DSP Interrupt Control for Handling Multiple Interrupts

The present invention pertains in general to a multiple processor system and, more particularly, to an interface between a system bus and a multiple processor system for handling interrupts therebetween.

In order to increase the processing capability of any system, multiple processor arrays have been utilized. Typically, these multiple processor arrays are comprised of independent central processing units (CPUs) which are disposed in an array with each having a local bus with local peripheral units disposed thereabout. The local buses of each of the CPUs is then interfaced with a global bus. In turn, the global bus is interfaced with a system bus. One type of system bus is a VME bus.

When handling data flow between the system bus and each of the CPUs in the array of processors, traffic must be routed over the global bus. In order for information to be transmitted either from the processors to the system bus or from the system bus to the processors, there must be some type of arbitration. Typically, a bus request is sent out to the global bus control system and then the bus request granted to that requesting device. Data can then be transmitted over the bus in the appropriate manner. This is a conventional operation. However, the global bus becomes the limiting factor in transfer of data between processors and the system bus, and even between adjacent processors on the global bus. This is due to the fact that only one device can occupy the global bus at a given time.

One type of CPU that has been used widely is a Digital Signal Processor (DSP). These processors execute instructions at a very high rate but, unfortunately, like most processors, the architecture of the processor will determine the limitations of that processor with respect to communicating with the global bus, communicating with other processors and handling interrupts. Typically, most DSPs are designed for single chip use and must be provided with another layer of infrastructure in order to be incorporated into an array of microprocessors.

One difficulty in dealing with a multiple processor array is that of handling interrupts between the System Bus and the array of processors. This is due to the fact that a device on the System Bus side of the global bus sends out an interrupt in a normal matter which must then be transmitted to one or more of the processors in the array to be serviced. This requires the global bus to be occupied for the interrupt period in order for anyone of the processors to, first, recognize the interrupt and, second, to then service the interrupt. This is difficult when dealing with multiple processors in that some scheme must be developed in order to define which of the processors is to service the interrupt. This can be difficult if an interrupt is to be serviced by more than one processor.

The present invention disclosed and claimed herein comprises an interrupt controller for mapping multiple interrupts and their associated interrupt vectors received from a system bus to a single external processor interrupt input to a processor. The controller includes an interrupt receiving device for receiving the multiple interrupts and a first storage device for storing an indication of the received interrupt. A second storage device is provided for storing the received interrupt vectors for each of the received interrupts and in association with the stored indications. A processor interrupt generating device is provided to generate an external interrupt when the first storage device has stored therein an indication of a received interrupt. The processor is operable to service the stored interrupts and interrupt vectors in response to receiving the external processor interrupt from the processor interrupt generating device.

In another aspect of the present invention, there is provided a plurality of processor nodes, each with a processor that is responsive to receiving a single external processor interrupt. The interrupt receiving device is operable to receive the multiple interrupts and store them in the first storage device with an indication of which processor is to service the interrupt. The processor interrupt generating device then generates an external interrupt for the processor that is to receive the interrupt. The processors are operable then to send information to the processor interrupt generating device to indicate to the processor interrupt generating device which interrupts they are to service.

In a further aspect of the present invention, the interrupt controller provides an interrupt acknowledge manager for generating and sending an interrupt acknowledgment signal to the system bus whenever the processor services any of the stored interrupts. An inhibit device is provided for inhibiting receipt of an interrupt whenever the second storage device has stored therein an interrupt vector corresponding to a newly generated interrupt. The inhibit device inhibits receipt of the interrupt vector until such pending interrupt is serviced by the processor.

For a more complete understanding of the present invention and the advantages thereof, reference is now made to the following description taken in conjunction with the accompanying Drawings in which:

FIG. 1 illustrates an overall block diagram of a multi-processor system;

FIG. 2 illustrates a more detailed block diagram of the multi-processor system of FIG. 1 detailing the processor section;

FIG. 3 illustrates a memory map for the paging scheme;

FIG. 4 illustrates a block diagram of the interrupt control;

FIG. 5 illustrates a more detailed block diagram of the interrupt controller;

FIG. 6 illustrates a block diagram of the distribution logic of the interrupt controller;

FIG. 7 illustrates a flow chart for the System Bus interrupt servicing operations;

FIG. 8 illustrates a flow chart depicting the operation of the Eternal Interrupt to the CPU;

FIG. 9 illustrates a flow chart for the interrupt service operation on the CPU side;

FIG. 10 illustrates a block diagram of an interrupt mapping scheme for a single CPU;

FIG. 11 illustrates a block diagram of the DPSRAM structure;

FIG. 12 illustrates a diagrammatic view of the global address space illustrating the dedicated space for the DPSRAM of all processor nodes;

FIG. 13 illustrates a prior art global shared memory;

FIG. 14 illustrates the shared memory aspect of the dual port memory of the present invention;

FIGS. 15 and 16 illustrate flow charts depicting the semaphore operation for interfacing with the dual port memory;

FIG. 17 illustrates a diagrammatic view of an alternate memory access technique;

FIG. 18 illustrates a flow chart depicting the operation of the embodiment of FIG. 17;

FIG. 19 illustrates a flow chart depicting an alternate access technique utilizing a block counter;

FIG. 20 illustrates a block diagram of the system illustrating the paging feature;

FIG. 21 illustrates a more detailed block diagram of the paging registers and how they are multiplexed to the global bus; and

FIG. 22 illustrates a diagrammatic view of the generated paging address.

Referring now to FIG. 1, there is illustrated a block diagram of a processing system for interfacing between one of N processor cores 10 and an External Bus 12. The processor cores 10 are coupled to a global bus 14, which global bus 14 has associated therewith some type of global memory 16. The External Bus 12 is interfaced to the global bus 14 through an External Interface 18. In the preferred embodiment, the External Bus 12 is a Versatile Planar Bus (VME) which is a conventional bus described in "IEEE Standard for a Versatile Planar Bus; VMEbus STD 1014-1987", which is an IEEE standard for this bus. The External Interface 18 for the VME Bus is provided by typical chipset, one of which is termed the VIC64 part manufactured by Cypress Semiconductor.

As will described herein below, each of the processor cores 10 contains an internal Central Processing Unit (CPU) coupled to private memory and to I.backslash.O resources. Each of the CPUs has access to four types of I.backslash.O resources: a 32-bit; Bi-FIFO coupled, daughterboard expansion port supporting up to four hundred MB/s of I.backslash.O; a 32-bit control interface allowing the CPU random access to daughterboard resources; full MCBSP connectivity to the daughterboard; and global bus expansion through a dual port memory structure. The global bus 14, by comparison, contains shared resources, such as the global RAM 16, that are accessible from the External Bus 12 side or the processor cores 10.

Processing Node

Referring now to FIG. 2 there is illustrated a more detailed block diagram of the system of FIG. 1 illustrating in more detail one of the processor cores 10, this being labeled "Processor #0", there being additional processor cores 10 associated therewith. At the heart of the processor core 10 is a CPU 20 which comprises a 200 MHZ 320c 6201 processor, manufactured by Texas Instruments. The CPU 20 is interfaced with a local bus 22, local bus 22 having memory interfaced therewith. The memory comprises synchronous memory in the form of a synchronous DRAM (SDRAM) 24 and synchronous burst SRAM (SBSRAM) 26. The SDRAM 24 is a 16 MB memory configured as 4 M.times.32, with the SBSRAM 26 being 512 KB configured as 128 K.times.32. The local bus 22 is connected to an intermediate bus 30 through a bi-directional buffer 32 which has associated therewith a dual port SRAM (DPSRAM) 34 associated therewith, the bus 30 is input to one port of the DPSRAM 34, the other port thereof interfaced with the global bus 14. A bi-directional buffer 36 is connected between the bus 30 and the global bus 14. The intermediate bus 30 is also interfaced with an interprocessor bi-directional FIFO (Bi-FIFO 38) and also connected to an I.backslash.O expansion port 40 through a bidirectional buffer 42. The buffer 42 is designed to handle control transfers. With respect to communication with various peripheral devices, an I.backslash.O bi-directional FIFO (Bi-FIFO) 44 is provided which is disposed between the bus 30 and the I.backslash.O expansion port 40. There is also a register 46 provided for interface with the bus 30. The CPU 20 can also interface directly with the global bus 14 through a host port (H) through a bus 50 and a bi-directional buffer 52. The CPU 20 is also able to interface with a serial port 48 through a bi-directional serial buffer 50.

