公开(公告)号：JP2010271755A

公开(公告)日：2010-12-02

申请号：JP2009120575

申请日：2009-05-19

Applicant: Internatl Business Mach Corp ,  インターナショナル・ビジネス・マシーンズ・コーポレーションInternational Business Maschines Corporation

Inventor： YOSHIZAWA TAKERO ,  SHIMIZU SHUICHI ,  DOI ATSUSHI

IPC: G06F17/16

CPC classification number: G06F9/325

Abstract: PROBLEM TO BE SOLVED: To reduce accumulation of an error by calculating the error of prediction systematically in quickening a processing loop by speculative pipelining in a multi-core or multi-processor system.
SOLUTION: In an environment of a multi-core or a multi-processor system, the processing of each of the loops of a control block is suitably assigned to individual cores or processors as individual threads by a speculative pipelining method. A correction value is calculated by using the prediction variable value of the pertinent thread with respect to output based on the prediction input. The calculation of the correction value is applied as the gradient value as the approximate value of the primary differential coefficient of the prediction variable value. Generally, for a simulation system, there are several variable values, so that the primary differential coefficients are represented as a Jacobian matrix. Then, a matrix whose components are respectively the gradient value as the approximate value of the primary differential coefficient is called a Jacobian matrix. The calculation of the correction value is performed on the basis of the Jacobian matrix defined in this way.
COPYRIGHT: (C)2011,JPO&INPIT

Abstract translation: 要解决的问题：通过在多核或多处理器系统中通过推测流水线加快处理循环来系统地计算预测误差，减少误差累积。 解决方案：在多核或多处理器系统的环境中，通过推测流水线方法将控制块的每个循环的处理适当地分配给各个核或处理器作为单独的线程。 基于预测输入，通过使用相对于输出的相关线程的预测变量值来计算校正值。 将校正值的计算作为梯度值作为预测变量值的一次微分系数的近似值。 通常，对于仿真系统，存在几个可变值，使得初级微分系数被表示为雅可比矩阵。 然后，将其分量分别作为初级微分系数的近似值的梯度值的矩阵称为雅可比矩阵。 基于以这种方式定义的雅可比矩阵来执行校正值的计算。 版权所有（C）2011，JPO＆INPIT

公开(公告)号：JP2010134614A

公开(公告)日：2010-06-17

申请号：JP2008308660

申请日：2008-12-03

Applicant: Internatl Business Mach Corp ,  インターナショナル・ビジネス・マシーンズ・コーポレーションInternational Business Maschines Corporation

Inventor： CANEDO ARQUIMEDES ,  YOSHIZAWA TAKERO ,  KOMATSU HIDEAKI

IPC: G06F9/45 ,  G06F15/167

CPC classification number: G06F8/30 ,  G06F8/456

Abstract: PROBLEM TO BE SOLVED: To speed up iteratively executed processing when generating clusters in order to assign processing to individual processors in a multi-processor environment. SOLUTION: In a full vehicle simulation system, when clusters are formed based on strongly connected components, a single cluster identified as a fat cluster, in which there are at least a predetermined number of blocks or an expected processing time exceeds a predetermined threshold, is subjected to unrolling processing to make multiple copies of the processing of the fat cluster and to assign the copies to individual processors. Thus, the processing is executed by the multiple processor devices in a pipelined manner, so that the processing can be quickly achieved. When one time processing result of the fat cluster to be iteratively executed depends on the previous processing results, an input value needed for execution is generated based on certain prediction, and the fat cluster is speculatively executed. COPYRIGHT: (C)2010,JPO&INPIT

Abstract translation: 要解决的问题：为了在生成集群时加快迭代执行的处理，以便在多处理器环境中为单个处理器分配处理。 解决方案：在全车模拟系统中，当基于强连接的组件形成集群时，识别为脂肪群的单个群集（其中存在至少预定数量的块或预期处理时间）超过预定的 阈值，进行展开处理，以进行胖群处理的多个副本，并将副本分配给各个处理器。 因此，处理由多个处理器设备以流水线的方式执行，使得可以快速地实现处理。 当要迭代执行脂肪群的一次处理结果取决于先前的处理结果时，基于某些预测产生执行所需的输入值，并且推测性地执行脂肪群。 版权所有（C）2010，JPO＆INPIT

公开(公告)号：JP2013164657A

公开(公告)日：2013-08-22

申请号：JP2012026145

申请日：2012-02-09

Applicant: Internatl Business Mach Corp ,  インターナショナル・ビジネス・マシーンズ・コーポレーションInternational Business Machines Corporation

Inventor： YOSHIZAWA TAKERO ,  MAEDA KUMIKO ,  SHIMIZU SHUICHI

IPC: G06F9/45 ,  G06F17/12

CPC classification number: G06F8/35 ,  G06F8/41 ,  G06F8/451

Abstract: PROBLEM TO BE SOLVED: To provide a technique that can accelerate execution parallelized on multi processors or multiple cores, by well-balanced segmentation of a code of a model.SOLUTION: Blocks with internal states and blocks without any outputs among functional blocks included in a block diagram describing a model are focused on, and segments consisting of sets of blocks needed to calculate inputs to each block with an internal state and each block without any outputs are extracted by traversing a graph from the blocks with internal states and the blocks without any outputs in the opposite direction of the dependencies, where blocks can be duplicated among segments. Next, the segments are merged to reduce duplication, and the number of segments is reduced to a number for parallel execution. Here, the duplication between the segments is reduced by merging segments with many of the same blocks, and the number for parallel execution is typically the number of available cores or processors.

