JP3075543B2 - Scanning operation device in which propagation units for executing a scanning operation are arranged in a tree configuration - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a large-scale data array a.
_{1, a 2, a 3 ...} a 1 for _{a M, (a 1 Oa 2} ),
_{(A 1 Oa 2 Oa 3)} ..., a 1 Oa 2 Oa 3 O ... O
a _M , where O is called associative operator, +, V, Λ, M
A scanning operation (associative operation, such as associative operation,
The present invention relates to a scanning operation device in which propagation units for executing a prefix operation and a propagation operation are arranged in a tree configuration.

[0002]

2. Description of the Related Art A scanning operation is one of basic parallel operations. By increasing the speed of the scanning operation, the processing time of various processes in the parallel data processing device is reduced.

Conventionally, as a method for speeding up a scanning operation, reference [G. E. FIG. Bellaloc, "Scans as
Primitive Parallel Operat
i-ons "Proc.Int.Conf.Paral
ll Process. , Aug. 1987, pp.
355-362], a method has been known in which the processing is performed using an array of arithmetic units having a hierarchically connected hierarchical tree configuration.

As shown in FIGS. 29 and 30, one (or two)
) Arithmetic and logic unit ALU and a binary tree-like combination of an arithmetic unit 1 comprising an m-bit memory, 1) performing a predetermined associative operation between two inputs from the lower layer, Outputting the result to the upper layer and storing the input from the lower left arithmetic unit in its own memory (represented by the numerical value in the box indicating the arithmetic unit in the figure) in order from the lower layer to the upper layer (Up-sweep) (FIG. 29) and 2) a process of sequentially obtaining the offset of the lower right arithmetic unit 1 from the data held in the memory obtained by the up-sweep toward the lower layer (down-sweep) (FIG. 30) ), The scanning operation result is output from the lowermost operation unit 1. FIG.
As is clear from FIG. 30, the hierarchy of the arithmetic unit 1 is l
Since it is composed of og ₂ M stages, if one operation is required for each layer, scanning operation of the entire array can be executed in ₂ log ₂ M steps required for both up sweep and down sweep. Here, the offset of the operation unit 1 is a scan operation result immediately before the scan operation area covered by the operation unit 1. Therefore, by combining the offset and the local scan operation result of the operation unit 1, the scan operation result at that point can be obtained.

In this method, a dedicated ALU is provided for each of the up-sweep and the down-sweep, and a register is provided between hierarchies to pipeline the processing so that the required number of steps can be reduced to one equivalently. . However, in a general process in which the scanning operation is not repeated, the latency period of the pipeline is ₂ log ₂ M effectively.
It will take minutes of steps. This means that even if the required number of steps is extremely smaller than the incubation period, the contribution to speeding up is small.

In order to solve this problem, Bellaloc uses bit pipe lining (scanning operation is performed sequentially from the least significant digit of the scanned array data in bit units), although the required number of steps is increased to the word length of the scanned array data. This method is also proposed in this document 1. However, even with this method, 1) when handling array data having a short word length such as an image, the latency period cannot be ignored. 2) Although the bit width of the ALU is reduced to 1 bit, the memory capacity of each arithmetic unit for storing the associative operation result cannot be reduced, and the hardware scale of the arithmetic unit is not sufficiently reduced. is there.

On the other hand, as a method for avoiding the problem of the pipeline processing, the scanning operation is executed with the number of delay stages of the order of N (log _NM ) (the delay time per stage can be reduced to several ns or less). A scan operation device having a tree structure suitable for M LSI is executed without using a pipeline, in which M performs several hundred scan operations in several hundred ns or less, that is, several clock cycles or less. Arithmetic mechanism, IEICE Transactions C-II Vol.
C-II No. 5 pp. 388-397 May 1991
The month has been disclosed. ).

FIG. 31 shows a scanning operation device in which a propagation unit for executing this type of scanning operation described in the paper of the present applicant has a tree structure, and a processing element array (P).
E) shows a parallel data processing device configured in combination with E).

This device comprises a plurality of propagation units (PO
After the U) 10a are hierarchically connected in a tree shape, the lowermost-layer propagation unit arrangement is connected to the processing element (PE) arrangement 60, respectively.

As shown in FIG. 32, the propagation unit 10a comprises a cascade connection of N-1 propagation elements 30a, and a local area (N of the array to be scanned) is obtained by a propagation operation between the propagation elements 30a. It has an N-ary configuration in which a scanning operation is performed for each of a plurality of cascaded elements (sub-arrays).

The propagation element 30a is configured as shown in FIG. 33, and includes ALUs 32a and 33a for scanning operation, an AND gate 31 for controlling a propagation start point, and a selector (S).
EL) 34, 35, etc. Here, FL _i is propagation start point attribute data indicating that it is a propagation start point when the scanning area is divided, and DI _i is data to be scanned. fd _i is the offset of the scanning operation result to be output to the unit immediately below. The propagation start point attribute data FL _i and the scanned operation data DI _i are stored in a unit or processing element (PE) array immediately below (reference numeral 6 in FIG. 31).
0).

