive SRAM to support the DSP.
`
`SUMMARY OF THE INVENTION
`
`From the foregoing, it can be appreciated that a computer
`architecture which provides for a digital signal processor to
`operate as a co-processor with a CPU over a common
`memory bus without requiring expensive static random
`access memory would be greatly advantageous. Accord(cid:173)
`ingly, it is an object of the present invention to provide a
`mechanism and method for arbitrating the memory bus
`bandwidth to efficiently allow the use of a digital signal
`processor and a CPU over a common memory bus sharing
`the system's dynamic random access memory subsystem
`without requiring an expensive block static random access
`memory. It is further an object of the present invention to
`provide an arbitration scheme in which, in addition to the
`CPU and digital signal processor, a DMA controller and a
`peripheral card expansion bu~ controller might also access
`the memory bus with all bus masters receiving sufficient
`
`Page 7 of 25
`
`

`
`5,546,547
`
`4
`FIG. 3 illustrates a state diagram of the arbitration scheme
`for assigning bandwidth slots to the various components of
`the preferred embodiment computer architecture.
`FIG. 4 illustrates a more detailed logical representation of
`the arbiter for implementing the arbitration scheme of the
`present invention.
`FIG. 5 illustrates a timing diagram showing a simple state
`transition sequence in accordance with a preferred embodi(cid:173)
`ment implementation of the present invention.
`
`DETAILED DESCRIPTION OF THE
`INVENTION
`
`10
`
`20
`
`3
`bandwidth over the memory bus to provide for real-time
`isochronous data processing.
`These and other objects. of the present invention are
`provided by a computer architecture in which a CPU and
`digital signal processor (DSP) as well as other memory bus
`masters are provided on a common memory bus with the
`system' s dynamic random access memory subsystem. An
`application specific integrated circuit (ASIC) for arbitrating
`between bus masters is provided which implements an
`arbitration scheme for sharing the bandwidth of the system's
`memory bus to provide the DSP with sufficient access to the
`DRAM to carry out real-time, isochronous data processing
`while still allowing the system's I/0 controller to satisfy
`Ethernet communication requirements over the memory bus.
`The arbitration scheme of the present invention is an adap- 15
`tive arbitration scheme that varies access to the memory bus
`as a function of time and depends upon what operations the
`various bus masters are reque sting. The next state of the bus
`arbiter depends on the present state, in addition to the
`amount of bus traffic and history of prior requests.
`The arbitration scheme is tuned to maximize accessibility
`of the memory bus to the DSP which has by far the greatest
`bandwidth requirements. A cycle in the arbitration scheme is
`divided into 10 time slices of which 5 slices are available for
`the DSP to assert the memory bus. However, the total 25
`amount of time in a given cycle that the DSP is allotted the
`memory bus is limited by a watchdog timer in the arbitration
`module so as not to choke off the other potential bus masters.
`During the arbitration cycle, each of the potential bus
`masters may signal a bus request over the memory bus and
`a priority scheme defined by a state diagram determines
`which bus master will next have access to the memory bus.
`In the preferred embodiment of the present invention, to
`save a pin on the CPU, the CPU does not utilize the means
`for requesting the memory bus. Instead, the default state for
`the arbiter is to assign the memory bus to the CPU when no
`other bus master is requesting it. Further, at the end of each
`arbitration cycle, the last frame is reserved for the CPU
`which may then carry out any desired operation.
`Finally, the present invention is implemented in such a
`way that the clock rate of the DSP need not be synchronized
`with the rest of the modules in the computer system. In this
`way, enhancements in DSP technology will allow faster
`digital signal processors to be incorporated in computer 4s
`systems without requiring total redesign of the architecture.
`The arbiter ASIC effectively has a separate set of state
`machines devoted to the time domain of the digital signal
`processor which operate in conjunction with the state
`machines of the modules operating in the time domain of the 50
`rest of the computer system. A brief amount of time is
`required for resynchronizing to the different time domains
`on the memory bus which. is justified by the increased
`flexibility in the design of the computer architecture.
