US RE41904 E1
Techniques are described for providing mechanisms of data distribution to and collection of data from multiple memories in a data processing system. The system may suitably be a manifold array (ManArray) processing system employing an array of processing elements. Virtual to physical processing element (PE) identifier translation is employed in conjunction with a ManArray PE interconnection topology to support a variety of communication models, such as hypercube and such. Also, PE addressing nodes are based upon logically nested parameterized loops. Mechanisms for updating loop parameters, as well as exemplary instruction formats are also described.
1. An apparatus for performing virtual identification (VID) to physical identification (PID) translation for data elements to be accessed within local memory of a processing element (PE) whereby a direct memory access (DMA) controller can access PE local memories according to their VIDs, the apparatus comprising:
an array of multiple PEs each having local PE memory;
a DMA controller; and
a memory maintained in the DMA controller for storing a processing element VID-to-PID table mapping processing element VIDs to processing element PIDs utilized by the DMA controller to access local memories according to their VIDs.
2. The apparatus of
3. The apparatus of
4. The apparatus of
5. The apparatus of
6. The apparatus of
7. The apparatus of
8. The apparatus of
9. The apparatus of
a local memory interface unit for each processing element (PE) storing a VID for each PE.
10. The apparatus of
11. The apparatus of
12. The apparatus of
13. A processing apparatus comprising:
a plurality of processing elements (PEs) communicatively connected by a bus, each PE comprising a register storing a virtual identification number (VID) identifying the PE; and
a direct memory access (DMA) controller connected to the bus for accessing local data memory of the PEs, each data access at least partially identified by a VID;
wherein during a common data to access multiple PEs, a PE responds to the data access if the VID stored in the register matches the VID of the data access.
14. The processing apparatus of
15. The processing apparatus of
16. The processing apparatus of
17. The processing apparatus of
18. An apparatus for accessing local memory of a plurality of processing elements (PEs), the apparatus comprising:
a transfer controller running a process containing a set of nested loops, the set of nested loops having a plurality of parameters to be specified by a transfer instruction, the plurality of parameters, when assigned, control PE selection and address generation for accessing a memory location in local memory of each selected PE; and
a means for receiving the transfer instruction for transferring data between system memory and local memory of the plurality of PEs, the transfer instruction having fields which specify values for the plurality of parameters, the transfer instruction indicating an addressing mode, the addressing mode specifying a particular pattern of accessing local memory of the plurality of PEs, wherein the transfer controller decodes the transfer instruction to assign values to the plurality of parameters, the process generating addresses for accessing a memory location in local memory of each selected PE in a particular pattern, wherein the particular pattern is based on the assigned parameters.
19. The apparatus of
20. The apparatus of
21. The apparatus of
22. The apparatus of
23. The apparatus of
24. The apparatus of
25. The apparatus of
26. A method of accessing local memory of a plurality of processing elements (PEs), the method comprising:
receiving a transfer instruction for transferring data between system memory and the local memory of a plurality of processing elements (PEs);
running a process containing a set of nested loops, the set of nested loops having a plurality of parameters to be assigned values of fields carried in the transfer instruction;
decoding the transfer instruction to assign field values to the plurality of parameters;
assigning the field values to the plurality of parameters in order to control PE selection and address generation for accessing a memory location in local memory of each selected PE; and
generating addresses to access local memory of each PE in a defined pattern.
27. The method of
28. The method of
29. The method of
30. The method of
More than one reissue application has been filed for the reissue of U.S. Pat. No. 6,453,367. The reissue applications are application Ser. No. 10/819,885 and which is the present divisional reissue application.
The present application is a division of U.S. application Ser. No. 09/472,372 filed Dec. 23, 1999, now U.S. Pat. No. 6,256,683, which in turn claimed the benefit of U.S. Provisional Application Ser. No. 60/113,637 entitled “Methods and Apparatus for Providing Direct Memory Access (DMA) Engine” and filed Dec. 23, 1998 which is incorporated by reference in its entirety herein.
