Performance libraries are the building blocks of high
performance applications. These libraries, architecture-dependent, are
mostly hand-tuned.  On-going research efforts have led to the design
of automatic library generators such as ATLAS for linear algebra functions,
SPIRAL or FFTW for signal processing and DFT functions. These generators,
domain-specific, automatically tune the code according to the target
architecture features.

In this talk I will present a new approach for the generation of performance
libraries.  This approach is not application-specific: it relies on
performance measures of source code kernels and on a very simple
performance model in order to build library functions from these building
blocks. Performance results will be compared with ATLAS and vendor
libraries on Itanium2 and Pentium architectures. Challenges of this approach,
in particular of the performance model, will be discussed.