`
`A method and apparatus are described for the arbitration
`of a number of bus masters over a memory bus in a computer
`system architecture supporting a digital signal processor.
`Throughout this detailed description numerous details are
`specified such as I/0 devices and network standards, in order
`to provide a thorough understanding of the present inven(cid:173)
`tion. To one skilled in the art, however, it will be understood
`that the present invention may be practiced without such
`specific details. In other instances well-known control struc(cid:173)
`tures and gate level circuits have not been shown in detail in
`order not to obscure unnecessarily the present invention.
`Particularly, the arbitration module for implementing the
`arbitration scheme of the present invention, in the preferred
`embodiment, is implemented in an application specific inte(cid:173)
`grated circuit (ASIC). With today's manufacturing technol-
`30 ogy, the development of AS[Cs generally does not reqvire
`the rendering of fully detailed circuit diagrams. The defini(cid:173)
`tion of logic functionality and state diagrams allow com(cid:173)
`puter aided design techniques to design the desired inte(cid:173)
`grated <Circuit. Accordingly, the present invention will be
`35 described primarily in terms of functionality to be imple(cid:173)
`mented by an ASIC. The software used for developing the
`ASIC chip from the logic defined for the preferred embodi(cid:173)
`ment will be attached as an appendix. Those of ordinary skill
`in the art, once given the following descriptions of the
`40 various functions to be canied out by the present invention
`will be able to implement the necessary logic in an ASIC or
`other technology without undue experimentation.
`A portion of the disclosure of this patent document
`contains material which is subject to copyright protection.
`The copyright owner has no objection to the facsimile
`reproduction by anyone of the patent disclosure, as it
`appears in the Patent and Trademark Office patent files or
`records., but otherwise reserves all copyrights whatsoever.
`The present invention concerns a computer architecture in
`which a digital signal processor (DSP) operates as a true
`co-processor in the computer system. That is, an arbitration
`technique and mechanism are implemented which allows a
`DSP to reside on the system's CPU or memory bus and share
`the memory bus resources with the other potential bus
`55 masters on the memory bus. The scheme is .implemented
`such that the DSP is provided with sufficient bandwidth to
`perform real-time digital signal processing using the sys(cid:173)
`tem's dynamic random access memory (DRAM) and not
`requiring the incorporation of an expensive block of static
`random access memory (SRAM). DRAM is far less expen(cid:173)
`sive than SRAM and the elimination of a block of SRAM
`·greatly reduces the cost of computer systems. The DSP is
`provided with sufficient bandwidth on the memory bus to
`perform real-time isochronous signal processing, yet at the
`same time the arbitration scheme of the present invention
`prevents the other bus masters from being starved from the
`bandwidth they require in carrying out their operations.
`
`BRIEf DESCRIPTION OF THE DRAWINGS
`
`The objects, fe'atures and advantages of the present inven(cid:173)
`tion will be apparent from the following detailed description
`in which:
`FIG. 1 illustrates a prior art computer architecture in
`which a digital signal processor requires an expensive static
`random access memory block.
`FIG. 2 illustrates a block diagram of a computer archi- 65
`lecture incorporating
`the present invention arbitration
`scheme.
`
`60
`
`Page 8 of 25
`
`

`
`5,546,547
`
`5
`While the present invention arbitration scheme will be
`described in term of a particular implementation of a com(cid:173)
`puter system, it should be understood that the broad concept
`of providing for a DSP on a memory bus and sharing the
`system's DRA.\.1 may be extended to other more compli-
`cated systems, or generalized to simpler systems.
`
`Overview of the Present Invention Computer
`System
`
`s
`
`40
`
`6
`computer systems and share common resources such as
`community printers, etc. As will be described further herein,
`the provision of an Ethernet connection to the present
`invention computer system introduces constraints for the
`bandwidth sharing of the memory bus.
