GB2393288A - method of generating an interleave pattern for work loads in an array of processing elements. - Google Patents

Abstract

A method of determining an interleave pattern for n lots of A and y lots of B, when n plus y equals a power of two such that the expression 2<z>-n may be used to represent the value of y, is comprised of generating a key comprised of the reverse bit order of a serially indexed count from 0 to 2<z>. An interleave pattern can be generated from the key in which all values less than n are replace by A and all other values are replaced by B. In cases where n plus y does not equal a power of two, the method is comprised of selecting a value of 2<z> where, preferably, (n + y) < 2<Z> < 2(N + Y). A list is created in which the entries are comprised of the reverse bit order of a serially indexed count from 0 to 2<z>. A portion of the list is selected and renumbered to form a key. An interleave pattern can be generated from the key in which all values in the key less than n are replaced by A and all other values in the key are replaced by B. In both cases, the keys can be used to generate a table that contains all possible combinations of values of A and B. The table can then be stored such that an interleave pattern can be automatically selected based on either the number of lots of A or the number of lots of B. The pattern may be used in load balancing for an array of processing elements.

Description

( METHOD OF OBTAINING INTERLEAVE INTERVAL

FOR TWO DATA VALUES

BACKGROUND OF TIJE INVENTION

[00021 The present invention is directed generally to interleaving and, more particularly, to the interleaving of two data values.

100031 Conventional central processing units ("CPU's"), such as those found in most personal computers, execute a single program (or instruction stream) and operate on a single stream of data. For example, the CPU fetches its program and data from a random access memory ("RAM"), manipulates the data in accordance with the program instructions, and writes the results back sequentially. There is a single stream of instructions and a single stream of data (note: a single operation may operate on more than one data item, as in X = Y + Z. however, only a single stream of results is produced). Although the CPU may determine the sequence of instructions executed in the program itself, only one operation can be completed at a time. [because conventional CPIJs execute a single program (or instruction - 1

( stream) and operate on a single stream of data, conventional CPUs may be referred to as a single-instruction, single data CPU or an SISD CPU.

100041 The speed of conventional CPUs has dramatically increased in recent years.

Additionally, the use of cache memories enables conventional CPUs faster access to the desired instruction and data streams. However because conventional CPUs can complete only one operation at a time, conventional CPUs are not suitable for extremely demanding applications having large data sets (such as moving image processing, high quality speech recognition, and analytical modeling applications, among others).

100051 Improved performance over conventional SISD CPUs may be achieved by building systems which exhibit parallel processing capability. Typically, parallel processing systems use multiple processing units or processing elements to simultaneously perform one or more tasks on one or more data streams. For example in one class of parallel processing system, the results of an operation from a first CPU are passed to a second CPU for additional processing, and from the second CPU to another CPU, and so on. Such a system, commonly known as a "pipeline", is referred to as a multiple-instruction, single-data or MISD system because each CI,U receives a different instruction stream while operating on a single data stream. Improved performance may also be obtained by using a system which contains many autonomous processors, each running its own program (even if the program running on the processors is the same code) and producing multiple data streams. Systems in this class are rclerred to as a multiple-instruction, multiple-data or MIMD system.

100061 Additionally, improved performance may be obtained using a system which has multiple identical processing units each performing the same operations at once on different data streams. I he processing units may be under the control of a single sequencer running a single program. Systems in this class arc referred to as a single-instruction, multiple data or SIMD system. When the number of processing units in this type of system is very large (e.g., hundreds or thousands), the system may be referred to as a massively parallel SIMD system.

100071 Nearly all computer systems now exhibit some aspect of one or more of these types of parallelism. For cxamplc, MMX extensions are SIMD; multiple processors (graphics processors, etc) are MIMI); pipelining (especially in graphics accelerators) is MISI).

Furthermore, techniques such as out of order execution and multiple execution units have been used to introduce parallelism within conventional CPUs as well.

