Solving a large scale radiosity problem on GPU-based parallel computers

详细信息    查看全文

• 作者：Eduardo D&rsquo ; Azevedo^a ; ^{dazevedoef@ornl.gov" class="auth_mail} ; Zhiang Hu^b ; ^{zhianghu@math.cu.edu.hk" class="auth_mail} ; Shi-Quan Su^c ; ^{shiquansu@hotmail.com" class="auth_mail} ; Kwai Wong^c ; ^{kwong@utk.edu" class="auth_mail}
• 关键词：Radiosity ; View factor calculation ; Cholesky decomposition ; Out-of-core algorithm ; Hybrid multicore/GPU system
• 刊名：Journal of Computational and Applied Mathematics
• 出版年：November 2014
• 年：2014
• 卷：270
• 期：Complete
• 页码：109-120
• 全文大小：1892 K

文摘

The radiosity equation has been used widely in computer graphics and thermal engineering applications. The equation is simple to formulate but is challenging to solve when the number of Lambertian surfaces associated with an application becomes large. In this paper, we present the algorithms to compute the view factors and solve the set of radiosity equations using an out-of-core Cholesky decomposition method. This work details the algorithmic procedures of the computation of the view factors and the Cholesky solver. The data layout of the radiosity matrix follows the block cyclic decomposition scheme used in ScaLAPACK. The parallel computation of the view factors on the GPUs extends the algorithms based on a serial community code called view3d. To handle large matrices that exceed the device memory on GPU, an out-of-core algorithm for parallel Cholesky factorization is implemented. A performance study conducted on Keeneland, a hybrid CPU/GPU cluster at the National Institute for Computational Sciences, composed of 264 nodes of multicore CPU and GPU are shown and discussed.