Showing papers by "Erdal Oruklu published in 2021"
01 Aug 2021
TL;DR: This work provides a summary of the dataflow processor’s instruction set and details how to write programs for this processor, and presents 12 building block programming patterns that are expected to benefit future compilers.
Abstract: In this era where the growth of computer performance is significantly flattening, new parallel computer architectures are needed to continue to fuel that growth. One candidate architecture to fuel this growth is a recently proposed dataflow processor that is meant to be tiled together into a fabric to create a polymorphic computer (i.e. a reconfigurable computer whose architecture can be changed to match the software) where program instructions can be independently migrated in a fine-grained manner around the fabric to logically change the computer architecture. This work provides a summary of the dataflow processor’s instruction set and details how to write programs for this processor. The techniques that are presented to write programs for this processor are diagramming and 12 building block programming patterns. The 12 patterns are: Operations with a Constant, Simple Copy, Pitch (Discard), Counters, Critical Sections, Funnels (Gathers), Distribution Tree (Scatters), Lookup Tables, Single Value Selection, Single Value Steering, Switch (Multi-Value Steering), and Loops. These patterns are expected to benefit future compilers. In addition, the ideal timing results of several test algorithms that are implemented in this processor and that utilize the above patterns are presented.
3 citations
11 Sep 2021
TL;DR: The Grid Based Localization (GBL) algorithm as mentioned in this paper is a regression method which uses deep learning to find the bounding box for the grid structure. But it is not suitable for combining with generalized DL classifier algorithms such as Meta Learning.
Abstract: In many NDE applications, accurate localization of Ultrasonic echoes can be an important first step for further characterization tasks, including estimation, detection and classification. This is especially critical for a generalized classifier pipeline which can handle multi-class flaw analysis without additional supervised training steps. Therefore, the purpose of this study is to find a localization algorithm which can localize and find a bounding box for reflections of a void (flaw) in raw ultrasonic B-scan data. The proposed Grid Based Localization (GBL) algorithm is a regression method which uses Deep Learning (DL) to find the bounding box for the grid structure. Results demonstrate more robust performance than conventional contour-based techniques and the algorithm is suitable for combining with generalized DL classifier algorithms such as Meta Learning.