`Another potential bus master coupled to the system bus
`no of the pn~ferrecl emhndiment computer system is a
`peripheral card expansion bus controller. The peripheral
`card expansion bus controller illustrated in the figure is the
`10 NuBus controller 40. While the present illustration uses a
`NuBus controller, other peripheral card expansion bus pro(cid:173)
`tocols are generally known. The NuB us controller 40 is used
`to provide communication between the various memory bus
`masters and the system's expansion bus 130, called NuBus.
`IS The NuBus provides a fast backplane for coupling such
`things as video controllers 131, RAM expansion cards, mass
`storage device controller cards, or other communications
`devices 132. Again, the preferred embodiment computer
`system is constrained by some of the NuBus requirements in
`20 the bandwidth sharing of the memory bus 110.
`Also, in PlG. 2 there is illustrated another potential bus
`master, the digital signal processor (DSP) 20. Unlike prior
`art computer systems, the present invention provides for the
`DSP 20 to reside on the system's memory bus and operate
`25 from the computer system's main memory subsystem 14. In
`implementing the prese.nt invention this greatly reduces
`system cost by eliminating the need for an expensive block
`of SRAM. In the preferred embodiment implementation, the
`DSP 20 is an AT&T DSP3210 which provides an internal 8K
`30 SRAM cache. This is an off-the-shelf DSP which is highly
`programmable and has a fairly well defined operating sys(cid:173)
`tem. The DSP can be programmed to carry out such func(cid:173)
`tions as speech processing, audio channel control, modem
`emulation, image processing and the like. Many of these
`35 functions are real-time operations and require a tremendous
`amount of the memory bus bandwidth between the DSP and
`the DRAM of the main memory subsystem 14. For reasons
`that will .be described further herein, the DSP 20 is not
`constrained to operating at the same clock speed as the rest
`f o · the components of the computer system. This will require
`to be
`for various operations
`some resynehronizing
`described, but provides for Oexibility as newer technology
`and faster DSPs are developed.
`Finally, in FlG. 2 there is coupled to the memory bus llO
`the memory controller and arbiter (MCA) 200. In the
`preferred embodiment implementation, the arbiter 200 is an
`applications specific integrated circuit (ASIC) for arbitrating
`the memory bus 110 between the various bus masters subject
`to the constraints each imposes to pro,•ide optimal band(cid:173)
`width for each, particularly the DSP which is responsible for
`a significant amount of real-time signal processing. In an
`alternative embodiment, the arbiter logic could be designed
`in some other form of logic.
`
`Referring now to FIG. 2, a preferred embodiment com(cid:173)
`puter architecture implementing the present invention is
`shown. The constituents of the computer architecture are
`shown coupled to a CPU or memory bus 110. The memory
`bus 110 provides the signal paths for the exchanging. of data
`between the various clements on the memory bus. Further
`provided by the memory bus are control lines for such things
`as bus requests and bus granting signals and other system
`level control signals. The signals required for implementing
`the prcs.cnt invention will be described further herein. As
`with the prior art computer systems, the architecture illus(cid:173)
`trated in FIG. 2 has the CPU 10, the ROM 12 and the
`system's main memory subsystem 14 coupled to the
`memory bus 110. ll should be understood that in the illus(cid:173)
`tration the main memory subsystem 14 coupled to the
`memory bus refers to the system's DRAM and the controller
`required for writing to and reading from the DRAM based
`on a requested transaction. The CPU 10 may be considered
`a potential bus master as contrasted with the memory
`componems 12 and 14 which are considered bus slaves. The
`ROM 12 and main memory subsystem 14 will not drive the
`memory bus for transactions on their own behalf.
`The prefe.rred embodiment computer architecture of AG.
`2 has a number of other potential bus masters coupled to it.
