Differences Between Generated Code and MATLAB Code
To convert MATLAB® code to efficient C/C++ code, the code generator introduces optimizations that intentionally cause the generated code to behave differently, and sometimes produce different results, than the original source code.
Here are some of the differences:
Output Argument Validation of Conditionally-Assigned Outputs
Size of Empty Array That Results from Deleting Elements of an Array
Growing Variable-Size Column Cell Array That is Initialized as Scalar at Run Time
Binary Element-Wise Operations with Single and Double Operands
MATLAB Classes in Nested Property Assignments That Have Set Methods
Converting Strings with Consecutive Unary Operators to double
These differences are applicable for:
MEX and standalone C/C++ code generation by using the
codegen
command or the MATLAB Coder™ app.Fixed-point code acceleration by generating MEX using the
fiaccel
(Fixed-Point Designer) command.MATLAB Function (Simulink) block simulation using Simulink®.
When you run your generated fiaccel
MEX, C/C++ MEX or standalone
C/C++ code, run-time error checks can detect some of these differences. By default,
run-time error checks are enabled for MEX code and disabled for standalone C/C++ code.
To help you identify and address differences before you deploy code, the code generator
reports a subset of the differences as potential differences.
Functions that have Multiple Possible Outputs
Certain mathematical operations, such as singular value decomposition and eigenvalue decomposition of a matrix, can have multiple answers. Two different algorithms implementing such an operation can return different outputs for identical input values. Two different implementations of the same algorithm can also exhibit the same behavior.
For such mathematical operations, the corresponding functions in the generated
code and MATLAB might return different outputs for identical input values. To see if a
function has this behavior, in the corresponding function reference page, see the
C/C++ Code Generation section under Extended
Capabilities. Examples of such functions include svd
and eig
.
Passing Input Argument Name at Run Time
Suppose that foo
is a function that uses name-value argument
validation. When you call foo
from another function
bar
, the code generator must be able to determine the names
that you provide to foo
at compile time.
If the argument names are passed at run time, code generation fails in most situations. See Names Must Be Compile-Time Constants.
In certain situations, the code generator assigns the name that you passed to an optional positional or repeating input argument. In such situations, code generation succeeds with a warning and the generated code might produce results that are different from MATLAB execution. For example, consider this function:
function out = myNamedArg_warns(a,b) out = local(a,b); end function out = local(varargin,args) arguments (Repeating) varargin end arguments args.x args.y end if isfield(args,'x') && isfield(args,'y') out = args.x / args.y; elseif isfield(args,'x') out = args.x; else out = varargin{1}; end end
Behavior of MATLAB Execution
If you call myNamedArg_warns
with 'x'
as
the first input argument, MATLAB matches it against the first name-value argument of the function
local
.
myNamedArg_warns('x',5)
ans = 5
By contrast, if you call myNamedArg_warns
with
'z'
as the first input argument (that does not match with
either name-value argument of local
), MATLAB assigns the inputs into elements of
varargin
.
myNamedArg_warns('z',5)
ans = 'z'
Behavior of Generated Code
Attempt to generate a MEX by running the codegen
command.
Specify the type of the first argument to be a character scalar and the second
argument to be a double scalar. Code generation succeeds with a warning.
codegen myNamedArg_warns -args {'x',2}
Warning: This argument is not constant, and therefore does not match against a name-value argument inside 'myNamedArg_warns/local' during code generation. Code generation might fail or produce results that do not agree with MATLAB if a name passed at a call site is not known during code generation. Warning in ==> myNamedArg_warns Line: 2 Column: 13 Code generation successful (with warnings): View report
Irrespective of whether you pass 'x'
or
'z'
as the first input argument, the generated MEX
assigns it to the first cell of varargin
.
myNamedArg_warns_mex('x',5)
ans = 'x'
myNamedArg_warns_mex('z',5)
ans = 'z'
Workaround. To enable the code generator to match the first input against the
name-value arguments of the function local
, declare the
first input to be a compile-time constant with value 'x'
.
You can do this by using the coder.Constant
function with
the -args
option of the codegen
command.
codegen myNamedArg_warns -args {coder.Constant('x'),2}
Code generation successful.
