A Modular MATLAB Compilation

Infrastructure Targeting Embedded Systems

Ricardo Nobre

Faculdade de Engenharia da Universidade do Porto (FEUP)

Departamento de Engenharia Informática, Rua Dr. Roberto Frias s/n,
4200-465 Porto, Portugal
Email: [email protected]

Abstract. Dynamic programming languages make extremely difficult

for static compilation to achieve efficient results. MATLAB is one of
the programming languages offering dynamic features and abstractions
to easily perform matrix operations. These features, the vast portfolio
of domain-specific libraries, and the tools to simulate MATLAB sys-
tem specifications, make MATLAB a widespread programming language,
ranging from embedded to scientific computing domains. However, MAT-
LAB programs are usually seen as models and require aggressive special-
izations in order to achieve efficient embedded computing implementa-
tions. For example, the same MATLAB model may be implemented us-
ing double, single, and/or fixed-point precision data types. This kind of
specialization needs analysis steps exploring different alternatives, mea-
suring the impact on accuracy and precision errors, before deciding to
a specific implementation. Furthermore, a selection of a particular im-
plementation may have to be done according to the target embedded
computing architecture. This paper presents a MATLAB compiler in-
frastructure able to generate different implementations from the same
MATLAB input code. Specifically, our MATLAB compiler is modular
and can be easily extended by using a strategic programming approach,
and specific implementations can be derived by aspect-oriented infor-
mation specified by the user. We also present the use of the current
version of the compiler for translating MATLAB codes to C code imple-
mentations for embedded computing systems using different input type
specifications, thus simulating different target scenarios.

Keywords: modular, MATLAB, C, compilation, specialization

1 Introduction

MATLAB [2] is a de facto standard high-level language and an interactive numer-

ical computing environment for many areas of application, including embedded
computing. It is widely used by engineers to quickly develop and evaluate their
solutions [3]. By including extensive domain-specific and visualization libraries,
MATLAB substantially enhances engineers productivity [3].
 The flexibility of MATLAB, however, comes at the cost of interpretation
(and JIT1 compilation) and lack of type/shape information. In MATLAB the
same identifier can be used to hold various data types and array shapes (i.e. the
number of elements of each dimension of a multi-dimensional array) throughout
the code execution. This makes it very handy for quick program prototyping but
it is not ideal for static analysis, and in many cases precludes the application of
advanced program transformations. The flexibility of MATLAB programs results
in lower execution performance. This lack of performance is typically addressed
by the development of an auxiliary reference implementation once the MATLAB
code has been validated. This amplifies the maintenance costs as the programmer
must now deal with multiple language versions of the application.
An alternative approach relies on a compilation tool to perform advanced
analyses and to generate a reference C code directly from MATLAB, thus avoid-
ing the lengthy user intervention. This automatic approach, however, is also
fraught with obstacles as there are inherent limits to what information static
analyses can extract. Rather than pushing the limits of static analysis - some-
times too conservative due to the lack of complementary information about the
problem - we have adopted the pragmatic approach of allowing the user to con-
trol some of the aspects of the translation process at a very high-level. In our
approach the compiler is aware of separate type and shape specifications pro-
vided by the user as a vehicle to convey information to the compiler regarding the
types and array shapes of the used variables. The compiler uses the user-provided
information and complements/checks its consistency against the information it
can derive from its own analyses.
As to the compiler infrastructure, we rely on the concept of strategic pro-
gramming2 to accomplish a modular and flexible compiler framework. The in-
frastructure uses Tom3 to achieve a built-in rewriting system for analysis, trans-
formations, and code generation. The end result is a synergy between compiler
analysis and the user that allows the compiler to generate very high quality
code from MATLAB specifications and the possibility to generate different C
code versions, important when targeting different embedded systems. In addi-
tion, the use of a well-known and mature strategic programming system as Tom
allows our compiler infrastructure to be easier to extend in comparison to an
approach using proprietary and specific data-structures and implementations.
This paper is organized as follows. Section 2 describes the compilation in-
frastructure. Section 3 presents some experimental results. Section 4 describes
related work. Finally, Section 5 draws some conclusions and presents future work.