The global bus 14 includes global resources in the form of a flash memory 60, a non-volatile RAM (NVRAM) 62 and static RAM (SRAM) 64. The system will boot from the flash memory 60 which is an 8 MB global bus resident flash memory 60 with the SRAM 64 being 2 MB in size. The two provide a shared memory space for data transfers. A global bus slave 66 provides I.backslash.O expansion support. The system bus 12 is interfaced with global bus 14, as described hereinabove, through the VME interface 18, which VME interface 18 is of the type VIC64, manufactured by Cypress Semiconductor, the operation of which is described in the "VIC068NVAC068A User's Guide" which is published by Cypress Semiconductor, and which is incorporated here in my reference. The system bus 12 is also interfaced with the various expansion devices as indicated by block 68.

The CPU 20 utilizes the SBSRAM 26 as the processor core's zero wait state memory resource. After an initial time penalty to prime the pipe of the device, random accesses can occur to the memory structure at one transfer per clock yielding a throughput of 800 MB/s. The DPSRAM 34 functions as the data pipe for the processor core 10, allowing data to pass in or out of the processor core 10 without the need for arbitration on the global bus 14 or the double accesses associated with a shared memory structure (i.e., VME transfers data into global RAM 16, and the CPU 20 transfers the data from the global RAM 16 into the core, halving global bus throughput), this being described in more detail hereinbelow. With this memory structure, the CPU 20 can continue to execute instructions while a VME master loads or retrieves data from the memory structure.

The processor core 10 is provided with two ports through which data can be passed to or from the outside world. The first is through the I/O expansion port 40 and the other is the master/slave global bus gateway interface utilizing the DPSRAM 34. Both of these interfaces are designed to provide a high throughput datapipe structure that allows the CPU 20 to continue operating while data is loaded to or removed from the core, while additionally providing a random access interface for control functions. The I.backslash.O expansion port 40 provides the interface through the Bi-FIFO 44 to facilitate a high throughput streaming data interface, address, data, control, interrupts and both MCBSPs from the CPU 20. The streaming interface provided by the Bi-FIFO 44 with associated logic allows access to be performed at a 100 MHz rate, yielding a transfer rate up to 400 MB/s. The Bi-FIFO 44 additionally provides a mailbox function between the mother and daughter board, the daughter board being connected to the I.backslash.O expansion port 40. In connecting the I.backslash.O expansion port 40 in a direct point-to-point to the Bi-FIFO minimizes signal integrity issues. The global bus gateway provides a mechanism allowing the CPU 20 direct access to global resources and the VME transfers, while offering a streaming data path via DPSRAM 34. The DPSRAM 34 allows block transfer data to be directly transferred into a processor core 10 while the CPU 20 continues to execute instructions and eliminates the double accesses associated with a traditional shared global memory structure.

In one embodiment, the CPU 20 is capable of executing eight instructions in parallel and can produce 1600 MIPs/400 MMACs at 200 MHz. The CPU 20 offers an extensive set of memory and peripheral resources including one Mbit of on-chip SRAM, a 16-bit host port, to T1/E1 compatible Multi-channel Buffered Serial Ports (McBSP), a 4-channel bootable DMA, 2-general purpose timers and an advanced memory interface capable of directly interfacing to the SDRAM 24, the SBSRAM 26.

The SDRAM is a synchronous DRAM which is 16 MB in size configured as 4 M.times.32 to provide a large storage resource capable of providing a peak data throughput of 400 MB/s. Unlike the SBRAM 26 that can provide continuous random access at 800 MB/s, the user must exercise care in utilizing the SDRAM 24 as there is a non-trivial 60 nS row access penalty whenever a page boundary is crossed. However after the row access penalty, all subsequent accesses within the same page will require only two clocks offering 400 MB/s of throughput with 32-bit transfers. The SDRAM 24 is accessed through the 16 Mbyte CE2 memory window of the CPU 20.

In order for the array to allow communication between various processor cores 10, each processor core 10 is able to efficiently exchange blocks of data with other processor cores 10. The Bi-FIFO 30 is provided that is a switched bank FIFO scheme that allows interprocessor communication to occur of up to 400 MB/s per path along with an ability to exercise multiple paths concurrently. Additionally, this scheme passes directly between the CPUs 20 eliminating the delays associated with arbitration and traffic for a traditional global bus based scheme. A programmable flag is provided which allows users to specify their relevant block size and whether they will be interrupted or poll to determine transfer status. This scheme is linearly scalable supporting a peak interprocessor transfer rate of 3.2 GB/s in a complex of eight CPUs 20.

The DPSRAM 34 is a 256 KB memory, configured at 64 K.times.32, to provide an interface between processor core #0 and the VME bus 12. Not only does the DPSRAM 34 provide a high throughput memory device that minimizes global bus traffic, but the DPSRAM 34 additionally provides for eight mailbox resources (four each direction) between the CPU 20 and the global bus 14. Since the memory block is dual ported between the processor core 10 and the global bus 14, a single access is all that is required to transfer a word of data between the VME bus 12 and each processor core 10. This offers a greater than 100% performance improvement over traditional global shared memory architectures that required data to be transferred to the global resident SRAM, another arbitration cycle performed, and then transferred out of the global memory. This not only requires twice the time, but ties up the global bus 14 for other CPUs 20 and interprocessor communication. Secondly, due to the larger depth of the DPSRAM 34 utilized, the distributed structure significantly increases the size of the available global memory. As this is a switched bank architecture, there may need to be an arbitration cycle to obtain the structure prior to access. The mailboxes offer an efficient means to transfer messages between the VME bus 12 and each processor core 10.

The processor expansion port 40 provides a high throughput streaming data interface with a FIFO scheme, described hereinabove, that allows blocks of data to pass between the motherboard and the daughterboard at up to 400 MB/s. A 2.times.1 K.times.32 Bi-FIFO operating at 100 MHz acts as the interface between the daughterboard connector and the intermediate bus 30 of the processor core 10. The daughterboard will have access to all of the necessary control and flags for the Bi-FIFO 44. This interface also provides a microprocessor style interface allowing interface to buffered versions of the intermediate bus 30 signals. The daughterboard has access to sixteen bits of address, thirty two bits of data, transfer control signals, an interrupt and a clock synchronized to the processor 20. This interface is accessed via the CPU 20 memory space and is visible regardless of the contents of a global paging register (described hereinbelow).

CPU Addressing Overview

The CPU 20 has four memory windows with up to 52 MBytes of addressable space for resources outside the CPU 20. However, this amount of memory is only possible if three of the windows contain a multiplexed address resource, this being the SDRAM 24. The entire multiple-processor system contains only one such window, and, therefore, the maximum total amount of external memory available through direct addressing is 28 MBytes. Each of these windows is dedicated to a specific memory type, with the three choices being: synchronous DRAM (SDRAM) 24, synchronous burst SRAM (SBSRAM) 26 and asynchronous memory. The memory type and data width are fixed at boot time, based on the state of certain inputs to the CPU 20. The following table 1.0 illustrates the programming of these windows for the multiple-processor system.

The multi-processor architecture employs a hybrid addressing scheme that allows users to trade off the highest possible performance in favor of ease of programming. The CE3 and CE2 memory windows are not affected by this hybrid scheme as the CPU 20 is a direct connect to the memory resources. The CE1 memory window is the less complicated of the remaining two windows, as the memory space may have either SBSRAM or Async timing, and a Switch-Bank (SB) architecture is employed. The trade off considerations for selecting the interface's timing as well as a detailed description of the operation of the CE1 address space is described hereinbelow with respect to the inter-processor communications. The Switch-Bank architecture is employed to assure optimal performance of the inter-processor/VME bus FIFO paths. As the timing of the SBSRAM 26 does not afford logic an adequate amount of time to perform an address decode on the fly, a pre-decode is performed through the programming of the processor core control register (described hereinbelow). Bits 15 and 16 of the miscellaneous control register control which FIFO block is logically connected to the CPU 20. In SBSRAM 26, this structure allows each of the FIFOs to operate at a peak throughput of 400 MB/s. As a decode in the Async space would also force a wait state access, the SB architecture is also useful with this timing.

The CPU 20 offers only an address space of 4 MB (max) in its asynchronous window (CE0). To accommodate for this, a paging scheme is employed to allow the addressing of over 2 Gbytes of address space necessary for the CPU 20 to address all of the system's resources and the VME bus 18, as will be described hereinbelow. Additionally, since some resources, including private registers associated with the CPU 20, the local side of the CPU 20 DPSRAM 34, and the random access daughterboard interface need to be accessible at all times, this local address space is divided into two sections, a local section and a global section. This allows global accesses to exist in the lower 2 MB and private resources associated therewith to exist in the upper 2 MB. This is determined by the address A21 being set equal to "0" for the global access and being set equal "1" for the local access. This division facilitates an extremely quick and simple decode. At reset, the paging register is set to "0000" such that the CPU 20 will automatically point at the flash memory 60 for booting. The CE0 address map is illustrated in FIG. 3.