IOOOXI Parallel processing is also used in active memory applications. An active memory refers to a memory device having a processing resource distributed throughout the memory structure. The processing resource is most often partitioned into many similar

( processing elements (PEs) and is typically a highly parallel computer system. By distributing the processing resource throughout the memory system, an active memory is able to exploit the very high data bandwidths available inside a memory system. Another advantage of active memory is that data can be processed "on-chip" without the need to transmit the data across a system bus to the CPU or other system resource. Thus, the work load of the CPU may be reduced to operating system tasks, such as scheduling processes and allocating system resources. 100091 A typical active memory includes a number of interconnected PEs which are capable of simultaneously executing instructions sent from a central sequencer or control unit.

The PEs may be connected in a variety of different arrangements depending on the design requirements for the active memory. For example, PEs may be arranged in hypercubes, butterfly networks, one-dimensional stringsAoops, and two-dimensional meshes, among others. 100101 In typical active memories, load balancing (i.e. having each PE perform the same number of tasks) is important to maximize the effectiveness of the active memory. By balancing the load, the amount of time that one or more PEs is idle while waiting for one or more other PEs to complete their assigned tasks is minimized. For load balancing, there is a requirement to distribute T tasks across P PEs or, more generally, P processors. This would give a mean number of tasks M = T/P on each processor. However, in general T/P is not an integer. To preserve the number of tasks T. some processors will be assigned A tasks and some processors will be assigned B tasks, where A = truncated (T/P), and B = A + 1, and A and n are integers. If the number of processors with A tasks is X, then r = A.X + B.(P - X).

100111 In certain circumstances, it may be desirable to interleave the processors with A tasks and with B tasks. l bus, a need exists for a method of evaluating an interleave pattern for n occurrences of A and y occurrences of B. BRIEr SUMMARY OF THE INVENTION

100121 The present invention is directed to methods of determining an interleave pattern for n lots of A and y lots of B. when n plus y equals a power of two and when it does not.

When n plus y equals a power of two, the expression 2Z-n may be used to represent the value of y. The method is comprised of generating a key comprised of the reverse bit order of a serially indexed count from O to 2Z An interleave pattern can be generated from the key in which all values less than n are replaced by A and all other values are replaced by B.

( 100131 In cases where n plus y does not equal a power of two, the method is comprised of selecting a value of 2Z where, preferably, (n + y) < 2Z < 2(n + y). A list is created in which the entr. ies are comprised of the reverse bit order of a serially indexed count from 0 to 2Z. A portion of the list is selected and renumbered to fomm a key. An interleave pattern can be generated from the key in which all values in the key less than n are replaced by A and all other values in the key are replaced by B. 100141 In both cases, the keys can be used to generate a table that contains all possible combinations of values of A and B. The table can then be stored such that an interleave pattern can be automatically selected based on either the number of lots of A or the number of lots of B. The method of the present invention, although simple to implement, provides a near optimal interleave pattern for most cases. The present invention contemplates hardware, e.g., memory containing an ordered set of instructions, for carrying out the disclosed methods.

Those advantages and benefits, and others, will be apparent from the Detailed Description of

the Invention herein below.

BRIEF DESCRIPTION OF THE DRAWINGS

100151 For the present invention to be easily understood and readily practiced, the present invention will now be described, for purposes of illustration and not limitation, in conjunction with the following figures, wherein: 100161 FIG. I is a block diagram illustrating an active memory according to an embodiment of the present invention; 100171 FIG. 2 is a block diagram of a processing element for the active memory illustrated in FIG. I according to an embodiment of the present invention; 100181 FIG. 3 illustrates an array of the processing elements illustrated in FIG. 2 arranged in a line according to an embodiment of the present invention; 100191 FIGs. 4A - 4E illustrate the creation of a key in which the entries of FIG. 4E are the reverse bit order of a serially indexed count from 0 to 2Z of FIG. 4A; 100201 FIG. 5 is an example of the interleave pattern produced by the key of FIG. 4E for the example A = 4, B = 12; 100211 FlGs. 6A - 6D illustrate an example of how to construct a key when the n occurrences of A plus the y occurrences of B does not equal a power of two; 100221 FIG. 7 illustrates the resulting interleave pattern produced by the key of FIG. 6D for the example A = 4, B = 8; and - 4

100231 FIG. 8 is a table constructed using the key of FIG 6D illustrating the interleave patterns for all values of B from 0 to 11.