Now, the behavior of the generated MEX agrees with MATLAB, although the MEX is unable to accept
any value other than
'x'
for the first input.
myNamedArg_warns_mex('x',5)
ans = 5
myNamedArg_warns_mex('z',5)
Constant function parameter 'a' has a different run-time value than the compile-time value. Error in myNamedArg_warns_mex
Empty Repeating Input Argument
In code generation, if a repeating input argument (that is declared in an
arguments
block) is empty at run time, the size of that argument is
0x0
. By contrast, in MATLAB execution, the size of an empty repeating input argument is
1x0
.
For example, consider this function:
function out = testVararginSize out = local; end function out = local(varargin) arguments (Repeating) varargin end out = size(varargin); end
Running testVararginSize
in MATLAB returns [1 0]
. If you generate a MEX for
testVararginSize
and run the generated MEX, you get
[0 0]
. However, iterating over elements of
varargin
by using length(varargin)
or
numel(varargin)
produces the same behavior across MATLAB and code generation.
Output Argument Validation of Conditionally-Assigned Outputs
The code generator validates an output argument if the argument is assigned a type during code generation. By contrast, MATLAB execution validates an output argument if the argument is assigned a value when the MATLAB function returns.
In most situations, this underlying behavioral difference does not cause your generated code to behave differently than MATLAB. Here is an example function for which you do see this difference:
function outerFunc(in) innerFunc(in); end function out = innerFunc(inputVal) arguments (Output) out {mustBePositive} end if inputVal out = inputVal; end end
In MATLAB, the execution of func
succeeds for
all double inputs. If the input is positive,
out
is assigned this positive value and the validator
mustBePositive
runs without assertion. If the input is
negative or zero, out
is not assigned and is not
validated.
Attempt to generate code for func
. Specify the input type to be
a double scalar.
codegen outerFunc -args 0
Variable 'out' is not fully defined on some execution paths. Error in ==> outerFunc Line: 7 Column: 5 Code generation failed: View Error Report
Because the variable out
is assigned a double scalar value on
one execution path, code generation assigns a double scalar type to
out
at compile time. The code generator then attempts to
perform validation on out
and discovers that it is not fully
defined if the if
condition fails.
Writing to ans
Variable
When you run MATLAB code that returns an output without specifying an output argument,
MATLAB implicitly writes the output to the ans
variable. If the variable
ans
already exists in the workspace, MATLAB updates its value to the output returned.
The code generated from such MATLAB code does not implicitly write the output to an ans
variable.
For example, define the MATLAB function foo
that explicitly creates an
ans
variable in the first line. The function then implicitly
updates the value of ans
when the second line executes.
function foo %#codegen ans = 1; 2; disp(ans); end
Run foo
at the command line. The final value of
ans
, which is 2
, is displayed at the
command line.
foo
2
Generate a MEX function from foo
.
codegen foo
Run the generated MEX function foo_mex
. This function
explicitly creates the ans
variable and assigns the value
1
to it. But foo_mex
does not implicitly
update the value of ans
to 2
.
foo_mex
1
Logical Short-Circuiting
Suppose that your MATLAB code has the logical operators &
and |
placed inside square brackets
([
and ]
). For such code patterns, the
generated code does not employ short-circuiting behavior for these logical
operators, but some MATLAB execution employs short-circuiting behavior. See Tips and Tips.
For example, define the MATLAB function foo
that uses the &
operator inside square brackets in the conditional expression of an
if...end
block.
function foo if [returnsFalse() & hasSideEffects()] end end function out = returnsFalse out = false; end function out = hasSideEffects out = true; disp('This is my string'); end
The first argument of the &
operator is always
false
and determines the value of the conditional expression. So,
in MATLAB execution, short-circuiting is employed and the second argument is not
evaluated. So, foo
does not call the
hasSideEffects
function during execution and does not display
anything at the command line.
Generate a MEX function for foo
. Call the generated MEX
function foo_mex
.
foo_mex
This is my string
In the generated code, short-circuiting is not employed. So, the
hasSideEffects
function is called and the string is displayed
at the command line.
