Patent Grant
8990741
This application is based on and claims the benefit of priority from Japanese Patent Application No. 2012-108308, filed in Japan on May 10, 2012, the content of which is incorporated herein by reference in its entirety.
The present invention relates to a circuit design support device which supports circuit design.
More specifically, the present invention relates to a circuit design support device which supports semiconductor design employing high level synthesis (behavior synthesis) which automatically generates register transfer level from behavior description code (also called simply as “behavior description”, hereinafter).
In a conventional semiconductor integrated circuit design, the register transfer level (RTL) which describes behaviors of registers and combined circuits between registers included in a circuit is designed using hardware description language.
The circuit scale of the integrated circuit has been increased in recent years, which takes large amount of time to design RTL, resulting in a problem.
Then, a high level synthesis technique which generates automatically RTL using C language, C++ language, System C language, and so on that are high level language of which the level of abstraction is higher than RTL are proposed, and a high level synthesis tool implementing the same is commercially available.
Patent Literature 1 discloses that, as a pre-processing of input to the high level synthesis tool, in order to eliminate redundant memories, array description part which may cause generation of redundant memories is detected, and the array is automatically deleted, thereby obtaining an integrated circuit with a small circuit scale.
However, according to Patent Literature 1, only an array in which an index is expressed by only a loop variable is to be deleted, and thus there still exists a problem that the undeleted array requires memories for elements of the array.
Main object of the present invention is to solve the above problem; the invention aims to effectively eliminate redundant buffers.
According to the present invention, circuit design support device includes: a code inputting unit that inputs a behavior description code which describe behavior of a circuit which is a target of high level synthesis using a write access array to be accessed to write and a read access array to be accessed to read; an access order determining unit that analyzes the behavior description code, and determines an order of using each write access address when the behavior description code is executed and an order of using each read access address when the behavior description code is executed; and an access order changing unit that performs either one of a write access order changing process to change the order of using the write access addresses when the behavior description code is executed based on the order of using the read access addresses determined by the access order determining unit and a read access order changing process to change the order of using the read access addresses when the behavior description code is executed based on the order of using the write access addresses determined by the access order determining unit.
According to the present invention, an order of using write access addresses or an order of using read access addresses is changed, and the order of write accesses can be matched with the order of read accesses, thereby reducing the buffer.
The present invention will become fully understood from the detailed description given hereinafter in conjunction with the accompanying drawings, in which:
In describing preferred embodiments illustrated in the drawings, specific terminology is employed for the sake of clarity. However, the disclosure of the present invention is not intended to be limited to the specific terminology so selected, and it is to be understood that each specific element includes all technical equivalents that operate in a similar manner and achieve a similar result.
The present embodiment explains a semiconductor design support device which obtains a code which is capable to reduce the number of memories or registers located between circuits and to output an operation result with low latency.
More specifically, the semiconductor design support device extracts a dependency between operations and changes an order of operations.
Then, according to the semiconductor design support device related to the present embodiment, it is possible to obtain a hardware configuration (architecture) with high performance and small-circuit scale in a short time, independently from a capability of a designer.
In
The semiconductor design support device 1 corresponds to a circuit design support device.
A high level synthesis device 2 receives a behavior description code and generates RTL.
The semiconductor design support device 1 provides the behavior description code on which high level synthesis can be done to the high level synthesizing device 2, thereby obtaining targeted RTL.
A processing part 3 inputs the behavior description code written in the high level language which is a target of high level synthesis, detects access dependency of array, and changes an order of access to the arrays so that the device should be with low delay and a small circuit scale.
The processing part 3 corresponds to a code inputting unit, an access order determining unit, and an access order changing unit.
A table storage part 4 stores tables showing processed result of the processing part 3 and tables to which the processing part 3 refers.
A file storage part 5 stores the behavior description code inputted by the processing part 3.
Further, the file storage part 5 may also store the behavior description code after the processing part 3 changes the behavior description code.
Next, operation of the first embodiment will be explained.
The function of the flowchart is implemented by the processing part 3, and data are exchanged between the table storage part 4 and the file storage part 5 when necessary.
Behavior description code that is shown in
Here, the code shown in
When the description is inputted directly to the high level synthesizing device 2, N buffers (memories or registers) are required between the first loop and the next loop for storing N pieces of data.
There is a problem that it takes time to output the result “out”, since the processing cannot be moved from the first loop to the next loop unless the processing on the first loop has been finished.