With reference to FIG. 3, the CPU 20 external memory illustrating how the CE0 is provided with two sections, an upper section 21 and a lower section 23. This constitutes the space that is addressable from the CPU 20, i.e., this address space residing directly in the address space of the CPU 20.

The lower section 23 provides access to the global resources and the VME bus 18. The paging scheme allows the lower 2 MB section 23 to be mapped into the global bus memory map, which is illustrated as having 4 GBytes of addressable space which has a portion thereof associated with the VME A32 master access space, the VME A24 master access space, the VME A16 master access space, the daughterboard global bus resources and the motherboard global resources.

For each CPU 20, there is provided a private 16-bit paging register. This paging register is provided for in the FPGA. The upper twelve and lower two bits of the paging register contain the paging field with the remaining two bits reserved. The paging field is combined with the CPU 20 buffered address to form the 32-bit global bus address. The top bit of the paging register (bit 11) determines whether the CPU 20 is performing a VME access or not. If this bit is set, a VME access is performed, and the remaining three bits of the paging register are combined with the CPU 20 address that is enabled on the global bus 23 and presented to the VME. This allows the CPU 20 to have full access to the VME A32 address.

After a CPU 20 has been granted the global bus 23, the contents of the paging register and buffered versions of the CPU 20 address and byte enables will be combined to drive the bus. The lower 21 bits of the global bus address are formed from buffered address bits 20:2 of the CPU 20 and the encoded processor byte enables. The 13 bits of the paging register drive the upper 11 bits global address, 31:21, and the lower two, 1:0.

The local bus memory map is set forth in table 2.0, which shows the portion of the memory map that is associated with the CE0, the CE1, internal program memory and internal peripheral memory, the SDRAM (CE2), the SBSRAM (CE3) and some internal data memory. It is noted that private control registers are associated with the address range "0028 0000-002B FFFF".

The global bus memory map, as it is viewed by the CPU 20, is viewed as a 4 GByte address space that includes all of the local resources and all of VME space for master accesses. As noted above, this space is divided up into 2 MB pages, as that is all that is available with the CPU 20 address space after assignment of the remaining space to the core private resources. This global bus memory map is illustrated in Table 3.0.

With respect to the VME A32, A24 and A16 slave memory maps, they are viewed by the external bus, as illustrated in Tables 4.0, 5.0 and 6.0 as follows:

Processor Private Control Register Definition

Each of the processor cores 10 has associated therewith a number of control registers. These are contained within an FPGA, which is mapped to the address locations 00280000h-002FFFFFh. The specific registers associated with each of the processor cores 10 are set forth in the memory map of Table 7.

There are a number of interrupt registers for handling interrupts for the interprocessor communications generated by the Bi-FIFOs 38 for each of the processor cores 10. These will basically be the interrupt enable registers "0" and "1" and the interrupt flag register "0" and "1". In addition, there are provided two interrupt status registers. The function of each of these is set forth as follows:

Interrupt Enable Register 0, IER0

This register contains the enables for the interrupt sources associated with one of the two inter-processor communication Bi-FIFOs. Each processor 20 has a unique version of this register that is presented to the processor 20 at the above address. A given processor 20 can only access its own register. The register contains the enables which allow the associated flags to set the specified interrupt. If a given enable is set, a flag condition becoming active will result in the interrupt being set.

Interrupt Enable Register 1, IER1

This register contains the enable bits for the interrupt sources associated with second of the two inter-processor communication Bi-FIFOs and the processor private expansion I/O Bi-FIFOs. Each processor 20 has a unique version of this register that is presented to the processor 20 at the above address. A given processor 20 can only access its own register. The register contains the enables which allow the associated flags to set the specified interrupt. If the associated enable is set, a flag condition becoming active will result in the interrupt being set.

Interrupt Flag Register 0, IFR0

This register contains the flag bits for the interrupt sources associated with the first of the two inter-processor communication Bi-FIFOs. Each processor 20 has a unique version of this register that is presented to the processor 20 at the above address. A given processor 20 can only access its own register. The flag bit is a latched version of the status bit that becomes set when an individual status bit makes a transition from inactive to active. A flag bit will remain set until it is cleared by writing a 0 to the specific bit position. Note: A flag can not be directly set by writing a 1 to a specific register bit position.

Interrupt Flag Register 1, IFR1

This register contains the flag bits for the interrupt sources associated with second of the two inter-processor communication Bi-FIFOs and the processor private expansion I/O Bi-FIFOs. Each processor 20 has a unique version of this register that is presented to the processor 20 at the above address. A given processor 20 can only access its own register. The flag bit is a latched version of the status bit that becomes set when an individual status bit makes a transition from inactive to active. A flag bit will remain set until it is cleared by writing a 0 to the specific bit position. Note: A flag can not be directly set by writing a 1 to a specific register bit position.

Interrupt Status Register 0, ISR0

This register contains the status bits for the interrupt sources associated with the first of the two inter-processor communication Bi-FIFOs. Each processor 20 has a unique version of this register that is presented to the processor 20 at the above address. A given processor 20 can only access its own register. The status bit is a buffered version of the interrupting source and may be monitored by software to verify that an interrupting condition has been properly cleared.

Interrupt Status Register 1, ISR1

This register contains the status bits for the interrupt sources associated with second of the two inter-processor communication Bi-FIFOs and the processor private expansion I/O Bi-FIFOs. Each processor 20 has a unique version of this register that is presented to the processor 20 at the above address. A given processor 20 can only access its own register. The status bit is a buffered version of the interrupting source and may be monitored by software to verify that an interrupting condition has been properly cleared.

For each enable, flag and status register, Table 8 illustrates the connection path for the XX path and the YY path.

Miscellaneous Interrupt & VME Control Register, MIVR

This register contains the interrupt control bits for the remaining interrupt sources associated with each processor core. The sources include the dual port memory bank, the Vbus expansion daughterboard interrupt, and the global daughterboard interrupt. Each CPU 20 has a unique version of this register that is presented to the CPU 20 at the above address. A given CPU 20 can only access its own register. The enable, flag and status bits operate the same as those for all other interrupt sources.

Miscellaneous Control Register 0, MCR0

This register contains control bits associated with logic functions within the processor core 10. The FIFO select bits contain which of the three Bi-directional FIFO paths will be selected when accessing the processing core's BI-FIFO memory space. The mapping of codes to the FIFO selected is illustrated in table 9.0. The FIFO resets allow the three Bi-directional FIFOs connected to the CPU 20 to be individually reset. Setting the bit will result in forcing a FIFO into reset and holding it there until it is cleared. The LED bits are general purpose bits used to illuminate the processor cores LEDs. Writing a one to a bit position will cause the associated LED to illuminate.

Board Configuration Register, BCR

This register contains configuration and status bits that provide the CPU 20 with information regarding the configuration of the board/system including the board's VME base address, indicators whether daughterboards are attached to the processor core 10 and/or the global bus 14, the CPU's 20 number and the quantity of CPUs 20 contained on the card. All of the bits contained in this register are Read Only and may be regarded as static signals. Each CPU 20 has a private copy of this register.

Global Bus Paging Register, GBPR

The global bus paging register contains control bits that are used whenever a CPU 20 performs a master access on the global bus 14. When a CPU 20 is granted master status on the global bus 14 by the bus arbiter, the contents of this register are used to drive the upper 12 and lower two bits of the global bus address. These bits allow the limited external address space of the CPU 20 to be augmented such that each CPU 20 can access all global resources including the entire 2 GB address space of the VME Bus. Each CPU 20 has a private copy of this register.

Global bus Semaphore Register, GBSR

The global bus semaphore register is each processor cores 10 gateway to the four bank semaphores associated with the GDSS block that separates each processor core 10 from the global bus 14, and the twelve general purpose global bus semaphores. The table at the bottom of the page specifies the protocol for obtaining and then releasing any of the 16 semaphores accessible by the CPU 20. Note: it is important to exercise care in the use of the dual port semaphores as writes to banks of memory not owned by the CPU 20 will be ignored and reads will return non-deterministic data. Each CPU 20 has a private copy of this register. This register provides a local view of semaphores, and can allow a CPU 20 to see semaphores for its banks of memory in the DPSRAM 34, and not the semaphores for the other DPSRAMs 34.