Loop Index Overflow
Suppose that a for
-loop end value is equal to or close to
the maximum or minimum value for the loop index data type. In the generated code, the last
increment or decrement of the loop index might cause the index variable to overflow. The
index overflow might result in an infinite loop.
When memory integrity checks are enabled, if the code generator detects that the loop index might overflow, it reports an error. The software error checking is conservative. It might incorrectly report a loop index overflow. By default, memory-integrity checks are enabled for MEX code and disabled for standalone C/C++ code. See Check for Issues in MATLAB Code Using MEX Functions and Generate Standalone C/C++ Code That Detects and Reports Run-Time Errors.
To avoid a loop index overflow, use the workarounds in this table.
Loop Conditions Causing the Potential Overflow | Workaround |
---|---|
| If the loop does not have to cover the full range of the integer type, rewrite the loop so that the end value is not equal to the maximum value of the integer type. For example, replace: N=intmax('int16') for k=N-10:N for k=1:10 |
| If the loop does not have to cover the full range of the integer type, rewrite the loop so that the end value is not equal to the minimum value of the integer type. For example, replace: N=intmin('int32') for k=N+10:-1:N for k=10:-1:1 |
| If the loop must cover the full range of the integer type, cast the type of the loop start, step, and end values to a bigger integer or to double. For example, rewrite: M= intmin('int16'); N= intmax('int16'); for k=M:N % Loop body end M= intmin('int16'); N= intmax('int16'); for k=int32(M):int32(N) % Loop body end |
| Rewrite the loop so that the loop index in the last loop iteration is equal to the end value. |
Indexing for
Loops by Using Single Precision Operands
Suppose in your MATLAB code, you are indexing a for
loop that has a colon
operator, where at least one of the colon operands is a single type operand and the
number of iterations is greater than flintmax('single') =
16777216
. When all these conditions
are true, code generation might generate run-time or compile-time errors because the
generated code calculates different values for the loop index variable than the
values that MATLAB calculates.
For example, consider this MATLAB code:
function j = singlePIndex n = flintmax('single') + 2; j = single(0); for i = single(1):single(n) j = i; end end
This code snippet executes in MATLAB, but it causes a compile-time or run-time error because the value of
the loop index variable, i
, is calculated differently in the
generated code. The code generator displays a compile-time or run-time error and
stops code generation or execution to prevent this discrepancy.
To avoid this discrepancy, replace the single type operands with double type or integer type operands.
For more information on run-time errors, see Generate Standalone C/C++ Code That Detects and Reports Run-Time Errors.
Index of an Unentered for
Loop
In your MATLAB code and generated code, after a for
loop execution
is complete, the value of the index variable is equal to its value during the final
iteration of the for
loop.
In MATLAB, if the loop does not execute, the value of the index variable is stored as [ ] (empty matrix). In generated code, if the loop does not execute, the value of the index variable is different than the MATLAB index variable.
If you provide the
for
loop start and end variables at run time, the value of the index variable is equal to the start of the range. For example, consider this MATLAB code:function out = indexTest(a,b) for i = a:b end out = i; end
Suppose that
a
andb
are passed as1
and-1
. Thefor
loop does not execute. In MATLAB,out
is assigned [ ]. In the generated code,out
is assigned the value ofa
, which is1
.If you provide the
for
loop start and end values before compile time, the value of the index variable is assigned [ ] in both MATLAB and the generated code. Consider this MATLAB code:function out = indexTest for i = 1:-1 end out = i; end
In both MATLAB and the generated code,
out
is assigned [ ].
Character Size
MATLAB supports 16-bit characters, but the generated code represents characters in 8 bits, the standard size for most embedded languages like C. See Encoding of Characters in Code Generation.
Order of Evaluation in Expressions
Generated code does not enforce the order of evaluation in expressions. For most expressions, the order of evaluation is not significant. For expressions that have side effects, the generated code might produce the side effects in a different order from the original MATLAB code. Expressions that produce side effects include those that:
Modify persistent or global variables
Display data to the screen
Write data to files
Modify the properties of handle class objects
In addition, the generated code does not enforce order of evaluation of logical operators that do not short circuit.