The present invention relates generally to improvements in array processing, and more particularly to advantageous techniques for providing improved mechanisms of data distribution to, and collection from multiple memories often associated with and local to processing elements within an array processor.
Various prior art techniques exist for the transfer of data between system memories or between system memories and I/O devices.
The DMA controller 160 provides a mechanism for transferring data between processor local memory and system memory or I/O devices concurrent with uniprocessor execution. DMA controllers are sometimes referred to as I/O processors or transfer processors in the literature. System performance is improved since the host uniprocessor can perform computations while the DMA controller is transferring new input data to the processor local memory and transferring result data to output devices or the system memory. A data transfer is typically specified with the following minimum set of parameters: source address, destination address, and number of data elements to transfer. Addresses are interpreted by the system hardware and uniquely specify I/O devices or memory locations from which data must be read or to which data must be written. Sometimes additional parameters are provided such as element size. One of the limitations of conventional DMA controllers is that address generation capabilities for the data source and data destination are often constrained to be the same. For example, when only a source address, destination address and a transfer count are specified, the implied data access pattern is block-oriented, that is, a sequence of data words from contiguous addresses starting with the source address is copied to a sequence of contiguous addresses starting at the destination address. Array processing presents challenges for data collection and distribution both in terms of addressing flexibility, control and performance. The patterns in which data elements are distributed and collected from processing element local memories can significantly affect the overall performance of the processing system. With the advent of the ManArray architecture it has been recognized that it will be advantageous to have improved techniques for data transfer which provide these capabilities and which are tailored to this new architecture.
As described in detail below, the present invention addresses a variety of advantageous methods and apparatus for improved data transfer control within a data processing system. In particular we provide improved techniques for: distributing data to, and collecting data from an array of processing elements (PEs) in a flexible and efficient manner; and PE address translation which allows data distribution and collection based on PE virtual IDs.
Further aspects of the present invention are related to a virtual-to-physical PE ID translation which works together with a ManArray PE interconnection topology to support a variety of communication models (such as hypercube and mesh) through data placement based upon a PE virtual ID. This result can be accomplished in a DMA controller by translation, through a VID-to-PID lookup table or through combinational logic, where the resulting PID becomes an addressing component on the DMA bus to PE local memories. This result can also be achieved at the PE local memories within the interface logic, where a VID available to the interface logic is compared to a VID presented on the DMA bus. A match at a particular memory interface allows that memory to accept the access. The present invention also addresses the provision of PE addressing modes based on generating data access patterns from logically nested parameterized loops. Varying assignments of loop parameters to nesting level allows flexible data access patterns to be generated. Providing varying mechanisms for updating loop parameters provides greater flexibility for generating complex-periodic access patterspatterns, such as select-index modes which provide a table of index-update values which are used when the index loop parameter is updated; select-PE modes which provide a table of bit-vector control values, each of which specifies the PEs to be accessed for an iteration through the “PE update loop” (i.e., the loop which PE update is assigned); and select-index-PE modes which provide both select-index and select-PE update capability and combine to form the most flexible mode for generating complex-periodic data access patterns. Further, the invention addresses the design of a looping mechanism to be reentrant thereby allowing any addressing mode to be restarted after completing a specific number of element transfers, by just loading or reloading a new transfer count and continuing the transfer. This result is accomplished by initializing addressing parameters at instruction load time, and only updating them after a loop exits.
These and other advantages of the present invention will be apparent from the drawings and the Detailed Description which follow.
