Merge pull request #246 from vmagnin/fortran90org

Transferring fortran90.org "Fortran Best Practices" into a mini-book

Author: awvwgk

_data/learning.yml

@@ -53,6 +53,23 @@ books:
       - link: /learn/os_setup/ides
       - link: /learn/os_setup/tips
 
+  - title: Fortran Best Practices
+    description: This tutorial collects a modern canonical way of doing things in Fortran.
+    category: Getting started
+    link: /learn/best_practices
+    pages:
+        - link: /learn/best_practices/style_guide
+        - link: /learn/best_practices/floating_point
+        - link: /learn/best_practices/integer_division
+        - link: /learn/best_practices/modules_programs
+        - link: /learn/best_practices/arrays
+        - link: /learn/best_practices/multidim_arrays
+        - link: /learn/best_practices/element_operations
+        - link: /learn/best_practices/allocatable_arrays
+        - link: /learn/best_practices/file_io
+        - link: /learn/best_practices/callbacks
+        - link: /learn/best_practices/type_casting
+
 # Web links listed at the bottom of the 'Learn' landing page
 #
 reference-links:

learn/best_practices.md

@@ -0,0 +1,131 @@
+---
+layout: book
+title: Allocatable Arrays
+permalink: /learn/best_practices/allocatable_arrays
+---
+
+The ``allocatable`` attribute provides safe way for memory handling.
+In comparison to variables with ``pointer`` attribute the memory is managed
+automatically and will be deallocated automatically once the variable goes
+out-of-scope. Using ``allocatable`` variables removes the possibility to
+create memory leaks in an application.
+
+They can be used in subroutines to create scratch or work arrays, where
+automatic arrays would become too large to fit on the stack.
+
+```fortran
+real(dp), allocatable :: temp(:)
+allocate(temp(10))
+```
+
+The allocation status can be checked using the ``allocated`` intrinsic
+to avoid uninitialized access
+
+```fortran
+subroutine show_arr(arr)
+  integer, allocatable, intent(in) :: arr(:)
+
+  if (allocated(arr)) then
+    print *, arr
+  end if
+end subroutine show_arr
+```
+
+To allocate variables inside a procedure the dummy argument has to carry
+the ``allocatable`` attribute. Using it in combination with ``intent(out)``
+will deallocate previous allocations before entering the procedure:
+
+```fortran
+subroutine foo(lam)
+  real(dp), allocatable, intent(out) :: lam(:)
+  allocate(lam(5))
+end subroutine foo
+```
+
+The allocated array can be used afterwards like a normal array
+
+```fortran
+real(dp), allocatable :: lam(:)
+call foo(lam)
+```
+
+An already allocated array cannot be allocated again without prior deallocation.
+Similarly, deallocation can only be invoked for allocated arrays. To reallocate
+an array use
+
+```fortran
+if (allocated(lam)) deallocate(lam)
+allocate(lam(10))
+```
+
+Passing allocated arrays to procedures does not require the ``allocatable`` attribute
+for the dummy arguments anymore.
+
+```fortran
+subroutine show_arr(arr)
+  integer, intent(in) :: arr(:)
+
+  print *, arr
+end subroutine show_arr
+
+subroutine proc
+  integer :: i
+  integer, allocatable :: arr
+
+  allocate(arr(5))
+
+  do i = 1, size(arr)
+    arr(i) = 2*i + 1
+  end do
+  call show_arr(arr)
+end subroutine proc
+```
+
+Passing an unallocated array in this context will lead to an invalid memory access.
+Allocatable arrays can be passed to ``optional`` dummy arguments, if they are unallocated
+the argument will not be present. The ``allocatable`` attribute is not limited to
+arrays and can also be associated with scalars, which can be useful in combination
+with ``optional`` dummy arguments.
+
+Allocations can be moved between different arrays with ``allocatable`` attribute
+using the ``move_alloc`` intrinsic subroutine.
+
+```fortran
+subroutine resize(var, n)
+  real(wp), allocatable, intent(inout) :: var(:)
+  integer, intent(in), optional :: n
+  integer :: this_size, new_size
+  integer, parameter :: inital_size = 16
+
+  if (allocated(var)) then
+    this_size = size(var, 1)
+    call move_alloc(var, tmp)
+  else
+    this_size = initial_size
+  end if
+
+  if (present(n)) then
+    new_size = n
+  else
+    new_size = this_size + this_size/2 + 1
+  end if
+
+  allocate(var(new_size))
+
+  if (allocated(tmp)) then
+    this_size = min(size(tmp, 1), size(var, 1))
+    var(:this_size) = tmp(:this_size)
+  end if
+end subroutine resize
+```
+
+Finally, allocations do not initialize the array, the content of the uninitialized
+array is most likely just the bytes of whatever was previously at the respective address.
+The allocation supports initialization using the source attribute:
+
+```fortran
+real(dp), allocatable :: arr(:)
+allocate(arr(10), source=0.0_dp)
+```
+
+The ``source`` keyword supports scalar and array valued variables and constants.

