JP3583032B2 - Vector diffusion processor - Google Patents

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a vector spread processing apparatus in a parallel computer, and more particularly to a vector spread processing apparatus that realizes high-speed vector spread processing.
[0002]
[Prior art]
2. Description of the Related Art Conventionally, vector operation processing is a technique used to speed up processing of a computer, and is employed in almost all types of supercomputers.
[0003]
In a large-scale computer such as a supercomputer, a large amount of time is spent in a process of repeatedly performing the same procedure for each data, such as a matrix calculation or a Fortran DO loop. In order to speed up such repetitive processing, the vector operation processing is to collectively process each matrix or data as a vector and execute a matrix operation with one instruction, one for each element of the matrix. Since there is no need to perform calculations, processing can be performed at high speed. This high-speed processing is performed by dividing an instruction for each element into "call of instruction", "decoding of instruction", "address calculation", and the like, and simultaneously executing the instructions in parallel by pipeline control.
[0004]
In addition, in many cases, each data grouped into a vector is subjected to individual processing on the way. This may include a conditional operation such as an IF statement in a Fortran DO loop.
[0005]
In the state where each data is put together as a vector, the processing of such a conditional operation becomes complicated.However, the method of the conditional vector operation, which is a method of performing this operation, includes a method using a vector mask, There are many different methods, such as using compression (or collection) and spreading (or decompression, expansion).
[0006]
A conditional vector operation method using a vector mask uses a vector of a bit string of “0” or “1” having the same number of bits as the number of elements of a vector to be operated as a vector mask. Is set to "1" for an element that satisfies the condition, and "0" is set for an element that does not satisfy the condition. Then, the condition of the vector to be operated is satisfied by referring to this vector mask. This is a processing method of storing the processing result of the THEN statement only in the element to be executed. Each bit constituting the vector mask is called a mask bit.
[0007]
In the method using the vector mask, after the operation processing on the THEN statement side is actually executed for all elements, the operation result is stored for elements for which the condition is not satisfied (the mask bit is "0"). In this method, only processing is suppressed. Therefore, in this method, useless calculations are performed for elements whose conditions are not satisfied, and if the conditions are not likely to be satisfied, a large amount of useless time is required for processing.
[0008]
As shown in FIG. 7, a conditional vector operation method using compression / diffusion is an element that satisfies a condition from an operation target vector using the vector mask described above (an element that performs processing on the THEN statement side). By extracting and arranging only the vectors, a compressed new vector is created, the process on the side of the THEN statement is executed on the compressed vector, and the result is again referred to the vector mask to obtain the original format. This is a processing method for returning (spreading).
[0009]
In the method using compression / diffusion, since the arithmetic processing on the side of the THEN statement is executed only for the elements for which the conditions are satisfied, the probability that the conditions are satisfied is low, or a sparse matrix (many of the elements of the matrix are "0"). It is suitable for use in the calculation of (a matrix with "").
[0010]
However, in the above-described conditional vector operation method using compression / spreading, since special compression / spreading processing is performed in addition to the normal operation, the compression / spreading may be performed at a high speed with few steps. desired.
[0011]
Japanese Patent Application Laid-Open No. 09-054769 discloses a technique for performing diffusion processing at a high speed, without going through processing for expanding a vector compressed in a register before storing an operation result to a vector length before compression. , A technique of designating an address after diffusion for each element of a vector and directly storing an operation result.
[0012]
This is explained with reference to the example of FIG. 7. Prior to Japanese Patent Application Laid-Open No. 09-054769, an operation is performed on a compressed vector to obtain vectors “B0, B4, B6, B8,. Before storing the result and obtaining the vectors "B0, A1, A2, A3, B4, A5, B6, A7, B8,...", Once "B0, 0, 0, 0, B4, 0, B6" , 0, B8,... "To the vector length form before compression, but in JP-A-09-054769, the operation result is directly stored without passing through this conversion. is there.
[0013]
Further, the above-described method of conditional vector calculation using compression / spreading includes redundant processing for spreading processing.
[0014]
This redundant processing is performed by diffusing the elements in order to check the position of the mask bit validity (that is, the mask bit having "1") of the vector mask which is the restoration position of each element of the compressed vector. When writing, all the mask bits of the vector mask are read regardless of whether they are valid or invalid.