DETAILED DESCRIPTION OF THE INVENTION

100241 As discussed above, parallel processing systems may be placed within one or more classifications (e.g., MISD, MIMD, SIMD, etc.). For simplicity, the present invention is discussed in the context of a STMD parallel processing system. More specifically, the present invention is discussed in the context of a SIMD active memory. It should be noted that such discussion is for clarity only and is not intruded to the limit the scope of the present invention in any way. The present invention may be used for other types and classifications of parallel processing systems.

100251 FIG. I is a block diagram illustrating an active memory 10 according to an embodiment of the present invention. It should be noted that the active memory 10 is only one example of a device on which the methods of the present invention may be practiced and those of ordinary skill in the art will recognize that the block diagram of FIG. 1 is an overview of an active memory device 10 with a number of components known in the art being omitted for purposes of clarity.

[00261 Active memory 10 is intended to be one component in a computer system.

Processing within active memory 10 is initiated when the active memory 10 receives commands from a host processor (not shown), such as the computer system's CPU. A complete processing operation (i.e., data movement and processing) in the active memory 10 may consist of a sequence of many commands from the host to the active memory device 10.

100271 Active memory 10 is comprised of a host memory interface ("IMP") 12, a bus interface 14, a clock generator 16, a task dispatch unit ("TDU") 18, a DRAM control unit ("DCU") 20, a DRAM module 22, a programmable SRAM 24, an array control sequencer 26, and a processing element array 28, among others.

1002f] The IIMI 12 provides an input/output channel between the host (such as a CPU, not shown) and the DRAM module 22. In the current embodiment, the lIM1 12 receives command (cmd), address (addr), and data signals (among others) from and sends data and ready (ray) signals (among others) to the host. The MIMI 12 approximates the operation of a standard non-active memory so that the host, without modifications, is compatible with the active memory 10.

- 5

l0029l The HMI 12 may be similar in its operation to the interface of a synchronous DRAM as is know in the art. Accordingly, the host must first activate a page of data to access data within a DRAM module 22. In the current embodiment, each page may contain 1024 bytes of data and there may be 16,384 pages in all. Once a page has been activated, it can be written and read through the HMI 12. The data in the DRAM module 22 may be updated when the page is deactivated. The HMI 12 also sends control signals (among others) to the DCU 20 and to the processing element array 28 via the task dispatch unit 18.

100301 The HMI 12 may operate at a frequency different than that of the frequency of the master clock. For example, a 2x internal clock signal from clock generator 16 may be used.

Unlike a traditional DRAM, the access time for the HMI 12 uses a variable number of cycles to complete an internal operation, such as an activate or deactivate. Thus the ready signal (ray) is provided to allow the host to detect when a specific command has been completed.

100311 The bus interface 14 provides and input/output channel between the host and the TDU 18. For example, the bus interface 14 receives column select (cs), write command (w), read command (r), address (addr), and data signals (among others) from and places interrupt (intr), flag, and data signals (among others) onto the system bus (not shown). The bus interface 14 also receives signals from and sends signals to TDU 18.

I0032l The clock generator 16 is operable to receive an external master clock signal (xl) and operable to provide the master clock signal (xl) and one or more internal clock signals (x2, x4, x8) to the components of the active memory. It should be apparent to one skilled in the art that other internal clock signals may be produced by the clock generator 16.

100331 The TDU 18 communicates with the bus interface 14, the HMI 12, the programmable SRAM 24, the array control sequencer 26, and the D('U 20. In the current embodiment, the 'I'DU 18 functions as an interface to allow the host to issue a sequence of commands to the array control sequencer 26 and the DCU 20. Task commands from the host may be buffered in the TDU's FIFO buffers to allow a burst command to be issued.

Commands may contain information on how the tasks in the array control sequencer 26 and the DCU 20 should be synchronized with one another, among others.

100341 The DCU 20 arbitrates between the 'I'DlJ 18 and the l f M1 12 and sends commands to the DRAM modules 22 and the processing element array 28. The DCU 20 also schedules refreshes within the DRAM modules 22. In one embodiment, the DRAM modules 22 of the active memory 10 may be comprised of sixteen 64k x 128 eDRAM (or embedded DRAMf) cores. Each eDI/f core may be connected to an array of sixteen Plus, thus providing 256 (16 x 16) PEs in all.