1
Just-in-time compilation is a method that compiles code at run-time in order to
improve the runtime performance of computer programs.
2
Strategic programming is “a generic programming idiom for processing compound
data such as terms or object structures”, relying on the “separation of two concerns:
basic data-processing computations vs. traversal schemes”. Traversals are composed
by passing the former as arguments to the latter [4].
3
Tom is a language extension that provides pattern matching facilities for manipu-
lating tree structures and XML documents [5,6].
2 Compilation Infrastructure

Our compiler infrastructure is organized according to the component diagram

depicted in Fig. 1: a single front-end (mat2tir ) and two back-ends (tir2mat and
tir2c). The compiler is modular enough to allow back-ends for other output
programming languages.

Front end
MATLAB XML IR

MAT2TIR
(JavaCC + Tom)

Type/Shape
Specification

Back end Optimized

C Code

TIR2MAT TIR2C
(Tom) Optimized
MATLAB Code

Fig. 1. Component diagram of MATLAB Compilation infrastructure.

The mat2tir component uses a parser generated from an LL MATLAB gram-

mar 4 with JavaCC5 to parse MATLAB M-files. The abstract syntax tree (AST)
generated by the parser is then converted to a Tom [5,6] intermediate repre-
sentation (IR) exported in XML format. This mat2tir component is internally
structured as a traditional front-end parser and a Tom IR creator step.
The IR AST grammar is declared by Gom [5,6]. Gom provides a syntax to
define ASTs. The Tom IR structure is built from the JavaCC AST by recursively
visiting each node, respecting the specification declared using Gom syntax.
The Tom IR represents the input MATLAB code as expression trees directly
obtained by the AST produced by the parsing front-end. The IR also represents
the line numbers of MATLAB statements, code annotations, and information
about types and shapes of the MATLAB expressions. Information about types
and shapes is unknown until an inference engine is executed. Annotations are
embedded in the MATLAB comments (they start with %@ ) and are used in our
approach to identify specific MATLAB code locations (join points), or are inside
a separated file loaded by the backend.
4
LL grammars are a subset of contex-free grammars parsable from left to right, pro-
ducing the leftmost derivation of the input.
5
JavaCC is a parser generator for use with Java applications [8]
 Table 1 depicts MATLAB code for a recursive factorial, part of the corre-
spondent Tom IR and the generated C code specifying the input variable n as
scalar and integer.

Table 1. (a) MATLAB code, (b) XML IR for the factorial function and C code gen-
erated from MATLAB with the following type and shape information input to the
compiler: n:int32:1x1.

1 function x = factorial(n) 1 <Start> 1 #include ”tensor.c”

2 if n == 1 2 <FunctionMFile> 2 void factorial(int n, int∗ x) {
3 x = 1; 3 <ConcIdentifier> 3 if(n==1) {
4 else 4 <Identifier> 4 (∗x) = 1;
5 x = factorial(n−1); 5 ”x” ”1” 5 } else {
6 x = n∗x; 6 </Identifier> 6 factorial(e minus(n, 1),
7 end 7 ... &(∗x));
8 ”1” 7 (∗x) = e times(n,(∗x);
9 </FunctionMFile> ”1” 8 }
10 </Start> 9 }
a) b) c)

A tir2x form component then takes the Tom IR as input and, by combining
it with the type/shape specifications written by the user, applies specific IR-
based transformations or code generation steps. We have developed two backend
applications, namely tir2mat and tir2C.
The MATLAB Generator backend (tir2mat ) is used for validation purposes
and generates MATLAB code from the IR. This MATLAB code is helpful as it
gives the user the possibility to compare results obtained by using transformed
and/or specialized code with the original code. This can be a way to evaluate
and analyze accuracies in the case of type and word-length transformations, and
to trace and debug in the case of inserting code for monitoring.
The C Generator backend (tir2C ) allows us to generate C code from the
MATLAB input code with the possibility to use the complementary information
added by the user.
The transformations in the Tom IR use Tom [5,6] capabilities to manipu-
late tree structures. Pattern matching is used to find specific patterns of data-
structures in the Tom IR where transformations are applied. The use of Tom
allows us to achieve a flexible compiler infrastructure as transformations rules
are described using Tom specifications.
Regarding the back-end, and when translating typeless/shapeless languages
such as MATLAB to C, a key step is the definition of specific types and in
particular array shapes. To this extent we have implemented a limited form
of data-flow analysis [9] as well as an external user-provided mechanism for
static determination of types and shapes of variables as described next. The idea
is that just declaring information about the argument variables of a function,
the external knowledge gained by the inference engine is sufficient to statically
determine the shapes and types of all, or at least almost all variables.

