| Version 5 (modified by nakasato, 10 years ago) (diff) |
|---|
Y.Suzuki
Fast N-Body Calculation Implemented by OpenCL with Vectorization
I compared the performance of N-body simulations on CPU and GPU with a several optimization techniques. Each program is written in OpenCL which standardizes APIs for GPU, and an important optimization technique in OpenCL is a vectorization. It enables us to utilize multiple variables as one variable. As a result, the program which utilized 4 variables as one was the best performance. I optimized the program using shuffle function. I found the calculation of N-body problem using shuffle function was about 1.3 times faster than without it. I also found Intel SDK had an ability to efficiently vectorize the kernel program.
file:///home/committee/aac/Thesis2011/s1140123
K.Kamijima
Performance Evaluation of the Octree Method on GPU
The purpose of my research is to examine the performance of graphic processing unit (GPU) with a numerical algorithm for particle simulations. Specifically, I adopt the Octree method that requires many branch instructions. In general, GPU is intrinsically not good at dealing with branch instructions. I have implemented the Octree method in single and double precision on Cypress GPU. With the GPU, the peak performance of single precision operations is five times better than that of double precision. Although it was expected that the performance of the Octree method with double precision is much slower, I found that it is not the case. On GPU, I found that the performance of the Octree method is constrained by not the computing power of the GPU but the performance penalty due to branch instructions.
file:///home/committee/aac/Thesis2011/s1150062