VME Interrupt Control Registers

IPL Gating Control Register, IGCR

The CPU 20 uses this register to decide which IPL levels that the CPU 20 will respond to. Setting a bit in this register will allow a given IPL level to set the associated flag in the interrupt pending register. Whether this will result in an interrupt being generated to the CPU 20 is determined by the state of the IPL's associated enable bit in the IPL Enable Register. Each CPU 20 has a private copy of this register.

VME Interrupt Pending Register

The VME IPL Interrupt Pending Register is used by the CPU 20 to determine the cause of EXT.sub.-- INT4 assertion or to poll the status of IPL conditions that are not enabled to cause interrupts to the processor but are still of interest to the CPU 20. The interrupt acknowledge logic uses the state of all of the processors registers to determine when an interrupt has been fully serviced and another may be acknowledged at that level. If a given bit position is clear for all CPUs 20 the interrupt acknowledge controller is free to acknowledge interrupt at that level. A bit may be set in this register if the associated IPL bit is enabled in the gating register and that IPL condition becomes active. Once a bit becomes set, it is cleared by either reading the vector for the associated IPL level or writing a one to this register in the associated position. Each CPU 20 has a private copy of this register.

IPLx Interrupt Vector Registers, IVR1-IVR7

This register is generally accessed by the CPU 20 in response to data received from the interrupt pending register. One eight bit register is assigned to each IPL level. The register contains the vector for the last IPL acknowledged to the VIC64 and currently being serviced by the CPUs 20. Each CPU's 20 version of this register is loaded after an IPL level is acknowledged to the VME interface and the interrupt pending bits updated based on the contents of the IPL gating register. The act of reading a given IPL vector register will clear the associated bit position in the interrupt pending register. Each CPU 20 has a private version of this register.

Interrupt Structure

The basic system described hereinabove has the ability to handle interrupts to all CPUs 20 from multiple on-board requesters and from the VME bus 12, as well as being able to generate interrupts to the VME bus 12. Multiple sources mapped to each interrupt require that the CPU 20 poll the interrupt status register to determine the source (s). Interrupt sources may be individually enabled to cause an interrupt or poll condition (latched and unconditioned versions)to determine status. The structure is expandable with the ability to support up to eight CPUs 20. Interrupts to the VME bus 12 are handled through the VME interface controller 18.

The CPU 20 has four input pins on which it can receive external inputs, EXT.sub.-- INT <7.. 4>. All four interrupts are edge sensitive (programmable edge). Each interrupt is assigned to a logical function which has multiple interrupt sources behind it. The mapping of logical function to interrupt is illustrated in the following Table 11. In order to ascertain the source of the interrupt, the interrupt handler needs to poll the Interrupt Flag Register.

Each motherboard resident interrupt source is associated with three control register bits that affect it's operation. These are the interrupt enable, the interrupt flag, and the interrupt status control bits. The enable bit associated with the interrupt allows the source to generate an interrupt when its status becomes active. When the bit is set, an interrupt will be generated by the condition. The interrupt flag is a latched version of the source that will latch and hold the condition (even after it is rescinded) until the CPU 20 clears a bit by writing a "1" to it. Condition status bit is a raw (unlatched) version of the interrupting condition. The bit will be set when the condition is active (i.e., halffull *-"0"). The status bit is read only, while the other two bits are read/write.

The EXT.sub.-- INT7 interrupt to each CPU 20 is associated with the inter-processor Bi-FIFO 38 that connects the four processor cores 10 to each other and the I/O expansion port 40 (or alternatively IP connection to the DB resident processors). Each CPU 20 may be interrupted based on the status of its side of the three Bi-FIFOs 38 attached to it and the state of the EXT.sub.-- INT7 enabled field of the Interrupt Enable Register. There are nineteen potential sources for the EXT.sub.-- INT7 interrupt which are illustrated in Table 12. This table illustrates the bit position of the interrupt enable, the condition flag and the condition status for each source. For the daughterboard interrupt, additional polling may be required if multiple daughterboard interrupt sources exist for the single MvPBus interrupt. The status for interrupt is contained in the interrupt status registers (ISR) 0 and 1, found at 0.times.0028 0010 and 0028 0014 in the memory map.

The EXT.sub.-- INT6 interrupt to each CPU 20 is associated with the Dual Port Memory, DPSRAM 34, mailbox logic. When one of the mailboxes is written to by the current global bus master, the interrupt will be sent to the CPU 20 if it is enabled in the miscellaneous Interrupt and VME Control Register. Each mailbox (four total) has an enable, flag and status bit associated with it. The mailboxes and their associated control register are contained within the DPSRAM 34 that connects the global bus to each of the CPU's 20 local bus. The control information associated with EXT.sub.-- INT6 is illustrated in Table 7. There are two sets of registers associated with the VME Mailbox logic. The first set of registers is read/write accessible and represents each CPUs 20 outgoing data being passed to the VME bus 12. Each CPU 20 has a private copy of these registers. The second set of registers that is read only represents the data coming in form the VME bus 12. Each CPU 20 has a private copy of these registers.

The EXT.sub.-- INT5 interrupt of each processor 20 is associated with the inter-processor mailbox logic. When one of the mailboxes is written to, the interrupt will be sent to the processor 20 if it is enabled in the Interrupt Control Register. Each mailbox (four total) has an enable, flag and status bit associated with it. The mailbox is contained within one of the three Bi-FIFOs attached to the processor complex. Two of the Bi-FIFOs are connected to other processor complexes and the third is associated with the I/O expansion site. The control information associated with EXT.sub.-- INT5 is illustrated in Table 10. There are two sets of registers associated with the IP Mailbox logic. The first set of registers is read/write accessible and represents outgoing data being passed to the other processor 20. The second set of registers that is read only, represents the data coming from the other processor 20.

The EXT.sub.-- INT4 interrupt to each processor 20 is associated with interrupts that cause the state of the VME interface's IPL lines to change. These sources include VME interrupts, local Interrupts, VME interface resident mailboxes, error conditions, and general status (such as DMA completion). Additionally, unlike the other interrupt sources that have a unique source for each processor 20, the interrupt source is common to all CPUs 20, and has the ability to interrupt any or all of the processor cores 10 (8 possible). Because of this, significant additional logic is required to control the IPL destinations and the potential multiple interrupt acknowledge cycles coming from the processor cores 10.

Referring now to FIG. 4, there is illustrated a block diagram of the architecture for handling interrupts between the VME Bus 12 and the processor cores 10 for each of the associated CPUs 20. Each of the CPUs 20 have a separate interrupt line 80 connected to the EXT.sub.-- INT4 interrupt input. The interrupt lines 80 are driven by an interrupt controller 82, which interrupt controller 82 interfaces with the CPU local bus 22 for each of the processor cores 10. The interrupt controller 82 also interfaces with the global bus 14 and also directly with the VME Interface 18 for receiving interrupt information therefrom.

As will be described in more detail herein below, the interrupt controller 82 is basically a steering device. Whenever an interrupt is received from the system over the VME Bus 12, the VME Interface 18 in conjunction with interrupt controller 82 treats the CPUs 20 as if there were a single CPU 20 on the global bus of the VME Interface 18. Each of the CPUs 20 are programmed such that it downloads information to the interrupt controller 82 as to what interrupt(s) it will handle. The interrupt controller 82 is operable then to receive the interrupt from the VME Interface 18 and steer this interrupt to the appropriate one or more of the CPUs 20. The interrupt controller 82 will send back a single acknowledgment signal when the interrupt has been serviced. The interrupt controller 82 will therefore monitor the CPU 20 side of the global bus 14 to determine if the CPUs 20 have in fact serviced the interrupt directed thereto.

Referring now to FIG. 5, there is illustrated a high-level block diagram of the interrupt controller 82. The interrupt controller 82 sits between 19 potential interface interrupt sources on the VME Bus 12 and the EXT.sub.-- INT4 inputs of each of the CPUs 20. The interrupt sources are input on an interrupt line 100 to an interrupt-to-IPL Mapper Block 102. This block handles the matching of interrupt sources to the various IPL levels. The output of the Mapper Block 102 is input to a Distribution Logic Block 104, which is contained in the VME Interrupt Control Array, this Block 104 handling the conversion of individual IPL codes to an edge interrupt that can be presented to each (or all) CPU's EXT.sub.-- INT4 Interrupt Control Logic. The output of this will drive Enable Logic Blocks 106 which are common to all of the interrupt inputs, this controlling whether the conditioned interrupt source reaches a given one of the processor cores 10. Additionally, there is provided an acknowledge block 110 which is operable to generate an interrupt acknowledge signal back to the mapper block 102 which indicates that the interrupts have been serviced.