CD _i and fu _i shown in FIG. 33 are the results of the logical sum of the attribute data of the propagation starting point and the scanning operation between the propagation elements 30a.

[0013] The DUI is directly above the propagation unit 10a.
Offset received from

As is apparent from the tree-like coupling configuration of the propagation unit 10a and the configuration of the propagation element 30a,
Each of the lowermost propagation units 10a transmits the scanned data DI in the local area covered by the scanned operation array data.
_{For i} , the scanning operation is performed in a propagating manner in a series of the ALU 32a of the propagation element 30a and the selector 34, and the resulting CD _N and fu _N are transferred to the propagation unit 10
a, the start point attribute data FL _i and the scanned data DI _i
Output as Here, the starting point attribute data FL from the lower layer
_{When i} indicates active (“0” in this configuration) (meaning that the scanned data DI _i is the data of the propagation start point), the selector 34 is controlled to obtain the result from the preceding stage. The propagation is interrupted, and the propagation element is newly set as the propagation start point.

Similarly, in the second and subsequent stages, CD _N , f
u _N is obtained and output as propagation start point attribute data FL _i and scanned data DI _i for the higher-order propagation unit 10a.

As a result, the scanning operation results of the local area are accumulated in the propagation unit 10a of the uppermost layer. On the other hand, if “0” is given as the scanned data DUI for the uppermost layer, the ALU 33a of the propagation element 30a and the selector 3
5, the configuration of the AND gate 31, from fd _i, offset with respect to the propagation unit 10a immediately below is output. Propagation unit 10a immediately below that described above, from the received offset, so calculates and outputs the offset of the further lower layer, the terminals of fd _i propagation units the lowermost, with respect to sequence data immediately before the terminal as an offset A scanning operation result is obtained.

The propagation units 1 other than the uppermost layer on the way
In 0a, if the propagation starting point attribute data FL _i is "0", the later, as the offset from the upper layer is ignored, the scanning operation is to begin a new from the propagation starting point.

[0018] In this configuration, since the scanning operations in a tree-like hierarchy, the number of layers of the required computing unit via is reduced an order (Nlog _N M) degree. Here, M represents the array size of the array data to be scanned.

The number of transit layers can be equivalently further reduced by introducing a selective propagation method into the propagation operation system of the propagation unit 10a (see Japanese Patent Application Laid-Open No. 63-193232 by the present inventor). Data processing device "). In the selective propagation method, before the input from the preceding stage is determined, the propagation operation for all combinations of the signal values of the input is performed in advance, and when the input is determined, the preceding calculation result is selected. This is a method of realizing propagation only. However, in the above-described selective propagation method applied to each propagation unit, even if propagation is limited to 1-bit signals, the number of input combinations is two, and propagation arithmetic systems of the number of combinations are provided in parallel. Must be provided. Therefore, the number of arithmetic units and selectors is required to be at least twice that in the case where selective propagation is not used. Further, the configuration of the propagation element becomes complicated.

As described above, in the conventional tree-type scanning operation device, when the size of the array to be scanned is large, the latency period when passing through a multi-level tree is long even when pipelined, and the degree of parallelism is high. Cannot obtain the scanning operation performance corresponding to the above.

An arithmetic unit,
A memory, a pipeline register, and the like must be incorporated, and the same amount of hardware as the processor unit is required.

In order to realize the function of setting the starting point of scanning at an arbitrary position (necessary for performing the scanning operation in parallel for each local region of the array), it is necessary to add a starting point setting control logic to each propagation unit. is there. And other issues were unresolved.

As a similar technique of the present invention for performing high-speed acceleration between two data, Japanese Patent Laid-Open Publication No.
337 or a principle of speeding up by selecting a preceding operation result, for example, the document “Carry-select”
Adder ", IRE TRANSACTIONS
ON ELECTRONIC COMPUTERS, J
Une, page 340 to 344. The configuration is similar to the scanning operation device of the present invention, but differs in the following points. In other words: 1) To achieve both high speed and reduced hardware scale,
A characteristic characteristic of carry is also used, and a regular tree structure for controlling a two-input selector with an offset as in the present invention has not been conventionally taken.

2) While the present invention is directed to a scanning operation for each bit of a large number of element data constituting an array, a carry select adder basically performs addition between two data. It is targeted.

As for 1), although the symmetry of the tree is slightly lowered, an adder using a tree composed only of selectors has been proposed as in the present invention. However, it is a method that realizes addition only with up sweep,
Although there is a possibility that the speed can be further increased, there are disadvantages such as low tree regularity and an increase in hardware scale in proportion to the square of the word length.

