The Development & Optimization Of Stream Application For Scientific Computation

Stream processor integrates lots of ALU and on-chip memory controlled by software and could gain high performance on data-intensive media application. It has many good peculiarities including low frequency,low power,high data parallelism,high peak performance, so attempting to use stream processor in the large scale scientific and engineering computation is a hot topic nowadays.But the development and optimization of stream application are very different from conventional method. It's a critical issue to study the development and optimization of scientific stream application to apply stream processor in scientific computation domain.In this paper, we develop many typical scientific stream applications after analyzing the stream programming model and language deeply. And based on the peculiarity of stream architecture, we study the optimization method for scientific application. At last, we make many tests for these stream applications and evaluate the efficiency of optimization method on the X stream processor developed by our institute.Main works include:1. By analyzing the characteristic of stream programming model, we summarize the method for scientific stream programming and develop many typical scientific stream applications.2. Develops some optimization methods for stream program based on loop transformation, owing to that loop is the main structure in scientific application.3. Performs some tests for typical scientific kernel and whole application on the X stream processor. The kernels includes Fast Fourier Transform algorithm,Laplace transform,Jacobi Iteration on dense matrix,Jacobi Iteration on sparse matrix. The applications includes Spec2000 171.Swim,optical application Capao and Partial Differential Equation. The result shows that the development and optimization methods in this paper could fully exploit the performance potential of the stream processor.