Vector processor using buffer for preparing vector data

20~376 NE-225 I _ TLE OF THE INVENTION 2 "Vector Prooessor Using Buffer For Preparing Vector Data" 3 BACKGROUND OF THE INVEN'IION 4 The present invention relates generally to vector prooessors, and more S specifically to a vector processor having improved data loading circuitry.

6 Conventional vector processors include a main memory in which 7 instructions and vector data are stored, a memory controller that accesses 8 the memory a calculation circuit to perform vector calculations on data read 9 out of the memory, and an execution controller which interprets instructions 10 from the memory and monitors available resources, or vector registers to 1 1 control the pipelined components of the pr~oessor so that vector calculations 12 are performed at high speeds.

The operating performance of a conventional 13 vector processor can be evaluated by allowing to execute the following 14 vector instructions:

VLD Vl (1) 16 VADD VZ ~ S + V1 (2) 17 VLD V3 (3) 1 8 VMPY V4 ~ V2 x V3 (4) 19 Instruction (I) is a "vector load" instruction that directs the loading of 2 0 vector data from the main memory into a vector register V1, instruction (2) is 21 an addition instruction-that directs the vector data in register V1 to be 22 summed with a scalar value S and directs the result to be loaded into a 2 3 vector register V2, and instruction (3) is a second vector load instruction that 24 directs vector data to be loaded from the memory intn a register V3.

2 S Instruction (4) is a multiplication instruction that directs multiplication of 2 6 data in register V2 with data in register V3 and directs the result into a ;~ , . . .

; ~ .

.

.

NE-225 2000376 - 2 register V4.

Fig. 1 depicts a series of events involved with the execution of 2 instructions (1) through (4) using the prior art vector processor.

The 3 execution controller of the processor first analyzes instruction (1) as it is 4 loaded into the instruction register and directs the memory controller to S access the memory to load vector data into register Y1.

As the vector 6 register V1 is being loaded, the execution controller directs the calculation 7 circuit to execute instruction (2) by adding a scalar value S to the data loaded 8 into register V1 and loading the result into register V2.

In response to 9 instruction (3), the execution controller directs the memory controller to 10 access the memory to load vector data into register V3 and waits until the 11 result of the addition is loaded into register V2, whereupon it urges the 12 calculation circuit to execute instruction (4).

13 However, if it takes a substantial amount of time to access vector data 14 in the memory, the calculation circuit would have to wait until the data is l S ready for calculation since the vector load instruction is not issued to the16 memory controller until the next instruction is supplied to the calculation 17 circuit Therefore, a system slowdown can occur as a result of access delays.

18 S~RY OF THE INVENIION 19 It is therefore an object of the present invention to provide a vector 2 0 processor which is free from prior art disadvantages caused by delays in 21 accessing vector data.

2 2 This object is attained by loading vector data from the memory into a 2 3 buffer in advance of the time the next instruction is issued to a calculation 2 4 circuit if a vector register is available for calculation.

2 s According to the present invention, an instluction stack is connected to 2 ~ an instruction register into which instructions are fetched from a memory. A ...;,. ..

' :, ....

., . . ~ NE-225 2000376 - 3 - decoder is connected to the instruction register for generating a load request 2 signal when the instruction stored in the instruction register is a vector load 3 instruction and causing the instruction to be transferred from the register to4 the instruction stack. A resource manager has a plurality of flags associated 5 respectively with vector registers provided in a calculation circuit and 6 constantly updates the flags in accordance with contents of the associated 7 vector registers. A contention detector is connected to the resouroe manager 8 and the instruction stack for detecting a contention therebetween.

The 9 contention detector is responsive to a vector load instruction being loaded 10 into the stack for generating a proceed-to-transfer signal if no contention is 1 1 detected between the vector load instruction and a corresponding flag in the12 resource manager. A buffer is provided in a memory controller for loading 13 vector data from the memory in response to the load request signal from the 14 decoder and loading vector data from the buffer into the calculation circuit 15 in response to the prooeed-to-transfer signal from the contention detector.

16 BRIEF DESCRIPIION Ol~ THE DRAWINGS 17 The present invention will be described in further detail with referenoe 18 to the accompanying drawings, in which:

19 Fig. 1 is a timing diagram of vector calculations performed by a prior 2 0 art vector processor;

21 Pig 2 is a block diagram of a vector processor of the present invention;2 2 Fig. 3 is a block diagram of the memory controller of Fig. 1;

2 3 Fig. 4 is a block diagram of the resource manager of Fig. 1;

2 4 Fig 5 is a circuit diagram of the contention detector of Fig. l; and Fig. 6 is a timing diagram of vector calculations performed by the 2 6 vector processor of the present invention.

