COMPILER TECHNIQUES FOR MATLAB PROGRAMS ... - CiteSeerX

A structural inference mechanism was designed to improve code performance by enhancing
library selection by the code generator.
Suppose for example, that in a matrix
multiplication, both variables are known to be matrices, but one of them can be inferred
to be a triangular matrix. In this case, instead of using a library function to perform a
general matrix multiplication operation, an optimized function that takes into consideration
the triangular structure of the variable could be used. This structural inference mechanism
has not been implemented in the current version of the compiler.
Sources of Information for the Static Inference
The static inference mechanism extracts information from four main sources: input les
program constants operators and functions. If the program reads data from an external
le, the system requires a sample le to be present during the compilation.
From this
le, the compiler extracts the initial intrinsic type and rank 2 of the variables being loaded 3 .
Variable shapes are not extracted from the input les because they are much more likely than
intrinsic type and rank to dier between runs. However, with the rank information from the
loaded variables, the compiler can propagate partial (or even complete) shape information
represented in terms of values obtained when vectors (and scalars) are loaded.
The second source of information is program constants, from which intrinsic type, rank,
and shape can be inferred and propagated.
MATLAB operators are the third source of
information. From them, we can extract type information in the case of logical operators,
and rank, shape, and structural information by taking into consideration the conformability
requirements imposed by theMATLAB operators.
Finally, the fourth source of information is the high-level summary information from the
MATLAB intrinsic functions. Built-in functions can provide information for both forward
and backward inference.
An example of this is the function rcond, for the conditional
2 The use of this initial rank information can be turned o by a compiler ag. In this case, all loaded
variables are assumed to be two-dimensional allocatable arrays.
3 For clarity of presentation, a MATLAB comment (%) is used to indicate which variables are being loaded
in the pseudo-code examples using the load function. This comment is not required by the compiler.
23