[0026]

SUMMARY OF THE INVENTION The present invention solves the above-mentioned problems, and has a tree structure in which a propagation unit having a high scanning operation performance and executing a scanning operation capable of effectively reducing the scale of hardware is implemented. It is an object of the present invention to provide an arrayed scanning operation device.

[0027]

In order to solve the above-mentioned problems, the present invention is directed to a transmission unit which is divided into groups for each layer and arranged, and the transmission units are provided one by one corresponding to each group. And a propagation unit array having a hierarchically-connected hierarchical structure in a tree structure, wherein each propagation unit in the propagation unit array is connected to at least one cascaded unit.
The ith cascaded propagation element E _i has input signals LA _i-1 and LB _i-1 from adjacent propagation elements E _i-1 as control signals. Receiving the input signal D from one layer below
First and second selecting means SLA and SLB for selecting A _i and DB _i and outputting output signals LA _i and LB _i respectively;
Receiving an input signal U _j from propagation unit layer on one level to be connected to a transmission unit including the i-th propagation element E _i as the control signal, the input signal LA from the propagation elements E _i-1 to the adjacent _i-1, LB _i-1 respectively selected, a third selection means SLD for outputting an output signal U _i to the propagation unit layer of the lower layer of the result as an offset
And the last cascade-connected propagation element is configured to output the output signals LA _p and LB _p as the input signals DA _j and DB _j for the propagation unit in the layer one level higher, respectively. and connected to said propagation units belonging to the lowest layer of the propagation unit sequence consists of a sequence of a plurality of arithmetic units, the first l-th arithmetic unit _{0Oa 1 Oa 2 O ... Oa r} , 1Oa 1 Oa 2 O ... O
Arithmetic means for outputting a _r as output signals DA _l and DB _l to the corresponding propagation units belonging to the lowest layer of the propagation unit array, where i is an integer of 2 or more;
l, r is each an integer of 1 or more, O is an arbitrary operator sequence of _{a 1, a 2, a 3} ... a r are aligned in the l-th computing unit of the computing means, the A row of elements of a region shared by the scanning array.

[0028]

According to the present invention, each propagation unit of a propagation unit array which is connected in a regular tree and performs a scanning operation sequentially and hierarchically has cascade-connected propagation elements consisting of three 2-input selectors. The signal input from the lower layer to the propagation element is sequentially and progressively operated from the first propagation element to the last propagation element, thereby performing high-speed scanning operation on the array data.

[0029]

Embodiments of the present invention will be described below in detail. FIG. 1 shows the configuration of a SIMD-type parallel data processing device incorporating a tree-structured scanning operation device comprising a tree-type propagation unit array 25 and an interface element array 50 as a first embodiment of the present invention. In this parallel data processing device, a propagation unit array configured by connecting propagation units (POUs) 10 in a tree shape is connected to an interface element (IE) array 50 via each propagation unit 10 in the lowest layer. The IE array 50 is connected to the processing element (PE) array 60, and the IE array 50 and the PE array 60 are connected to the control unit 70. The entire apparatus is controlled by the control unit 70.

FIG. 2 is a diagram showing a detailed configuration of the propagation unit (POU) 10. Here, the propagation element E
_i 30 is an A-side input when the control unit signal input is 0,
In the case of 1, the selector 41A (SL
A), selector 41B (SLB), selector 41D (S
ED).

The propagation element 30 includes a signal U from an upper layer propagation unit and a signal DA _i ,
DB _i is input. The propagation element 30 outputs signals DA _j and DB _j to an upper layer propagation unit and an output signal U _i to a lower layer propagation unit.

FIG. 3 shows the configuration of a basic interface element (IE20) that can be used for scanning operations such as logical sum, logical product, and copy.
It is composed of an LU 32 and a selector 41D. Each of the two ALUs has a DI supplied from an element processor.
The associative operation designated by the control unit 70 is performed between the data and "0" and "1", and the results are output as DA and DB, respectively. Further, the selector 41D selects the ALU output from the offset with respect to the IE, and determines the scanning operation result at this point.

Next, the operation of the tree configuration arithmetic device having the above-described configuration will be specifically described.

First, the arithmetic operation (up sweep) from the lower layer to the upper layer will be described in order from FIG. PE array 60
Thereafter, as in the conventional example, the scanned operation data DI is input to the interface element IE array 50. In response to the input, in each interface element, as shown in FIG. 3, the control unit 70 controls both the case where "0" propagates from the previous stage and the case where "1" propagates.
Performs an associative operation specified by
B is transmitted to the upper layer propagation unit (POU) 10.