- . ' NE-225 2000376 DETAILED DESCRIPIION 2 Referring now to Fig. 2, there is shown a vector processor according to3 an embodiment of the present invention.

The vector processor of the 4 invention generally comprises an execution controller 1, a memory s controller 2 having a buffer 2a for storing data read out of a main memory 3 6 in which data and instructions are stored, and a calculation circuit 4.

7 Execution controller 1 includes an instruction register 11 connected to a 8 common bus (not shown), an instruction stack 12, a contention detector 13, a g resouroe manager 14, and a decoder 15.

An instruction fetched by the CPU (not shown) from the main memory 1 1 3 is supplied through the common bus and stored into the instruction register 12 11 for temporary storage.

Decoder 15 examines the instruction stored in 13 register 11 to determine if there is a "vector load" instruction in the 14 instruction register 11 and, if there is none, it causes the instruction stack 12 15 to read that instruction from the register 11.

If there is one, decoder 15 16 causes a "load requestn signal to be supplied to the memory controller 2, l 7 while causing the instruction stack 12 to read an instruction from the register l 8 11.

In response to the "load request" signal, the memory controller 2 19 accesses the memory 3 to begin a process for storing vector data into the 2 0 buffer 2a.

21 Resource manager 14 receives a proceed-to-transfer signal and a 2 2 proceed-to-calculate instruction from the contention detector 13, 23 instructions from the stack 12, and start~f-transfer and end-o~-transfer 24 signals from the memory controller 2.

Based upon these signals and 2s instructions, the resource manager 16 updates the various resources 2 6 available in the calculation circuit 4, such as vector registers, to produce ~' ' ~ ', ' .

' ~ ' ' .

NE-225 2000376 resource management data, which are to be supplied to the contention 2 detector 13.

3 As shown in Fig. 3, the memory controller 2 comprises a request 4 manager 200 to which the "load request" is applied from the decoder 15. C)n s receiving this signal, the manager 200 checks to see if there is one that is 6 being processed.

If there is one, the request is queued.

If there is none, the 7 manager 200 sends access information such as "main-memory requestn and 8 addresses to the memory 3, while enabling a read/write manager 201.

9 Memory 3 responds to the "main-memory request" by forwarding vector 1 0 data and an appropriate address to buffer 2a and read/write manager 201.

1 1 Read/write manager 201 specifies a location of the buffer 2a in response to 1 2 the address it received from the memory 3 to store the vector data therein.

l 3 On the other hand, the main-memory request issued from the request 1 4 manager 200 is passed through a timing circuit 202 to the count-up input of 1 5 an up-down counter 203.

The purpose of timing circuit 202 is to introduce a 1 6 delay so that the counter 203 is timed to increment its count by one with the 1 7 transfer of corresponding data to the buffer 2a.

Therefore, the content of 1 8 counter 203 indicates the number of granted "load requests" whose 1 9 corresponding vector data have been put in the buffer 2a.

2 o A nonzero detector 204 is connected to the counter 203.

If the counter2 1 203 is not zero, detector 204 supplies a logical 1 to a gate 205 to enable it to 2 2 open in response to a logical 1 at input 208 from a timing circuit, not shown.

2 3 If zero is reached in counter 203, the output of detector 2û4 switches from 2 4 logical 1 to logical 0 to disable, or close the gate 205.

Proceed-to-transfer 2 5 signals from the contention detector 13 are reoeived by a stack 206 and each 2 6 proceed-to-transfer signal is read out of the stack 206 by a read circuit 207.

- , . . . : . .

: . . . ~.

NE-225 -6- 2000376 If gate 205 is open, the output of read circuit 207 is applied to the resource 2 manager 14 as a start-of-transfer signal.

3 The proceed-to-transfer signal from the contention circuit 13 contains 4 information that indicates to which one of many vector registers a given S data item is to be transferred.

Thus, the start-of-transfer signal from the 6 memory controller 2 also bears information as to which vector register is 7 ready to receive data.

This start-of-transfer signal is also fed to the 8 read/write manager 201 and to the count-down input of up-down counter 9 203 to decrement it by one.

Read/write manager 201 responds to this start10 of-transfer signal by generating addresses to successively read data items l 1 out of buffer 2a into the appropriate vector registers of calculation circuit 4.

12 The start-of-transfer signal is also applied to a timing circuit 209.

This 13 timing cir~uit produces a delayed pulse which coincides with the timing at 14 which all data items have been transferred to appropriate vector registers of1 5 the calculation circu* 4.

The delayed pulse is applied to a gate 210 to which a 16 periodic timing signal is also supplied from a timing circuit, not shown.

l 7 Gate 210 is thus opened at periodic intervals to pass the delayed pulse fr~m18 the timing circuit 209 as an end-of-transfer signal to the resource manager 19 16.