Further details of a presently preferred ManArray core, architecture, and instructions for use in conjunction with the present invention are found in U.S. patent application Ser. No. 08/885,310 filed Jun. 30, 1997, now U.S. Pat. No. 6,023,753, U.S. patent application Ser. No. 08/949,122 filed Oct. 10, 1997, now U.S. Pat. No. 6,167,502, U.S. patent application Ser. No. 09/169,255 filed Oct. 9, 1998, U.S. patent application Ser. No. 09/169,256 filed Oct. 9, 1998, now U.S. Pat. No. 6,167,501, U.S. patent application Ser. No. 09/169,072 filed Oct. 9, 1998, now U.S. Pat. No. 6,219,776, U.S. patent application Ser. No. 09/187,539 filed Nov. 6, 1998, now U.S. Pat. No. 6,151,668, U.S. patent application Ser. No. 09/205,558 filed Dec. 4, 1998, now U.S. Pat. No. 6,173,389, U.S. patent application Ser. No. 09/215,081 filed Dec. 18, 1998, now U.S. Pat. No. 6,101,592, U.S. patent application Ser. No. 09/228,374 filed Jan. 12, 1999, now U.S. Pat. No. 6,216,223, U.S. patent application Ser. No. 09/238,446 filed Jan. 28, 1999, U.S. patent application Ser. No. 09/267,570 filed Mar. 12, 1999, U.S. patent application Ser. No. 09/337,839 filed Jun. 22, 1999, U.S. patent application Ser. No. 09/350,191 filed Jul. 9, 1999, U.S. patent application Ser. No. 09/422,015 filed Oct. 21, 1999, U.S. patent application Ser. No. 09/432,705 filed Nov. 2, 1999, U.S. patent application Ser. No. 09/471,217 filed Dec. 23, 1999, now U.S. Pat. No. 6,260,082, as well as, Provisional Application Ser. No. 60/139,946 entitled “Methods and Apparatus for Data Dependent Address Operations and Efficient Variable Length Code Decoding in a VLIW Processor” filed Jun. 18, 1999, Provisional Application Ser. No. 60/140,245 entitled “Methods and Apparatus for Generalized Event Detection and Action Specification in a Processor” filed Jun. 21, 1999, Provisional Application Ser. No. 60/140,163 entitled “Methods and Apparatus for Improved Efficiency in Pipeline Simulation and Emulation” filed Jun. 21, 1999, Provisional Application Ser. No. 60/140,162 entitled “Methods and Apparatus for Initiating and Re-Synchronizing Multi-Cycle SIMD Instructions” filed Jun. 21, 1999, Provisional Application Ser. No. 60/140,244 entitled “Methods and Apparatus for Providing One-By-One Manifold Array (1×1 ManArray) Program Context Control” filed Jun. 21, 1999, Provisional Application Ser. No. 60/140,325 entitled “Methods and Apparatus for Establishing Port Priority Function in a VLIW Processor” filed Jun. 21, 1999, Provisional Application Ser. No. 60/140,425 entitled “Methods and Apparatus for Parallel Processing Utilizing a Manifold Array (ManArray) Architecture and Instruction Syntax” filed Jun. 22, 1999, Provisional Application Ser. No. 60/165,337 entitled “Efficient Cosine Transform Implementations on the ManArray Architecture” filed Nov. 12, 1999, and Provisional Application Ser. No. 60/171,911 entitled “Methods and Apparatus for Loading of Very Long Instruction Word Memory” filed Dec. 23, 1999, respectively, all of which are assigned to the assignee of the present invention and incorporated by reference herein in their entirety.
The following definitions of terms are provided as background for the discussion of the invention which follows:
A “transfer” refers to the movement of one or more units of data from a source device (either I/O or memory) to a destination device (I/O or memory).
A data “source” or “destination” refers to a device from which data may be read or to which data may be written which is characterized by a contiguous sequence of one or more addresses, each of which is associated with a data storage element of some unit size. For some data sources and destinations there is a many-to-one mapping of addresses to data element storage locations. For example, an I/O device may be accessed using one of many addresses in a range of addresses, yet it will perform the same operation, such as returning the next data element of a FIFO, for any of them.
A “data access pattern” is a sequence of data source or destination addresses whose relationship to each other is periodic. For example, the sequence of addresses 0, 1, 2, 4, 5, 6, 8, 9, 10, . . . etc. is a data access pattern. If we look at the differences between successive addresses, we find: 1,1,2, 1,1,2, 1,1,2, . . . etc. Every three elements the pattern repeats.