The mapper block 102 is responsible for encoding the 19 possible interrupt sources onto the IPL outputs that are presented to the IPL gate logic 104. The interrupt to IPL Mapper can encode each interrupt or interrupt group to any of seven levels, as programed by the Interrupt Handler Register (IHR) contained in the VME Interface 18. The contents of the these registers are loaded by CPU 20 in processor core 10 (#0) after reset. Once programmed, the IPL lines will reflect the state of the highest ranked interrupt source. Table 17 illustrates the 19 possible interrupt sources and their respective ranking.

The IPL pending and Distribution Logic Block 104 converts the IPL code received from the VME Interface 18 into a pulse input that drives the EXT.sub.-- INT4 interrupt controller Logic for each CPU 20. The logic performs gating functions based on the contents of each CPU's 20 Gating Register (IGR=0.times.0028 0100) and the state of the Interrupt Acknowledge Controller. The format of each IPL Gating Register is illustrated in Table 18. Each CPU 20 has a unique copy of this register located at the same address, and can only access its copy of this register. Each bit location within the byte is assigned to a given IPL code. If the bit associated with that IPL code is set, then the output of the distribution logic block 104 to that CPU 20 is pulsed when the appropriate IPL code is presented and no other interrupts at the IPL level are outstanding. This results in the associated flag being set in that CPU's 20 IPL Interrupt Pending Register (logically equivalent to Flag Register). It is noted that multiple processor enables for the CPUs 20 maybe set for a given IPL code. If no outstanding interrupts for a given IPL level are present, the Interrupt Acknowledge Logic will fetch and latch the interrupt vector from the VME Interface 18. Once the vector is latched, the interrupt will be sent off to the Processor Enable Logic.

Due to the ability to support multiple interrupts and due to limitations of hardware, there is a window of opportunity where the actual interrupt pulse may be missed by the CPU 20. This situation occurs between the time that an interrupt has occurred but not been acknowledged. If a new interrupt occurs at this time, the CPU 20 will miss it. However, the interrupt service routing can handle this condition and essentially recover the missing interrupts. During the interrupt acknowledge cycle the CPU 20 will read the contents of its Interrupt Pending Register. The contents of this register will indicate if multiple outstanding interrupts exist, even though the CPU 20 is only responding to one. After servicing the highest priority interrupt, the CPU 20 can re-read the IPR to determine if the bits associated with it are "0." If not, the CPU 20 can service the next interrupt and repeat until it is "0." Because the pending interrupt information is latched, and subsequent interrupts are gated via the Controller 100, interrupts will not be lost.

The Acknowledgment Manager 110 is responsible for servicing interrupt acknowledgments and vector passing between the multiple CPU 20 and the VME interface 18. Additionally, through interface with the Distribution Logic 104, the interrupting acknowledgment Manager Block 110 effectively throttles the interrupts heading to the CPUs 20. The initial problem needing to be solved is that the VME Interface 18 is only expecting a single processor to respond to an interrupt. Once an acknowledge is sent back to the VME Interface 18, the vector associated with it is lost forever. Therefore, a latch is employed for each IPL level to retain the state of a vector until all of the processors 20 have a chance to respond. No additional interrupts for a given IPL level can be passed until the previous interrupt at that level has been fully responded to. Otherwise, interrupt overrun would occur, resulting in lost interrupts.

Because an interrupt vector is lost once a single interrupt acknowledge cycle has been performed to the VME Interface 18, the Interrupt Acknowledge Manager 110 must latch the value presented to the VME Interface 18 during the first interrupt acknowledgment for the eight possible interrupt acknowledge cycles coming from the CPUs 20. Once the Control Logic determines that an IPL code can be acknowledged, it asserts a global bus request. After it gains control of the global bus 14, it will perform an interrupt acknowledge cycle, latch the interrupt vector byte for the specific IPL code into its associated Interrupt Vector Register, update the associated Interrupt Pending Register Byte, and then release the global bus 14.

The Interrupt Acknowledge Manager 110 operates based on the contents of eight 8-bit read only registers that are bit-wise auto-loaded with the contents of the IGR registers when an IPL code is acknowledged. These registers, Interrupt Pending Registers (1 per CPU 20), contain information concerning how many CPUs 20 have responded to each IPL level. Seven bits in each register (the lower bit of each register is reserved) are each associated with an IPL code. The organization of this register is illustrated in the Table 19. When the register is initially loaded at the time of the interrupt acknowledge, the bit will be loaded with a "1" if the associated bit is set in that CPU's IGR. As the CPU 20 performs an interrupt acknowledge cycle by reading the Interrupt Vector Register (IVR, 1 per IPL Code ), the bit position associated with that CPU 20 and IPL code is reset as CPUs poll this register.

The value of IPR is additionally utilized by the Interrupt Acknowledge Controller to determine when a new IPL level can be acknowledged. If the value of this bit (1 per IPR) for a specific IPL code is non-zero when that code is presented to the IPL gate logic, it indicates that the previous interrupt at that IPL level has not been acknowledged by all enabled CPUs 20. An acknowledge for the new interrupts cannot occur or Interrupt overrun (and loss) would occur. If the bits associated with an IPL is zero when the IPL is presented again, the Controller 82 will arbitrate for the global bus 14, the interrupt will be acknowledged, the vector loaded and the IPR bits associated with the IPL code loaded with contents of the associated IGR bit. In this manner, multiple CPUs 20 can respond to multiple outstanding interrupts without interrupt or vector loss.

Each IPL code will have two control bits associated therewith, the enable bit and the pending bit. These control bits are contained in two 8-bit registers accessible through Distribution Logic Block 104. The format of the control registers is illustrated below in Table 15.

Referring now to FIG. 6, there is illustrated a block diagram of the interrupt controller 82. Which describes the servicing of the interrupt from both the system bus side and the CPU side. For illustrative purposes, only one CPU 20 is illustrated, it being understood that multiple CPUs can be incorporated into the array, and alternatively, a single CPU 20 could be utilized. Each of the CPUs 20 receives on the input thereof an external interrupt, EXT.sub.-- INT4. This interrupt is one of multiple interrupts at the CPU 20 will accommodate. This interrupt is generated by interrupt control logic 120. The interrupt generated by this interrupt control logic block 120 indicates to the CPU 20 that there has been an interrupt generated on the system bus side of the array that is to be serviced by the CPU 20.

The CPU 20 determines which interrupts it will service. Once this is determined, it will load information regarding these interrupts into a gating register 122, as described hereinabove, which has associated therewith one bit for each interrupt level. Once it determines which interrupts it will service, it sets these particular bits it sets its gating register 122, it being understood that there is one gating register per CPU 20. The gating register 122 is an addressable register that is accessed from the local bus 22.

The CPU 20 not only determines which interrupts it will service, but also which interrupts it desires to examine. The gating register 122 will have the bits set therein that are associated with all interrupts that it both service and will poll. For interrupts that it will service, it will subsequently run an Interrupt Service Routine (ISR) the ones that it will poll, it does not run the ISR. In order to determine which interrupts it will actually service with the ISR, an enable register 124 is provided which is interfaced with the local bus 22 to allow the CPU 20 to store therein the interrupts that it will service, there being one bit for each interrupt and there being one enable register 124 per CPU 20. The contents of the enable register are input to the interrupt control logic block 120. In addition, there are provided status lines 126, one for each interrupt level, which are compared to contents of the enable register 124. The status lines 126 indicate the that there is an interrupt that is to be serviced or polled by the CPU 20, as indicated by the contents of the gating register 122. This will be described in more detail herein below.

Whenever an interrupt is received, it is received on a 3 bit bus 130, this providing for a binary encoding of a seven interrupt levels. This is decoded by an interrupt decode block 132 to provide seven decode output lines which are then input to interrupt distribution logic block 134. The interrupt distribution logic block 134 receives an input from each of the gating registers 122 for each of the CPUs 20. For the CPU 20 illustrated, the gating register provides an output on a bus 136 for input to the interrupt distribution logic block. A group of buses 140 receive the instructions from gating registers of other CPUs 20. The interrupt distribution logic block 134 provides on the output thereof a status bit. Theses status bits are provided on status lines 126 for the illustrated CPU 20. These are input to the interrupt control logic block 120, as described herein above, and also to a flag register 142. The flag register 142 will have the flag set for each status bit that is set by the interrupt distribution logic block 134. The interrupt distribution logic block 134 will only set the status bits associated with a received interrupt that compares with the serviceable or pollable status bits defined in the gating register 122 associated with that CPU 20, which status bits are delivered to the other CPUs 20 through a group of status lines 137. This flag register 142 is addressable from the local bus 122 side. Therefore, the CPU 20, after receiving its external interrupt EXT.sub.-- INT4, will then examine the contents of the flag register 142 in order to determine which interrupt it is to service. The CPU 20 will service the interrupts in a priority based upon the interrupt number, the highest priority interrupt being serviced first.

