JPH0311479A - Method and device for detecting completion of stream arithmetic operation - Google Patents

Abstract

PURPOSE:To perform cumulative addition in a data processor which links a memory part to an arithmetic part and controls arithmetic sequence with a data driving system by using the addition instruction of a stream, attaching a termination mark on the last position of the stream, and performing the arithmetic operation of the stream when performing the arithmetic control of a token. CONSTITUTION:A data processor is operated according to the token inputted via a bus token input part 11, and a result is stored in a link table 12. A data part 64 having a data value to be processed, a link table address 63 for token identification, and the termination mark 62 to represent the termination of the stream are provided at the table 12, and plural tokens having the same link table address are outputted as keeping sequence with each other. Also, the token is fetched in a table address to be provided after the arithmetic operation, an address to refer to a stream arithmetic operation completion detector 1, and the table 12 before receiving the arithmetic operation at a processing unit 16.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a data flow processor characterized in that a memory section and an arithmetic section are connected through a pipeline bus, and the order of operations is controlled by a data-driven method. The present invention relates to a method and apparatus for detecting the end of stream computation that performs 1--kun computation control.

[Conventional technology]

Conventionally, as a data flow one-type processor, Japanese Patent Application Laid-open No. 5
There is a technology described in Japanese Patent No. 8-70360, and a commercialized version of this technology is μPD7281, a data flow processor manufactured by NEC Corporation.

This μPD 7281 has the configuration shown in FIG. 6.Tokens, which are the units of data input from the external bus to the data flow processor, contain a data value and a link table address for referring to the link table 92 after input. and a module number indicating the data flow processor on which the token should be processed. The token input unit 91 inputs the token internally when the module number of the token passing through the external bus matches the number of the data flow processor, and otherwise outputs the token to the external bus as it is through the token output unit 97. The input token refers to the link table 92 using the link table address of the token, and obtains the function table address for referencing the function table 93 and the link table address for referencing the link table 92 next time. After that, it is sent to the function table 93.

The token is referenced by the function table address in the function table 93, and there the management information in the data memory 94 is referenced and updated, and at the same time, the processing code indicating the processing content in the processing unit 96 and the access address of the data memory 94 are stored. get. The token is sent from there to data memory 94;
Therefore, waiting for the other operand of a binary operation or reading a constant for a constant operation is performed as necessary.

The queue memory 95 is a memory for temporarily holding tokens when the processing unit 96 is processing the previous token and cannot input the next token. When the processing unit 96 is not busy, the tokens are processed from the queue memory 95. Depending on the processing code, the unit 96 performs one of the following operations, such as addition, subtraction, and multiplication of integer data, logical operations, shift comparisons, bit inversion, priority encoding, shunting, numerical generation, copying, and operations using internal registers. undergo processing.

The processed token is sent to the link table 92 and referenced again using the link table address.

Thereafter, the internal ring bus is similarly circulated until the external bus output command is executed, and the data value undergoes necessary processing.

When a token is to be output to an external bus by an external bus output command after being processed by the processing unit 96, after the last arithmetic processing by the data flow processor is received by the processing unit 96, the token is first output from the link table 92. The link table address that should be held as an external bus token is fetched from the function table 93, and the module number and processing code indicating the output that should also be held as the external path token are fetched from the function table 93, and then input to the queue memory 95 via the data memory 94. . If the processing code of the token at the exit of the queue memory 95 indicates output, the token is sent from the queue memory 95 to the token output section 97, and the module number and link table address obtained as described above are sent from the queue memory 95 to the token output section 97. After being transformed into an external bus token using

In this way, in conventional data flow processors, multiple tokens that should receive the same processing are given the same token identifier (
A method is used in which a link table address (called a link table address) is given and the data is streamed continuously. Such continuous flow is called a stream.

In conventional processors, when performing cumulative addition of stream data, the following method is used as shown in FIG.

That is, in the method shown in FIG. 7(a), a register is provided in the arithmetic unit (processor unit), stream data is poured in, and the register performs cumulative addition every time a token is input.

In the method shown in FIG. 7(b), addition is broken down into combinations of binary additions, programmed in advance in a tournament format so that the sum is determined at the end, and then stream data is distributed and input.

