JPS5927352A - Data flow computer controlling system - Google Patents

Abstract

PURPOSE:To suppress an environment tag to a fixed bit, and to realize generation and erasion of a tag at a high speed, by using a memory address of an idle area of a tag control table, as a new environment tag bit, and storing an old environment tag in the idle area. CONSTITUTION:A processing element PE has K-pieces of PEs 11-1k. As for an operand of a result operated by the PEs 11-1k, PE of a transfer destination is determined in accordance with a PE mapping method, and is sent to an operand network 2. The network 2 executes routing by seeing a transfer destination PE number added to the operand, and sends the operand to a prescribed PE. Inherent device number is given to I/O controllers 31-3m for controlling an input/ output apparatus, etc., in the same way as the PE. When controlling the I/O controller from the PE, a destination I/O controller number is added to a control data, it is subjected to routing in the network 2, and is sent to a prescribed I/O controller. Plural input/output apparatuses 4 are connected to the I/O controller.

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field of the Invention The present invention relates to a method for efficiently managing tag generation and deletion in a data flow computer using a tagged token control method.

Technical background Traditionally, computers that have been called 7-on-Neumann type have
Basically, only one instruction can be executed at a time, and many different instructions can be executed in parallel at once, since the instructions at the address specified by a single program counter are sequentially retrieved from memory and executed. It was essentially difficult to do so. Computers using the data flow control method (data flow computers) differ from 7on Neumann type computers in that they do not have a program counter. Instruction execution is controlled according to data-driven rules such that the command is executable. According to this data-driven rule, υ, if there are multiple instructions that can be executed with the same data, they can be executed in parallel at the same time.
It can achieve higher processing power than a von Neumann computer. An example of instruction execution control in a data flow computer is shown below.

In Figure 1, C=(,4+1)*(f-1, D=(A-
E) A data flow graph obtained by converting the /B program using a compiler. Each node corresponds to an instruction from the data flow computer, and data (
This is called a token) is moved.

For example, in Figure 1, the initial value is Am8, B=2
Assuming that, when the data of Am8 is input to the node town, the node town is activated and outputs the calculation result 8+1=9 on the output arc. On the other hand, node n is 2 of Am8 and B=2.
It is activated when the data for 8-2=
Outputs 6. As a result, the two input data for node town (9 and 6) and the two input data for node n4 (6 and 2) are now available, so both nodes are activated in parallel, and node n
, outputs 9*6=54, node n outputs 6+2=3, and the operation ends.

Regarding the booter lobe 27 in FIG. 1, in reality, the type of operation of an instruction node and the connection relationship between the nodes are stored in memory as a program, and the machine outputs the operation result (data) of a certain instruction using the specified connection. This is accomplished by transferring to the memory where the next instruction resides according to the relationship.

Here, when the program in Figure 1 is simultaneously called by two different programs (PIQ) and executed in parallel, two types of programs corresponding to the two programs p+Q are placed on the same arc in the data flow graph in Figure 1. Token exists. As a method to perform calculations under such conditions, we add identification information (environment tag) to each token indicating which PrQ it was called from, and perform calculations between tokens with the same tag. A ``tagged token method'' has been proposed.

PRIOR ART AND PROBLEMS If the conventional tagged token method is directly implemented in a data flow computer, the following problems will arise.

For example, program P1 calls program P,
When program P, calls program P1, the environment tag pit length of the token used in P is the P1 identification code length, and the environment tag bit length of the token used in P is P1 identification code length ten P. This is the identification code length. In other words, the length of the environment tag pit becomes variable during execution depending on the depth of the nesting of the program, which significantly complicates the control of the hardware, and when the nesting is deep, the length of the environment tag pit becomes extremely large. However, it has been difficult to provide an efficient data flow calculator.

Purpose of the Invention In order to eliminate these drawbacks, the present invention provides a tag management table for storing environment tags in memory, uses a memory address pointer of an empty area of the tag management table as a new environment tag, and stores environment tags in the empty area. By storing the old environment tag in the area, the environment tag can be suppressed to a fixed pit, and the environment tag can be generated and deleted quickly and easily.Examples of the present invention will be described in detail below using drawings. explain.