When the interrupt is received on the interrupt bus 130, a vector distribution logic block 150 is operable to transfer interrupt vectors from a data bus 152, which transfer the vectors from the VME bus 12 through the global bus 14 to a plurality of vector registers 152. The vector registers 152 are registers that store a vector for a single interrupt level. If there are seven interrupt levels, then there will seven vector registers 152. These vector registers are accessible by each of the CPUs 20. There can be a separate set of vector registers 152 for each of the CPUs 20 or there can be one global set. However, whenever an interrupt is generated from the system side, the vector distribution logic block 150 will route the vector to the appropriate one of the vector registers 152 and latch it therein. In the illustrated embodiment, there are provided a plurality of distribution buses 154 that are operable to transfer the information from the vector distribution logic block 150 to the appropriate vector register 152. It being understood that these can essentially be one bus 154 that can address the multiple registers 152. These buses 154 also go to vector registers 152 associated with the other CPUs 20.

In order to perform an acknowledge of the servicing of an interrupt back to the system bus, an interrupt acknowledge block 160 is provided which generates an interrupt acknowledge signal on a line 162, which is delivered back to the system side of the VME interface 18. The interrupt acknowledge block 160 examines the contents of a plurality of interrupt pending registers 162. There is provided a single interrupt pending register (IPR) 162 for each of the interrupt levels. The contents of the pending registers are set whenever an interrupt designated for a particular CPU 20, as defined by the contents of the gating register 122, is to be serviced by that CPU 20. The contents of the gating register 122 are compared to the presence of the logic state on the status lines 126 to generate the control bits for the interrupt pending registers 162, the control bits transferred thereto on lines 170. Since the interrupt pending registers 162 have a bit associated with each processor, the interrupt acknowledge block 160 can determine if any of the bits in any of the registers 162 are still set. For example, if all CPUs in an array were to service IPL7, then each CPU's 20 interrupt pending register bit associated with IPL7 would be bits set therein. The interrupt acknowledge block 160 would receive both indication on lines 176 from the output to the decode block 132 indicating the interrupt that was generated and also would receive the pending bits on lines 178 from the interrupt pending registers 162. Since the interrupt acknowledge block 160 is aware that an IPL7 was generated, it will not generate the acknowledge signal on line 162 until the all bits in the IPL7 portion of the interrupt pending registers 162 have been cleared. These bits are cleared whenever the contents of the vector registers 152 are read by the CPU 20, the vector register 152 interfaced with the local data bus 22. Each of the vector registers 152 outputs a Clear signal on lines 180 to the interrupt pending registers 162.

Referring now to FIG. 7, there is illustrated a flow chart depicting the operation of servicing the interrupt on the system bus side. This initiated at a block 190 and then proceeds to a decision block 192 to determine if an interrupt has occurred on the system bus side. If not, the program will flow back to the input thereof and, if so, the program will flow to a function block 194 to decode the interrupt and then to function block 196 to check the interrupt pending register. As described herein above, the interrupt pending registers will determine if any pending interrupts are present for that given interrupt. This is determined in a decision block 198. If there is a existing pending interrupt, the program will flow back to the input of decision block 192 in order to continue checking the associated IPR 162. When all pending interrupts have been cleared for that interrupt, the program will flow from decision block 198 to an acknowledge block 200 in order to acknowledge that the interrupt has been received and then to a function block 202 in order to retrieve the interrupt vector for latching in the vector register 152. The program will then flow to a function block 204 to load the vector register and then to a function block 206 to set the pending bits based on the gate register values. The program will flow to a return block 208.

Referring now to FIG. 8, there is illustrated a flow chart depicting the operation for generating the external interrupt to the CPU 20, the interrupt EXT.sub.-- INT4 is initiated at block 220 and then proceeds to a decision block 222 to determine if an interrupt has been received from the system bus. If not, the program will flow back to the input of decision block 222 and, as so, the program will flow to a function block 224, where the contents of the associated gate register 122 are read. The program will then flow to function block 226 to distribute the status bits. The program will flow to a function block 228, wherein the flag register bits will be set for later access by the CPU 20. The interrupt control logic block 120 will then check the enable bit, as to find in a function 230 and then flow to decision block 232 in order to determine whether the enable bit for the particular activated status line 126 is set. If so, the program will flow along a "Y" path to function block 234 in order to set the external interrupt, EXT.sub.-- INT4. The program will flow then to return block 236. If the enable for the associated active status line 126 is not set, the program will flow from the decision block 232 to the return block 236.

Referring now to FIG. 9, there is illustrated a flow chart depicting the CPU service operation for the interrupt when the external interrupt, EXT.sub.-- INT4, is received. This flow chart is initiated a block 240 and then proceeds to a decision block 242 wherein it is determined whether the external interrupt, EXT.sub.-- INT4, has occurred. If not, the program will flow back to the input of the decision block, and, if so, the program will flow to a function block 244 wherein the enable register 124 will be polled. Thereafter, the flag register will be polled, as indicated by a function block 246 and then the priority of the interrupt determined is indicated by function block 248. The flag register will indicate all of the interrupts that exist wherein the enable register 244 will indicate the ones of the flag register bits that are to be serviced. Thereafter, the highest priority vector will be retrieved by the CPU 20 from the associated vector register 152, as indicated by function block 250. This reading of the vector register 122 will clear the related bit position of the IPR. The program then flows to a function block 252 where the CPU 20 performs the Interrupt Service Routine (ISR) and then the condition is cleared. The program then flows to function block 254 wherein the flag bit is cleared and then to a decision block 256 to determine if additional flags have been set. If so, the program will flow along the "Y" path back to the input of function block 244. When all flags have been reset, then the program will flow to a return block 260.

Referring now to FIG. 10, there is illustrated a block diagram illustrating the use to the interrupt controller with a single CPU 20. If a single CPU 20 is provided, the purpose of the interrupt controller would be to receive multiple inputs on interrupts lines 300 and map those interrupts through an interrupt mapping block 302 into a single interrupt line 304, for input to the CPU 20. The interrupt mapping block 302 will contain information as to which interrupt generated a single interrupt which can be read by the CPU 20 through a bus 306. The CPU 20, once determining the interrupt that is responsible for generating the single interrupt, will then read the contents of the vector latch 310 which receives and latches vectors from a data bus 312. The interrupt mapping block 302 and vector latch 310 will interface with an interrupt acknowledgment block 314 which will generate an acknowledgment signal on line 316 whenever the contents of the vector latch have been read, indicating that the CPU 20 has serviced the interrupt, and will prevent.

Dual Port SRAM

Referring now to FIG. 11, there is illustrated a detailed block diagram regarding the operation of the DPSRAM 34. In general, the CPU 20 in a conventional mode will receive information from the global bus 14 and from one of multiple global resources, as illustrated by block 300 on the global bus 14, through the host port in the CPU 20. This is performed in a conventional operation which requires in general two accesses. The first access by the global resources 300 in the conventional mode will require writing of the data to the global RAM 16. The second access is for the CPU 20 to then access the information in the global RAM 16. Of course, during storage of data in the global RAM 16, the system must undergo some type of addressing of the CPU 20 to ensure that it is that CPU 20 that constitutes the destination CPU. In this mode, the global RAM 16 is essentially a "scratchpad" memory device. With this type of operation, there is only one region of the global address space that will be designated as the common area for writing to any of the CPUs 20 in the system.

In the present invention, the DPSRAM 34 allows each CPU 20 to have a designated portion of the global address space associated therewith. Therefore, whenever any of the global resources 300 desires to write information to the CPU 20, it need only write directly to the address space, such that the CPU 20 now constitutes an addressable entity on the global bus 14 within its address space. The DPSRAM 34 facilitates this feature while isolating the global address space from the local address space of a particular CPU 20. The DPSRAM 34 is basically a single bank of memory of 256 Kbytes, on each side, arranged in a 64 K.times.32 configuration on each side, one port on the global side and one port on the local side. The performance specifications for the DPSRAM 34 are as follows:

DPSRAM, local side

Size: (per processor)

Standard: 256 Kbytes, 4 Banks, 64K.times.32

Arbitration: None, but may need to obtain semaphore

CPU 20 Access: 0 Wait State, 4 clock access, 200 MB/s @32 bits

VME Access: No

DPSRAM, global side

Size: (per processor)

Standard: 256 Kbytes, 4 Banks, 64K.times.32

Arbitration: None, but may need to obtain semaphore

CPU 20 Access: 32 bit, 6 clock access after global bus arbitration

VME Access: Yes, A32: D32, D32 Word, D32 Byte, D32 Block, D64 Block

A24: D32, D32 Word, D32 Byte

A16: Not accessible

The DPSRAM 34 is of the type 1017028, manufactured by IDT, which is interfaced with the CPU 20 via an interrupt line 302. The interrupt line 302 will send an interrupt to the CPU 20 anytime that data is written into the DPSRAM 34 mailbox register by any global resource. Once written thereto, the CPU 20 can service the information stored in the DPSRAM 34 from the local side via the intermediate bus 30.