Each of the upper propagation units (POU) 10
As shown in FIG. 2, the propagation unit (POU) 10
(In the case where the offset is both 0 and 1, which of DA and DB of the operation result rising from the lower layer is true is sequentially propagated between the propagation elements by selecting the selectors 41A and 41B. And the resulting propagation elements E _N 30 at the left and right ends
Are output to the propagation unit (POU) 10 in the upper layer as operation results DA and DB.

Therefore, the output signals DA and DB of each propagation unit (POU) 10 forming the tree are determined from the lower layer to the upper layer, and each propagation unit (POU) of each layer is determined.
In, the propagation results for both the offsets “0” and “1” are sequentially obtained. When DA and DB are determined up to the top layer of this tree by the above upsweep,
This time, the operation operation (down sweep) from the upper layer to the lower layer becomes meaningful.

The input U _i from above to the propagation unit (POU) 10 in the uppermost layer, that is, the offset with respect to the entire tree, has a logical value (“0” in addition, logical sum, etc.) In a logical product or the like, "1") is added in advance as shown in FIG.
_As can be seen from FIG. 2, _i is one of the two output signals of the preceding propagation element determined by the up-sweep, which is selected by the selector 41D, and the true offset of the propagation unit (POU) 10 in the lower layer. U _i is output to the lower layer. The propagation element E _i 30 of each propagation unit (POU) 10 includes a true offset U _i transmitted from the upper layer and an output signal of the propagation element E _i that has been previously determined by the arithmetic processing when going from the lower layer to the upper layer. , Sequentially, the true offset U _{i of the} lower layer propagation unit (POU) 10
Is calculated and output toward the lower layer.
The final scanning operation result is obtained as the scanned data output fd (FIGS. 3, 4, 5, and 6).

The reason why the offset for the entire tree in FIG. 1 is given in advance as either “0” or “1” in accordance with the content of the operation is that the scan operation is performed in accordance with the present embodiment of the first embodiment. This is because it is assumed that the parallel data processing device incorporating the scan operation device of the present invention is executed independently, and there is no scan operation region immediately before.
On the other hand, the scan operation device of the present invention in an embodiment to be described later can be incorporated as a partial tree in a larger scan operation device (in fact, the third and fourth embodiments of the present invention described below It can be said that this is a scanning operation device in which the first or second embodiment of the present invention is incorporated in a part of a tree, and in this case, the entire scanning operation device has a regular structure as in the first and second embodiments. In such a case, a tree structure may not be obtained.) In such a case, the scan operation area immediately before exists, and the scan operation result must be given as an offset.

FIG. 7 shows a configuration of the propagation unit (POU) 10 in which the redundant part of FIG. 2 is deleted. The difference from the configuration of Figure 2, the propagation elements E _i of the head that was placed for operation principle described is that it is replaced by a mere connection. This connection is, the input signal from the preceding stage with respect to the propagation element E _i is "0", a constant "1", the selector S
Since the selection function of LA41A, SLB41B, and SLD41D is fixed to one side, it is easily derived.

Here, the contents of the scanning operation will be described in more detail in the case where the operator is OR or addition. However, as can be compared with the conventional example shown in FIGS. 29 and 30, the scanning operation device is a binary propagation unit (POU) 1 shown in FIG.
The tree structure shown in FIG. 9 is adopted using 0 '. First, a scan operation of a logical sum will be described. In this operation, since each digit can be executed independently, it is sufficient to show that only one digit of the scanning operation can be executed. Therefore, the output D
As I, a 1-bit data array {0,0,1,0,1,0,0} given from the PE array 61 to the IE array 51
A scan operation of a logical sum from left to right with respect to 1,1} will be described as an example. In this case, the interface element IE shown in FIG. 3 is used as it is, and the function of the ALU is set to a logical sum by the designation of the control unit 70 (omitted in FIG. 9). Based on the above conditions and the assumption that each two-input selector operates to select the A-side input when the control signal is 0 and the B-side input when the control signal is 1, the interface element IE51 and the propagation unit (POU) 10 ′
Are sequentially obtained, the tables shown in FIGS. 10 and 11 are obtained for up sweep and down sweep. By comparing the output DI of the up sweep with the output fd of the down sweep, it can be seen that the scanning operation of the OR is realized. Here, the input U from above to the propagation unit POU at the uppermost stage in the down sweep is an offset with respect to the entire tree in FIG. 9. In this case, the scan other than the area covered by the tree in FIG. Since he is not thinking, he inputs 0.