Embodiment of the Invention FIG. 2 is a system configuration diagram of a data flow computer according to the present invention, which has 11 to 1□IdK processing elements (PR), which corresponds to the processor (memory ). In addition to executing instruction operations, each PE has functions specific to data flow control, such as waiting for the arrival of operands and detecting executable instructions. The operands resulting from the calculations performed by the PEs 1° to 1K are sent to the operand network 2 after a transfer destination PE is determined according to a certain rule (this is called a PE mapping method). The operand network 2 performs routing by looking at the transfer destination PK number added to the operand, and sends the operand and do to a predetermined PH. 3. ~3. is a large 10 controller that controls input/output devices and connects PE1 via operand network 2.
．． Connected to ~1K. Each 10 controllers include
A unique device number is assigned like the PE, and the P
When controlling 10 controllers from E, a destination 10 controller number is added to the control data (operand 2), which is routed within the operand network 2 and sent to the IO controller in Shingu. 4 is multiple input/output devices, each 10 controller #1 (3,)
``2(3*)+..., *m(Am) are connected to ``1*'' M r, m, and rm input/output devices, respectively. A path cup 2 is used to connect each PE1. to 1X to a common memory controller 6 and a common memory 7. 8 is a common path.

FIG. 6 is a detailed diagram illustrating the relationship between the common memory controller 26 and the common memory 7, which is an embodiment of the present invention.

Here, reference numeral 9 denotes a reception data register, and the environment tag update and restoration requests sent from each PE via the common bus 8 are temporarily stored in this reception data register 9. The reception data register 9 includes a control order field 9, an environment tag field 9. as shown in FIG. It is composed of many. Reference numeral 10 denotes a 9M address register, which includes a transmitting device number field 101 for specifying the transmitting device of received data and a receiving device number field 10° for specifying the receiving device. Reference numeral 11 denotes a sequence control unit that decodes the control order in the received data register 9 and generates various timing signals and control signals to the arithmetic/test circuit 12 and various registers. 16 is an 0M address register that stores an address for accessing the common memory (chr) 7; 14tj is a data register group that stores intermediate results of operations and various control information; 15 is an operand path;
is the result pass. 17 is connected to PE via common path 8
Transmit data register for returning answer data to 1
8 is the transmit address register, and receive address register 1
Similarly to 0, the transmitting device number field 181 and the receiving device number field 18. It is composed of many.

First, we will explain how to generate a new environment tag.

Step 1: An environment tag generation request is transmitted to the common memory controller 56 when a certain PE executes a special instruction (for example, a procedure call instruction) for generating a tag. For example, when a tag generation request command is executed in PE#1 (1□), PE*1 (1t) is the PE in Figure 4 consisting of the "address" in Figure 4 α and the "data" in Figure 4 As shown in the information transferred from to the common memory controller 6, the information necessary for new tag generation (the control order of the new tag request and the current environment tag pit) is transferred to the memory controller 6 via path coupler #1 (5,). Receive data register 9
0 control order field 9° and environment tag field 9.
and the transmitting device number and receiving device number in the transmitting device number field 10. in the receiving address register 10. and receiving device number field 10. Send it to. Control order 1 in FIG. 4 means a tag generation request.

Step 2 The error control unit 11 inputs the received data register 90 control order field 9. When it is decoded and found to be a new tag generation request (=1), hashing is performed using the arithmetic/test circuit 12 using the environmental tag pit as a key. In other words, X is the key, hash function A(z
), for example, if A(z) = (modulo 256) and the entry size of the hash table (tag management table) prepared on the common memory 7 is 256, then the fourth
From the figure, let's say z-24"r Z2= 24" -576
However, since the calculation is performed modulo 256, this value is A
(Z) ~ 24” = 576 → 576-256 x 2 =
64 (modulo 256 operation $#), and this value is finally output from the arithmetic/test circuit 12. In order to use this group as the address for the next staple, it is temporarily assigned to the CM address register 13. As shown in FIG. 5, the hash table occupies an area from address 0 to address 255 in the common memory 7, and each entry is composed of an empty/full field and an environment tag field. In this example, address 64 of the hash table is read.