- 6

100351 The progrararnable SRAM 24 functions as a program memory by storing commands issued by the TDU 18. For example, the TDU 18 may transmit a "write program memory address" command which sets up a start address for a write operation and a "write program memory data" command which writes a memory location and increments the program memory write address, among others. The programmable SRAM 24, in the current embodiment, has both an address register and a data output register.

100361 The array control sequencer 26 may be comprised of a simple 16 bit minimal instruction set computer (16-MISC). The array control sequencer 26 communicates with the TDU 18, the programmable SRAM 24, and the DCU 20, and is operable to generate register file addresses for the processing element array 28 and operable to sequence the array commands, among others.

100371 The processing element array 28 is comprised of a multitude of processing elements ("PEs") 30 (see FIG. 2) connected in a variety of different arrangements depending on the design requirements for the processing system. For example, processing units may be arranged in hypercubes, butterfly networks, one-dimensional strings/loops, and two dimensional meshes, among others. In one embodiment, the processing elements 30 are arranged in a loop (for example, see FIG. 3). The processing element array 28 communicates with the DRAM module 22 and executes commands received from the programmable SRAM 24, the array control sequencer 26, the DCU 20, and the IlM1 12. Each PE in the processing element array 28 includes dedicated 11-registers for communication with the HMI 12. Control of the H-registers is shared by the MIMI 12 and the DCU 20.

10038] Referring now to FIG. 2, a block diagram of a PE 30 according to one embodiment of the present invention is illustrated. I'E 30 includes an arithmetic logic unit ("ALU") 32, Q-registers 34, M-registers 36, a shift control and condition register 38 (also called "condition logic" 38) , a result register pipeline 40, and register file 42. The PE 30 may also contain other components such as multiplexcrs 46 and logic gates (not shown), among others. 100391 In the current embodiment, the Q-rcgisters 34 arc operable to merge data into a floating point format and the MRegisters 36 are operable to de-merge data from a floating point format into a single magnitude plus an exponent format. 'I he ALU 32 is a multiplier adder operable (among others) to receive information from the Q-registers 34 and M-registers 36, execute tasks assigned by the TDU 18 (see FIG. 1), and transmit results to the shift control and condition logic 38 and to the result register pipeline 40. The result register pipeline 40 is operable to communicate with the register file 42, which holds data for transfer

into or out of the DRAM modules 22 via a DRAM interface 44. Data is transferred between the PE and the DRAM module 22 via a pair a registers, one register being responsive to the DCU 20 and the other register being responsive to the PE 30. The DRAM interface receives command information from the DCU 20. The DRAM interface 44 also permits the PE 30 to communicate with the host through the host memory access port 46.

100401 In the current embodiment, the H-registers 42 are comprised of synchronous SRAM and each processing element within the processing element array 28 contains eight H registers 42 so that two pages can be stored from different DRAM locations, thus allowing the interleaving of short do bursts to be more efficient. Result register pipeline 40 is also connected to one or more neighborhood connection registers ("X-register") (not shown). The X-register links one PE 30 to its neighboring PE's 30 in the processing element array 28. The reader desiring more information about the hardware shown in FIGs. I and 2 is directed to UK Patent application (serial no. not yet assigned) entitled "Control of Processing Elements in Parallel Processors" filed 17 September 2002, (Micron no. 02- 1604) which is hereby incorporated by reference.

100411 FIG. 3 is a simplified diagram showing the interconnections of an array of PEs 30 (as illustrated in FIG. 2) arranged in a line 50 according to an embodiment of the present invention. In the current embodiment, line 50 is comprised of eight (8) PEs (i.e., PEr, where r = 0, 1, 2 7) which are interconnected via their associated X-register links. It should be noted that the number of PEs 30 included in line 50 may be altered while remaining within the scope of the present invention. It should further be noted that although the current embodiment is discussed with respect to a single line of PEs, the present invention is applicable to other arrangements as well. I;or example, the present invention may be employed for PEs arranged in NxN e-dimensional arrays, one-dimensional loop arrays, hypercubes, butterfly networks, two-dimensional meshes, etc. while remaining with the scope of the present invention. In the current embodiment, each PE 30 in line 50 is operable to receive instructions t'rom TDU 18 as discussed in conjunction with FI(,. l.

l0042l As illustrated in FIG. 3, each 1'1., has a local number of tasks (v,) associated therewith. For example, PEo, PET, PE2, PE7 each have local number of tasks v0 = 2, v/ = 2, v2 = 2 V7 = 2, respectively, associated therewith. Because each of l'E, through PE7 have the same number of local tasks, the line 50 is said to be balanced.