For more predictable results, it is good coding practice to split expressions that depend on the order of evaluation into multiple statements.
Rewrite
A = f1() + f2();
as
A = f1(); A = A + f2();
so that the generated code calls
f1
beforef2
.Assign the outputs of a multi-output function call to variables that do not depend on one another. For example, rewrite
[y, y.f, y.g] = foo;
as
[y, a, b] = foo; y.f = a; y.g = b;
When you access the contents of multiple cells of a cell array, assign the results to variables that do not depend on one another. For example, rewrite
[y, y.f, y.g] = z{:};
as
[y, a, b] = z{:}; y.f = a; y.g = b;
Name Resolution While Constructing Function Handles
MATLAB and code generation follow different precedence rules for resolving
names that follow the symbol @
. These rules do not apply to
anonymous functions. The precedence rules are summarized in this table.
Expression | Precedence Order in MATLAB | Precedence Order in Code Generation |
---|---|---|
An expression that does not contain periods, for example
@x | Nested function, local function, private function, path function | Local variable, nested function, local function, private function, path function |
An expression that contains exactly one period, for example
@x.y | Local variable, path function | Local variable, path function (Same as MATLAB) |
An expression that contains more than one period, for example
@x.y.z | Path function | Local variable, path function |
If x
is a local variable that is itself a function handle,
generated code and MATLAB interpret the expression @x
differently:
MATLAB produces an error.
Generated code interprets
@x
as the function handle ofx
itself.
Here is an example that shows this difference in behavior for an expression that contains two periods.
Suppose that your current working folder contains a MATLAB namespace x
, which contains another namespace
y
, which contains the function z
. The
current working folder also contains the entry-point function foo
for which you want to generate code.
This is the definition for the file foo
:
function out = foo x.y.z = @()'x.y.z is an anonymous function'; out = g(x); end function out = g(x) f = @x.y.z; out = f(); end
This is the definition for function z
:
function out = z out = 'x.y.z is a namespace function'; end
Generate a MEX function for foo
. Separately call both the
generated MEX function foo_mex
and the MATLAB function foo
.
codegen foo
foo_mex
foo
ans = 'x.y.z is an anonymous function' ans = 'x.y.z is a namespace function'
The generated code produces the first output. MATLAB produces the second output. Code generation resolves
@x.y.z
to the local variable x
that is
defined in foo
. MATLAB resolves @x.y.z
to z
, which is
within the namespace x.y
.
Termination Behavior
Generated code does not match the termination behavior of MATLAB source code. For example, if infinite loops do not have side effects, optimizations remove them from generated code. As a result, the generated code can possibly terminate even though the corresponding MATLAB code does not.
Size of Variable-Size N-D Arrays
For variable-size N-D arrays, the size
function might return a
different result in generated code than in MATLAB source code. The size
function sometimes returns
trailing ones (singleton dimensions) in generated code, but
always drops trailing ones in MATLAB. For example, for an N-D array X
with
dimensions [4 2 1 1]
, size(X)
might return
[4 2 1 1]
in generated code, but
always returns [4 2]
in
MATLAB. See Incompatibility with MATLAB in Determining Size of Variable-Size N-D Arrays.
Size of Empty Arrays
The size of an empty array in generated code might be different from its size in MATLAB source code. See Incompatibility with MATLAB in Determining Size of Empty Arrays.
Size of Empty Array That Results from Deleting Elements of an Array
Deleting all elements of an array results in an empty array. The size of this empty array in generated code might differ from its size in MATLAB source code.