learn/best_practices/allocatable_arrays.md

@@ -0,0 +1,200 @@
+---
+layout: book
+title: Arrays
+permalink: /learn/best_practices/arrays
+---
+
+Arrays are a central object in Fortran. The creation of dynamic sized arrays
+is discussed in the [allocatable arrays arrays](./allocatable_arrays.html).
+
+To pass arrays to procedures four ways are available
+
+1. *assumed-shape* arrays
+4. *assumed-rank* arrays
+2. *explicit-shape* arrays
+3. *assumed-size* arrays
+
+The preferred way to pass arrays to procedures is as *assumed-shape* arrays
+
+```fortran
+subroutine f(r)
+  real(dp), intent(out) :: r(:)
+  integer :: n, i
+  n = size(r)
+  do i = 1, n
+    r(i) = 1.0_dp / i**2
+  end do
+end subroutine f
+```
+
+Higher dimensional arrays can be passed in a similar way.
+
+```fortran
+subroutine g(A)
+  real(dp), intent(in) :: A(:, :)
+  ...
+end subroutine g
+```
+
+The array is simply passed by
+
+```fortran
+real(dp) :: r(5)
+call f(r)
+```
+
+In this case no array copy is done, which has the advantage that the shape and size
+information is automatically passed along and checked at compile and optionally at
+runtime.
+Similarly, array strides can be passed without requiring a copy of the array but as
+*assumed-shape* discriptor:
+
+```fortran
+real(dp) :: r(10)
+call f(r(1:10:2))
+call f(r(2:10:2))
+```
+
+This should always be your default way of passing arrays in and out of subroutines.
+Avoid passing arrays as whole slices, as it obfuscates the actual intent of the code:
+
+```fortran
+real(dp) :: r(10)
+call f(r(:))
+```
+
+In case more general arrays should be passed to a procedure the *assumed-rank*
+functionality introduced in the Fortran 2018 standard can be used
+
+```fortran
+subroutine h(r)
+  real(dp), intent(in) :: r(..)
+  select rank(r)
+  rank(1)
+  ! ...
+  rank(2)
+  ! ...
+  end select
+end subroutine h
+```
+
+The actual rank can be queried at runtime using the ``select rank`` construct.
+This easily allows to create more generic functions that have to deal with
+differnet array ranks.
+
+*Explicit-shape* arrays can be useful for returning data from functions.
+Most of their functionality can be provided by *assumed-shape* and *assumed-rank*
+arrays but they find frequent use for interfacing with C or in legacy Fortran
+procedures, therefore they will be discussed briefly here.
+
+To use *explicit-shape* arrays, the dimension has to be passed explicitly as dummy
+argument like in the example below
+
+``` fortran
+subroutine f(n, r)
+  integer, intent(in) :: n
+  real(dp), intent(out) :: r(n)
+  integer :: i
+  do i = 1, n
+    r(i) = 1.0_dp / i**2
+  end do
+end subroutine
+```
+
+For high-dimensional arrays additional indices have to be passed.
+
+``` fortran
+subroutine g(m, n, A)
+  integer, intent(in) :: m, n
+  real(dp), intent(in) :: A(m, n)
+  ...
+end subroutine
+```
+
+The routines can be invoked by
+
+``` fortran
+real(dp) :: r(5), s(3, 4)
+call f(size(r), r)
+call g(size(s, 1), size(s, 2), s)
+```
+
+Note that the shape is not checked, therefore the following would be valid code
+with will potentially yield incorrect results:
+
+```fortran
+real(dp) :: s(3, 4)
+call g(size(s), 1, s)  ! s(12, 1) in g
+call g(size(s, 2), size(s, 1), s)  ! s(4, 3) in g
+```
+
+In this case the memory layout is preserved but the shape is changed.
+Also, *explicit-shape* arrays require contiguous memory and will create temporary
+arrays in case non-contiguous array strides are passed.
+
+To return an array from a function with *explicit-shape* use
+
+``` fortran
+function f(n) result(r)
+  integer, intent(in) :: n
+  real(dp) :: r(n)
+  integer :: i
+  do i = 1, n
+    r(i) = 1.0_dp / i**2
+  end do
+end function
+```
+
+Finally, there are *assumed-size* arrays, which provide the least compile and runtime
+checking and can be found be found frequently in legacy code, they should be avoided
+in favour of *assumed-shape* or *assumed-rank* arrays.
+An *assumed-size* array dummy argument is identified by an asterisk as the last dimension,
+this disables the usage of this array with many intrinsic functions, like ``size`` or
+``shape``.
+
+To check for the correct size and shape of an *assumed-shape* array the ``size`` and
+``shape`` intrinsic functions can be used to query for those properties
+
+```fortran
+if (size(r) /= 4) error stop "Incorrect size of 'r'"
+if (any(shape(r) /= [2, 2])) error stop "Incorrect shape of 'r'"
+```
+
+Note that ``size`` returns the total size of all dimensions, to obtain the shape of
+a specific dimension add it as second argument to the function.
+
+Arrays can be initialized by using an array constructor
+
+```fortran
+integer :: r(5)
+r = [1, 2, 3, 4, 5]
+```
+
+The array constructor can be annoted with the type of the constructed array
+
+```fortran
+real(dp) :: r(5)
+r = [real(dp) :: 1, 2, 3, 4, 5]
+```
+
+Implicit do loops can be used inside an array constructor as well
+
+```fortran
+integer :: i
+real(dp) :: r(5)
+r = [(real(i**2, dp), i = 1, size(r))]
+```
+
+In order for the array to start with different index than 1, do:
+
+```fortran
+subroutine print_eigenvalues(kappa_min, lam)
+  integer, intent(in) :: kappa_min
+  real(dp), intent(in) :: lam(kappa_min:)
+
+  integer :: kappa
+  do kappa = kappa_min, ubound(lam, 1)
+    print *, kappa, lam(kappa)
+  end do
+end subroutine print_eigenvalues
+```