• 検索結果がありません。

Stream-Driven Processor for Image Processing “SDP-i” and its Application to Hough Transform

N/A
N/A
Protected

Academic year: 2021

シェア "Stream-Driven Processor for Image Processing “SDP-i” and its Application to Hough Transform"

Copied!
3
0
0

読み込み中.... (全文を見る)

全文

(1)

要 旨

画像処理向きストリーム駆動プロセッサ

SDP-i

とその

Hough

変換への応用

田原 匡浩

近年, IoT(Internet of Things)の普及に伴い, 様々な用途に実時間画像認識システムが導 入されている. 認識精度向上のため, 高フレームレート化や高精細化, アルゴリズムの高度 化が進んでおり, 計算量が増加傾向にある. そのため, 画像処理専用プロセッサで高速処理 する必要性が高まっている. また, 本研究室では, 大規模なデータ構造(ベクタ, 行列)を効 率よく処理可能なストリームフロープログラミング (SFP : Stream-Flow Programming)

モデル, 及び, SFP モデルのハードウェア実装であるストリーム駆動プロセッサ(SDP :

Stream-Driven Processor) の研究を行っている. SDPは、オペランドデータのストリーム を生成・吸収するStream MemorySMem)と、要素単位でストリームを操作するデータ 駆動プロセッサ(DDP : Data-Driven Processor)で構成される。

本研究では、画像処理向きストリーム駆動プロセッサ(SDP-i : SDP for Image Process- ing)のプロセッサコアであるDDPアーキテクチャを提案する。提案アーキテクチャでは、

演算のスループットを向上させるために、DDP内の拡張パイプライン型演算器アレイの各 ステージに並列化したALUユニット,またはスカラデータメモリを導入する。また, オペラ ンドのシャッフル, コピーなどの新しい命令を追加し、2つの32ビットオペランドを含むパ ケット形式を定義する。提案したDDPアーキテクチャを65nm CMOSセルライブラリを 用いて論理合成,評価を行った。その結果、Hough変換は、dual DDP1つのコアでρ−θ 空間への変換が、もう一方のコアで投票処理がリアルタイムに実行できることを確認した。

キーワード データ/ストリーム駆動プロセッサ, パイプライン型演算器アレイ, 並列処理

i

(2)

Abstract

Stream-Driven Processor for Image Processing “SDP-i” and its Application to Hough Transform

Masahiro TABARA

With the diverse spread of Internet of Things (IoT) technology, real time image recognition has been introduced into various application systems. In order to improve the recognition accuracy, higher frame rate, higher definition, and more advanced al- gorithm are progressing and its computational complexity thus tends to increase. This trends demand high-performance processor dedicated to image processing in place of general-purpose processors. Therefore, our laboratory has studied a stream flow pro- gramming (SFP) model that can efficiently process large-scale data structures (vec- tors, matrices), and a stream-driven processor (SDP) which is a hardware implementa- tion of the SFP model. The SDP is composed of a stream memory (SMEM) produc- ing/consuming streams of operand data and a data-driven processor (DDP) operating streams in element-by-element.

In this paper, we propose a DDP architecture of SDP for image processing (SDP-i).

In the proposed architecture, we introduce a pair of ALU units or a scaler data memory into every stage of the FParray in the DDP in order to improve throughput of operations in the DDP. For this architecture, we additionally adopt some of new instructions such as shuffle or copy of operands and define a new packet format including two pairs of 32 bit operands. After that, we conducted post-synthesis evaluation of the proposed DDP architecture by using 65 nm CMOS cell library. As a result, we verified that

ii

(3)

Hough transform can be executed in real time on dual DDP’s, where the transform to the rho-theta space is executed on one DDP core and the voting process is done on the other core.

key words Data/Stream-Driven processor, Pipelined array of FP’s, LUT and DMem, Parallel Processing

iii

参照

関連したドキュメント

In the second computation, we use a fine equidistant grid within the isotropic borehole region and an optimal grid coarsening in the x direction in the outer, anisotropic,

Keywords: continuous time random walk, Brownian motion, collision time, skew Young tableaux, tandem queue.. AMS 2000 Subject Classification: Primary:

This can be seen even more clearly from the discrete transforms: the famous uncertainty principles of Balian-Low for the discrete Gabor transform [Bali81, Daub90] and Battle for

The main purpose of the present paper is a development of the fibering method of Pohozaev [17] for the investigation of the inhomogeneous Neumann boundary value problems

Correspondingly, the limiting sequence of metric spaces has a surpris- ingly simple description as a collection of random real trees (given below) in which certain pairs of

Using the batch Markovian arrival process, the formulas for the average number of losses in a finite time interval and the stationary loss ratio are shown.. In addition,

We formalize and extend this remark in Theorem 7.4 below which shows that the spectral flow of the odd signature operator coupled to a path of flat connections on a manifold

The configurations of points according to the lattice points method has more symmetry than that of the polar coordinates method, however, the latter generally yields lower values for