2.1 Pipeline architecture and Strategies

The compiler is structured as a set of distinct engines attached to a single

pipeline. An engine has both accesses to the compiler internal variables, like
symbol tables, and the IR representing the input code at each step.
Fig. 2 depicts a diagram representing the main classes backend.

Pipe
+_engines: List<Engine>
+execute(program:Program)
1

1..* Program
Engine -_callables: Map<String, Callable>
-_actionSelector: ActionSelector -_callGraph: Map<String, Node>
-_mainCallable: String
+Engine()
+execute(program:Program) +Program(mainCallable:String)
+createCallGraph(): void
1

1..*

Callable
-_iR: IRAbstractType
+_symbolTable: SymbolTable
+_code: String[]
+_name: String
+Callable(iR:IRAbstractType,name:String)
+isRecursiveQ(): boolean
+dimensionsKnownQ(lang:int): boolean
+typesKnownQ(lang:int): boolean

Fig. 2. Main classes composing a backend instance.

The main class of the tir2c backend, Tir2C, instantiates a single pipeline
(Pipe class) and attaches engines (extend Engine) in the order they should exe-
cute. Engines can manipulate the IRs representing the input MATLAB files, can
change the internal state of the compiler (e.g. symbol tables, generated code), or
both. For instance, the C generator engine, CGeneratorEngine class, is respon-
sible for generating the C code from the imported IR, using the type and shape
information about variables inferred by the inference engine (InferenceEngine-
Filter class) with information directly specified in external specification files, in
case this files exist and contain type or shape declarations. The script inliner
engine (ScriptInlinerEngine) is an example of an engine that inlines IRs for
multiple MATLAB script files into a single IR.
The pipeline is executed after the call graph creation method, which returns
an instance of the Program class holding information about all M-Files, each
represented by an instance of the Callable class. A Callable object holds all
compiler internal variables, symbol tables and IR structure representing an M-
File. When executing (execute() method) the pipeline the Program object is
passed as argument, and then each engine attached to the pipeline is executed
(execute() method) with it passed as argument. Engines have access to all the
internal variables (e.g. IRs and symbol tables) of the compiler.

2.2 Engine creation methodology

The execution of an engine is performed by passing each Callable from the

Program object to a new instance of a class extending an abstract class called
ActionSelector. The class extending the ActionSelector is always described using
primitives from the Tom language extension in a NAMEActionSelector.t file,
where NAME represents the name of the engine.
The logic to traverse the IR is encapsulated in NAMEActionSelector.t by a
Tom strategy that for each Tom IR structure found in the IR calls a method from
a class extending the ActionSelection class, which encapsulates the only methods
that are allowed to change the compiler internal state for a given Callable passed
as argument.
The ActionSelector of an engine acts as a selector for choosing the appro-
priate method to call from the ActionSelection component. Having the logic for
the traversal of the IR completely separated from the actions to perform during
the traversal of each IR node in separated files allows interesting software devel-
opment techniques as the ActionSelector and the ActionSelection parts of each
engine require different levels of understanding about the compiler architecture.
Extending the ActionSelector to create a new engine only requires knowledge
about the IR and the Tom primitives used to implement strategies for traversing
the IR. Extending the ActionSelection neither requires knowledge about how to
specify Tom strategies nor how the IR is structured, as the ActionSelection part
only deals with updating the symbol table and the IR and all the methods needed
are automatically created when extending the abstract class. Transformations are
only specified, using a strategic programming approach, in the class extending
the ActionSelector abstract class. These classes are represented in 3.

2.3 C Generation

In terms of the implementation, the generated C code relies on a very flexible

data structure called tensor. The tensor is used to represent, in C, all array
variables. This structure has its dimensions and base type (integer, real and
complex) dynamically allocated, thus relying on run-time tests to determine
which operation to apply to the elements of the multi-dimensional array, only in
case the precise dimensions are not known at compilation time (e.g., when such
information is not determined by the static shape analysis stage). The number
of elements of the whole structure is also stored for efficiency. We also developed
a C tensor library with all the structures and functions to support the most
common MATLAB matrix built-in functions.
 Fig. 3. Action Selection Framework abstract classes.

2.4 User-provided Type and Shape Information

To address the limitations of a static type/shape inference analyses and/or to

force implementations according to user knowledge about the target implemen-
tations, our MATLAB compiler supports the specification of types and shapes
of variables as a form of annotations in a separate file. This mechanism com-
plements the capabilities of the type/shape-inference and allows the compiler to
generate code using more accurate type information and/or specialized imple-
mentations.
The code generated by our compiler when considering the specification of
the type and shape of the arguments of the top function results in a specialized
code without type checks which gives an efficient C implementations, both in
terms of performance and memory, whenever such specialization can be used in
the final implementation. The generated code is clean when compared to human
made code.