An “address mode” or “addressing mode” refers to a rule that describes a sequence of addresses, usually in terms of one or more parameters. For example, a “block” address mode is described by the rule: address[i]=base_address+i where i=0, 1, 2, . . . etc. and where base_address is a parameter and refers to the starting address of the sequence.
Another example is a “stride” address mode which may be described by the rule:
An “address generation unit (AGU)” is a hardware module that generates a sequence of addresses (a data access pattern) according to a programmed address mode.
“EOT” means “end-of-transfer” and refers to the state when a transfer execution unit (described in the following text) has completed its most recent transfer instruction by transferring the number of elements specified by the instruction's transfer count field.
The term “host processor” as used in the following description is any processor or device which can write control commands and read status from the DMA controller and/or which can respond to DMA controller messages and signals. In general, a host processor interacts with a DMA controller to control and synchronize the flow of data between devices and memories in the system in such a way as to avoid overrun and underrun conditions at the sources and destinations of data transfers.
The present invention provides a set of flexible addressing modes for supporting efficient data transfers to and from multiple memories, together with methods and apparatus for allowing data accesses to be directed to PEs according to virtual as opposed to physical IDs. This section describes an exemplary DMA controller and a system environment in which the present inventions may be effectively used. The following sections describe PE memory addressing, virtual-to-physical PE ID translation and its purpose, and a set of PE memory addressing modes or “PE addressing modes” which support numerous parallel algorithms with highly efficient data transfer.
In this representative system, the DMA controller also connects to two system busses, a system control bus (SCB) 235 and a system data bus (SDB) 240. The DMA controller is designed to transfer data between devices on the SDB 240, such as a system memory 250 and the DSP 203 local memories 210-215. The SCB 235 is used by an SCB master such as the DSP 203 or a host control processor (HCP) 245 to program the DMA controller 201 with read and write addresses and registers to initiate control operations and read status. The SCB 235 is also used by the DMA controller 201 to send synchronization messages to other SCB bus slaves such as the DSP control registers 225 and a host I/O block 255. Some registers in these slaves can be polled by the DSP and HCP to receive status from the DMA. Alternatively, DMA writes to some of these slave addresses can be programmed to cause interrupts to the DSP and/or HCP allowing DMA controller messages to be handled by interrupt service routines.
Each transfer controller within a ManArray DMA controller is designed to fetch its own stream of DMA instructions. DMA instructions are of five basic types: transfer; branch; load; synchronization; and state control. The branch, load, synchronization, and state control types of instructions are collectively referred to as “control instructions”, and distinguished from the transfer instructions which actually perform data transfers. DMA instructions are typically of multi-word length and require a variable number of cycles to execute although several control instructions require only a single word to specify. Although the presently preferred embodiment supports multiple DMA instruction types as described in further detail in U.S. patent application Ser. No. 09/471,217 filed Dec. 23, 1999, now U.S. Pat. No. 6,260,082, and incorporated by reference in its entirety herein, the present invention focuses on instructions and mechanisms which provide for flexible and efficient data transfers to and from multiple memories.
Referring further to system 400 of
A “transfer-system-inbound” (TSI) instruction moves data from the SDB 470 to the IDQ 405 and is executed by the STU. A “transfer-core-inbound” (TCI) instruction moves data from the IDQ 405 to the DMA Bus 425 and is executed by the CTU. A “transfer-core-outbound” (TCO) instruction moves data from the DMA Bus 425 to the ODQ 406 and is executed by the CTU. A “transfer-system-outbound” (TSO) instruction moves data from the ODQ 406 to the SDB 470 and is executed by the STU. Two transfer instructions are required to move data between an SDB system memory and one or more SP or PE local memories on the DMA bus, and both instructions are executed concurrently: a TSI, TCI pair or a TSO, TCO pair.