As shown in Fig. 4, resource manager 13 essentially comprises load 21 wait management flags 300, load busy management flags 301, read busy 22 management flags 302, and write busy management flags 303.

These 2 3 management flags have corresponding vector registers in the calculation 2 4 circuit 4 and are constantly updated with the proceed-to-transfer signal and25 proceed-to-calculate signal supplied from the contention detector 13 in 2 6 response to each of the start-of-transfer and end-of-transfer signals from , ... .

-' . .

NE-225 2000376 -7 the memory controller 2.

2Specifically, load wa* management flags 300 are set in response to a 3proceed-to-transfer signal from the contention detector 13 and reset in 4response to a start-of-transfer signal from the memory controller 2.

Load 5busy management flags 301 are set in response to the proceed-to-transfer 6signal and reset in response to the end-of-transfer signal.

Read busy 7management flags 302 are set in response to the proceed-to-calculate 8instruction and reset in response to an end-of-read signal.

Write busy 9management flags 303 are set in response to a proceed-to-calculate l 0 instruction and reset in response to an end-of-write signal.

1 1 A time management circuit 304 Is provided to receive from the stack 12 12 a count of vector elements contained in each instruction and determines the l 3 amount of time elapsed from the reception of a proceed-to-calculate 14 instruction to the time at which the reading of a given vector register is 15 complete and further determines the amount of time elapsed from the 16 reception of the proceed-to-calculate instruction to the time at which the 17 writing of the given vector register is complete.

Time management circuit 18 304 generates an end-of-read signal at the end of a vector read operation 19 and generates an end-of-write signal on completing a vector write 2 0 operation.

Signals designated LDW (load wait), LDBSY (load busy), RBSY 21 (read busy) and WBSY (write busy) indicate the set and reset conditions of 2 2 the flags and are supplied to the contention detector 13.

2 3 In Fig. 5, contention detector 13 comprises a decoder 400 and a gate 2 4 array formed by gates 401 through 408.

If the first of a series of instructions 2 5 stored in the stack 12 is a "vector Ioad" instruction which directs the loading 2 6 of a vector register VRi with vector data, a signal VLDj identifying that . ~,. ~ .......... . ..

NE-225 20~0376 -- 8 vector register is issued from the decoder 400 and supplied to gates 401 and 2 402 to which load-busy (LDBSYj) and write-busy (WBSYi) signals are also 3 supplied from the resource manager 14.

These signals respecffvely indicate4 whether the vector register VRi is "load busy" or "write busy." If none ofthese conditions exists, the signal VLDi is passed through gate 403 and 6 applied to the memory controller 2 as a proceed-to-transfer signal.

7 If the first stack instruction is one that involves reading the content of a 8 vector register VRj, calculating it, and writing the result of the calculation 9 into a vector register VRk, then a signal VRrj is issued from the decoder 400 1 0 for reading the content of vector register VRj and supplied to gates 404 and 1 1 405 and, at the same time, a signal VRwkis issued from the decoder 400 forl 2 writing data into a vector register VRk and supplied to gates 406 and 407.l 3 The signals supplied to the gates 404, 405, 406 and 407 respectively are l 4 matched with signals from the resource manager 14.

To gate 404 is supplied 1 5 a signal RBSY; indicating if the content of vector register VRj is being read l 6 out or not.

The signal supplied to gate 405is LDWj which indicates if vector1 7 register VRj is w~iting to be loaded or not.

The signal fed to gate 406 is 1 8 WBSYk indicating if vector register VRk is being written or not, and the one1 9 fed to gate 407iS LDBSYk which indicates whether the vector register VRkis 2 0 being loaded or not.

If none of these busy conditions exists, a proceed-to2 1 calculate instruction will be issued from gate 408 to the calculation circuit 4.

22 The operation of the vector processor of this invention will be best 2 3 visualized with reference to Fig. 6.

For pulposes of comparison with the 2 4 aforesaid prior art, the following instr~ctions are used, which are identical 2 5 to those described with referenoe to the prior art:

~. .

NE-225 2000376 VIAD Vl (1) 2 VADD V2 ~S ~V1 (2) 3 VLD ~73 (3) 4 V~Y V4 ~ V2 x V3 (4) S With a vector load instruction (1~ being stored in the instruction 6 register 11, decoder 15 issues a load request a to memory controller 2 to 7 cause it to start reading a first data item from memory 3 into buffer 2a.

8 Instruction (1) is subsequently transferred to the instruction stack 12 and a 9 calculaffon instruction (2) is stored in register 11, interpreted by decoder 15 and transferred to instruction stack 12. A vector load instruction (3) is then 11 stored into register 11.