`The 1/0 interface 30 is used for coupling the memory bus to
`the I/0 bus 120 for providing the computer system with a
`number of 110 capabilities. The 110 interface 30 of the
`present invention is also a DMA controller which controls
`the transactions between the main memory system 14 and
`1/0 devices without requiring the resources of the CPU 10
`for these transactions. The DMA controller of the preferred
`embodiment implementation is described in co-pending
`U.S. patenL applications: Ser. No. 081189,139, entitled "Dual
`Bus Concurrent Multi-Channel Direct Memory Access Con(cid:173)
`troller and Method", Ser. No. 08/189,132, entitled "Multiple
`Register Set Direct Memory Access Channel Architecture",
`and Scr. No. 08/189,131, entitled "Direct Memory Access
`Channel Architecture and Method for Reception of Ethernet
`Packets-", assigned to the assignee of the present invention
`and filed concurrently herewith.
`The UO bus 120 of the preferred embodiment implemen(cid:173)
`tation as illustrated in FIG. 2 provides the avenue for a
`number of different 1/0 devices to be incorporated in the
`computer system. For example, the SCSI pon 121 is coupled
`through the 110 bus 120. SCSI stands for small computer 55
`system interface which is one of the computer industry's
`standards for coupling IJO devices such as the hard disk
`drive 122 or floppy drives or other storage media and other
`l/0 devices to today's microcomputer systems. The com(cid:173)
`puter system's serial port 125 is also coupled to the 110 bus 60
`120. The serial pon can be used for attaching a modem or
`printer or other serial devices. Finally, the l/0 bus 120 is
`used to couple the computer system to a local area network
`through the Ethernet port 128·. The Ethernet port 128 is used
`for attaching the computer system of the present invention to 65
`a local area network such as Ethernet or other LAN 129.
`This allows the computer system to communicate with other
`
`45
`
`50
`
`Arbitration Constraints
`
`Because the primary motivation of the present invention
`is to incorporate a digital signal processor as a co-processor
`on a computer system's memory bus while providing it
`sufficient bandwidth to utilize the system's DRAM rather
`than an expensive block of SRAM, it is necessary to talk
`about the various modes of operation of the DSP, particu(cid:173)
`larly for carrying out real-time processing. As was indicated,
`the DSP 20 is programmable and has an 8-Kbyte internal
`SRAM cache. Ideally, software written for the DSP should
`be segmented in such a way that blocks may be loaded into
`
`Page 9 of 25
`
`

`
`5,546,547
`
`25
`
`30
`
`8
`7
`times require the memory bus. When this occurs, the
`the cache of the DSP allowing the DSP to run as much of the
`resource propagates a bus request signal over the memory
`time as possible from its internal cache. Thus, one mode in
`bus 110 to the arbiter 200. A figure referred to further herein
`which the DSP will utilize the memory bus 110 is to read a
`large block memory from the DRAM 14 into its internal
`wilJ indicate more clearly what signals are required for the
`SRAM. This first mode of operation on the memory bus can 5 various bus requests. When the memory bus is available for
`be referred to as a "block read". Another mode of operation
`assignment, the arbiter 200 will issue a bus grant signal to
`one of the resources according to a priority scheme which
`concerns the handling of data that has already been pro-
`takes into eonslderatlon all of the consrrainrs described
`cessed by the DSP. In many cases it will be necessary to push
`above as well as various previous states of the system and
`that data back out to the DRAM so that some other parts of ·Jo the present state of the system.
`the computer system can utilize it. Thus, the capability of
`f
`·
`d
`hi h
`Referring now to FIG. 3, a state diagram is illustrated
`bursting data out is a second mo eo operation w c may
`which is used to derive the logic of the arbiter ASIC 200.
`be referred to further herein as a "block write".
`.
`.