Next, the addition scanning operation will be described.
In this calculation, it is necessary to perform the calculation sequentially from the lower digit. Considering this point and the fact that it can be compared with the conventional example, as the output DI, the same as the scan target array data of FIG.
A scanning operation of addition from 0, 4, 1, 1, 3} from left to right will be described. In this case, the interface element IE having the configuration shown in FIG. 12 is used to handle a carry required for addition. 3 is replaced with two ALUs 32 'and 32' capable of performing 1-bit addition, and carry outputs CO from the two ALUs are respectively provided.
_A and CO _B are selected by the input U, and the selector SL is used to make it available as a carry input of the next upper digit.
C41C and carry register 45 are added. FIG.
FIG. 20 to FIG. 20 summarize the values of the input and output signals of the up sweep and the down sweep in the case of performing the addition scanning operation in order from the lower digit after clearing the carry register 45 to 0 in advance. Here, fd of the IE output of the down sweep is the scanning operation result of each digit, and
If converted to base number, $ 3,4,6,6,10,11,1
2,15 °. This result is the final scanning operation result, not the offset for each processing element PE as in the conventional example of FIGS. 29 and 30. Therefore,
As in the prior art, it is not necessary to further add DI at each processing element PE to obtain the final result.

In the first embodiment, it is not possible to set the scanning start point in the middle of the array data to be scanned. A second embodiment of the present invention in which the setting of the scanning start point in the middle can be realized will be described below. Since the configuration of the scanning operation device of this embodiment and the parallel data processing device incorporating the same are the same as those of the first embodiment except for the interface element IE20 ', only the configuration of the interface element IE is shown in FIG. The difference from the interface element IE of the first embodiment in FIG.
LU 32b is ALU 32a and SEL 34 shown in FIG.
Has the function of the combination of That is, when the scanning start point attribute data FL input from the processor element indicates the activation, the interface element IE 20 ′ outputs the input DI as it is as DA and DB, and when the scanning start point attribute data FL does not indicate the activation, the interface element IE 20 ′ shown in FIG. Similar to the interface element IE20, the result of performing the associative operation designated by the control unit 70 with "0" and "1" is represented by D
Output as A and DB respectively. As is apparent from the function of this interface element IE20 ', FL
By giving the data indicating the activity as, the function of the scanning start point for outputting the input DI to the interface element IE20 'as DA and DB irrespective of the offset is realized. In addition, tree-shaped propagation unit array 2
As is apparent from the function No. 5, a new scanning operation from this scanning start point to the next scanning start point is performed hierarchically. Here, the case of performing the scan operation regarding the logical sum according to the second embodiment will be described in detail. The scanning operation device has the configuration shown in FIG. The difference from FIG. 9 of the first embodiment is that a quaternary propagation unit (POU) shown in FIG. 22 is used for the first layer in the first layer, and the tree has a two-layer structure. This is to show that the present invention can utilize a propagation unit (POU) of any size. Since the operation content of the interface element IE is a logical sum, the function of the ALU 32b is set to a logical sum, and the IE 20 'of FIG. 4 is used as it is. The array of DI of the scan target data is the first
The same as the one used in the scan operation example of the logical sum of the embodiment,
{0,0,1,0,1,0,1,1} is used. Also,
The array of the attribute data FL of the scanning start point corresponding to each element of the DI array is {1, 0, 0, 0, 0, 1, 0, 0}.
And In this case, the value of FL indicates that 1 is active and 0 indicates inactive. Under the above conditions, each propagation unit (POU) 1 and (POU) in up sweep and down sweep
FIGS. 23 and 24 show the results obtained by summarizing the input / output signal No. 2 and the signal values of the propagation element outputs LA and LB in each propagation unit in a table. As is clear from the table of the down sweep, the scanning result is {0, 0, 1, 1, 1, 0, 1, 1}, which indicates that the setting control of the scanning start point is well realized. Although the description is omitted here, in order to realize the scanning start point setting control function for the addition scanning operation, when the FL is active, the interface element IE22 shown in FIG. _A function of directly outputting DI as outputs DA and DB and CO _A and C
0 it is necessary to add a function of outputting as O _B.

As is clear from the above description, the advantage of the scanning operation device of the present invention is that firstly the propagation element (E) 3
0 is composed of only three 2-input selectors and does not require an arithmetic unit, a memory, a pipeline register, etc. unlike the conventional method. Second, the target array is individually scanned for each partial area. Even if it is desired to execute an operation, the function of the scan start point of the scan area is realized by sending the same operation data to the two propagation operation systems, as in the second embodiment. There is no need to add hardware for setting the scanning start point to the unit array. Accordingly, the hardware scale of the tree-type propagation unit array can be reduced to 1/4 or less as compared with the case where the selective propagation is applied to the conventional high-speed and small-scale hardware system shown in FIG. This is because the two-input selector can be configured with a gate number of 1/4 or less as compared with a 1-bit arithmetic unit having functions of addition and logical operation required in the conventional method.

