ゼミナール発表

日時: 1月21日(月)3限 (13:30-15:00)


会場: L1

司会: 米田友和
王昊 1151209: M, 2回目発表 中島康彦,井上美智子,嶋田創,姚駿

title: An Instruction Translating Method for A Multi-Mode Accelerator by Exploiting GCC Vectorizer

abstract: To acquire high processing performance and to hide difference between hardware structures, general-purpose graphics processing unit (GPGPU), which includes many function units and employs programming language that specifying parallel processing explicitly like CUDA. However, in order to pull out the desired performance, considerable hardware tuning cost and detail understanding of hardware structure is necessary. On the other hand, we have proposed a linear array pipeline processor (LAPP), which characterized by implementing a structure that includes a number of combinations of local memory and FUs to achieve the balance between reduction of power consumption and improving processing performance. By inserting the pre-fetch information into the existing VLIW instruction sequence, LAPP realizes high performance on executing loops with no dependency between iterations. However, instead of this advantage, LAPP has constraint on the high-speed executable loops. Also, it’s hardly adapted to a general processor in which has a different instruction set without redesigning the accelerator portion of LAPP. In this paper, we describe a new structure of accelerator to alleviate the constrains while following the idea of LAPP, and an instruction translating method for generating instructions oriented to the accelerator by exploiting the GCC vectorizer. Currently, We are now implementing control flow, data flow and memory access patterns analysis based on the information of UNCPROP. Comparing to the LAPP, we get a reduction of 65% on average in FUs array stages for some simple programs.


language of the presentation: English

発表題目: GCCvectorizerを利用した演算器アレイ向け命令変換手法

発表概要: 多数の演算ユニットを備えるGPGPUでは,CUDA 等明示的な並列処理の記述が必要なプログラミング言語を採用することにより,ハードウェアの差異を隠蔽することと,処理の高速化を両立している.ただし,所望の性能を引き出すためには,ハードウェア構造の理解と,相当のチューニングコストが必要である.一方,我々は,演算速度向上と消費電力低減の両立を目的として,演算器とローカルメモリの組を多数配置する構成の演算器アレイ型アクセラレータ(LAPP)を提案してきた.しかし,従来のLAPPには,既存のVLIW 命令列にプリフェッチ情報を挿入するだけで,イタレーション間に依存関係のないループを高速実行できる利点がある代わりに,適用可能なループに制約がある.また,命令セットが異なる基本プロセッサに適用するためには,アクセラレータ部分を新たに設計する必要がある.本発表では,LAPP の実行方式を踏襲しつつ従来の制約を緩和する新たなアクセラレータ構成方式,および,GCC vectorizer を利用する命令生成方式について述べる.現在,Uncprop 情報に基づき,コントロールフロー解析,データフロー解析,および,メモリアクセスパターン解析を行い,簡単な構造のループに対して,アクセラレータ用命令列を生成できる段階にある.簡単なプログラムに対して適用したところ,LAPP に比べて,平均65%の命令行数を削減できることがわかった.また,32行構成を仮定した場合,行数の削減により生じた空き演算器を使用すると,LAPP に比べて,2倍から8倍の性能向上を期待できることがわかった.

 
濱田龍之介 1151085: M, 2回目発表 池田和司,杉本謙二,久保孝富
title: Applying Nonparametric Bayesian Approach to Multiple Time Series towards Prediction of Driving Operations
abstract: Prediction of driving behaviors is important problem in developing the next-generation driving support system. In order to take account of diverse driving situations, it is necessary to deal with multiple time series data considering commonalities and differences among them. In this study we utilize the beta process autoregressive hidden Markov model (BP-AR-HMM) that can model multiple time series considering common and different features among them using the beta process as a prior distribution, and model multiple driving operation time series data. We applied the BP-AR-HMM to actual driving operation data to estimate VAR process parameters that represent the driving behaviors, and with the estimated parameters we predicted the driving operations of unknown test data. The result suggests that it is possible to predict driving operations in actual environment with BP-AR-HMM.
language of the presentation: English
 
金周 1151212: M, 2回目発表 加藤博一,横矢直和,武富貴史,宮崎純

title: An indoor positioning system for supporting maintenance using a tablet device.

abstract: For managin servers in data centers, I suggest a system that for easily checking maintenance lists using a markerless Augmented Reality(AR) system. This system uses only a built-in camera, an accelerometer and a gyroscope to calculate positions in indoor locations in which GPS(Global Positioning System) signals cannot be received. First, positions are calculated in indoor situations using accelerometer and gyroscope sensors. Then to compute the positions more accurately, image processing techniques with the built-in camera are used. For this system, it is important to switch between these two techniques automatically. For this purpose, a system for managing servers, that combines Kourogi's indoor positioning technique and a Panorama Mapping and Tracking System, is proposed.


language of the presentation: Japanese