The address parameter of STU transfer instructions TSI and TSO refers to addresses on the SDB while the address parameter of CTU transfer instructions refers to addresses on the DMA bus to PE and SP local memories.
While there are six memories 210, 211, 212, 213, 214, and 215 shown in
The ManArray architecture supports a unique interconnection network between processing elements (PEs) which uses PE virtual IDs (VIDs) to support useful single-cycle communication paths, for example, torus or hypercube paths. In some array organizations, the PE's physical and virtual IDs are equal. The VIDs are used in the architecture to specify the pattern for data distribution and collection. When data is distributed according to the pattern established by VID assignment, then efficient inter-PE communication required by the programmer becomes available. As an example, if a programmer needs to establish a hypercube connectivity for a 16 PE ManArray processor, the data will be distributed according to a VID assignment in such a manner that the physical switch connections allow data to be transferred between PEs as though the switch topology were a hypercube even if the switch connections between physical PEs do not support the fill hyper-cube interconnect. The present invention describes two approaches whereby the DMA controller can access PE memories according to their VIDs, effectively mapping PE virtual IDs to PE physical IDs (PIDs). The first uses VID-to-PID translation within the CTU of a transfer controller. This translation can be performed either through table-lookup, or through logic permutations on the VID. The second approach associates a VID with a PE by providing a programmable register within the PE or the PE local memory interface unit (LMIU),
VID to PID Translation within the DMA Controller
With this approach, a PE VID-to-PID table is maintained in the DMA controller so that data may be distributed to the ManArray according to a programmer's view of the array. In the preferred embodiment, this table is maintained in the CTU of each transfer controller.
The approach of
In the presently preferred embodiment, a lookup table is used to perform the VID-to-PID translation. Two approaches are provided for initializing the translation table. The first is through a DMA instruction 800, shown in FIG. 8. When executed, DMA instruction 800 loads a PETABLE register 900 which is illustrated in FIG. 9. The second approach is through a direct write of the PETABLE register 900 via the SCB.
PE Virtual IDs Stored in Local Memory Interface Units
The second approach to directing data access according to PE VID relies on distributing the PE VIDs to each PE local memory interface unit (LMIU). The VID for each PE might reside in a register either in the PE itself or in its LMIU. In this case, there is no translation table or logic in the DMA lane controllers. In common with the preceding approach, there is a PE ID component of the DMA bus which is driven by the transfer controllers and used by the LMIUs to compare for a match with the locally visible PE VID. When a match is detected in a PE, then it accepts the access which may be either a write or a read request. Means for updating the VIDs stored locally in the LMIUs may be provided through the use of registers visible in the PE register address space, or through a PE instruction which broadcasts the table to all PEs, who then select their VID using their hard-coded PID stored locally. This approach has advantages when VIDs are used for other purposes than just data distribution and collection by a DMA controller.
CTU Addressing Modes
A CTU 408 shown in
Flexible PE Addressing Modes through Parameterizable Logical Loops
Many algorithms which are distributed across multiple PEs require complex data access patterns to achieve peak efficiency. The basis for our loop-based PE addressing modes is a logical view of data access consisting of a set of nested loops in which one component of the PE memory address is assigned to be updated at the end of each loop. As stated above, a PE memory address consists of three components called “address components”, a PE virtual ID (VID), a base value (Base) and an index value (Index). This model requires the following: a mechanism for assigning address components to logical loops; a mechanism for initializing address components; and a mechanism for updating address components; and a mechanism for indicating a loop's exit condition.
Assignment of an address component to a loop specifies the order in which the three address components are updated. In an embodiment which uses a three-loop model, there are six possible orders for updating address components (i.e. six ways to re-order VID, Base and Index). The base and index components are defined to be ordered in this embodiment so that the index is always updated prior to the base, which reduces the number of possible orderings to three, since base and index are summed to form an offset into PE memory, allowing loop assignments that update the base before the index is redundant. An exemplary loop assignment is: update VID on inner loop; update index on middle loop; and update base on outer loop.