3 Experimental Results

We carried out a series of experiments to evaluate our MATLAB compiler and the
impact of the type and shape information on the performance of the generated
C code.
In our experimental evaluation we used a set of 11 code kernels originally
written in MATLAB. We used our compiler to automatically derive C codes.
These kernels are drawn from simple arithmetic operations in some cases us-
ing linear algebra operations such as dense matrix-vector multiplications, from
digital signal processing applications, and from an industrial application.
One of the set of kernels used is an industrial code delivered by Honeywell for
the REFLECT project [7,22]. The code is related to a Three-dimensional (3D)
Path Planning algorithm. This kernel plans a 3D path between the current posi-
tion of an autonomous vehicle and a predetermined goal position. In this paper
we consider the most computationally intensive task of the 3D path planning
algorithm: the gridIterate function. This function consists of 4 nested for type
loops and uses 3D arrays. In the experiments we consider a partial map of the
environment of size 32 ∗ 64 ∗ 16.
We then compared the execution time of each sample code using different
type and shape declaration files. To compare the performance of the base types
the programmer can specify, we carried out experiments where the different base
types corresponding to distinct precision requirements were used. These include
the specification of double (default), float, and fixed-point precision.
Preliminary tests comparing the number of cycles needed to run executables
generated with Mathworks MCC tool6 and the C code generated by our com-
pilation infrastructure shown that, as expected, the compiled C code is much
faster than Mathworks JIT approach (between 2 and 4 times faster). These tests
were performed on an x86 microprocessor (in our case the Intel Core 2 Duo),
not at all a processor for embedded systems, as it is one of the few platforms
compatible with executables generated by MCC (execution requires run-time
compiled libraries). Therefore MCC can not be considered as an alternative to
our compiler for use with embedded systems.

3.1 Methodology
Experiments were conducted using the SimpleScalar/ARM simulator [10,11] to
execute the object output by gcc7 with the -O3 flag from the C code generated
by our compiler from MATLAB input files. The simulator models Intel Stron-
gARM SA-11xx microprocessors, has been validated against a large collection of
workloads, and is believed to be within 4% of real hardware for performance es-
timation [11]. The StrongARM family of microprocessors implements the ARM
V4 instruction set architecture. This simulator reports the cycle count with high
accuracy.
Note that we only declared the type and shape of the parameters of each
function in the M-files, thus asserting that for all call-sites the same information
holds. The information about all other variables is inferred at compilation time
by evaluating the expressions where the input parameters are present. Further-
more, for these examples, we relied on type inference to determine the type and
shape of all other local variables. This is clearly not the most flexible situation
and we will continue to explore scenarios where type/shape specialization makes
sense. In addition to the impact of shape information we also tested the perfor-
mance improvement gained by using single precision floating point rather than
the standard double precision.
6
The MATLAB compiler “prepares MATLAB file(s) for deployment outside of the
MATLAB environment, generates wrapper files in C or C++, optionally builds stan-
dalone binary files, and writes any resulting files into the current folder, by default.”
[23]
7
The GNU Compiler Collection [24]
3.2 Results and Discussion

We show in Fig. 4 the results obtained when using the SimpleScalar/ARM sim-
ulator.
The results consider the execution time improvements between float vs. dou-
ble, fixed vs. double, and fixed vs. float implementations. The last group of
columns (labeled “average”) corresponds to the average speedup for our set of
code kernels.

Fig. 4. Speedups for different implementations of the benchmarks.

As the performance results reveal there is on average an increase in per-

formance of 1.16 times and 1.09 times when using the fixed-point arithmetic
representation versus double and float precision, respectively. This is mainly due
to the use of integer operations, that are faster than instructions operating on
float and double data types for the specific ARM architecture used.
The fact that the performance relation between double and float is not consis-
tent between benchmarks (mainly because of float-to-double and double-to-float
conversions) is a proof that the specification of types and shapes can be valuable,
even when the target architecture does not change. For instance, when consid-
ering fixed vs. float implementations, a speedup of 1.57 times is achieved for the
dirich code and a slowdown is achieved for the fir code.