The method shown in FIG. 7(C) uses the μPD7281 co
As seen in the "nvo instruction", the stream data input to the operation node and the partial sums calculated so far are divided into odd and even numbers, and the addition instruction below is used to add two terms at a time. After performing the operation of returning the result to the input arc a number of times corresponding to the length of the stream given in advance, the cumulative addition result is output.

By such a method, cumulative addition of stream data is performed in the data flow processor.

[Problem to be solved by the invention]

However, each method of cumulatively adding stream data in a data flow processor has the following problems.

The method shown in FIG. 7(a), that is, the method using registers, uses registers in the processing unit, so it is not possible to perform cumulative addition at multiple locations at the same time, which impedes speedup and requires program control. It becomes necessary.

In the method shown in FIG. 7(b), that is, the tournament method, it is necessary to prepare a program that defines a data flow so that tournament-type addition can be performed on data of a predetermined length.

The method shown in FIG. 7C, that is, the method using the convo instruction, cannot be used unless the length of the stream is known in advance using the convo instruction in this case as well, and lacks flexibility.

In this way, with conventional data flow processors, when stream data of any length is passed through one arc in a data flow diagram and the cumulative addition is performed, it can be used in parallel at multiple points to handle any length. There is no means for cumulative addition, and no matter which method is used, the problem is that the processing is inefficient. This is because there is no means to discover the end of a stream of arbitrary length input to a conventional data flow processor and to detect the end of its operation.

The present invention provides a method for speeding up cumulative addition of a variable-length stream by flowing a token with an end mark as the end of stream data and detecting the end, and a data flow processor that implements the method. It is intended to.

[Means to solve the problem]

The first invention performs processing by sending a token, which has a link table address that is an identifier for distinguishing each other, data to be processed, and an end mark indicating the end of the stream, to an internal bus as a unit of data. , in a stream operation completion detection method for detecting the end of an operation on a data stream consisting of a plurality of tokens in a data flow processor equipped with a data memory, for each input token, the terminal state obtained as a result of the immediately preceding determination process is detected. a value of a flag and 1, a value of a write flag that stores whether or not writing was performed to the data memory immediately before;
The terminal mark of the token input this time is input, and from the combination, it is determined whether or not to write the data value of the input token to the data memory as an operation for the input token this time, and the result of the judgment process this time. The value of the termination state flag as the result of this judgment process, the value of the write flag as the result of this judgment process, the value of the link table address of the token to be output as a result, the value of the end mark of the token to be output as a result, and the value of the end mark of the token as a result. , determine the operation code that specifies the operation after the token to be output, and
It is characterized by determining the end of a stream operation by giving different link table addresses during and at the end of the stream operation.

The second invention has an internal memory that performs processing by flowing tokens, which have identifiers for distinguishing each other, data to be processed, and an end mark indicating the end of the stream, onto an internal bus as a unit of data. The stream operation end detection device of the data flow processor includes an input token register that temporarily holds incoming tokens, an operation code field that is referenced by an address in the token register and specifies an operation, and an operation code field that specifies an operation for accessing the internal memory. a code table having a data memory address field; and a state holding section that is accessed by an address in the token register and that holds and stores a terminal state flag field and a write flag field as internal states for realizing a terminal detection operation. , an end detection controller that receives the contents of the input token register, the code table, and the state holding section as input, updates the contents of the state holding section, and determines the value of the output token; and a signal from the end detection controller. and an incrementer that generates a link table address field of an output token from a value of a link table address field of the input token register.

[Operation] When using the stream end detection method of the present invention, an end mark is given in advance to the last token of the stream in which cumulative addition is to be performed.

When there are two consecutive token inputs, the stream end detection device writes the former to memory and reads and adds them when the latter arrives.

When a stream is input, two adjacent pieces of stream data are added together, and the partial sum is output as a token and returned as an input stream to the original instruction, which is repeated. , the length of the input stream decreases by one for each addition.

The stream calculation end detection device holds the value of the end mark of the token processed last time in the end state flag, and both this and the end mark of the token input this time are “1゛′”
Continue the above operations until . If both become "'1", it is determined that the stream length has become 1, and the data at that time has a different link table address as a result of cumulative addition of all the original stream data. (i.e. from somewhere else).