Thus, as PE addresses are generated, the VID component updates first (inner loop). When all VIDs have been used (VID loop exit condition has been reached), then the VID is reinitialized, the index is updated, and the VID loop is reentered. This looping continues until the number of index updates is exhausted (Index loop exit condition has been reached) at which point the index is reinitialized, the base is updated, the index loop is reentered, then the VID loop is reentered. This further looping continues until the transfer count is exhausted.
Updating an address component is performed by selecting a new value for the component either based on the old value (e.g. new=old+1) or by some other means, such as by table lookup. A loop exit condition specifies what causes the loop to exit to the next-most outer loop in the model.
In summary, three different aspects of loop control are used to vary the sequence in which PE memories may be accessed. These are:
The following aspects of the loop formulation are noted. When the requested number of accesses are made (TC in
The functions used to update an address (see UpdateAddress( ) in
The function used to update the loop control variable, UpdateLoopControl( ), may be performed as part of the address update or as a separate operation as shown in
The function used to check for loop termination simply tests the loop termination variable for an end of loop condition. This condition may be a particular count value or the state of a mask register.
The initialization of address parameters (see Initialize( ) function:
The following discussion addresses instruction formats and describes PE addressing modes for one embodiment of the invention. It will be recognized other instruction encodings may be used consistent with the teachings of the present invention. In the preferred embodiment, a transfer controller reads transfer instructions from a local memory and decodes them. Transfer instructions come in two types, those for the STU and those for the CTU. The STU transfer instructions specify the addressing mode and transfer count for accesses to the system data bus while CTU transfer instructions specify the addressing mode and transfer count for accesses to the DMA bus and all SP and PE memories. The instruction formats addressed below are only those instructions which control special PE memory addressing for the CTU. Instruction mnemonics are used to indicate the instruction type and addressing mode. “TCI” stands for “transfer, core-inbound”, while “TCO” stands for “transfer, core-outbound”. “TCx” stands for either TCI or TCO. The following PE addressing modes are described as illustrative of the present invention: PE Block-Cyclic, PE Select-Index, PE Select-PE, and PE Select-Index-PE.
PE Block-Cyclic Addressing
PE blockcyclic addressing provides the basic framework for all of the PE addressing modes. A Loop parameter specifies the assignment of address components to loops: BIP, BPI, or PBI.
The operation of the PE select-index address mode is similar to the PE blockcyclic address mode except that rather than updating the index component of the address by adding a constant to it, the instruction specifies a table of index update values which are used sequentially to update the index.
An index select parameter allows finer-grained control over a sequence of index values to be accessed. In the example, this is done using a table of eight 4-bit index-update (IU) values. Each time the index loop is updated, an IU value is added to the effective address. These update values are accessed from the table sequentially starting from IU0 for IUCount updates. After IUCount updates, the index update loop is complete and the next outer loop (B or P) is activated. On the next entry of the index loop, IU values are accessed starting at the beginning of the table.
PE Select-PE Addressing
The operation of the PE Select-PE address mode is similar to the PE blockcyclic address mode except that rather than updating the PE VID component of the address by adding 1 to it, the instruction specifies a table of bit vectors, where each bit vector specifies the PE's to select for access. A bit set to “1” in a bit vector indicates, by its bit position, the VID of the PE to access. Bits in each bit vector are scanned from right to left (least to most significant when viewed in a first instruction format such as instruction format 1900 of FIG. 19). When there are no more “1” bits in a vector, the PE loop exits. The next iteration of the loop uses the next bit vector in the table.
The PE select fields together with the use of the PE translate table allow out of order access to PEs across multiple passes through them.
PE Select-Index-PE Addressing
This addressing mode combines both select-index and select-PE addressing. An exemplary instruction format 2100 is shown in FIG. 21. This form of addressing provides for complex-periodic data access patterns. An exemplary access pattern table 2200 for the PE-select-index-PE address mode is shown in FIG. 22.