Referring now to FIG. 12, there is illustrated a diagrammatic view of the global address space. It can be seen that each processor node has associated therewith a defined region in the global address associated with its DPSRAM 34. Therefore, whenever a global resource 300 desires to send information to any CPU 20 at any of the processor nodes 10, it need only generate the address for the particular DPSRAM 34. Since this occupies an address space in the global address space, a single Write operation will write the information to the DPSRAM 34 for the designated CPU 20 and the next access cycle on the global bus 14 can write information to the DPSRAM 34 of another CPU 20 or to that CPU 20. The access cycle involves first writing to the DPSRAM address the block of data. Thereafter, the global resource writes to the mailbox in the DPSRAM 34, which action causes the interrupt EXT.sub.-- INT6 to be generated. Further, in another aspect of the operation of the DPSRAM 34 in conjunction with the CPU 20, the CPU 20 itself can access another CPU 20 via that CPU's DPSRAM address. This is facilitated through the buffer 32, which is the path that is utilized by the CPU 24 to gain access both to the VME or other global resources in the block 300. In this mode, a CPU 20 on another processing node constitutes a global resource to a given CPU 20 when that CPU 20 is acting as master. In general, the buffer 32 is utilized for data that is being transferred to and from the global bus 14 by the associated CPU 20. Without the DPSRAM 34, the CPU 20 would have to access another CPU via conventional methods, wherein the data would first have to be stored in the global RAM 16 and then instructions sent to another CPU 20 in order for it to access the data from the global RAM 16 designated thereto.

Data can be transferred from the CPU 20 to the global resource 300. This is facilitated in the same manner as transferring data from the global resource 300 to the CPU 20. Data is first transferred to the DPSRAM 34 on the local side thereof by the CPU 20. On the local side of the DPSRAM 34 in the local address space, a designated resource on the global bus must have a defined address in the local address space. By writing to this address, the global resource to which it is directed can retrieve the data. After writing the data, then the mailbox for that resource, a dedicated mailbox, is written to. This generates an interrupt for the global resource, which global resource can then access the global side of the DPSRAM 34 for retrieval of the data therefrom.

Referring now to FIG. 13, there is illustrated a block diagram of a prior art system for providing a shared or distributed global memory operation. In this operation, a CPU 310 is interfaced with a local bus 312, which is separated from a global bus 314, the global bus 314 being the same as the global bus 14 described hereinabove. Interfaced with the global bus 314 is a global resource 316. This global resource 316 can be a VME system bus, it can be global RAM or it can be another CPU. In order for there to be some memory transfer between the global resource 316 and the CPU 310, there needs to be some type of memory. This is provided with a shared memory 318, which is interfaced with an intermediate bus 320. The intermediate bus 320 is interfaced with the local bus 312 through a buffer 322 and with the global bus 314 through a buffer 324. An arbiter 326 is provided for determining when the data is transferred between the CPU 310 and the shared memory 318 or between the global resource 316 and the shared memory 318. Although not shown, both the CPU 310 and the global resource 316 have a control connection with the arbiter 326. The shared memory 318 is to be distinguished from local resources in a block 328 which are directly addressable by the CPU 310 through its local bus 312. In a local access, there is no need to request access through a global memory address, as the local resource 328 resides in the local memory space of the CPU 310. However, when the CPU 310 desires to transfer information to the global bus 314 and to the global resource block 316, one way to achieve this is to transfer data to the shared memory 318 after gaining access thereto. When the CPU 310 has access to shared memory 318, the global resource 316 cannot have access thereto. After data is transferred to the shared memory 318, the global resource 316 will then be apprised of the presence of data in the shared memory 318 and will then access the shared memory 318 at a defined address. A shared memory block 318 can be provided for in multiple processor nodes on the global bus 314. This allows the global resource to access the multiple shared memories at different times and, when writing data to another of the shared memories 318, data previously written to memory 318 can be simultaneously accessed for a Read or a Write operation.

One disadvantage to the distributed memory of FIG. 13 is that, even though each shared memory block 318 will have a separate address in the global address space, it must first have dedicated access to the shared memory 318, transfer data thereto and then provide an indication to the CPU 310 that there is data to be retrieved. Therefore, the global resource 316, or the CPU 310, must have complete access to the shared memory 318 to the exclusion of the other of the CPU 310 or the global resource 316.

Referring now to FIG. 14, there is illustrated a simplified block diagram of the DPSRAM 34 as implemented in the embodiment of FIG. 13, wherein like numerals refer to like parts in the two FIGUREs. With the DPSRAM 34 disposed between the local bus 312 and the global bus 314, the global resource block 316 has a means to communicate with the CPU 310. However, it is important to note that the difference between the shared memory block 318 and the DPSRAM 34 is that the global resource 316 has dedicated access to the DPSRAM 34 for transferring data thereto or retrieving data therefrom. Although there is a semaphore that must be retrieved before it can actually write to or read from the DPSRAM 34, the read/write operation is independent of the read/write operation of the CPU 310 in the local address space thereof. Therefore, each of the DPSRAMs 34 occupy a defined portion of the global address space, which defined portion of the global address space is independently accessible by the global resources 316 from the CPU 310.

Referring now to FIG. 15, there is illustrated a flow chart depicting the semaphore operation to determine how the CPU 20 will access the DPSRAM 34 for a read/write operation and also how the global resource side of the global bus 14 will access the DPSRAM 34. In general, in the DPSRAM 34, there is a single memory cell for each storage location that is accessible from either side of the dual port memory. However, since there is only one cell, there can only be one side accessing at a given time for a write operation. The dual port memory has associated therewith logic to prevent simultaneous access thereto to ensure that there is no contention problem. There is provided global semaphore registers to facilitate this.

The global bus 14 has twelve general purpose semaphores that on a bit by bit basis are shared by all of the processors and the VME bus. The semaphores are accessed by the CPU 20 locally by each CPU 20 through the local bus semaphore register. The semaphore bits are contained in the lower twelve bits of this register. The VME bus 12 accesses these semaphores through the global semaphore register. The semaphore bits are contained in the lower twelve bits of the register.

A semaphore is assigned to each of the banks that make up each processor cores 10 DPSRAM 34 bank. The semaphores control whether the global bus or the local bus can access a specific bank within the structure. Accesses to regions of memory that are not owned by the accessing party will be ignored in the case of Writes and will return non-deterministic data in the case of Reads. These semaphores are accessed by the CPU 20 through the local bus semaphore register. The four semaphores associated with the banks that are locally connected are accessed through the upper four bits of this register. These semaphores are accessed globally through the DPSRAM semaphore register.

In all cases, each semaphore bit operates as follows. Reading a zero indicates that the semaphore is available. If it is available, a "1" is written to the bit to request it. This is followed by a read to confirm that the semaphore has been granted to the CPU 20 (automatically resolved by hardware during contention). If set, the CPU 20 owns the semaphore. To release the semaphore, the owning CPU 20 writes a 0 to that bit. At reset, all semaphores are set to be available. There is also a dual port semaphore provided on in a dual port semaphore register, as follows:

Referring further to FIG. 15, the read/write operation is initiated at a block 340 and then proceeds to a block 342. At block 342, the ownership of the semaphore is verified. The program proceeds to a block 344 to determine if the requestor is the owner of the semaphore. If not, the program will flow along the "N" path to a block 346 to request the semaphore and back to the input of block 344. When ownership is determined, the program will flow to function block 348 to perform the operation, either a Read or a Write operation, and then to a block 350 to release the semaphore and then to an END block 352.

Referring now to FIG. 16, there is illustrated a flow chart for depicting the semaphore control, which is initiated at a block 354 and then proceeds to a decision block 356 to determine if the semaphore has been requested. Until a request has been lodged, the program will flow back to the input thereof. Once a request has been lodged, the program will flow to a decision block 358 to determine if the semaphore is in use. If in use, the program will flow back to the input of decision block 358. Once it is determined to be free, the program will flow from decision block 358 to a function block 360 to deliver the semaphore and then to a function block 362 in order to set the requestor's ownership bit. The program will then flow to a decision block 364 to determine if the semaphore has been released. Until it released, the program will loop back to the input of decision block 364 and, once released, the program will flow back to the input of decision block 356.

