Incompatibilities with MATLAB in Variable-Size Support for Code Generation

Incompatibility with MATLAB for Scalar Expansion

Scalar expansion is a method of converting scalar data to match the dimensions of vector or matrix data. If one operand is a scalar and the other is not, scalar expansion applies the scalar to every element of the other operand.

During code generation, scalar expansion rules apply except when operating on two variable-size expressions. In this case, both operands must be the same size. The generated code does not perform scalar expansion even if one of the variable-size expressions turns out to be scalar at run time. Therefore, when run-time error checks are enabled, a run-time error can occur.

Consider this function:

function y = scalar_exp_test_err1(u) %#codegen
y = ones(3);
switch u
    case 0
        z = 0;
    case 1
        z = 1;
    otherwise
        z = zeros(3);
end
y(:) = z;

When you generate code for this function, the code generator determines that z is variable size with an upper bound of 3.

If you run the MEX function with u equal to 0 or 1, the generated code does not perform scalar expansion, even though z is scalar at run time. Therefore, when run-time error checks are enabled, a run-time error can occur.

scalar_exp_test_err1_mex(0)
Subscripted assignment dimension mismatch: [9] ~= [1].

Error in scalar_exp_test_err1 (line 11)
y(:) = z;

To avoid this issue, use indexing to force z to be a scalar value.

function y = scalar_exp_test_err1(u) %#codegen
y = ones(3);
switch u
    case 0
        z = 0;
    case 1
        z = 1;
    otherwise
        z = zeros(3);
end
y(:) = z(1);

Incompatibility with MATLAB in Determining Size of Variable-Size N-D Arrays

For variable-size N-D arrays, the size function can return a different result in generated code than in MATLAB^®. In generated code, size(A) returns a fixed-length output because it does not drop trailing singleton dimensions of variable-size N-D arrays. By contrast, size(A) in MATLAB returns a variable-length output because it drops trailing singleton dimensions.

For example, if the shape of array A is :?x:?x:? and size(A,3)==1, size(A) returns:

• Three-element vector in generated code

• Two-element vector in MATLAB code

Workarounds

If your application requires generated code to return the same size of variable-size N-D arrays as MATLAB code, consider one of these workarounds:

• Use the two-argument form of size.

  For example, size(A,n) returns the same answer in generated code and MATLAB code.

• Rewrite size(A):

  B = size(A);
  X = B(1:ndims(A));

  This version returns X with a variable-length output. However, you cannot pass a variable-size X to matrix constructors such as zeros that require a fixed-size argument.

Incompatibility with MATLAB in Determining Size of Empty Arrays

The size of an empty array in generated code might be different from its size in MATLAB source code. The size might be 1x0 or 0x1 in generated code, but 0x0 in MATLAB. Therefore, you should not write code that relies on the specific size of empty matrices.

For example, consider the following code:

function y = foo(n) %#codegen
x = [];
i = 0;
while (i < 10)
    x = [5 x];
    i = i + 1;
end
if n > 0
    x = [];
end
y = size(x);
end

Concatenation requires its operands to match on the size of the dimension that is not being concatenated. In the preceding concatenation, the scalar value has size 1x1 and x has size 0x0. To support this use case, the code generator determines the size for x as [1 x :?]. Because there is another assignment x = [] after the concatenation, the size of x in the generated code is 1x0 instead of 0x0.

This behavior persists while determining the size of empty character vectors which are denoted as ''. For example, consider the following code:

function out = string_size
out = size('');
end

Here, the value of out might be 1x0 or 0x1 in generated code, but 0x0 in MATLAB.

For incompatibilities with MATLAB in determining the size of an empty array that results from deleting elements of an array, see Size of Empty Array That Results from Deleting Elements of an Array.

Workaround

If your application checks whether a matrix is empty, use one of these workarounds:

• Rewrite your code to use the isempty function instead of the size function.

• Instead of using x=[] to create empty arrays, create empty arrays of a specific size using zeros. For example:

  function y = test_empty(n) %#codegen
  x = zeros(1,0);
  i=0;
  while (i < 10)
      x = [5 x];
      i = i + 1;
  end
  if n > 0
      x = zeros(1,0);
  end
  y=size(x);
  end

Incompatibility with MATLAB in Determining Class of Empty Arrays

The class of an empty array in generated code can be different from its class in MATLAB source code. Therefore, do not write code that relies on the class of empty matrices.

For example, consider the following code:

function y = fun(n)
x = [];
if n > 1
    x = ['a' x];
end
y=class(x);
end 

fun(0) returns double in MATLAB, but char in the generated code. When the statement n > 1 is false, MATLAB does not execute x = ['a' x]. The class of x is double, the class of the empty array. However, the code generator considers all execution paths. It determines that based on the statement x = ['a' x], the class of x is char.

Workaround

Instead of using x=[] to create an empty array, create an empty array of a specific class. For example, use blanks(0) to create an empty array of characters.

function y = fun(n)
x = blanks(0);
if n > 1
    x = ['a' x];
end
y=class(x);
end

Incompatibility with MATLAB in Matrix-Matrix Indexing

In matrix-matrix indexing, you use one matrix to index into another matrix. In MATLAB, the general rule for matrix-matrix indexing is that the size and orientation of the result match the size and orientation of the index matrix. For example, if A and B are matrices, size(A(B)) equals size(B). When A and B are vectors, MATLAB applies a special rule. The special vector-vector indexing rule is that the orientation of the result is the orientation of the data matrix. For example, if A is 1-by-5 and B is 3-by-1, then A(B) is 1-by-3.

