Performance
| -O levels | |
| Memory layout first | |
| Whole-array wins | |
| Measure with gprof concept |
-O levels
Optimization levels trade compile time for run speed. Level zero means no optimization and fast builds, level two is a balanced default for release runs.
Start development with no optimization and checks on, then measure again with -O2 for speed. Extra architecture flags stay as text in build notes until needed.
gfortran -O2 main.f90 -o main_fast
main_fast runs the same results with less time
Memory layout first
Fortran stores arrays in column major order, so elements with neighboring first indices sit next to each other in memory.
Loop over the first index in the inner loop. This walks memory in order and keeps the cache busy with useful data.
program layout_demo
implicit none
integer, parameter :: n = 4
real :: a(n, n)
integer :: i
integer :: j
do j = 1, n
do i = 1, n
a(i, j) = real(i + j)
end do
end do
print *, a(1, 1), a(n, n)
end program layout_demo
Whole-array wins
Whole-array statements express math on all elements at once and let the compiler pick a fast loop.
Prefer array assignment and intrinsics like sum and matmul for clear code before writing manual loops.
program array_demo
implicit none
real :: b(4)
real :: c(4)
real :: a(4)
b = [1.0, 2.0, 3.0, 4.0]
c = [5.0, 6.0, 7.0, 8.0]
a = b + c
print *, a, sum(a)
end program array_demo
Measure with gprof concept
Profiling shows where time really goes. The classic gprof idea is simple: build with profiling marks, run a typical workload, then read the flat report.
Optimize only the top routine from the report. Guessing without numbers usually tunes code that hardly matters.
gfortran -pg main.f90 -o main_prof
run workload, then read gmon report for hot routines
Next: Modern Fortran tour: 2008, 2018, 2023
Article author: Arthur Isaev