Due to the above-described effects, according to the present invention, by detecting the end of the stream, efficient stream cumulative addition with variable length can be realized.

〔Example〕

Embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a basic configuration diagram of a stream calculation completion detection device of the present invention, FIG. 2 is a configuration diagram of a data flow processor having a stream calculation completion detection device, and FIG. 3 is provided for explanation of the present invention. It is a figure which shows the format of a token and a table.

The data flow processor shown in FIG. 2 performs processing by flowing tokens, which have identifiers for distinguishing each other, data to be processed, and an end mark indicating the end of the stream, onto an internal bus as data units. A data flow processor having an internal memory, which includes a path token input section 11, a link table 12, a stream operation completion detection device 1, a data memory 14, a buffer queue 15, a processing unit 16, and a path token output section. 17.

First, the operation of the data flow processor using the present invention will be explained using FIG. 2 showing the configuration of the data flow processor and FIG. 3 showing the token format inside the data flow processor. This data flow processor consists of units shown by rectangles in FIG. 2, and each unit operates by exchanging tokens in synchronization with a pipeline clock supplied from the outside.

In FIG. 2, the data flow processor operates according to tokens input from the outside via the pass token human input unit 11. The token in the link table 12 is
As shown in FIG. 3(a), a data section 64 containing data values to be processed has a link table address 63 for identifying the token and an end mark 62 for indicating the end of the stream. A stream is a stream in which a plurality of tokens having the same link table address flow while maintaining their mutual order, and arithmetic processing is performed using the fact that the relative positional relationship between the tokens in the stream is unchanged.

The end mark has a value of "'1" for the last token and "'0" for other tokens.

The token is the processing unit 1 shown in FIG.
Before receiving the calculation in step 6, the link table address to be held after the calculation and the code table address 73 for referring to the stream calculation end detection device 1 are fetched from the link table 12. By this reference, the token having the format shown in FIG.
, and undergoes a stream calculation completion detection operation as described below. Thereafter, the token passes through the buffer queue 15 to the processing unit 16 along with the two operands obtained by memory reading in the data memory 14 and the operation code used in the processing unit 16 (FIG. 3(d)), if necessary. It is sent there, where it undergoes arithmetic processing according to the operation code.

After the arithmetic processing, the link table address is sent to the link table I2 via the signal line 103 and the same operation is continued, or the pass token output unit 17
is output to the outside of the data flow processor as a result token.

Next, the configuration and operation of the stream calculation end detection device 1 will be explained using FIGS. 1 and 3.

This stream operation completion detection device I includes an input token register 4 that temporarily holds incoming tokens, an operation code 30 that is referenced by an address in the token register 4 and specifies an operation, and a data memory address for internal memory access. A code table 5 having 32 and a terminal state flag 3 as an internal state for realizing the terminal detection operation.
6 and a write flag 37, which is accessed by the address in the token register 4;
, an end detection controller 8 that receives the contents of the input token register 4, code table 5, and state holding section 6 as input, updates the contents of the state holding section 6, and determines the value of the output token; and an end detection controller. The incrementer 9 generates the link table address 71 of the output token from the value of the link table address of the input token register 4 in response to the signal 8.

In such a stream calculation end detection device 1, the signal line 1 is
The token input to the stream operation completion detection device 1 via 01 is latched into the input token register 4 of the stream operation completion detection device 1 shown in FIG. 1 in synchronization with the pipeline cycle of the data flow processor. . In FIG. 1, the registers are divided to correspond to each bit field within the token. The code table 5 and state holding unit 6 are referenced by the code table address 73 of the input token.

The code table 5 is a table memory having a plurality of fields shown in FIG. 3(e), and the row referenced here is the stream operation end detection bag W1 of the token. Data memory 14 and processing unit 16 shown in FIG.
Specifies the behavior in The operational code 30 shown in FIG.
This is an instruction code indicating an operation in the stream operation end detection device 1 and the processing unit 16 shown in FIG. 2. In addition to the stream addition instruction described in detail below, the stream operation completion detection device 1 executes operations such as token waiting and data memory read/write, which are provided by the data flow processor described in the prior art section. 1st
The data memory address 32 shown in the figure provides an address when accessing the data memory with a stream addition instruction.