However, in the entire scanning operation device, it can be considered that the operation unit has just moved into the interface element of the IE array 50 which is not required in the conventional method, and the hardware scale is not reduced. I can see. However, the interface element IE array is 1
Since this is a regular array of dimensions and a high degree of integration is obtained, the effective hardware scale is greatly reduced. Further, as in the first and second embodiments, the scanning operation device is S
When incorporated in an IMD type parallel data processing device, the interface element IE is integrated with the processing element (PE),
Since the function of E can be emulated by the processing element PE and the arrangement of the interface element (IE) can be effectively omitted, the hardware scale can be reduced to almost 1/4 even in the whole scanning operation device.

Further, since the scanning operation device of the present invention is configured on the premise of scanning in bit units, the feasible scanning operation is limited to one that can be divided into scanning operations in bit units. However, important scanning operations frequently used in parallel processing such as addition, logical operation, MIN, and MAX can be decomposed into bit-unit scanning operations, and this restriction does not pose a problem in practical use.

The interface element (IE)
The array 50 has a configuration in which two types of scanning operations are performed on only one element of the array data, but there is also a configuration method in which two types of scanning operations are performed on a plurality of elements. . The interface element shown in FIGS. 5 and 6 is an extension of the interface element IE of the first and second embodiments to this configuration method. With these interface elements IE, DI, FL, f
The set of signals of d enters and exits from one processing element each. This configuration is different from the configurations of FIGS. 3 and 4 in that the scanning operation mechanism is mounted on an LSI.
It can be realized on the same scale in terms of area.

Next, a third embodiment will be described. Except for the first propagation unit and the tree-type propagation unit array, the configuration and operation are the same as those of the first or second embodiment. Therefore, only the first propagation unit 11 and the tree-type propagation unit array 26 are shown in FIG. 25 and FIG. The present embodiment focuses on the fact that the head is always the scanning start point in the first and second embodiments, and simplifies the head propagation unit of each layer. That is, the first unit at the bottom of the tree-type propagation unit array is
Since it always becomes the starting point of scanning, there is no need to carry out propagation calculations for two different offsets in this unit in advance. The inputs DA ₁ and DB _{1 at} the left end of the unit are equal to each other because they are the scanning start points. From these conditions, the head propagation unit in the lowermost layer has a configuration shown in FIG. Each propagation element S ₁ is composed of only one 2-input selector (SL) 41. Also, since there is only one propagation system, there is only one output for the upper layer of this unit. This means that the propagation unit at the head of the second layer from the bottom does not need to perform the propagation operation for the two types of offsets in advance. Accordingly, the head propagation unit having the same configuration as the head propagation unit in the lowermost layer can be used. Similarly, the same unit can be used as the head unit in each of the third and higher layers from the bottom, and the tree-type propagation unit arrangement shown in FIG. 26 is derived. In the configuration of the third embodiment, the hardware scale is reduced because the configuration of the head propagation unit is simplified.

The fourth embodiment is an example of a scanning operation device which is important when incorporated in a parallel processor. In this embodiment, the propagation unit 12, the tree-shaped propagation unit array 2
Except for Example 7, the configuration is the same as that of the first and second embodiments. Therefore, the propagation unit 12, the tree-shaped propagation unit array 2
7 is shown in FIGS. 27 and 28. The first of this embodiment,
The difference from the second embodiment is that the output of the last propagation element of the propagation unit in the uppermost layer is fed back as an offset at the top. This is for looping the scanning operation path. As a result, the entire scanning operation device has a uniform structure without edges. Conversely, an end can be set at an arbitrary position by appropriately setting the scanning start point. However, in this embodiment, in order to make the whole structure uniform, as in the third embodiment, a fixed scanning start point is assumed, and the propagation unit located at the start point of each layer is simplified. Cannot be transformed. Another difference of this embodiment from the first and second embodiments is that the propagation system of the uppermost propagation unit is integrated into one set. This is because the same DA and DB inputs sequentially transmitted from the lowermost layer unit including the scanning start point always exist in the uppermost layer propagation unit.
This is because the two systems A and B in 7 have exactly the same propagation and one can be omitted. The loop structure of the fourth embodiment is very convenient in that the uniformity of functions is not disturbed when the scan operation device of the present invention is incorporated in a parallel data processing device having a structure for looping both ends of a processor array. Good. Also, the point that the propagation unit on the top layer can be simplified
When the scanning operation device is implemented by a plurality of LSIs each having a smaller scanning operation device, the propagation unit in the uppermost layer, which cannot be installed in the LSI, can be easily realized by TTL or PAL. Is effective in

The propagation unit in the uppermost layer may have a structure in which two propagation systems are used as they are, as in the first and second embodiments. However, the hardware scale increases accordingly.

In the above description of the scanning operation, all the functions of each interface element IE have been assumed to be the same. However, the function of the propagation unit (POU) is irrelevant to the calculation function of the interface element IE, and even in a scanning calculation in which the calculation content changes during scanning, the interface element I
This can be dealt with by switching the arithmetic function of the arithmetic logic unit ALU in E.