The code generator applies the same matrix-matrix indexing rules as MATLAB. If A and B are variable-size matrices, to apply the matrix-matrix indexing rules, the code generator assumes that size(A(B)) equals size(B). If, at run time, A and B become vectors and have different orientations, then the assumption is incorrect. Therefore, when run-time error checks are enabled, an error can occur.

To avoid this issue, force your data to be a vector by using the colon operator for indexing. For example, suppose that your code intentionally toggles between vectors and regular matrices at run time. You can do an explicit check for vector-vector indexing.

...
if isvector(A) && isvector(B)
    C = A(:);
    D = C(B(:));
else
    D = A(B);
end
...

The indexing in the first branch specifies that C and B(:) are compile-time vectors. Therefore, the code generator applies the indexing rule for indexing one vector with another vector. The orientation of the result is the orientation of the data vector, C.

Incompatibility with MATLAB in Vector-Vector Indexing

In MATLAB, the special rule for vector-vector indexing is that the orientation of the result is the orientation of the data vector. For example, if A is 1-by-5 and B is 3-by-1, then A(B) is 1-by-3. If, however, the data vector A is a scalar, then the orientation of A(B) is the orientation of the index vector B.

The code generator applies the same vector-vector indexing rules as MATLAB. If A and B are variable-size vectors, to apply the indexing rules, the code generator assumes that the orientation of B matches the orientation of A. At run time, if A is scalar and the orientation of A and B do not match, then the assumption is incorrect. Therefore, when run-time error checks are enabled, a run-time error can occur.

To avoid this issue, make the orientations of the vectors match. Alternatively, index single elements by specifying the row and column. For example, A(row, column).

Incompatibility with MATLAB in Matrix Indexing Operations for Code Generation

The following limitation applies to matrix indexing operations for code generation:

• Initialization of the following style:

  for i = 1:10
      M(i) = 5;
  end
  

  In this case, the size of M changes as the loop is executed. Code generation does not support increasing the size of an array over time.

  For code generation, preallocate M.

  M = zeros(1,10);
  for i = 1:10
      M(i) = 5;
  end
  

The following limitation applies to matrix indexing operations for code generation when dynamic memory allocation is disabled:

• M(i:j) where i and j change in a loop

  During code generation, memory is not dynamically allocated for the size of the expressions that change as the program executes. To implement this behavior, use for-loops as shown:

  ...
  M = ones(10,10);
  for i=1:10
      for j = i:10
          M(i,j) = 2*M(i,j);
      end
  end
  ...

  Note

  The matrix M must be defined before entering the loop.

Incompatibility with MATLAB in Concatenating Variable-Size Matrices

For code generation, when you concatenate variable-size arrays, the dimensions that are not being concatenated must match exactly.

Differences When Curly-Brace Indexing of Variable-Size Cell Array Inside Concatenation Returns No Elements

Suppose that:

• c is a variable-size cell array.

• You access the contents of c by using curly braces. For example, c{2:4}.

• You include the results in concatenation. For example, [a c{2:4} b].

• c{I} returns no elements. Either c is empty or the indexing inside the curly braces produces an empty result.

For these conditions, MATLAB omits c{I} from the concatenation. For example, [a c{I} b] becomes [a b]. The code generator treats c{I} as the empty array [c{I}]. The concatenation becomes [...[c{i}]...]. This concatenation then omits the array [c{I}]. So that the properties of [c{I}] are compatible with the concatenation [...[c{i}]...], the code generator assigns the class, size, and complexity of [c{I}] according to these rules:

• The class and complexity are the same as the base type of the cell array.

• The size of the second dimension is always 0.

• For the rest of the dimensions, the size of Ni depends on whether the corresponding dimension in the base type is fixed or variable size.

  • If the corresponding dimension in the base type is variable size, the dimension has size 0 in the result.

  • If the corresponding dimension in the base type is fixed size, the dimension has that size in the result.

Suppose that c has a base type with class int8 and size:10x7x8x:?. In the generated code, the class of [c{I}] is int8. The size of [c{I}] is 0x0x8x0. The second dimension is 0. The first and last dimensions are 0 because those dimensions are variable size in the base type. The third dimension is 8 because the size of the third dimension of the base type is a fixed size 8.

Inside concatenation, if curly-brace indexing of a variable-size cell array returns no elements, the generated code can have the following differences from MATLAB:

• The class of [...c{i}...] in the generated code can differ from the class in MATLAB.

  When c{I} returns no elements, MATLAB removes c{I} from the concatenation. Therefore, c{I} does not affect the class of the result. MATLAB determines the class of the result based on the classes of the remaining arrays, according to a precedence of classes. See Valid Combinations of Unlike Classes. In the generated code, the class of [c{I}] affects the class of the result of the overall concatenation [...[c{I}]...] because the code generator treats c{I} as [c{I}]. The previously described rules determine the class of [c{I}].

• In the generated code, the size of [c{I}] can differ from the size in MATLAB.

  In MATLAB, the concatenation [c{I}] is a 0x0 double. In the generated code, the previously described rules determine the size of [c{I}].