First, when the semiconductor design support device 1 starts the operation, the processing part 3 receives a file of the behavior description which is a target of high level synthesis, the behavior description is the code shown in
Next, the processing part 3 reads out the code stored in the file storage part 5 (
This extraction is made by writing a loop ID (Identifier) and its description part (the number of lines) for each “for” sentence in the table storage part 4 as shown in
The table of
Next, the processing part 3 analyzes the behavior description code (
Here the table in which the above data are written is referred to as an array table.
The name of array is an array variable name described in the code.
As the accessing method, “write” is set if the corresponding array is in the left-hand side in the statement, and “read” is set if the corresponding array is in the right-hand side.
The number of accesses shows how many times the corresponding array has been accessed.
The number of accesses is the number of accesses for each of repeated times.
In case of the code in
The total number of accesses shows how many times the corresponding array has been accessed during “for” loop.
Here, the array of the write access is also called as a write access array; the array of the read access as a read access array.
In
Further, an element of array in the write access array is also called as a write access address, and an element of array in the read access array is also called as a read access address.
The write access address shows an address to be accessed by each write access, and the read access address shows an address to be accessed by each read access.
In the code in
Similarly, the value of each i in dat[i] of the second “for” sentence is the read access address.
Next, it is checked if a preceding iteration affects the next iteration in repetition sentence (“for” sentence), for each loop ID in the loop tables generated at step ST1 (at step ST4: DETERMINE DEPENDENCY).
Specifically, after an arbitrary variable in the repetition sentence is referred, if assignment is done in the variable, dependency is determined to be present, and if not, absent.
In case of the code in
On the other hand, in the second “for” sentence, there is no variable like this, and thus dependency is absent in the second “for” sentence.
Here, if an order of using the write access addresses or the read access addresses included in the “for” sentence having dependency is changed, an operation of a circuit which is a target of high level synthesis may change.
Therefore, the order of using the write access addresses or the read access addresses included in the “for” sentence having dependency cannot be changed.
On the other hand, if an order of using the write access addresses or the read access addresses included in the “for” sentence without dependency is changed, the operation of the circuit which is the target of high level synthesis does not change.
Therefore, the order of using the write access addresses or the read access addresses included in the “for” sentence without dependency can be changed.
The presence or the absence of dependency that is extracted at the step ST4 is added to a loop table in the table storage part 4.
Here, in the behavior description code, there is a case in which an order of execution is desired to be decided beforehand as processing.
For example, image data may be inputted in an order of coordinates.
Therefore, information showing that an order of processing is unchangeable may be added to the behavior description code, or as an option, the processing part 3 may analyze it and may generate dependency information.
Here, the behavior description code in
Next, the processing part 3 refers to the array table (
More specifically, the processing part 3 extracts the write access array and the read access array having a common name, makes pairs of the write access array and the read access array having the common name, for each pair of the write access array and the read access array, an order of using the write access address at the time of executing the behavior description code and an order of using the read access address at the time of executing the behavior description code are determined.
In the example of
In the access delay table, a cycle number, a write access address, a read access address, an access delay, and an access cost are written.
The cycle number is time information; the cycle number is described by incrementing a cycle such as the 1st cycle, the 2nd cycle, and so on.
The upper limit is the number of total accesses to the loop.
The write access address is an index of the array for write access obtained by developing the loop.
The numerals shown in the columns of the write access address respectively represent write access addresses; an order of using the write access addresses when the behavior description code is executed is written in the columns of the write access address.
The read access address is provided for each array to be accessed by the read access.
In case of the code example in
The read access address is an index of the array for read access obtained by developing the loop.
The numerals shown in the columns of the read access address respectively show read access addresses; an order of using the read access addresses when the behavior description code is executed is written in the columns of the read access address.
Here, as for read limitation, when it is possible to read two pieces of data at one cycle, such limitation can be given to the processing part 3 as additional information.
The access delay shows relation between the cycle number of the write access address and the cycle number of read access address which have the common address values.
For example, the write access address 6 occurs in the cycle 7, and on the other hand, the corresponding read access address 6 occurs in the cycle 13.
The access delay of the read access address 6 is “7”, which is the cycle number of the corresponding write access address 6.
Further, for example, the read access address 15 occurs in the cycle 2, and on the other hand, the corresponding write access address 15 occurs in the cycle 16.
The access delay of the read access address 15 becomes “16” which is the cycle number of the corresponding write access address 15.
As for the access cost, if there are a plurality of reads, the value of the largest access delay is set to the access cost.
In the code of
Since the read access address 1 and the read access address 14 occur after the read access address 15, in the 17th cycle, out[1]=dat[1]+dat[15−1] in the loop ID 2 can be executed for the read access address 1 and the read access address 14.