100431 PEP through PE6 arc operable to communicate with both their let's and right neighboring PEs. I or cxamplc, PEP can communicate with PESO (i. e., PE''s left neighbor) and with I'E2 (i.e., I'l''s right neighbor). In the current embodiment, the line's 50 left end PE - 8

(i.e., PEo) is operable to communicate with its right neighbor (i.e., PE') , whereas the line's 50 right end PE (i.e., PE7) is operable to communicate with its left neighbor (i.e., PE6). It should be noted, however, that each PE on the end of line 50 (i.e., PEo and PE7) may also be operable to communicate with a PE from another line (e.g., to link two or more lines in an array) or to communicate with each other (e.g., to permit a wrap function).

100441 It should be noted that "line" refers to at least two serially connected PEs and it is intended to include PE's arranged in a linear array (e.g., rows, columns, diagonals, etc.) and other non-linear shapes. Serially connected arrays having unifomm and/or varied distances between one or more of the PEs are within the scope of the present invention. l [00451 FIGs. 4A through 4E illustrate the creation of a key in accordance with the method of the present invention. FIG. 4A illustrates a serially indexed count from O to 2Z.

FlGs. 4B through 4E illustrate a process of numbering positions so that adjacent numbers in FIG. 4A are as far apart as possible. Thus, in FIG. 4B, the numbers O and I are positioned at the beginning and mid point of the line, respectively. In FIG. 4C, the numbers 2 and 3 are positioned at the one quarter and three quarter locations, respectively. In FIG. 4D, the numbers 4 and 5 are positioned at the one-eighth and five-eighth positions, respectively, while the numbers 6 and 7 are positioned at the four-eighths and seven-eighths positions of the line, respectively. FIG. 4E illustrates the completed process with the line renumbered so that adjacent numbers are as far apart as possible. It can be seen from an examination of FIG. 4E, that if the numbers of FIG. 4A were illustrated in binary form, then the binary fomm of the numbers illustrated in 4E would be in the reverse bit order of their counterparts in FIG. 4A.

For example, the number one appearing in position two of FIG 4A has the binary form 0001 while the number 8 appearing in position two of FIG. 4E has the binary form 1000.

100461 The renumbered line of FIG. 4E may be thought of as a key comprised of the reverse bit order of a serially indexed count from O to 2Z An interleave pattern can be generated from the key by substituting A for all values less than n and B for all other values.

For example, if n is equal to four lots of A, then A is substituted in the key of l;lG. 4E for the values 0, 1, 2, and 3. B is substituted for all other values resulting in the interleave pattern illustrated in FIG. 5. It can be seen from FIG. 5 that the interleave pattcm is comprised of four sub-patterns each of which equals A, B. B. B. That sub-pattern approximates the mean value (3B +A)/4.

100471 If the sum of n lots of A plus y lots of B is not a power of two, then the next largest power of two is chosen for the initial set of positions. I lCi. 6A illustrates a line having twelve positions, which is not a power of two. The next highest power of two is 16, which is

( shown in FIG. 6B. The line of FIG. 6B is shown centered with respect to the line of FIG. 6A.

FIG. 6C shows a list of entries, each of which is the reverse bit order of its counterpart in the line of FIG. 6B. Because only twelve values are needed, FIG. 6D corresponds to a selected portion of FIG. 6C. That is, the line of FIG. 6D is only twelve positions long and it corresponds with the portion of the line of FIG. 6C centered on the line of FIG. 6A. More particularly, the line of FIG. 6D can be arrived at by alternately dropping entries from each end of the line of FIG. 6C. Thereafter, the positions in FIG. 6D are renumbered in ascending order. Thus, the position labeled I in FIG. 6C is labeled 0 in FIG. 6D, the position labeled 2 in FIG. 6C is labeled I in FIG. 6D, the position labeled 3 in FIG. 6C is labeled 2 in FIG. 6D, etc. to produce the key shown in FIG. 6D.