[0015]
Since the purpose of vector operation is to perform high-speed processing, the number of steps required for each processing is reduced as much as possible to shorten the processing time, and a system is configured. It is required to eliminate it.
[0016]
As a conventional technique for eliminating the redundant processing, Japanese Patent Application Laid-Open No. 03-006663 discloses a technique in which a vector mask is stored in a mask register and vector compression is performed, and then vector diffusion is performed. Prior to this, that is, while the arithmetic processing on the compressed vector is being performed, data of valid mask bit intervals is generated in advance from the vector mask, and diffusion processing is performed using this. is there.
[0017]
For this reason, when performing the spreading process, there is no need to perform a redundant process of reading a large number of mask bits regardless of whether they are valid or invalid. The position of the writing destination is obtained by performing the process of adding the data of (1). This technique is also used in the above-mentioned Japanese Patent Application Laid-Open No. 09-054769.
[0018]
[Problems to be solved by the invention]
As described above, the conventional vector diffusion processing apparatus has the following problems.
[0019]
The conventional vector spreading processing apparatus disclosed in Japanese Patent Application Laid-Open No. 09-054769 has a problem in that it is only necessary to directly specify a post-spread address for each element of a compressed vector. It is desirable to specify the storage position of the operation result in more detail by using the number of the pipe at the storage destination or a pointer in each pipe for speeding up.
[0020]
In the conventional vector spreading processing apparatus disclosed in Japanese Patent Application Laid-Open No. 03-006663, each time a read element of a compressed vector is read, data is sequentially added to data at valid mask bit intervals. In order to obtain the position of the destination, there is a problem in that a step of addition processing is required for each diffusion processing of each element. For high-speed processing, it is desirable that the number of required steps be as small as possible.
[0021]
A first object of the present invention is to realize a vector diffusion processing apparatus which realizes high-speed processing by realizing diffusion processing in conditional vector calculation of a vector computer with minimum steps and without imposing a load on other processing. Is to provide.
[0022]
A second object of the present invention is to provide a vector diffusion processing device which realizes efficient and high-speed processing in writing a compressed vector element to a diffusion position.
[0023]
[Means for Solving the Problems]
In order to achieve the above object, the vector diffusion processing apparatus of the present invention refers to a mask register that stores a vector mask indicating whether a branch process is valid or invalid for each element of a vector in a conditional vector operation, and data of the mask register. A diffusion position calculation unit that calculates a diffusion position of each element of the compressed vector; and a diffusion position storage unit that stores a calculation result of the diffusion position by the diffusion position calculation unit. Before performing the diffusion process, the diffusion position is calculated in advance by the diffusion position calculation unit and stored in the diffusion position storage unit, and when the compression process of the compressed vector is performed, the diffusion vector is stored in the diffusion position storage unit. Each of the elements of the compressed vector is written into a vector register of a diffusion destination with reference to the spread position.
[0024]
3. The vector spreading processing apparatus according to claim 2, wherein the data indicating the spreading position stored in the spreading position storage unit includes a pipe number of a pipe including a write destination vector register, and the write destination vector register. And a write destination pointer in the field.
[0025]
The vector spreading processing apparatus according to the third aspect of the present invention includes a data transfer control unit that controls data transfer between pipes for each element of the vector based on the spreading position stored in the spreading position storage unit. It is characterized by the following.
[0026]
5. The vector diffusion processing device according to claim 4, wherein, based on the diffusion position stored in the diffusion position storage unit, a write control in the vector register as a write destination for each element of the vector. And a writing control unit for performing the following.
[0027]
6. The vector transfer processing device according to claim 5, wherein the data transfer control unit controls data transfer between the pipes of the vector by referring to the pipe number, and the write control unit controls the write destination. The writing control in the writing destination vector register is performed by referring to the pointer.
[0028]
According to a sixth aspect of the present invention, the data transfer control unit and the write control unit of the vector spreading processing device execute individual controls in the spreading process independently and in parallel with each other.
[0029]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a vector diffusion processing device according to an embodiment of the present invention.