In addition, the size of the propagation unit (the number of constituent propagation elements) in each layer has been described as being the same.
As is evident from the principle of operation, different sized propagation units can be arranged as required.

[0052]

As described above, in the tree-structured scanning operation device according to the present invention, the tree-shaped propagation unit array at the core portion has a cascade connection of propagation elements consisting of three 2-input selectors without sacrificing high speed. Can be configured by a regular tree-like connection of the propagation units consisting of, so that the number of design and development steps can be significantly reduced as compared with the conventional method. When incorporated into a parallel data processing device having a processor array, the function of the interface element is emulated by each processing element, thereby effectively reducing the hardware scale of the entire scanning operation device to 1/4 or less. it can. Further, when the scanning operation device is configured by a plurality of LSIs, there is an advantage that the propagation unit of the uppermost layer which is difficult to be incorporated in the LSI can be configured by cascade connection of two-input selectors that can be easily realized by TTL or PAL. Therefore, it can be said that the present invention is extremely useful for realizing a parallel data processing device having high scanning operation performance in a compact and economical manner.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing one embodiment of a scanning operation device of the present invention.

FIG. 2 is a diagram showing an internal configuration of a propagation unit 10 shown in FIG.

FIG. 3 shows a basic interface element (IE2
The figure which shows the structure of 0).

FIG. 4 shows an interface element (IE2) shown in FIG.
The figure explaining other operation examples of the interface element (IE20 ') which receives the data FL which shows activity with respect to (0).

FIG. 5 is a diagram showing a configuration of an interface element (IE21) that performs two types of scanning operations on a plurality of elements.

FIG. 6 shows an interface element (IE2) shown in FIG.
The figure explaining other operation examples of the interface element (IE21 ') which receives the data FL which shows activity with respect to 0).

7 is a diagram showing a propagation unit 10 having a configuration in which a redundant portion of the internal configuration of the propagation unit 10 shown in FIG. 2 is deleted.

FIG. 8 is a diagram showing a configuration example of a binary propagation unit 10 'in the scanning operation device of the present invention.

FIG. 9 is a configuration diagram showing one embodiment of a scan operation device of the present invention for creating a binary tree configuration using the propagation unit shown in FIG. 5;

FIG. 10 is a table showing signal values in an up-sweep scan operation of a logical sum by the binary tree scan operation device shown in FIG. 9 and in a scan operation of the logical sum;

11 is a table showing signal values in a down-sweep logical OR scan operation in the logical OR scan operation by the binary tree-type scan arithmetic device shown in FIG. 9;

FIG. 12 is a diagram showing a configuration example of an interface element (IE22) capable of executing an addition scan operation in the binary tree scan operation device of the present invention.

FIG. 13 is a table showing signal values in an up sweep of a first digit (LSD) of a scanning operation of addition performed in order of one digit from the lower digit in the binary tree type scanning operation device of the present invention.

14 is a table showing signal values in a scan operation of addition of a down sweep of the first digit (LSD) of the scan operation in FIG. 13 in the binary tree-type scan operation device of the present invention.

FIG. 15 is a table showing signal values in a second-digit up-scoop of a scanning operation of addition performed in order of one digit from the lower digit in the binary tree-type scanning operation device of the present invention.

16 is a table showing each signal value in the second digit down sweep of the scan operation in FIG. 15 in the binary tree-type scan operation device of the present invention.

FIG. 17 is a table showing signal values in a third digit up-sweep of a scan operation of addition performed in order from the lower order digit by digit in the binary tree-type scan operation device of the present invention.

18 is a table showing signal values in the third digit down sweep of the scan operation in FIG. 17 in the binary tree scan operation device of the present invention.

FIG. 19 is a table showing signal values in a fourth digit up-sweep of an addition scanning operation performed sequentially one digit at a time from the lower digit in the binary tree-type scanning operation device of the present invention.

FIG. 20 is a table showing signal values in the fourth digit down sweep of the scan operation in FIG. 19 in the binary tree scan operation device of the present invention.

FIG. 21 is a configuration diagram showing another embodiment of a tree-shaped scan operation device according to the present invention capable of setting a scan start point at an arbitrary position.

FIG. 22 is a configuration diagram of a quaternary propagation unit (POU) used in the scanning operation device shown in FIG. 21;

23 is a table showing signal values in an up-sweep of a scan operation of a logical sum in the tree-type scan operation device shown in FIG. 21;

24 is a table showing signal values in down sweep in a scan operation of a logical sum in the tree-type scan operation device shown in FIG. 21;

FIG. 25 is a diagram showing another configuration example of the propagation unit (POU) in the scanning operation device of the present invention.

FIG. 26 is a diagram showing a tree-type propagation unit array in which the propagation units (POU) shown in FIG. 25 are arranged as the first propagation unit in each layer.

