Using GPU to Improve the Drawing Performance of VLSI Layout

• Main Authors: Wei-Cheng Chiu, 邱偉誠
• Other Authors: Shih-Chien Su
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/87072570887990015146

碩士 === 銘傳大學 === 電腦與通訊工程學系碩士班 === 98 === With the advancement of semiconductor technology, it is common that tens of millions of layout components are integrated in a single IC chip. To draw these layout components on the screens for further processing, the operations taken by the interactive layout tools increase exponentially. This normally leads to interaction delay when using the interactive layout tools. In order to improve the delay time of layout drawing, Solomon[1] proposed mipmapping technique that utilizes the texture mapping function of OpenGL to improve the performance of layout drawing. In [1], different resolutions of layout images are generated and stored hierarchically in GPU’s texture memory to suppor fast response of zoom in/zoom out operations of layout drawing. This approach performs well for static layout data. When the layout elements are changed, it needs to rebuild the hierarchy of texture images. This will create intolerable delay for interactive layout tools. Our study incorporates from [1] the concept of using GPUs to enhance the performance of layout drawing. We modified quad tree by copying the layout data to GPU memory and using the multiple threads in GPU to cooperate with CPU to enhance the performance of region query. Besides, we use the vertex buffer object (VBO) of OpenGL to enhance the drawing of layout. We also design a simple heuristic to filter the objects whose areas are too small to display in layout tools’ windows. Compare with the approach proposed by Solomon, our approach can accelerate layout drawing while maintains dynamic data insertions and deletions. We use the DEF files (the layout of ITC’99 benchmark) from VLSI lab of Yuan Ze University and a DEF file of a real design to test the performace of our approach. From our experimental results, when the objects is large enough, region query performances of large windows of our approach are improved twice compared with those from the traditionalregion queries on CPU. We also compare the perforamce of VBO with drawing API in X Window, XFillRectangles(). From the experimental results, the performace of VBO is twice better than that of XFillRectangles(). The simple filter heuristic could further reduce the drawing time by about 30% if applied in region queries on CPU only. If filter is used in region quries using GPU and CPU, filter could save at most around 7% of the region query time.