The read access address 2 and the read access address 13 occur further after one cycle, so that in the 18th cycle, out[2]=dat[2]+dat[15−2] in the loop ID 2 can be executed for the read access address 2 and the read access address 13.
In this manner, in the columns of the access cost of
As shown in
For example, as for the read access address 1 and the read access address 14, when the cycle number is 2, the write of the write access address 1 is done, and when the cycle number is 15, the write of the write access address 14 is done.
On the other hand, the processing of out[1]=dat[1]+dat[15−1] in the loop ID 2 is executed in the 17th cycle, so that the value of the write access address 1 should be retained in the buffer from the cycle number 2 to the cycle number 17, and the value of the write access address 14 should be retained in the buffer from the cycle number 15 to the cycle number 17 in order to absorb the timing difference between the write timing and the read timing.
Similarly, the value of the write access address 2 should be retained in the buffer from the cycle number 3 to the cycle number 18, and the value of the write access address 13 should be retained in the buffer from the cycle number 14 to the cycle number 18.
The same can be said for other address values, 16 buffers in total are required for retaining the written values for the read access address 0 to 15.
In this manner, when the code of
In the semiconductor design support device 1 according to the present embodiment, such mismatch of the order of using the write access and the order of using the read access can be suppressed.
Next, the processing part 3 changes the order of the write access or the read access (at step ST6: CHANGE ORDER OF ACCESS).
More specifically, the processing part 3 refers to the loop table (
Access Case 1: “write dependency: present” and “read dependency: present”
Access Case 2: “write dependency: absent” and “read dependency: present”
Access Case 3: “write dependency: absent” and “read dependency: absent”
Access Case 4: “write dependency: present” and “read dependency: absent”
Here, “write dependency: present” means if the order of using the write access addresses is changed, an operation of a circuit which is a target of high level synthesis changes, so that the order of using the write access addresses cannot be changed.
On the other hand, “write dependency: absent” means even if the order of using the write access addresses is changed, the operation of the circuit which is the target of high level synthesis does not change, so that the order of using the write access addresses can be changed.
Further, “read dependency: present” means if the order of using the read access addresses is changed, the operation of the circuit which is the target of high level synthesis changes, so that the order of using the read access addresses cannot be changed.
On the other hand, “read dependency: absent” means even if the order of using the read access addresses is changed, the operation of the circuit which is the target of high level synthesis does not change, so that the order of using the read access addresses can be changed.
In case of Access Case 1, since the orders of the write access and the read access should be fixed, the performance cannot be improved by the present embodiment, and the process terminates here.
In case of Access Case 2, since the order of the write access can be changed, the order of the write access can be changed according to the order of the read access.
Here, the processing to change the order of the write access to conform to the order of the read access is called as a write access order changing process.
In case of Access Case 3, since both have no dependency, one of the order of the write access and the read access is fixed, and the other is changed.
To determine which of the orders of the write access and the read access to be fixed, whichever the access cost is smaller can be chosen.
In case of Access Case 4, since the order of the read access is changeable, the order of the read access can be changed according to the order of the write access.
Here, the processing to change the order of the read access to conform to the order of the write access is called as a read access order changing process.
Here, Access Case 2 will be explained.
The write access array dat[i] of the first loop of
On the other hand, a read fixed option is given to the second loop.
This is an option to be given when the loop sentence itself has no dependency, but a design requires to execute the processing in an order as shown in the code such as 0, 1, 2, 3 . . . for “out” array.
The processing part 3 analyzes this option, and it is determined at the dependency determination, the dependency exists in the corresponding loop ID (“read dependency” is present in the second loop).
As discussed above, since there is no dependency in the write side (the first loop), the write access can be started from any i of “for” sentence.
Namely, the write access is implemented in the order requested by dat (read access array) of the read side (the second loop), and the result of the write access is transferred to dat of the read side, thereby improving the processing speed.
In this manner, in case of Access Case 2, the processing part 3 writes the changed write access addresses in the access delay table so as to implement the write in the order of read accesses.
Here, the contents of the second loop of the code of
In the access delay table of
Here, since the delay cost is 0, no buffer is necessary between the first loop and the second loop of
When the order of write accesses is changed as shown in
Further, as discussed above, since there is no timing difference between the write access and the read access, no buffer is necessary.
When the method of the present embodiment is not applied, if high level synthesis of the code of
N buffers are necessary, and the processing of the second loop (B1 in
When the method of the present embodiment is applied to change the order of write accesses, and the high level synthesis of the code of
According to the present embodiment, an intermediate buffer is unnecessary as shown in
Access Case 3: As for “write dependency: absent” and “read dependency: absent”, a processing with write dependency and without read dependency (Access Case 4) and a processing without write dependency and with read dependency (Access Case 2) are both carried out and either one with less access cost can be chosen between these two processings after processing these two cases.
