作者: Juan Luis Jerez , George A. Constantinides , Eric C. Kerrigan
DOI: 10.1109/TC.2013.162
关键词:
摘要: We consider the problem of enabling fixed-point implementation linear algebra kernels on low-cost embedded systems, as well motivating more efficient computational architectures for scientific applications. Fixed-point arithmetic presents additional design challenges compared to floating-point arithmetic, such having bound peak values variables and control their dynamic ranges. Algorithms solving equations or finding eigenvalues are typically nonlinear iterative, making these a nontrivial task. For types algorithms, bounding cannot be automated by current tools. focus Lanczos iteration, heart well-known methods conjugate gradient minimum residual. show how one can modify algorithm with low-complexity scaling procedure allow us apply standard derive tight analytical bounds all process, regardless properties original matrix. It is shown that numerical behavior implementations modified chosen at least good implementation, if necessary. The approach evaluated field-programmable gate array (FPGA) platforms, highlighting orders magnitude potential performance efficiency improvements moving form computation.