Case | Example Code | Size of Empty Array in MATLAB | Size of Empty Array in Generated Code |
---|---|---|---|
Delete all elements of an
m-by-n array by using the colon operator
(: ). |
coder.varsize('X',[4,4],[1,1]);
X = zeros(2);
X(:) = [];
| 0-by-0 | 1-by-0 |
Delete all elements of a row
vector by using the colon operator
(: ). |
coder.varsize('X',[1,4],[0,1]);
X = zeros(1,4);
X(:) = []; | 0-by-0 | 1-by-0 |
Delete all elements of a
column vector by using the colon operator
(: ). |
coder.varsize('X',[4,1],[1,0]);
X = zeros(4,1);
X(:) = []; | 0-by-0 | 0-by-1 |
Delete all elements of a column vector by deleting one element at a time. |
coder.varsize('X',[4,1],[1,0]); X = zeros(4,1); for i = 1:4 X(1)= []; end | 1-by-0 | 0-by-1 |
Growing Variable-Size Column Cell Array That is Initialized as Scalar at Run Time
In MATLAB execution, if you grow a scalar cell array by using
{end+1}
indexing, the cell array grows along the second
dimension and produces a row cell array. For example, define the function
growCell
:
function z = growCell(n, m) for i = 1:m n{end+1} = m; end z = n; end
Call growCell
with example inputs:
growCell({2}, 3)
ans = 1×4 cell array {[2]} {[3]} {[3]} {[3]}
By contrast, in code generation, suppose that:
You specify the cell array to be of variable-size column type (for example,
:Inf x 1
) at compile time, andInitialize this cell array as a scalar at run time.
In such situations, the generated code grows the scalar cell array
along the first dimension and produces a column cell array. For example, generate
MEX code for growCell
. Specify the input n
to
be a :Inf x 1
cell array with double as the underlying type.
Specify the input m
to be of double scalar type.
codegen growCell -args {coder.typeof({0}, [Inf 1], [1 0]), 1}
Code generation successful.
Run the generated MEX with the same inputs as before.
growCell_mex({2}, 3)
ans = 4×1 cell array {[2]} {[3]} {[3]} {[3]}
Binary Element-Wise Operations with Single and Double Operands
If your MATLAB code contains a binary element-wise operation that involves a single type operand and a double type operand, the generated code might not produce the same result as MATLAB.
For such an operation, MATLAB casts both operands to double type and performs the operation with the double types. MATLAB then casts the result to single type and returns it.
The generated code casts the double type operand to single type. It then performs the operation with the two single types and returns the result.
For example, define a MATLAB function foo
that calls the binary element-wise
operation plus
.
function out = foo(a,b) out = a + b; end
Define a variable s1
of single type and a variable
d1
of double type. Generate a MEX function for
foo
that accepts a single type input and a double type
input.
s1 = single(1.4e32); d1 = -5.305e+32; codegen foo -args {s1, d1}
Call both foo
and foo_mex
with inputs
s1
and d1
. Compare the two results.
ml = foo(s1,d1); mlc = foo_mex(s1,d1); ml == mlc
ans = logical 0
The output of the comparison is a logical 0
, which indicates
that the generated code and MATLAB produces different results for these inputs.
Floating-Point Numerical Results
The generated code might not produce the same floating-point numerical results as MATLAB in these:
When computer hardware uses extended precision registers
NaN and Infinity
The generated code might not produce exactly the same pattern of
NaN
and Inf
values as MATLAB code when these values are mathematically meaningless. For
example, if MATLAB output contains a NaN
, output from the
generated code should also contain a NaN
, but not necessarily
in the same place.
The bit pattern for NaN
can differ between MATLAB code output and generated code output because the C99 language
standard that is used to generate code does not specify a unique bit pattern for
NaN
across all
implementations. Avoid comparing bit patterns across different implementations,
for example, between MATLAB output and SIL or PIL output.
Negative Zero
In a floating-point type, the value 0
has either a positive
sign or a negative sign. Arithmetically, 0
is equal to
-0
, but some operations are sensitive to the sign of a 0
input. Examples include rdivide
, atan2
,
atan2d
, and angle
. Division by
0
produces Inf
, but division by
-0
produces -Inf
. Similarly,
atan2d(0,-1)
produces 180
, but
atan2d (-0,-1)
produces -180
.
If the code generator detects that a floating-point variable takes only integer
values of a suitable range, then the code generator can use an integer type for the
variable in the generated code. If the code generator uses an integer type for the
variable, then the variable stores -0
as +0
because an integer type does not store a sign for the value 0
. If
the generated code casts the variable back to a floating-point type, the sign of
0
is positive. Division by 0
produces
Inf
, not -Inf
. Similarly,
atan2d(0,-1)
produces 180
, not
-180
.