Referring now FIG. 17, there is illustrated a block diagram of the DPSRAM 34 illustrating an alternate method for accessing the DPSRAM 34 for transfer of data thereto and retrieval of data therefrom. With the use of the semaphore operation, there are multiple cycles that are required. In general, this is a four cycle operation which requires the semaphore to be queried and the grant followed by the write operation and then the release operation. However, this is only acceptable for large blocks, as the transfer of small blocks of data by this technique will result in a relatively large amount of overhead as a percentage of the operation. Alternately, as illustrated in the block diagram of FIG. 17, the accessing side of the DPSRAM 34 can directly access the DPSRAM 34 without making a determination as to whether it actually has exclusive access to this resource on the global bus. As described hereinabove, the only detriment to this type of access is that there may be a situation wherein one side tries to access a cell that the other side is writing to. In this situation, there will be a "collision" which will be dealt with by the DPSRAM 34. The DPSRAM 34 will lock out a request to access by the other side when one side is actually writing to the local or the similar situation with the read operation. The DPSRAM 34 will only provide the Ready signal to the appropriate resource whenever the other side is not accessing that particular cell. Therefore, when either the CPU 20 or a global resource attempts to access a DPSRAM 34 and the other side has already accessed it, the Ready signal will not be generated for that side and they will be forced into a Wait state. For small blocks of data, this will facilitate the transfer of data in the overall system and increase throughput.

As noted hereinabove, after writing of information to the DPSRAM 34, an interrupt must then be generated to the other side to indicate that data is present for a read operation. To facilitate this, mail boxes are provided. These mailboxes allow information to be transmitted in the form of control or command information to the other side. The other side will have logic associated therewith for generating an interrupt whenever information is in the mailbox to be read and then the receiving side will service that interrupt in a conventional manner to recognize that it is required to read the contents of the mailbox. This will indicate that there is information there to be read. With respect to this mailbox, many blocks of data can be stored in the DPSRAM 34. It is only important that the receiving side know what data is to be read. It could, of course, merely read all of the contents and compare them with the previous data transfer. However, it is anticipated that the mailbox will be utilized to store the actual location of the data that is to be read or the block size that is to be read. Any information about the data stored therein as to location or size can be contained within the mailbox and transferred thereto by the transmitting side.

Referring now to FIG. 18, there is illustrated a flow chart depicting the operation of writing information to the DPSRAM 34 and generating an interrupt therefrom. The program is initiated at a block 370 for a Write operation and then proceeds to a block 372 to address the DPSRAM 34, this assuming it already has access thereto through ownership of the semaphore. The program will then flow to a function block 374 to write the information thereto and then to a function block 376 to transfer the command information to the mailbox. Once this command information is stored in the mailbox, an interrupt will be generated, as indicated by a function block 378. The program then flows to an End block 380.

In an alternate embodiment of the present invention, a block counter is utilized. In this operation a counter value is stored in a register (not shown) which is operable to store an initial count value and decrement this count value for each block stored in the DPSRAM 34. Once all data has been transferred to the DPSRAM 34, the interrupt will be generated. This operation is described in the flow chart of FIG. 19.

In the flow chart of FIG. 19, the program is initiated in a block 382 and then proceeds to a function block 384 in order to load the block counter with the transfer size that is to transferred from the transmitting side, this being the blocks to be transmitted. The program will then flow to a function block 386 to transfer data thereto in the form of blocks. The program flows to a decision block 388 to determine if a block has been transferred. Once the data has been transferred, the program will flow to a function block 390 to decrement the counter value and then to a decision block 392. Decision block 392 determines if the count value is set equal to zero. If not, the program will loop back around to the input of function block 386. Once the count value is set equal to zero, indicating that all data have been transferred thereto, the program will flow to a function block 394 in order to generate the interrupt to the receiving side and then to an End block 396. With the system utilizing the block counter, large data transfer operations can be performed with multiple data without requiring an interrupt to be generated via the mailbox operation.

Paging Operation

Referring now to FIG. 20, there is illustrated a block diagram of two CPUs 20 associated with two processor nodes 10, which interface with a global register file 410. The global register file 410 contains registers that are to be associated with a common function and also to provide an output to the global bus 14. For the paging operation described hereinabove, the global register file 410 contains a global register (not shown) for each processor node 10 and for each associated CPU 20. A global bus arbiter 412 is provided which is interfaced with the global bus 14. The arbiter 412 operates conventionally in that it determines what global resource has access to the global bus 14 at any time. This is a conventional operation. The arbiter 412 will control the global register file 410 and allow any portion thereof to have access to the global bus 14.

As will be described hereinbelow, each CPU 20, when acting as a bus master, can generate an address within its 4 Mbytes of local address space and, with the use of the its associated paging register, it can add the upper bits required to address the much larger, 4 Gbytes, address space for the global bus 14. However, the difficulty in performing this operation is that intermediate bus 30 cannot carry a large enough address for output to the global bus 14.

Referring now to FIG. 21, there is illustrated a block diagram of the global register file 410 which is associated with the paging operation. A paging register 314 is associated with each of the CPUs 20 at each of the processor nodes 10, only three being illustrated in FIG. 21. Each paging register 314 is interfaced with the CPU 20 through the intermediate bus 30. The CPU 20, when desiring to access the global bus 14, will generate a paging address for storage in its associated paging register 314. As described hereinabove, twelve bits of the register contain the paging field which is then combined with the CPU 20 buffered address to form the 33-bit global bus address. After the particular CPU 20 for a given one of the processor nodes 10 is granted access to the bus and constitutes the bus master, a multiplexer 316 then selects from the output of one of the paging registers 414 for output to the global address bus 14 as the upper twelve bits of the global bus 14 address. The multiplexer 316 is controlled by the arbiter 412. This is illustrated in FIG. 22.

In operation, the paging method allows for a plurality of processors to share a common memory space, the global address space, with each other, while retaining some local address space for local resources. In FIG. 3, for example, it was noted that each processor has associated therewith a memory space divided into two memory spaces, the local memory space 21 and the global memory space 23. The local memory space 21 is a memory space that is associated with memory on the local bus or intermediate bus 30 which is utilized for the various local resources associated with that particular CPU 20. Therefore, the CPU 20 having only 4 MB of addressability has that addressable space divided into two segments, a first segment associated with the local resources and accessible only by that CPU 20 and the other addressable space, the addressable space 23, associated with the shared resources on the global side. If any CPU 20 wants to access a particular segment of the global address space in a particular page in the global address space, it need only generate the particular paging address. However, since the global address space is shared by all processor cores 10, this means each processor core 10 has the ability to, through its CPU 20, address this common global address space in any segment in any page. For example, if a register location was required for the operation of a CPU 20 and this register location were disposed in the global address space, there would have to be another level of inhibit logic utilized in the software to prohibit other CPUs 20 from accessing that register. Therefore, the system of the present invention has core resources distributed among the CPUs, which core resources do not occupy any portion of the address space of the other CPUs 20. It is noted that, even though the DPSRAM 34 may be accessible as a global resource within the global address space by another CPU 20, the DPSRAM 34 is configured such that there is a global side and a local side wherein the local side occupies a portion of the local address space and the global side occupies a portion of the global address space. Although there is no reason for any CPU 20 to access its own DPSRAM 34, it is possible through the paging scheme to actually enter the address for the global side of the DPSRAM 34 associated with the addressing CPU 20 since it does occupy a portion of the global address space. However, the address for addressing the global side of the DPSRAM 34 is different than the address on the local side, as the global address occupies the portion 23 and the local side occupies the portion 21 of the local address space for the local bus or intermediate bus 30. In effect, therefore, the DPSRAM 34 would actually occupy two portions of the local address space within a given processor core, it being recognized that a paging address must also be generated in addition to the lower portion of the address being generated in the portion 23.

In summary, there has been provided a method for interfacing global resources associated with a global bus with one of a plurality of processor nodes that are disposed on the global bus. Each of the processors has associated therewith a dual ported memory structure that occupies a specific portion of the global bus address space. When data is written to this address space, an interrupt is generated for the particular CPU at the designated processor node to read this information. In this manner, a block of information can be transmitted to the global side of the memory and read on the local side of the memory structure by the associated CPU. In the manner, only a single access cycle is required to transfer information from a global resource to a processor node.

Although the preferred embodiment has been described in detail, it should be understood that various changes, substitutions and alterations can be made therein without departing from the spirit and scope of the invention as defined by the appended claims.