4 Related Work

Given the significance of MATLAB in early algorithm design and evaluation,

many other research efforts focused on translating MATLAB code to languages
that provide a more efficient execution environment (see, e.g., [12,13,14]). Due to
the importance to execute directly MATLAB programs there has been research
efforts to improve the execution of JIT MATLAB compilers. A recent example is
the compiler presented in [15] that performs function specialization based on the
runtime knowledge of the types of the arguments of the functions. DeRose and
Padua developed the FALCON research project [12] that translates MATLAB
to FORTRAN90 code with performance improvements that resemble the ones
we report here for codes of similar complexity. They leverage an aggressive use
of static and type inference for base types (double and complex) as well as shape
(or rank) of the matrices.
Other researchers have also relied on simple type inference approaches such
as ours. Banerjee et al [16,17] have used annotations to specify data types and
shapes, simple type inference analysis and target VHDL code specifications for
hardware synthesis onto Field-Programmable Gate Arrays (FPGAs). Later on,
this work resulted on the commercial AccelFPGA compiler which supports,
among others, annotations to define the shape of MATLAB variables.
As with the compiler to optimize Octave programs presented in [18] our com-
piler relies on a strategic programming approach. However, in our approach we
extend the strategic programming approach (using Tom in our case instead of
the Stratego approach in [18]) with user information regarding types and shapes.
One of the closest related work is the embedded MATLAB (a subset of MAT-
LAB) to C code translation existent in the MathWorks Real-Time Workshop
[19] which allows the user to embed annotations with MATLAB code to achieve
C code implementations for embedded systems [20].
The popularity of the MATLAB language is also reflected in the similar lan-
guages proposed. Examples of those languages are Scilab and Octave. Recently,
a Scilab to C translator [21], named Sci2C, has been proposed. The Sci2C trans-
lator focuses entirely on embedded systems and it is completely dependent on
annotations, to specify data sizes and precisions, embedded in the Scilab code.
Our compiler distinguishes from Sci2C as it is capable to generate C code even
without annotations and specialization of the generated C code can be achieved
without polluting the original code (MATLAB, in our case). Furthermore, Sci2C
requires that the size of arrays is fixed and statically known while our compiler
also produces C code (including calls to realloc()) when those sizes are not stat-
ically known.

5 Conclusion

This paper presented a compiler infrastructure for MATLAB. The infrastruc-

ture relies on the concept of strategic programming, and especially on the Tom
framework to analyse, transform and generate code from a Tom intermediate
representation of the input MATLAB program.
The architecture of the compiler infrastructure is sufficiently flexible and
modular to allow an easy integration of compilation stages by taking advantage of
Tom rewriting capabilities, in the form of what we called the “Action Selection”
framework.
 One important stage of the compiler is the generation of C code. In order
to achieve more efficient and/or specialized C code our approach allows users
to add information about types and array shapes. This additional information
is of paramount importance in most cases, especially in cases where advanced
type/shape inference analysis needs to be conservative, and in the cases where
one needs specialized implementations dependent on the target system and/or
specific concerns. Additionally, the compiler infrastructure is sufficiently mod-
ular to allow the integration of other output languages generators, other than
MATLAB and C.
Our ongoing work focuses on the use of an aspect-oriented language to spec-
ify more powerful type and shape information (possibly parametrically) as well
as high-level code transformations as recently proposed. These code transforma-
tions will allow users to easily explore specific code optimizations at the MAT-
LAB level.
The comparison of code generated by tir2c with code generated by Sci2C
(performance, memory, and readability-wise) is work in progress.
Future plans include a static analysis (possibly further extended with profil-
ing information) to identify automatically the variables in a MATLAB program
whose types/shapes are more important for the type/shape inference engine.

Acknowledgment
This research has been partially funded by the Portuguese Science Founda-
tion Fundao para a Ciência e Tecnologia (FCT) under research grant PT-
DC/EIA/70271/2006.
The author also acknowledges the support of the FP7 EU-funded project
REFLECT for the access to MATLAB codes related to industrial applications.
The author acknowledges support from João M. P. Cardoso, affiliated with
Faculdade de Engenharia da Universidade do Porto (FEUP) and Pedro Diniz,
affiliated with Instituto de Engenharia de Sistemas e Computadores, Investigação
e Desenvolvimento em Lisboa (INESC-ID.