By inputting this to the arithmetic/test circuit 12 and testing the empty/occupied display field, the sequence control section 11 detects that the empty/occupied display is written at address 64. This is a collision phenomenon in the hash method, and indicates that information is already stored at address 64 for another purpose.

Step 6: Various methods such as the open address method and the chain method are known as methods for avoiding the collision phenomenon and finding another vacant area, but the simplest open address method will be described here. That is, the sequence control unit 11 calculates the contents (~64) of the CM address register 13 (7) and sends it to the test circuit 1.
2, it is incremented by 1 and returned to the CM address register 16 again. As a result, the contents of address 65 of the common memory 7 are read out.

As shown in FIG. 5, the empty/full display field at address 65 indicates empty, and this
When detected, the sequence control unit 11 assigns the memory address (address 65) itself as a new environment tag by assigning the contents (-65) of the CM address register 16.
is stored in the transmission data register 17 via the arithmetic/test circuit 12 (the arithmetic is not executed, but simply passed through).

Furthermore, in order to send a new tag back to PE#1 (11) that requested new tag generation, the reception address register 10 (7)
Each field 10. , 10. Transmitting device number (=
J) Exchange E#1) and the receiving device number (-memory controller number) with each other in the arithmetic/test circuit 12,
The transmission address register 18 is reset. The result is
The transmitting device number in the transmitting address register 18=memory controller number, and the receiving device number=PE#1.

Step 4: Perform the following post-processing before sending the contents of the transmission address register 18 and transmission data register 17 to the common bus. That is, the sequence control unit 11 performs calculations and
In the test circuit 12, the environment tag field 9. of the receive data register 9. (value 24) and the field in which the bit position corresponding to the empty/occupied field is set to 1 (=occupied state) are combined and edited, and the result is sent to the common memory 7 via the result bus 16.
Store in 065'& (specified by CM address register 13). The resulting hashtable map is shown in Figure 6.

Step 5: The "address" shown in FIG.
By sending it back to #1, PE group 1. ~1 can be used by adding a new environment tag with a value of 65 to the operand data.

As shown above, each time a request to allocate a new environment tag occurs, by searching for an empty area in the table and allocating that address as a new environment tag, a large number of environment tag bits overlap with each other. It can be easily controlled not to.

Next, a method for restoring to the old Kashikyou tag (returning from the current environment tag to the old Kashikyou tag) will be explained.

Step 1: A tag recovery request is issued by a special command similar to a tag generation request, and the tag
R (for example, PA'#3 (1a)), Shiba Katsupura 58. “Address” in FIG. 8 α showing information transferred from the PR to the common memory controller 6 via the common bus 8;
Control information shown in "Data" in FIG. 8 is entered into the reception address register 10 and the reception data register 9. The sequence control unit 11 looks at the control order (~2 (tag recovery request)) shown in FIG.
．． (having a value of 65) is input to the arithmetic test circuit 12, passed through, and stored in the CM address register 15. This value (value 65 of the current environment tag) is the exact address of the common memory area that stores the old environment tag, as shown in Figure 6, so there is no need to go out of your way to search for the area that stores the old environment tag. , the old item boundary tag (value 24) can be read directly and quickly using the current environment tag. This is input into the arithmetic/test circuit 12, passed through, and stored in the transmission data register 17 via the result lot 16. The sequence control unit further calculates and calculates the transmitting device number (PEII5) of the receiving address register 1o and the receiving device number (memory controller number).
The test circuit 12 replaces each other and the transmission address register 17 is reset.

As described above, the transfer information between the processing element and the memory controller shown in "address" in FIG. 9 and "data" in FIG. 9 is edited.

Step 2: Perform the following post-processing before sending the contents of the transmission address register 18 and transmission data register 17 to the common bus. That is, the sequence control section 11 causes the arithmetic/test circuit 12 to output all zero data, and writes this to the address (address 65) indicated by the CM address register 16, thereby returning to the state shown in FIG. From then on, 6
Address 5 is regarded as a vacant area, and the value 65 can be assigned again when the next tag request occurs.

