Reputation: 1827
I'm working on a project where I need to write some existing data to disk as ascii. I have something that works, but the IO itself is quite expensive and I'd like to optimise it further.
The data is basically an array of reals, however some of the columns store encoded strings which need to be recast as character strings (don't ask!). The input and output of this problem are beyond my control, I am receiving this real array and need to write it out as ascii.
I know that writing the array in one go as an unformatted write is faster, but this doesn't deal with the string columns correctly. Any ideas?
Here is some example code:
program test
implicit none
integer(kind=4), parameter :: nrows = 5000
integer(kind=4), parameter :: ncols = 400
integer, parameter :: real_kind = 8
integer(kind=4) :: i,j, handle
character(len=256) :: value_str
character(len=1) :: delimiter
real(kind=real_kind) :: data(nrows,ncols)
delimiter = " "
data(:,:) = 999.999
! Some examples of the "string columns"
data(:,10) = transfer(' foo ',data(1,1))
data(:,20) = transfer(' bar ',data(1,1))
handle=10
open(handle,file="out.txt",status="replace", access="stream")
do i=1,nrows
do j=1,ncols
! If this column contains encoded strings then recast
if((j==10).or.(j==20))then
write(handle) delimiter
value_str = transfer(data(i,j),value_str(1:real_kind))
write(handle) trim(value_str)
else
write(value_str,*) data(i,j)
write(handle) trim(value_str)
endif
enddo
write(handle) new_line('x')
enddo
close(handle)
end program test
gfortran test.F90 -o test.x
time test.x
real 0m2.65s
user 0m2.24s
sys 0m0.04s
Edit: removed "if(j/=1)" from original test.F90 code sample in response to comment.
Upvotes: 1
Views: 126
Reputation: 4656
Use the free formatting and have the system handle more for you. In this proposition, I handle the transfer beforehand and use a single loop to write the data to file. This is handy if you have only few columns of character data like the 2 in your example.
Your code will look like this
program test
implicit none
integer(kind=4), parameter :: nrows = 5000
integer(kind=4), parameter :: ncols = 400
integer, parameter :: real_kind = 8
integer, parameter :: pos1 = 10 ! I like named constants
integer, parameter :: pos2 = 20 ! I like named constants
integer(kind=4) :: i,j, handle
character(len=256) :: value_str
character(len=1) :: delimiter
real(kind=real_kind) :: data(nrows,ncols)
character(real_kind), dimension(nrows,2) :: cdata ! two columns array for
delimiter = " "
data(:,:) = 999.999
! Some examples of the "string columns"
data(:,pos1) = transfer(' foo ',data(1,1))
data(:,pos2) = transfer(' bar ',data(1,1))
handle=10
open(handle,file="out.txt",status="replace", form="formatted")
! Transfer beforehand
cdata(:,1) = transfer( data(:,pos1), cdata(1,1) )
cdata(:,2) = transfer( data(:,pos2), cdata(1,1) )
do i=1,nrows
write(handle,*) data(i,1:pos1-1), cdata(i,1)&
, data(i,pos1+1:pos2-1), cdata(i,2)&
, data(i,pos2+1:)
enddo
close(handle)
end program test
and give this timing
time ./test.x
real 0m1.696s
user 0m1.661s
sys 0m0.029s
instead of
time ./test.x
real 0m2.654s
user 0m2.616s
sys 0m0.032s
On my computer
Upvotes: 2