References
1. Ricardo Nobre, Joo M. P. Cardoso, Pedro C. Diniz ’Leveraging Type Knowledge
for Efficient MATLAB to C Translation’, paper presented at CPC2010, Vienna
University of Technology, Vienna, Austria, July 9, 2010.
2. MATLAB the Language of Technical Computing, http://www.mathworks.com/
products/matlab.
3. Richard Goering, ”Matlab edges closer to electronic design automation
world” EE Times, 10/04/2004 http://eetimes.com/electronics-news/4050334/
Matlab-edges-closer-to-electronic-design-automation-world
4. R. Lämmel, E. Visser, and J. Visser, Strategic programming meets adaptive pro-
gramming, In Proc. 2nd Intl Conf. on Aspect-Oriented Software Development
(AOSD’03), Boston, Mass., March 17 - 21, 2003. ACM, New York, NY, USA, pp.
168-177.
5. Jean-Christophe Bach et al.: Tom Manual - Version 2.7, http://tom.loria.fr
6. Emilie Balland, Paul Brauner, Radu Kopetz, Pierre-Etienne Moreau, and Antoine
Reilles, Tom: Piggybacking rewriting on java, In 18th International Conference
on Term Rewriting and Applications (RTA07), Paris, France, June 26-28, 2007,
Springer LNCS 4533, pp. 36-47.
7. João M. P. Cardoso et al., REFLECT: Rendering FPGAs to Multi-Core Embedded
Computing, Book Chapter, Reconfigurable Computing, Springer.
8. The JavaCC Home, https://javacc.dev.java.net/
9. A. Aho, J. Ullman, M. Lam and R. Sethi, Compilers: Principles, Techniques and
Tools, Addison Wesley, 2006.
10. T. Austin, E. Larson, and D. Ernst, SimpleScalar: An infrastructure for computer
system modeling, Computer, 35(2), 2002, pp. 5967.
11. SimpleScalar Version 4.0, Test Releases, http://www.simplescalar.com/v4test.
html
12. De Rose, L. and Padua, D. Techniques for the Translation of MATLAB programs
into Fortran 90. ACM Trans. Program. Lang. Syst. 21, 2 (Mar. 1999), pp. 28623.
13. Joisha, P. G. and Banerjee, P. 2007. A translator system for the MATLAB lan-
guage: Research Articles. Softw. Pract. Exper. 37, 5 (Apr. 2007), pp. 535-578.
14. G. Almsi , D. Padua, MaJIC: Compiling MATLAB for Speed and Responsiveness,
In Proc. of the ACM 2002 Conf. on Programming Language Design and Implemen-
tation (PLDI’02), June 17-19, 2002, Berlin, Germany.
15. Amina Aslam, and Laurie Hendren, McFLAT: A Profile-based Framework for Loop
Analysis and Transformations, in the 23rd International Workshop on Languages
and Compilers for Parallel Computing (LCPC2010), October 7 - 9, 2010, Rice Uni-
versity, Houston, Texas, USA.
16. A. Navak, M. Haldar, A. Choudhary and P. Banerjee, Parallelization of MATLAB
Applications for a Multi-FPGA System, in Proc. of the 9th Annual IEEE Symp.
on Field-Programmable Custom Computing Machines (FCCM’01), Rohnert Park,
Calif., May, 2001.
17. P. Banerjee, D. Bagchi, M. Haldar, A. Nayak, V. Kim, R. Uribe, Automatic Con-
version of Floating Point MATLAB Programs, in proc. of the 11th Annual IEEE
Symp. on Field-Programmable Custom Computing Machines (FCCM03), Napa,
Calif., 2003.
18. K. Olmos, E: Visser, Turning dynamic typing into static typing by program special-
ization in a compiler front-end for Octave, in Proceedings Third IEEE International
Workshop on Source Code Analysis and Manipulation (SCAM03), 26-27 Sept. 2003,
pp. 141- 150.
19. Real-Time Workshop: Generate C code from Simulink models and MATLAB code,
http://www.mathworks.com/products/rtw/
20. Houman Zarrinkoub, and Grant Martin, Best Practices for a MATLAB to C Work-
flow Using Real-Time Workshop, MATLAB Digest - November 2009, http://www.
mathworks.com/company/newsletters/digest/2009/nov/matlab-to-c.html
21. Scilab 2 C - Translate Scilab code into C code, http://forge.scilab.org/index.
php/p/scilab2c/
22. REFLECT, FP7 EU Project: http://www.reflect-project.eu.
23. MATLAB R2011b Documentation, MATLAB Compiler. http://www.mathworks.
com/help/toolbox/compiler/mcc.html
24. GCC, the GNU Compiler Collection. http://gcc.gnu.org/