Step 6: Finally, the contents of the transmission address register 18 and transmission data register 17 are sent to the PE5 via the common bus 8.
6(1,) and notifies the old storage tag with a value of 24. PE group 1. 1, this old broken boundary tag can be added to the operand data and used thereafter.

Next, we will show that even when programs have a nested structure, consistent control is possible using fixed-length tag bits.

FIG. 10 shows how procedure P1 calls procedure P2, procedure P2 calls procedure P6P6, and procedure P6 calls procedure P4. In this example, upon execution of the procedure call instruction, a new environment tag is allocated for the new glossary, and upon execution of the procedure return instruction)
A reversion to the old environment tag is performed. The state of the hash table when procedure P4 in FIG. 10 and procedure P4 in FIG. 11 are executed in parallel is as shown in FIG. 12, for example, according to the above-described operating principle. That is,
For the nested structure in Figure 10, 24→65→129→
When one environment tag list is assembled, it can be realized with a fixed length of environment tag bits (in this embodiment, 8 bits obtained from the table size 256) regardless of the depth of nesting.

Also, the environment tag (=1) of procedure P4 in Figure 10
and the environment tag (-97) of procedure P4 in Figure 11.
are guaranteed to be assigned different values based on the search for empty areas in the hash table.

Therefore, even if both procedures are executed in parallel, the environment tags of the data generated by both glossaries are different, so even if both types of data are mixed in PE (1, ~ IK), they can be correctly identified and processed. .

The present invention described above is not limited to this embodiment, and any known hash function can be selected depending on the purpose and application, and any known technique can be applied to collision processing. In addition, in this embodiment, in order to simplify the explanation, the output of the hash function A(z) itself is stored in the hash table address, but the hash table can be realized no matter where the hash table is located in the common memory. In order to do this, the base address of the hash table is stored in, for example, 1201 data registers, and the hash function A (
This can be easily realized by adding the output of 2) and the pace address command.

Furthermore, any new tag generation method may be used as long as it allocates the memory address of an empty area as a tag.For example, empty areas are connected in a chain using a pointer list, and the address of the empty area is assigned in response to a tag generation request. Assign it as a new tag, store the old tag in the empty area,
In response to a request to return to an old tag, a method can also be applied in which the old tag is removed from the area of the address indicated by the tag currently in use and this area is returned to the chain of free areas.

As described in detail, the present invention uses a memory address in an empty area of the tag management table as a new tag pit, and stores the old tag pit in the empty area. Bits can always be of a constant length, and control of tag generation and erasure can be achieved quickly, easily, and economically.

[Brief explanation of drawings]

Figure 1 is an example of a data flow graph, Figure 2 is a system configuration diagram of a data flow computer using the present invention, Figure 5 is a detailed diagram of an embodiment of the memory controller and common memory of the present invention, and Group + α, b, Fig. 7 α, b to Fig. 9 α, b
5 and M6 are maps of tag management tables, and FIGS. 10 and 11 are transition diagrams showing procedure call return and transition 9 of remote tags, FIG. 12 is a map of the tag management table corresponding to FIGS. 10 and 11. 1, to 1...processing element, 2...
Operand network, 6□~6rn...10 controllers 2.4...I/O equipment, 5. ~5K...Path coupler, 6...Common memory controller, 7...
Common memory, 8... Common path, 9... Reception data register, 10... Reception address register, 11...
Sequence system 611 part, 12... Arithmetic/test circuit,
16... CM address register, 14... Data register group, 15... Operand path, 16... Result bus, 17... Transmission data register, 18...
Transmit address register. Patent Applicant Nippon Telegraph and Telephone Public Corporation Agent Patent Attorney Gobe Tamamushi (3 others) Figure 1 Figure 2 Figure 3 Figure 4 Cause b α 31!5 Day 6 Kaoru a Figure 7 b Figure 8 9 Figure 11 Figure 12

Claims (1)

[Claims] In a data flow computer using a tagged token control method, a tag management table for storing tags in memory is prepared, and when a new tag generation request occurs, an empty area in the tag management table is detected and the empty area is Fixed length A control method for a data flow computer characterized by controlling tagged tokens using tags.