Abstract translation: 要解决的问题：提供一种可以通过对模型的代码进行均衡分割来加速在多处理器或多个核心上并行化执行的技术。解决方案：具有内部状态和块的块，不包括在功能块中的任何输出 描述模型的框图集中在一起，并且由包含内部状态的每个块的输入计算所需的块组成的段和不具有任何输出的每个块的块被提取，通过从具有内部状态的块和没有任何块的块遍历图 输出方向与依赖关系相反，其中块可以在段之间复制。 接下来，段被合并以减少重复，并且段的数量被减少到用于并行执行的数量。 这里，通过将段与许多相同的块合并来减少段之间的复制，并行执行的数量通常是可用的核心或处理器的数量。

公开(公告)号：JP2011170732A

公开(公告)日：2011-09-01

申请号：JP2010035691

申请日：2010-02-22

Applicant: Internatl Business Mach Corp ,  インターナショナル・ビジネス・マシーンズ・コーポレーションInternational Business Maschines Corporation

Inventor： YOSHIZAWA TAKERO ,  KOMATSU HIDEAKI

IPC: G06F9/44

CPC classification number: G06F8/456

Abstract: PROBLEM TO BE SOLVED: To accelerate execution of a program by parallelization in a simulation system. SOLUTION: A functional block is divided into strands, and a strand with the largest calculation time is found from a strand set (a step 604). In the strand with the largest calculation time, a movable block is found (a step 606), the calculation time of the strand as a result of moving to a strand in an input direction or an output direction according to nature of the movable block, and moving of the block between strands in which reduction of calculation time with an original strand becomes maximum is performed (a step 614). Thus, processing returns to a step of finding a strand with the largest calculation time. When it reaches a state where the calculation time of the block no longer decreases even when the movable block is moved (a step 612), the processing is ended. The strand of the functional block obtained in this way is converted into a corresponding source code such as C, and is further complied to be assigned to an individual core or an individual processor in order to execute it. COPYRIGHT: (C)2011,JPO&INPIT

Abstract translation: 要解决的问题：通过仿真系统中的并行化来加速程序的执行。 解决方案：功能块被划分为股线，并且从股线组发现具有最大计算时间的股线（步骤604）。 在具有最大计算时间的股线中，发现可动块（步骤606），根据可移动块的性质，由于沿输入方向或输出方向移动到股线的结果，股线的计算时间， 并且执行块之间的块与原始链的减少计算时间变为最大的步骤（步骤614）。 因此，处理返回到找到具有最大计算时间的线的步骤。 当达到即使可移动块移动时块的计算时间也不再减小的状态（步骤612），处理结束。 以这种方式获得的功能块的链转换成相应的源代码，例如C，并且进一步被指定为分配给单独的核或单个处理器以执行它。 版权所有（C）2011，JPO＆INPIT

公开(公告)号：JP2011113449A

公开(公告)日：2011-06-09

申请号：JP2009271308

申请日：2009-11-30

Applicant: Internatl Business Mach Corp ,  インターナショナル・ビジネス・マシーンズ・コーポレーションInternational Business Maschines Corporation

Inventor： MURASE MASAKATA ,  DOI OKIHIRO ,  MAEDA KUMIKO ,  YOSHIZAWA TAKERO ,  KOMATSU HIDEAKI

IPC: G06F9/45

CPC classification number: G06F8/4441

Abstract: PROBLEM TO BE SOLVED: To provide a code generation technology can generate an executable code optimized as much as possible regarding resource utilization and execution speed, on hybrid systems comprising a plurality of computer systems which can be connected to one another via a network.
SOLUTION: An optimized executable code is generated by the processing of generating an optimization table about library components of a source code, automatic computing resource assignment using the resultant optimization table, and the execution of network embedding for hybrid systems which are connected to one another via the network.
COPYRIGHT: (C)2011,JPO&INPIT

Abstract translation: 要解决的问题：为了提供代码生成技术，可以在包括可以经由网络彼此连接的多个计算机系统的混合系统上生成尽可能优化的关于资源利用和执行速度的可执行代码 。 解决方案：通过处理生成关于源代码的库组件的优化表，使用所得到的优化表的自动计算资源分配以及用于混合系统的网络嵌入的执行来生成优化的可执行代码 彼此通过网络。 版权所有（C）2011，JPO＆INPIT