ifort: problem reading unformatted sequential files

[gavpol]

Hi. I've written some f95 code to do a taskwhere speed of execution is important. I've been using a free compiler (gfortran) but thought I'd try Intel's, anticipating that it would be faster. I need to read hundreds of image files, to do a tomographic reconstruction, one after another. Each file consists of a 100 byte header and 1024x1024 2-byte integers. Formatted access has been painfully slow in the past, when I've tried it, so I have used the following ugly fix - pretending the images are unformatted file:

INTEGER(KIND=2), DIMENSION(N) :: PV
CHARACTER(LEN=96) :: HEADER

OPEN(UNIT=1, FILE=VIEWNAME, STATUS="OLD",FORM="UNFORMATTED",ACTION="READ")
READ(UNIT=1) HEADER,PV
CLOSE(UNIT=1,STATUS="KEEP")


By defining the HEADER length to be 4 bytes short, I account for the 4-byte bookend I expect fortran to introduce.The problem is: this works fine in gfortran but ifort has a fit when it tries to read the files (severe error (39) I think). My questions are:

(1) Is there a way to get this to work with ifort?
(2) Or can someone suggest a better way to read my files with comparable speed.

Any help would be much appreciated.

Thanks

G.

[Steven_L_Intel1]

Are you aware that there is a 4-byte "bookend" on both ends of the record? This should be the same as gfortran.

I would recommend opening the file as ACCESS='STREAM' (keep the other things you have in the OPEN.) Then when you do unformatted reads, you'll read ONLY data and there will not be any "bookends". This is a standard Fortran 2003 feature. You'll want to redefine the length of HEADER to be *100.

[jimdempseyatthecove]

If you want maximum performance using IVF consider using Binary Sequential files (must be written in binary sequential to be read in binary sequential).

If you use a user defined type containing your 100 byte record plus the 1024x1024 2 byte integers pluspadd data to next 512 byte addressyou could potentially read this with 1 I/O operation directly into the user defined type variable. Cann't say if IVF would bypass reading to an internal I/O buffer when reading disk block sized and aligned binary data.

Jim

[gavpol]

Hey.

Thanks alot! The stream i/o works a treat. Which has made my day. Didn't know about that.

Thecode works. And it runs very fast.

G.