There are other contexts in which the generated code might treat
-0
differently than MATLAB. For example, suppose that your MATLAB code computes the minimum of two scalar doubles x
and y
by using z = min(x,y)
. The corresponding
line in the generated C code might be z = fmin(x,y)
. The function
fmin
is defined in the runtime math library of the C
compiler. Because the comparison operation 0.0 == -0.0
returns
true
in C/C++, the compiler's implementation of
fmin
might return either 0.0
or
-0.0
for fmin(0.0,-0.0)
.
Code Generation Target
The coder.target
function returns different values in
MATLAB than in the generated code. The intent is to help you determine
whether your function is executing in MATLAB or has been compiled for a simulation or code generation target.
See coder.target
.
MATLAB Class Property Initialization
Before code generation, at class loading time, MATLAB computes class default values. The code generator uses the values that MATLAB computes. It does not recompute default values. If the property definition uses a function call to compute the initial value, the code generator does not execute this function. If the function has side effects, such as modifying a global variable or a persistent variable, then the generated code can produce different results than MATLAB. For more information, see Defining Class Properties for Code Generation.
MATLAB Classes in Nested Property Assignments That Have Set Methods
When you assign a value to a handle object property, which is itself a property of another object, and so on, then the generated code can call set methods for handle classes that MATLAB does not call.
For example, suppose that you define a set of variables such that
x
is a handle object, pa
is an object,
pb
is a handle object, and pc
is a
property of pb
. Then you make a nested property assignment, such
as:
x.pa.pb.pc = 0;
In this case, the generated code calls the set method for the object
pb
and the set method for x
. MATLAB calls only the set method for pb
.
MATLAB Handle Class Destructors
The behavior of handle class destructors in the generated code can be different from the behavior in MATLAB in these situations:
The order of destruction of several independent objects might be different in MATLAB than in the generated code.
The lifetime of objects in the generated code can be different from their lifetime in MATLAB.
The generated code does not destroy partially constructed objects. If a handle object is not fully constructed at run time, the generated code produces an error message but does not call the
delete
method for that object. For a System object™, if there is a run-time error insetupImpl
, the generated code does not callreleaseImpl
for that object.MATLAB does call the
delete
method to destroy a partially constructed object.
For more information, see Code Generation for Handle Class Destructors.
Variable-Size Data
See Incompatibilities with MATLAB in Variable-Size Support for Code Generation.
Complex Numbers
Converting Strings with Consecutive Unary Operators to
double
Converting a string that contains multiple, consecutive unary operators to
double
can produce different results between MATLAB and the generated code. Consider this
function:
function out = foo(op) out = double(op + 1); end
For an input value "--"
, the function converts the string
"--1"
to double
. In MATLAB, the answer is NaN
. In the generated code, the
answer is 1
.
Display Function
Statements and expressions in MATLAB code that omit the semicolon implicitly invoke the
display
function. You can also explicitly invoke
display
as shown
here:
display(2+3);
5
The MEX code generated for MATLAB code that invokes the display
function preserves
calls to this function and shows the output. In standalone code generated for
targets that do not have access to MATLAB Runtime, implicit and explicit calls to display
are
removed. This includes calls to overridden class methods of
display
.
To display text in code generated for other targets, override the
disp
function in your MATLAB classes. For example:
%MATLAB Class classdef foo methods function obj = foo end function disp(self) disp("Overridden disp"); end end end %Entry-point Function function callDisp a = foo; disp(a); end
The generated code for the entry-point function is shown here:
/* Include Files */ #include "callDisp.h" #include <stdio.h> /* Function Definitions */ /* * Arguments : void * Return Type : void */ void callDisp(void) { printf("%s\n", "Overridden disp"); fflush(stdout); }
Function Handle Difference
Invoking display
through a function handle in MATLAB prints the name of the variable as well. For example, running this
function in MATLAB results in the following
output:
function displayDiff z = 10; f = @display; f(z) end
z = 10
However, the generated code for this snippet only outputs the value
10
.