Abstract
To be cost effective, embedded systems are shipped with low-end micro-processors. These processors are dedicated to one or few tasks that are highly demanding on computational resources. Examples of widely deployed tasks include the fast Fourier transform, convolutions, and digital filters. For these tasks to run efficiently, embedded systems programmers favor fixed-point arithmetic over the standardized and costly floating-point arithmetic. However, they are faced with two difficulties: First, writing fixed-point codes is tedious and requires that the programmer must be in charge of every arithmetical detail. Second, because of the low dynamic range of fixed-point numbers compared to floating-point numbers, there is a persistent belief that fixed-point computations are inherently inaccurate. The first part of this thesis addresses these two limitations as follows: It shows how to design and implement tools to automatically synthesize fixed-point programs. Next, to strengthen the user's confidence in the synthesized codes, analytic methods are suggested to generate certificates. These certificates can be checked using a formal verification tool, and assert that the rounding errors of the generated codes are indeed below a given threshold. The second part of the thesis is a study of the trade-offs involved when generating fixed-point code for linear algebra basic blocks. It gives experimental data on fixed-point synthesis for matrix multiplication and matrix inversion through Cholesky decomposition.