Decoder 15 issues a vector request signal b to the12 memory controller 2 to cause it to start reading a second data item from 13 memory 3 into buffer 2a. A calculation instruction (4) is then stored in 14 register 11, interpreted by decoder 13 and transferred to instruction stack 12.

l S In this way, instructions (1), (2), (3) and (4) are successively stored in l 6 shck 12 to be processed in the order named and the neoessary data items are 17 stored in buffer 2a.

Contention detector 13 first examines load vector 18 instruction (1) and proceeds to examine the status signals from resource 19 manager 14 to determine if vector register V1 is available in the calculation20 circuit 4.

If V1 is available, contention detector 13 supplies a proceed-to21 transfer signal c to memory controller 2 as well as to resource manager 14, 2 2 which updates the appropriate flag.

23 Contention detector 13 proceeds to the next vector add calculation 2 4 instruction (2~ which involves the use of data stored in vector register V1 by 2 S the load instruction (1).

At this moment, contention detector 13 knows that 2 6 vector register V1 is still not loaded with data neoessary for instruction ~2) by NE-225 -lO- 2000376 examining the resource status data, and puts instruction (2) in a queue.

2In response to the proceed-to-transfer signal c memory controller 2 3begins transferring data to vector register V1.

Concurrently, memory 4controller 2 sends a start-of-transfer signal d to resource manager 14.

On scompleting the data transfer, memory controller 2 sends an end{)f-transfer 6signal to resource manager 14 to update the status of flag associated with 7vector register V1.

8With vector register V1 being loaded, contention detector 13 is notified 9of this fact by resource manager 14 and sends a proceed-to-calculate 10 instruction to the calculation circuit 4 to cause it to execute the instruction (2) 1 1 waiting in the queue, as well as to resource manager 14 to update the status12 of the flag associated with vector register V1.

The result of the vector 1 3 addition is stored in a vector register V2.

14 Contention detector 13 then examines appropriate status data from 15 resource manager 14 and determines if vector register V3 is available since 16 the next instruction (3) is a vector load instruction.

If vector register V3 is 17 available, a proceed-to-transfeI signal e is applied to memory controller 2 as 18 well as to resource manager 14.

19 In response to proceed-to-transfer signal e, memory controller 2 starts 20 reading data out of buffer 2a into vector register V3.

Since this data has 21 already been written into buffer 2a from memory 3 in response to the load 2 2 request signal b. a data transfer from buffer 2a to vector register V3 can 2 3 begin immediately following the receipt of the proceed-to-transfer signal e.2 4 Concurrently, a start-of-transfer signal f is issued from memory controller 2 2 5 to resource manager 14 to update the flag associated with vector register 26 V3.

, ' NE-225 -11- 20~0376 After invoking instruction (3), contention detector 13 recognizes that 2 the next instructîon (4) is one that directs a multiplication to be performed on 3 the contents of vector registers V2 and V'3.

If these vector registers are not 4 available, contention detector 13 puts instruction (4) in a queue.

As soon asS they become available, contention detector 13 sends a proceed-to-calculate 6 instruction to the calculation circuit 4 to cause it to multiply the contents of 7 registers V2 and V3.

The result of the multiplication i5 then stored in a vector 8 register V4.

9 It is seen from the foregoing that since a load request signal is issued to 10 memory controller 2 from the execution controlier 1 as it detects a vector l l load instruction in the instruction stack 12, necessary data are stored buffer 12 2a well in advance of the time a calculation is actually performed thereon.

13 Therefore, the amount of time taken by the memory controller 2 to access an 14 item of vector data in the memory 3 can be reduced, preventing the vector 15 processor from degrading its performance by the otherwise long access time.

l 6 The foregoing description shows only one preferred embodiment of the 17 present invention.

Various modifications are apparent to those skilled in thel 8 art without departing from the scope of the present invention which is only 19 limited by the appended claims.

While mention has been made of an 20 embodiment in which the instructions stored in stack 12 are processed 2 l according to their order, the present invention could equally be applied to a 2 2 vector processor in which mutually unrelated instructions are stored in such a 2 3 stack by invoking a later instruction waiting in a queue before invoking an 2 4 earlier instruction.

In the foregoing description, since instruction (3) can be 2 5 executed independently of instructions (1) and (2), the processing time can be 2 6 reduced by executing instruction (3) prior to instruction (2).

If instruction (3) - .

NE-225 2000376 - 12 is assumed to be one that requests data to be loaded into vector register Vl 2 rather than into register V3, such an instruction (3) can not be invoked until3 instruction (2) is executed.

However, by loading necessary vector data into 4 buffer 2a in a manner described above, it is possible to prevent the vector s processor from slowing down its operating speed.

' .