[0030]
Referring to FIG. 1, a vector spreading processing apparatus according to an embodiment of the present invention includes a vector register 10a, 10b, 10c, 10d, 20a, 20b, 20c, 20d and a data transfer for transferring data between the vector registers. Unit 30, a mask register 50 for storing a vector mask mask, a diffusion position calculation circuit 60 for detecting the diffusion position of the compressed vector with reference to the vector mask in the mask register 50, and an output of the diffusion position calculation circuit 60 , A data transfer control circuit 80 for controlling the data transfer unit 30, and a write control circuit 90 for controlling the writing of the vector register to which the spread data is written.
[0031]
FIG. 2 is a block diagram showing a configuration of the mask register 50 and the diffusion position buffer 70 according to one embodiment of the present invention.
[0032]
Referring to FIG. 2, the configuration of the mask register 50 and the diffusion position buffer 70 according to an embodiment of the present invention is a configuration in which the number of pipes is four pipes of pipes 0 to 3, and the pipes 0 to 3 The compressed vector data in the vector registers 10a, 10b, 10c, and 10d refers to the mask register 50, and arranges each element at an element number position where mask bits are valid, thereby restoring the vector data to the form of vector data before compression. Then, the data is written to the vector registers 20a, 20b, 20c, and 20d of the pipes 0 to 3.
[0033]
Pipe 0 has vector registers 10a and 20a, pipe 1 has vector registers 10b and 20b, pipe 2 has vector registers 10c and 20c, and pipe 3 has vector registers 10d and 20d.
[0034]
The data transfer unit 30 transfers data between pipes of the vector register of each pipe by controlling the selection of transfer data between pipes by the data transfer control circuit 80.
[0035]
The mask register 50 stores a vector mask. Since each mask bit, which is an element of the vector mask, is associated with an element of the vector register to execute the compression / spreading process, it has the same number of vector elements as the vector register.
[0036]
After storing the vector mask to be processed in the mask register 50 and before executing the vector diffusion process, the diffusion position calculation circuit 60 refers to this vector mask and stores the compressed data in the vector registers 10a, 10b, 10c, and 10d. The diffusion position of each element of the vector to the destination vector register 20a, 20b, 20c, 20d is determined. This determination of the diffusion position is for obtaining a write destination pipe number and a write pointer, and the obtained write destination pipe number and write pointer are written to the diffusion position buffer 70.
[0037]
The spreading position buffer 70 stores the writing destination pipe number and the writing pointer, which are the data of the spreading position transmitted from the spreading position calculating circuit 60, and holds them until the end of the spreading process.
[0038]
The diffusion position buffer 70 outputs the data of the diffusion position every cycle when the vector diffusion process is executed, and transmits the write destination pipe number to the data transfer control circuit 80 and the write pointer to the write control circuit 90.
[0039]
The data transfer control circuit 80 generates a data transfer control signal from the data of the write destination pipe number received from the diffusion position buffer 70, and controls the data transfer unit 30 to select the transfer data between pipes.
[0040]
The write control circuit 90 generates a spread data write pointer to the vector register and a write signal based on the write pointer received from the spread position buffer 70, and transmits this to each of the vector registers 20a, 20b, 20c, and 20d. And execute write control. In each of the vector registers 20a, 20b, 20c, and 20d, upon receiving the write signal from the write control circuit 90, the vector element is written at the position designated by the spread data write pointer.
[0041]
1 and 2 showing an example of the present embodiment, an example in which the number of pipes is four is shown. However, the present invention is not limited to four pipes, and is applicable to a case of an arbitrary plurality of pipes. It is applied to
[0042]
Next, the operation of the present embodiment will be described in detail with reference to the drawings. FIG. 5 is a flowchart for explaining a process of calculating a diffusion position according to an embodiment of the present invention. FIG. 3 is a diagram illustrating an example of a configuration of a vector register, an example of a vector mask, and the like according to an embodiment of the present invention. FIG. 4 is a block diagram showing data stored in each register corresponding to FIG.
[0043]
Hereinafter, the process of calculating the diffusion position according to the present embodiment will be described for the example of the vector mask “100010101000...” Shown in FIG. 3 in the case of the four-pipe configuration shown in FIGS. .