Access Case 4: As for “write dependency: present” and “read dependency: absent”, the order of read accesses is changed so as to reduce the access cost.
The example of code of
When the code in
First, the processing part 3 groups two read access addresses which are used with synchronization when the code of
For example, since the read 1 access address 7 and the read 2 access address 8 are processed at the same time in the loop ID2, they become a pair.
Then, the processing part 3 defines as a reference read access address, a read access address having an address value identical to an address value of a write access address whose timing of use is later between two read access addresses of the pair, for each pair of the read access addresses. And, the processing part 3 defines as a reference read access address, the write access address having the identical address value to the reference read access address.
Further, the processing part 3 changes the order of using the read access addresses by a unit of a pair of the read access addresses so that the order of using the reference read access addresses should be matched with the order of using the reference write access addresses.
In the example of
Further, the read access addresses of (15, 14, 13, 12, 11, 10, 9, 8) become reference read access addresses.
In addition, the write access address having the same address value with the reference read access address become a reference write access address.
Then, the processing part 3 changes the order of using the read access address by a unit of a pair of the read access addresses so that the order of using the reference read access addresses should be matched with the order of using the reference write access addresses of (8, 9, 10, 11, 12,13,14,15).
As a result, the order of the read access addresses is changed to the order of (7,8), (6,9), (5,10), (4,11), (3,12), (2,13), (1,14), and (0,15) as shown in
The following shows a concept of the algorithm of changing the order of the read access addresses.
Since the order of using the write access address is fixed, the process which can be implemented first in the loop ID2 is the read access address 7 and the read access address 8 which can be processed directly after processing the write access address 7 and the write access address 8.
Therefore, a pair of the read access address 7 and the read access address 8 is assumed to be the first target of the read access.
In this case, it is possible to access the read access address 7 at the time of cycle 8 when the write access address 7 is accessed, so that the access delay of the read access address 7 is 8.
Further, it is possible to access the read access address 8 at the time of cycle 9 when the write access address 8 is accessed, so that the access delay of the read access address 8 is 9.
As a result of this, the access delay of the pair of the read access address 7 and the read access address 8 is 9.
Namely, at the time of the 9th cycle, since both values of the read access address 7 and the read access address 8 are available, out[7]=dat[7]+dat[15−7] in the loop ID2 can be executed.
For the other pairs, since both data are not available, the processing cannot be started.
In the next cycle, the write access address 9 is accessed, since the write access address 6 is already done in the cycle 7, at the time of cycle 10, the read access address 6 and the read access address 9 become able to be operated.
In this way, the order of processing the read access addresses is changed, so that the operation is processed sequentially from the process which becomes able to be operated.
When the order of the read access addresses is changed to the order shown in
Therefore, the processing time can be largely reduced compared with 23 (
Further, in case of the order of
Therefore, it is possible to largely reduce the number of buffers compared with 16 which is the number of buffers which are required when the behavior description code of
In this way, the order of the read access addresses is changed and the delay cost becomes small, the delay of the data output from the loop ID2 becomes small, and as a result, low latency processing can be implemented.
Namely, when the method of the present embodiment is not applied, N registers or memories whose number is identical to the number of pieces of data are necessary; however, the order of access is changed according to the present embodiment, thereby reducing the number of buffers to a half.
Further, when the method of the present embodiment is not applied, the processing of the loop ID2 can be started after the 16th cycle. The order of access is changed according to the present embodiment, so that the processing of the loop ID2 can be started after the 8th cycle, thereby implementing the low latency processing.
The capacity of buffer is reduced to a half, and further, the calculation of the loop 2 can be started while the calculation of the loop 1 is carried out, and thereby high-speed processing can be done.
As shown in
As has been explained above, in either case of Access Case 2 or Access Case 4, the order of the write access or the read access is derived, which enables the number of buffers which will be generated by the high level synthesis to decrease, compared with the case in which the behavior description code is used directly for the high level synthesis. The order of using the write access addresses or the read access addresses is changed according to the derived order of the write access or the read access.
Then, finally the processing part 3 generates the code in which the order of accesses is changed, and the code in which the order of accesses is changed is stored in the file storage part 5 (at step ST7: GENERATE CODE).
