Behaviour and limitations

What VecFor does at the edges, and what it does not do. Every statement of this page is checked by a program of the docs or by the sources.

Comparisons

• <, <=, >, >= between two vectors compare their lengths: two vectors of equal length and different direction are both <= and >= each other. Between a vector and a number, all six compare the length with the number.
• == between two vectors compares the lengths and then the components of the unit vectors; /= is its negation.
• All are exact: no tolerance. For computed vectors, compare normL2(a - b) with a tolerance of your own.

Tolerances of the predicates

is_collinear and is_concyclic take an optional tolerance, 0 by default (is_concyclic adds tiny, the smallest positive normal number). Without it the tests are exact, and points that are collinear or concyclic only up to rounding, as computed points are, pass or fail by chance: pass a tolerance scaled to your geometry. is_concyclic also needs the points in their order around the circle.

Zero vectors and degenerate geometry

Case	Result
normalize, normalized of the zero vector	the zero vector
normalized_checked of a vector not longer than min_norm	ok false, the zero vector
angle with a zero vector	0
a .paral. b, a .ortho. b with b zero	NaN
distance_to_line with pt1 = pt2	NaN
distance_to_plane and the projections, with three collinear points	0, and the point itself: silently
face_normal3, face_normal4 of a degenerate face	the zero vector, also with norm
barycentric_triangle, barycentric_tetrahedron of a degenerate cell	IEEE infinities or NaN
a / b with a zero component of b; a / 0	IEEE infinities or NaN, as for numbers

Input and output

• save_into_file, load_from_file and printf never stop the program on an I/O error: without iostat the error is silently ignored, and a vector that fails to load keeps its old value. Always pass iostat.
• fmt is a format string: '*' is not list-directed I/O.
• There is no rec= for direct access files: write the components.
• The I/O procedures and iolen are not pure.

Types and kinds

• Vectors of two kinds do not mix: no a4 + a8, no a8 = a4. Convert by components.
• vector and vector_R8P are two different types, although of the same kind; ex and ex_R8P likewise.
• vector_R16P is quadruple precision only when PENF and VecFor are compiled with -DPENF_R16P, as the GNU and Intel builds are. nvfortran has no 128-bit real kind: there vector_R16P is double precision (see Precision and kinds).
• The scalar arguments of the methods (angle=, tolerance=) and the matrices must have the kind of the vector; the numbers of the operators may have any kind.

Fortran itself

• A method needs a variable or a named constant: (a - b)%normL2() is not Fortran; use the free function, normL2(a - b).
• The user-defined operators bind more loosely than any intrinsic one: a .dot. b + c is a .dot. (b + c); parenthesise (see Precedence).
• The intrinsics sum, dot_product, matmul, maxval, ... do not apply to arrays of vectors.

Not elemental

rotate, mirror, rotation_matrix, mirror_matrix, .matrix. and the I/O procedures work on one vector at a time; loop over arrays. polygon_vector_area takes an array, the vertices of one polygon, and barycentric_triangle and barycentric_tetrahedron return an array, the coordinates of one point: they are pure, not elemental. Every other operator and function, and normalize, normalized_checked and closest_point_triangle, are elemental.

Not on the device

Only the routines of the device-callable API run inside an OpenACC kernel: the sum, the difference and the products of vectors, the products with an R8P number, the matrices, the geometric methods, the accurate products, the measures and the point location. The other operators, the comparisons and the input/output do not.

Not available

Vectors of two or of more than three components; non-Cartesian frames; a type for points distinct from the type for directions (VecFor uses one type for both); quaternions.