`This state diagram is the preferred embodiment arb!Lrallon
`Because of the nature of some software, it is not always
`scheme for the implemented computer system described
`guaranteed that the code is going to be divisible into discrete
`with respect to FIG. 2 which has, in addition to a CPU and
`blocks for block read operations. There may be times when 15 a DSP on the memory bus, an 1/0 interface for coupling the
`it is necessary to repeatedly access the DRAM, effectively
`system to a local area network and other 1/0 devices as well
`supporting a scheme where the DSP executes code directly
`as a N uBus interface witJ1 each imposing considerable
`from the DRAM. As contrasted with a block read, this might
`constraints. The state diagram illustrated in FIG. 3 is instru-
`bc considered a situation of "back-to-back single reads"
`mental in developing the VERlLOG code or equivalent in
`from the DRAM. The occurrence of these back-to-back 20
`the production of the arbiter ASIC 200. The code for the
`operations is recognizable and will be considered by the
`preferred embodiment arbitration scheme is set forth in
`arbiter scheduling algorithm to be described further herein.
`Appendix A.
`Finally, during the course of normal ope(ation there may be
`The bus arbiter 200 follows the priority scheme iliustrated
`situations in which the DSP requires only a single piece of
`by the state diagram of FIG. 3. The arbiter receives the bus
`data from the DRAM. These types of read operation arc
`request signals from the potential bus masters, and based on
`referred to as "scattered-single reads". Because the DSP will
`the present state and a specified priority order asserts an
`be processing many real-time operations as well as others,
`appropriate bus grant signal. In the preferred embodiment
`each of these modes of operation where the DSP must have
`implementation, to save a pin on the CPU, the CPU 10 docs
`access to the memory bus must be taken into consideration
`not issue bus request signals. Instead, the stale of the
`by the arbitration scheme of the present invention.
`memory bus assignment defaults to the CPU and remains
`parked on the CPU until other resources request the memory
`In addition to the DSP' s huge requirement for bandwidth
`bus. The CPU is also provided with one time slot in the
`on the memory bus, the arbiter 200 must also contend with
`constraints imposed by other resources on the memory bus.
`priority scheme to be described in which it is granted the
`Particularly. the 1/0 interface 30 provides the interconnec-
`memory bus.
`·
`lion of the computer system to an Ethernet local area
`Each circle in the arbiter state diagram of FIG. 3 repre-
`network. This Ethernet connection should never be choked
`sents a present state memory bus assignment and the arrows
`off because losing an Ethernet packet results in a tremendous
`represent the state transition from a current bus master to the
`penalty in time for recovery. Thus, the l/0 interface 30 must
`next bus master when the specified conditions are met.
`be guaranteed at least a predetermined amount of time for 40 Because the DSP has the largest bus bandwidth requirement,
`every arbitration cycle to support the system's Ethernet
`the system is optimized to meet its need and sup port its
`connection. The floppy disk drive, serial ports and audio
`real-time operations. rn order to optimize this requirement
`input/output DMA, in a similar fashion, also require some
`and still give a reasonable bandwidth to the other bus
`guaranteed memory bus bandwidth. Likewise, with all other
`masters, the DSP is assigned 5 time slots among a tot al of 10
`constraints taken into account, the system's SCSI perfor- 45
`in the arbitration loop. The arbitration loop starts at DSP _ 1
`mance should not be allowed to degrade below an accept-
`state 301 and ends at the CPU state 310. The CPU state 310
`able level.
`is the point of reference which also designates the comple-
`Similar to the 1/0 interface constraints, there are con-
`tion of the entire arbitration loop. The NuB us controller 40
`and the 1/0 interface 30 each own 2 time slots in the
`straints imposed by the computer system being coupled to an
`expansion bus through the NuBus controller40. Like the 110 50 arbitration loop while the CPU 10 owns only 1 s]ot. The
`interface, there arc certain latencies that NuBus introduces
`sequence through the arbitration loop is the following:
`which must be provided for thus requiring a guaranteed
`DSP _ 1--?NuB_ l --?DSP _2--?lOB_ l --?DSP _3--? NuB_
`2~DSP _4-)IQB_2~DSP _ 5--KJ>U. This will be the
`amount of memory bus availability each arbitration cycle for
`the NuBus controller. Finally, though the CPU does very
`order of memory bus ownership if all bus masters are
`little in the way of real-lime operation in a system having a 55 requesting the bus in every state, though with some restric-
`DSP, it still must have some access every arbitration cycle
`