PhD  for Juan Carlos Rojas

Title: Multimedia Macros for Portable Optimized Programs


Abstract:

Multimedia processor architectures offer a combination of long partitioned
registers and complex instructions that can speed up applications
significantly when programmed manually.  Optimized programs for these
architectures have been non-portable up to now, because of differences in
the instruction sets, register lengths, alignment requirements and
programming styles.  This dissertation presents a method that overcomes all
these obstacles by providing a virtual instruction set common to a group of
target architectures.  This virtual instruction set is implemented as a
library of C pre-processor macros called MMM.  The macros can emulate long
registers on architectures with short ones, and emulate complex instructions
that are missing in certain targets.

This research is the first to provide a general solution to the portability
of optimized multimedia programs.  No other method to date allows an
arbitrary program to take advantage of the complex partitioned operations
available in multimedia instruction sets, while remaining portable. 
MMM libraries were implemented for MMX & SSE, SSE2, AltiVec and TriMedia
TM1300 multimedia architectures.  Three examples from video compression were
implemented in a portable way using MMM.  The examples include IDCT, block
distance for motion estimation, and block distance with interpolation.  The
portable examples were automatically translated into optimized code for each
of the targets.  Their performance is comparable, and in several cases
better, than equivalent examples optimized by the processor vendors.