100481 FIG 7 illustrates the interleave pattern produced using the key of FIG. 6D for the example where n is equal to four lots of A and y is equal to eight lots of B. 100491 The keys produced according to the method of the present invention may be used to generate a table. For example, the key shown in FIG. 6D may be used to generate the table shown in FIG. 8 which illustrated the interleave pattern for all combinations of A and B. indexed from B equals 0 to B equal to 12. The highlighted row corresponds to the interleave pattern illustrated in FIG. 7. The table shown in FIG. 8 may be stored such that as an application is presented with different values for n lots of A and y lots of B. the table can be used to automatically select the appropriate interleave pattern. One of the benefits of the present invention is that an optimal, or near optimal, interleave pattern can be produced for most cases utilizing a very simple lookup mechanism, in cases where a table has been created and stored, or using a key to appropriately order the values of A and B. 100501 This invention is particularly useful in combination with load balancing methods.

For load balancing, there is a requirement to distribute T tasks across P PEs or, more generally, P processors. This would give a mean number of tasks M = T/P on each processor.

However, in general T/P is not an integer. To preserve the number of tasks T. some processors will be assigned A tasks and some processors will be assigned B tasks, where A = truncated ( I /P), and 13 = A + 1, and A and B are integers. If the number of processors with A tasks is X, then T = A.X + B.(P - X).

100511 While the present invention has been described in connection with preferred embodiments thereof, those of ordinary skill in the art will recognize that many modifications and variations are possible. The present invention is intended to be limited only by the following claims and not by the foregoing description which is intended to set forth the

presently preferred embodiment.

- 10

Claims (15)

What is claimed is:

1. A method of generating an interleave pattern for n lots of A and (2Z n) lots of B. comprising: creating a key comprised of the reverse bit order of a serially indexed count from O to 2Z; and generating an interleave pattern corresponding to said key in which all values in the key less than n are replaced by A and all other values in the key are replaced by B.

2. A method of generating an interleave pattern for n lots of A and y lots of B. where n plus y does not equal a power of two, comprising: creating a list in which the entries are comprised of the reverse bit order of a serially indexed count from O to 2Z; selecting a portion of the list; renumbering the selected portion of the list to form a key; and generating an interleave pattern corresponding to said key in which all values in the key less than n are replaced by A and all other values in the key are replaced by B.

3. The method of claim 2 wherein said selecting includes selecting a centered portion.

4. The method of claim 2 wherein said selecting includes dropping entries alternately from each side of the list.

5. The method of claim 2 wherein said renumbering includes renumbering in order of ascending value.

6. A method, comprising: creating a key comprised of the reverse bit order of a serially indexed count from O to 2Z; creating a table of interleave patterns for all values of n lots of A and (2Z n) lots of B based on said key; and storing said table.

7. The method of claim 6 additionally comprising automatically selecting an interleave pattern from said table based on one of the values n and (2Z n).

8. The method of claim 7 additionally comprising generating an interleave pattern based on said selecting.

9. A method, comprising: selecting a value of 2Z which is greater than the value of n lots of A plus y lots of T3, but less than twice that value; 11

creating a list in which the entries are comprised of the reverse bit order of a serially indexed count from O to 2Z; selecting a portion of the list; renumbering the selected portion of the list to form a key; creating a table of interleave patterns for all values of n lots of A and y lots of B based on said key; and storing said table.

10. The method of claim 9 wherein said selecting includes selecting a centered portion.

11. The method of claim 9 wherein said selecting includes dropping entries alternately from each side of the list.

12. The method of claim 9 wherein said renumbering includes renumbering in order of ascending value.

13. The method of claim 9 additionally comprising automatically selecting an interleave pattern from said table based on one of the values n and y.

14. The method of claim 13 additionally comprising generating an interleave pattern based on said selecting.

15. A memory device carrying a set of instructions which, when executed, perform a method comprising: creating a key comprised of the reverse bit order of a serially indexed count from O to 2Z; and generating an interleave pattern corresponding to said key in which all values in the key less than n are replaced by A and all other values in the key are replaced by 13 to generate an interleave pattern for n lots of A and (2Z n) lots of B. - 12