[0044]
Referring to FIG. 5, the diffusion position calculation processing according to the present embodiment is performed in advance before the diffusion processing. First, the diffusion position calculation circuit 60 sets the mask bit of the mask register 50 to 4 Reading is sequentially performed from word 0 for the pipes (step 501).
[0045]
Then, the diffusion position calculation circuit 60 calculates the pipe number corresponding to the mask bit valid part and the write pointer of the vector register (that is, the diffusion position of the compressed vector) from the read mask bits (step 502), and performs the diffusion. The data of the difference position is stored in the position buffer 70 (step 503).
[0046]
Here, the vector mask “100010101000...” Shown in FIG. 3 corresponds to “word 0 / pipe 0”, “word 1 / pipe 0”, “word 1 / pipe” for the first 12 mask bits. 2 "and" word 2 / pipe 0 "are determined as the diffusion positions.
[0047]
FIG. 4 is a block diagram showing a configuration of data storage in the diffusion position buffer according to one embodiment of the present invention.
[0048]
Referring to FIG. 4, the diffusion position buffer indicates the diffusion position of each element determined by the diffusion position calculation circuit 60 according to the write pointer (that is, the number of words 0 to 3) and the information of the pipe number by “word 1 pipe”. Store as 2 ".
[0049]
Further, as shown in FIG. 5, the information storage positions of the respective diffusion positions in the diffusion position buffer 70 are sequentially stored from word 0 to word 1 and thereafter, and within each word from pipe 0 to pipe 1 and thereafter. In other words, the information indicating the diffusion position of each element of the compressed vector is sequentially stored in the diffusion position buffer 70 at the position next to the pipe 3 of the word 1 after the pipe 3 of the word 0.
[0050]
The above diffusion position calculation processing is executed for all the mask bits of the vector mask (step 504).
[0051]
Next, the operation of the diffusion process according to one embodiment of the present invention will be described in detail with reference to the drawings. FIG. 6 is a flowchart for explaining the operation of the diffusion process according to the embodiment of the present invention.
[0052]
Referring to FIG. 6, first, when the spreading process for the compressed vector is started, from the word 0 of the spreading position buffer 70, the pipe number and the write pointer of the position information of the writing destination for 4 pipes every cycle are sequentially read. The reading is executed (step 601), and the pipe number is transmitted to the data transfer control circuit 80 and the write pointer is transmitted to the write control circuit 90.
[0053]
At the same time, the data transfer unit 30 reads compressed vector data for a total of four pipes from the vector registers 10a, 10b, 10c, and 10d of the pipes 0 to 3 (step 604).
[0054]
The data transfer control circuit 80 provides the data transfer unit 30 with the transfer destination of the transfer of each element of the vector data between the pipes based on the pipe number which is the position information indicating the write destination pipe received from the diffusion position buffer 70. By performing control such as selection, the data transfer unit 30 uses the compressed vector data read from the vector registers 10a, 10b, 10c, and 10d as spread data as the spread vector write destination vector register 20a of the pipes 0 to 3. , 20b, 20c, and 20d (step 605).
[0055]
The write control circuit 90 sends the spread data write destination vector registers 20a, 20b, 20c, and 20d based on the write pointer that is the position information indicating the write position in each pipe received from the spread position buffer 70. , Write control by transmitting a write pointer and a write signal.
[0056]
The vector register 20a, 20b, 20c, 20d to which the spread data is written stores each element of the vector received from the data transfer unit 30 under the control of the writing position of the writing control circuit 90 (Step 606).
[0057]
When the above-described operation of the diffusion process is executed for all the position information stored in the diffusion position buffer 70 by one vector mask (step 607), the diffusion process of the set of vectors by the four pipes is completed.
[0058]
As described above, according to the conventional technique, the vector mask is read every cycle irrespective of the validity / invalidity of the mask bit, and the mask bits corresponding to the number of pipes per cycle are read directly from the mask register 50, whereby the data transfer unit 30 According to the present embodiment, a write pointer is designated for the control and vector registers 20a, 20b, 20c, and 20d, and a process with a large number of steps for performing write control is performed. Beforehand, the detailed data of the diffusion position is calculated in advance, and this data is stored in the diffusion position buffer 70 in a format that can be easily referred to when writing by a combination of the pipe number of the write destination and the data of the write pointer. During the execution of the write process, the data transfer control circuit 80 and the write control circuit 90 respectively select the pipe to be written to and the pointer of the pointer of the write destination. Is possible diffusion process at high speed with less steps to be performed.