The state holding unit 6 is a table having a plurality of rows as shown in FIG. 3(f), and the referenced row holds the internal state of the stream calculation end detection device to which the token refers. The terminal status flag 36 basically indicates whether a terminal token was processed during the previous instruction execution of that line.

The write flag 37 indicates whether the data memory was last written or read.

The end detection controller 8 detects the operation code field of the line specified by the code table address 73 on the human-readable register 4 in the code table 5, the end mark 72 from the input token register 4, and the input of the state holding unit 6. Code table address 7 on token register
Input the state held in the line indicated by 3 and perform the state transition described above. The resulting state is written back to the same line of the state holding unit 6, and the end mark 82 of the output token from the stream operation end detection device 1. Data memory code 83. Each field of the operation code 85 is generated and output. It also outputs a signal 86 to the incrementer 9.

The input/output of the termination detection controller and the update of the state holding section are as follows:
Follow Table 1 below.

However, in Table 1, the first half of the "operation code" describes the operation in the data memory, and the second half describes the operation in the processing unit. write'“ writes to data memory,
read'” reads from data memory, “dm
“no p” is no action on data memory (NOP),
add” is the addition of two operands in the processing unit, n.

p′ is N of the first operand in the processing unit
The OP output, "ERROR" indicates a transition to an error state. When a write is performed in the data memory, the token disappears in the data memory, so the corresponding output token term is not written.When the incrementer signal 86 shown in FIG. 1 is "1", the incrementer 9 outputs the value obtained by adding "1" to the input, and in other cases outputs the input as is.

The stream calculation end detection device 1 forms an output token as described above, and sends it to the subsequent data memo 4 via the signal line 102 in FIG. The format of the token sent from the stream calculation end detection device 1 to the data memo 4 is shown in FIG. 3(C). The data section 80 has the same data as when the token was input from the synchronization queue, and the rest is determined by the operation of the stream operation end detection device 1.

The data memory code 83 (1) passes the input token without accessing the data memory; For the output token, the right operand is set to "0" (code dmn op), (II) data memory IJ14 with the value of data memory address 84 as the address.
The read value is read from the right operand 66.
(Code read), (■) Writes the data value that the token has in the data section 80 to the data memory using the value of the data memory address 84 as an address, and erases the token there (Code w
r ite), or. The data memory address 84 indicates the access address at which the token reads and writes memory in the data memory 14 shown in FIG.

The operation code 85 is fetched from the code table 5 of the stream operation completion detection device 1 shown in FIG.

The operation of cumulative addition of stream data, which is the object of the present invention, will be explained using FIG. 4 using the stream calculation end detection method of the first invention. As shown in FIG. 4, the data string to be cumulatively added is input as stream data to a node 20 that executes a "stream addition instruction". Only the last token T4 of the stream data is a token whose end mark is "1". The stream addition instruction has an end state flag and a write flag internally, and when a token is input manually, a judgment is made according to the flowchart in Figure 5 based on these flags and the end mark of the input token. Execute the determined action based on the results. However, in Figure 5, the rectangle with two output branches has a value of 1" or 0".
Indicates that a determination operation is performed to determine whether the The other rectangles describe the disconnected state of the state holding unit and the end mark of the output token among the operations performed as a result. "ESM", "ESF'", "WF"
represent the end mark of the person's computer, the end state flag of the state holding unit, and the write flag, respectively.

Below, this operation will be explained as follows:
The operation when a stream is input will be explained as an example. The initial value of the state holding unit in the stream operation completion detection device corresponding to the node that executes the stream addition instruction is:
The terminal status flag is “0” and the write flag is “0”.

(1) First, when the first token TI enters this node with the end mark "0", the code table address shown in FIG.
The state holding unit 6 is referred to. In response to this input, line a of Table 1, which is the operation table for the end detection controller shown above, is executed, and the data of that token is written into the data memory accordingly. The token disappears in data memory.

(2) In the second token T2, the end mark is “
0, but the write flag is "1", so the first line is executed, and the processing unit adds the data read from the data memory and the data held by the input token.

The result is output by token T5 with the given link table address, which joins the original arc 21 via arc 22 in FIG.

(3) The token T5 having the partial sum of tokens 1 and 2 obtained in the above (2) and the third token T3 of the original stream meet at the entrance of the node 20 shown in FIG. 4.