FIG. 27 is a diagram showing still another configuration example of the propagation unit (POU) in the scanning operation device of the present invention.

FIG. 28 is a diagram showing a tree-shaped propagation unit array in which the propagation units (POU) shown in FIG. 27 are arranged in the uppermost layer.

FIG. 29 is a diagram illustrating an arrangement of propagation units of a conventional tree-structured scanning operation device and illustrating an up-sweep process.

FIG. 30 is a diagram illustrating an array of propagation units in a conventional tree-structured scanning operation device and illustrating a down sweep process.

FIG. 31 is a block diagram showing an example of a conventional parallel data processing device.

32 is a block diagram showing an internal connection of the propagation unit 10a shown in FIG. 31.

FIG. 33 is a block diagram showing an internal configuration of a propagation element 30a shown in FIG. 31;

[Explanation of symbols]

 Reference Signs List 10 Propagation unit 30 Propagation element 41A First selector 41B Second selector 41D Third selector 50 Interface element array

──────────────────────────────────────────────────続 き Continuation of the front page (58) Field surveyed (Int. Cl. ⁷ , DB name) G06F 15/16 610 G06F 15/80 JICST file (JOIS)

Claims (4)

(57) [Claims] A transmission unit arranged in groups for each layer has a tree-like hierarchical connection configuration in which each group is connected to one transmission unit belonging to a layer one level higher. propagation unit has at least one propagation elements are cascade-connected, the cascaded i-th of said propagation element E _i is input from the propagation element E _i-1 adjacent signal LA _i-1 , LB
receiving the _i-1 respectively as a control signal, selects the input signal DA _i, DB _i from the layer lower layer, the output signal LA _i, L
First and second selection means SLA for each output B _i, S
LB and; receives input signals U _j from propagating unit of the i-th one level on the layer to be connected to a transmission unit including a propagation element E _i as a control signal, from the propagation elements E _i-1 to the adjacent Third selection means SLD for selecting the input signals LA _i-1 and LB _i-1 and outputting the result as an output signal U _i to the propagation unit in the next lower layer;
A propagation unit array in which the cascade-connected last propagation element outputs the output signals LA _p and LB _p as input signals DA _j and DB _j for the propagation unit in the layer one level higher, respectively; , connected to said propagation units belonging to the lowest layer of the propagation unit sequence consists of a sequence of a plurality of arithmetic units, the first l-th arithmetic unit _{0Oa 1 Oa 2 O ... Oa r} , 1Oa 1 Oa 2 O ... O
Arithmetic means for outputting a _r as output signals DA _l and DB _l to the corresponding propagation units belonging to the lowest layer of the propagation unit array, where i is an integer of 2 or more;
l, r is each an integer of 1 or more, O is an arbitrary _{operator, a 1, a 2, a} 3 ... a r becomes a sequence of data, the arrangement of the l-th computing unit of the computing means And the arrangement of elements in a region shared by the array to be scanned. 2. The first propagation unit in the same hierarchy in the propagation unit array has at least one cascade-connected propagation element, and the k-th cascade-connected propagation element S _k is an input signal from the adjacent propagation element S _k−1 ,
At the same time, it receives an input signal LL _k-1 which is an input of the propagation unit of the lower layer as a control signal, selects input signals DA _k and DB _k from the lower layer, and outputs an output signal LL _k.
Has a selection means SL for outputting, cascaded final th propagation element outputs an output signal LL _q as an input signal for the propagation unit one level on the layer, where k is 2 or more integer 2. The scanning operation device according to claim 1, wherein: 3. The uppermost layer unit of the tree-like propagation unit array has two or more cascade-connected propagation units, and the k-th cascade-connected propagation element S
_k is an input signal from the adjacent propagation element S _k−1 , and receives an input signal LL _k−1 as a control signal, and
Selecting input signals DA _k , DB _k from the lower layer,
It has a selection means SL for outputting an output signal LL _k, final-th propagation elements connected in cascade, the output signal LL _q
Is fed back as an input signal LL ₀ to a propagation unit located at the head of the same hierarchy. 4. A computing means consisting arrangement of said plurality of arithmetic units, data a _1, a _2, ... data FL _1, FL ₂ corresponding respectively to a _r, all ... FL _r indicates inactive in the case of _{data, 0Oa 1 Oa 2 O ... Oa} r,
1Oa ₁ Oa ₂ O ... Oa _r respectively output signal DA _j, DB
The output to the corresponding propagation units belonging to the lowest layer as _j, if any data FL _1, FL ₂ ... FL _r is the data showing the _{activity, a s Oa s + 1 Oa} s + 2 O
... outputs the Oa _r respectively output signals DA _j, as DB _j, where s denotes the maximum value of the index data FL shown by activity and scanning operation device as set forth in claim 1, wherein claims, wherein .