[0059]
Although the present invention has been described with reference to the preferred embodiments and examples, the present invention is not necessarily limited to the above embodiments and examples, and various modifications may be made within the scope of the technical idea. Can be implemented.
[0060]
【The invention's effect】
As described above, according to the vector diffusion processing device of the present invention, the following effects can be achieved.
[0061]
First, before the vector diffusion processing, the diffusion position of the element of the compressed vector is calculated, and the data at that position is stored in a buffer in a format that can be easily referred to when writing. Does not require processing for calculating a position, such as reading a vector mask and adding distance data, so that high-speed diffusion processing can be performed with few steps.
[0062]
Second, there are two processes required in the process of writing the elements of the compressed vector to the diffusion position, namely, the process of selecting the pipe to write to for each element and the process of selecting the write destination in each pipe. The data transfer control circuit 80 and the write control circuit 90 execute the processing for designating the pointer independently and in parallel, so that the diffusion processing can be performed at high speed.
[Brief description of the drawings]
FIG. 1 is a block diagram illustrating a configuration of a vector diffusion processing apparatus according to an embodiment of the present invention.
FIG. 2 is a block diagram showing a configuration of a mask register and a diffusion position buffer according to an embodiment of the present invention.
FIG. 3 is a block diagram illustrating a configuration of a vector register according to an embodiment of the present invention, an example of a vector mask, and data stored in each register corresponding thereto.
FIG. 4 is a block diagram showing a configuration of data storage in a diffusion position buffer according to an embodiment of the present invention.
FIG. 5 is a flowchart illustrating a process of calculating a diffusion position according to the embodiment of the present invention.
FIG. 6 is a flowchart illustrating an operation of a diffusion process according to an embodiment of the present invention.
FIG. 7 is a diagram for explaining control of each element of a vector in compression / spreading processing according to an embodiment of the present invention.
[Explanation of symbols]
10a, 10b, 10c, 10d Read-side vector registers 20a, 20b, 20c, 20d Write-side vector registers 30 Data transfer unit 50 Mask register 60 Diffusion position calculation circuit 70 Diffusion position buffer 80 Data transfer control circuit 90 Write control circuit

Claims (6)

In a vector data diffusion processing device in a vector computer,
In a conditional vector operation, a mask register that stores a vector mask that indicates whether branch processing is enabled or disabled for each element of the vector,
A diffusion position calculation unit that calculates a diffusion position of each element of the compressed vector with reference to the data of the mask register;
A diffusion position storage unit for storing the calculation result of the diffusion position by the diffusion position calculation unit,
Before performing the diffusion processing of the compressed vector, the diffusion position is calculated in advance by the diffusion position calculation unit and stored in the diffusion position storage unit,
When performing the spread processing of the compressed vector, the spread position stored in the spread position storage unit is referred to, and each element of the compressed vector is written to a vector register of a spread destination. Vector diffusion processor. As data representing the diffusion position stored in the diffusion position storage unit,
A pipe number of a pipe having a vector register to be written;
2. The vector spreading apparatus according to claim 1, further comprising a write destination pointer in the write destination vector register. The vector spreading apparatus according to claim 2, further comprising a data transfer control unit that controls data transfer between pipes for each element of the vector based on the spreading position stored in the spreading position storage unit. Processing equipment. A write control unit configured to perform a write control in the vector register as a write destination for each element of the vector based on the diffusion position stored in the diffusion position storage unit. The vector diffusion processing device according to claim 3. The data indicating the diffusion position includes the pipe number and the write destination pointer,
The data transfer control unit includes:
By referring to the pipe number, controlling the transfer of data between pipes of the vector,
The write control unit includes:
5. The vector spreading processing apparatus according to claim 4, wherein the writing control in the writing destination vector register is performed by referring to the writing destination pointer. The data transfer control unit and the write control unit,
The vector diffusion processing apparatus according to claim 5, wherein individual controls in the diffusion processing are executed independently and in parallel with each other.

Images