Regardless of which data arrives earlier, the data that arrives earlier is written into the data memory according to the rows of the first table a, and the data that arrives later is read out according to the rows of the first table and both data are added. This yields a token T6 with the partial sum up to the third of the stream, which is also arc 2
2 and is input from arc 21 to node 20.

(4) If the fourth token T4 of the stream arrives earlier than the token T6 which has the partial sum of these three data, it has the end mark "1", so the data is stored in the data memory according to row g of Table 1. At the same time, the terminal state flag becomes "1".

(5) Then, when the other token T6 arrives at the node 20 following token T4, row d of Table 1 is executed. The value of the data memory is read and added, and the end state flag of the device for detecting the end of stream calculation becomes "1". The end mark of token T7 with the addition result is “1”
゛°, token T7 passes through arc 22 and reaches arc 2
1 to the node 20.

(6) When the token T7 is input to the node 20, the first
The row of table g is selected, so that the data of token T7 is output without undergoing any calculations. However, the incrementer in the stream calculation end detection device operates, and "1" degree is added to the link table address of the output token. As a result, the output 1.-kun is output from a different arc 23 than in other cases, and this has the result of cumulative addition of four data of the stream.

In the above explanation, an assumption was made regarding the order in which two tokens arrive when they join, but the operation will work correctly even if the order is other than this. Further, although the length of the stream is 4 as an example, the same operation can be performed for streams having an arbitrary length of 2 or more, and correct cumulative addition results can be obtained.

〔Effect of the invention〕

As explained above, by using the "stream addition instruction" realized by the device of the present invention to perform the method of the present invention on a stream with an end mark attached at the end of the stream, stream data of arbitrary length can be generated. This has the effect of achieving cumulative addition of .

Since this method does not use registers, this instruction can be used in parallel at multiple locations, and it is also possible to process streams whose lengths are not known in advance.

[Brief explanation of the drawing]

FIG. 1 is a basic configuration diagram showing an example of a stream calculation completion detection device of the present invention, FIG. 2 is a configuration diagram of a Tsutator flow processor having the stream calculation completion detection device, and FIG. 3 is provided for explanation of the present invention. FIG. 4 is a data flow graph that describes the operation of the data flow processor when the present invention is used; FIG. 5 is a flow diagram that describes the stream calculation completion detection method of the present invention; FIG. 6 is a block diagram of a conventional data flow processor, and FIG. 7 is a flow graph showing a method for implementing cumulative addition of stream data in the conventional data flow processor. 1... Stream calculation end detection device 4...
Person register 5...Code table 6...State holding unit 8...Completion detection controller 9...Incrementer

Claims (2)

[Claims] (1) A data memory that performs processing by passing tokens, which have a link table address that is an identifier to distinguish each other, data to be processed, and an end mark indicating the end of the stream, to an internal bus as a unit of data. In a stream operation completion detection method that detects the end of an operation on a data stream consisting of multiple tokens in a data flow processor equipped with , the value of the write flag that stores whether or not data was written to the data memory immediately before, and the end mark of the token input this time, and from the combination, calculate the operation for the input token this time. As a result, output whether or not to write the data value of the input token to the data memory, the value of the terminal state flag as a result of this judgment process, the value of the write flag as a result of this judgment process, and the result. Determine the value of the link table address of the token, the value of the end mark of the token to be output as a result, and the operation code that specifies the operation after the token to be output as a result. A method for detecting the end of a stream operation, characterized in that the end of the stream operation is determined by giving a different link table address. (2) A data flow processor with internal memory that performs processing by flowing tokens, which have identifiers for distinguishing each other, data to be processed, and an end mark indicating the end of the stream, onto an internal bus as data units. An input token register that temporarily holds incoming tokens, an operation code field that is referenced by an address in the token register and specifies an operation, and a data memory address field for accessing the internal memory. a state holding unit accessed by an address in the token register that holds and stores a terminal state flag field and a write flag field as internal states for realizing an end detection operation; an end detection controller that inputs the contents of the register, the code table, and the state holding section and updates the contents of the state holding section and determines the value of the output token; and a signal from the end detection controller that controls the input token register. An incrementer that generates a link table address field of an output token from a value of a link table address field of a stream operation completion detecting device.