The changed description part is determined by the number of lines described in the loop table (
Here, the generation of code is done by developing the loop based on the access delay table (
Namely, the processing part 3 rewrites the description of the first loop sentence of
Further, the processing part 3 rewrites the description of the second loop sentence of
Here, instead of rewriting the behavior description code by the processing part 3, the processing part 3 can output information (example of information of order of using the write access addresses) showing the contents of the changed write access of
Similarly, the processing part 3 can output information (example of information of order of using the read access addresses) showing the contents of the changed read access addresses of
As discussed above, the present embodiment has been explained the semiconductor design support device.
The semiconductor design support device includes the processing part inputting the behavior description which describes the behavior of the design circuit which is a target of the high level synthesis, extracting the dependency relationship between a plurality of array variables from the inputted behavior description, and generating a code by which the order of write access or read access to the array is changed so as to reduce the number of memories or registers implementing the array or to decrease the delay.
At the final section, an example of the hardware configuration of the semiconductor design support device 1 will be explained by referring to
The semiconductor design support device 1 is a computer, and each element of the semiconductor design support device 1 is implemented by executing processes by pro grams.
As for the hardware configuration of the semiconductor design support device 1, an operation device 901, an external memory 902, a main memory 903, a communication device 904, and an input/output device 905 are connected to buses.
The operation device 901 is a CPU (Central Processing Unit) which executes programs.
The external memory 902 is, for example, a ROM (Read Only Memory), a flash memory, and a hard disk drive.
The main memory 903 is a RAM (Random Access Memory).
The communication device 904 carries out communication with the high level synthesizing device 2 and other devices.
The input/output device 905 is, for example, a mouse, a keyboard, a display device.
The programs are usually stored in the external memory 902, in a status of being loaded to the main memory 903, read sequentially and executed by the operation device 901.
The programs are to implement a function which has been explained as “the processing part 3” shown in
Further, the external memory 902 also stores an operating system (OS), at least a part of the OS is loaded to the main memory 903, while executing the OS, the operation device 901 executes the programs to implement the function of “the processing part 3”.
Further, information, data, signal values, variables showing the result of “determination of—”, “extraction of—”, “calculation of—”, “derivation of—”, “analysis of—”, “detection of—”, “set of—”, “registration of—”, “selection of—”, “generation of—”, “input of—”, “output of—” and so on are stored in the main memory 903 as files.
Further, an encryption/decryption key, a random number, or a parameter can be stored in the main memory 903 as files.
Here, the configuration of
Further, the high level synthesizing device 2 of the present embodiment can also be configured like the hardware configuration of
Further, a method of supporting the circuit design related to the present invention can be implemented according to the procedure which has been shown in the present embodiment.
Numerous additional modifications and variations are possible in light of the above teachings. It is therefore to be understood that, within the scope of the appended claims, the disclosure of this patent specification may be practiced otherwise than as specifically described herein.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2012-108308 | May 2012 | JP | national |
| Number | Name | Date | Kind |
|---|---|---|---|
| 6532584 | Nishida et al. | Mar 2003 | B1 |
| 20030028854 | Nishida et al. | Feb 2003 | A1 |
| 20040024993 | Parthasarathy | Feb 2004 | A1 |
| 20050125762 | Hattori et al. | Jun 2005 | A1 |
| 20050268271 | Gutberlet et al. | Dec 2005 | A1 |
| 20050273752 | Gutberlet et al. | Dec 2005 | A1 |
| 20060288337 | Nishi | Dec 2006 | A1 |
| 20070234267 | Kakui | Oct 2007 | A1 |
| 20070250303 | Imai et al. | Oct 2007 | A1 |
| 20080148215 | Gutberlet et al. | Jun 2008 | A1 |
| Number | Date | Country |
|---|---|---|
| 8-44773 | Feb 1996 | JP |
| 11-184895 | Jul 1999 | JP |
| 2001-209664 | Aug 2001 | JP |
| 2002-269162 | Sep 2002 | JP |
| 2004-54641 | Feb 2004 | JP |
| 2005-173648 | Jun 2005 | JP |
| 2005-189939 | Jul 2005 | JP |
| 2006-350849 | Dec 2006 | JP |
| 2007-272797 | Oct 2007 | JP |
| 2007-287044 | Nov 2007 | JP |
| 2008-501185 | Jan 2008 | JP |
| 2010-238054 | Oct 2010 | JP |
| Entry |
|---|
| Lai et al, “A Read-Write Aware DRAM Scheduling for Power Reduction in Multi-Core Systems”, Jan. 20-23, 2014, Design Automation Conference (ASP-DAC), 2014 19th Asia and South Pacific, pp. 604-609. |
| Number | Date | Country | |
|---|---|---|---|
| 20130305198 A1 | Nov 2013 | US |
