Increase speed to read text file and parse date time data.

I'm searching a faster way to read this file and convert the date and time to serial date number.

Datum  Tid  Värde
2015-10-12  00:02:16   23.399999619
2015-10-12  00:07:16   23.399999619
2015-10-12  00:12:16   23.399999619
2015-10-12  00:17:16   23.399999619
2015-10-12  00:22:17   23.399999619
2015-10-12  00:27:17   23.399999619
2015-10-12  00:32:17   23.399999619
2015-10-12  00:37:17   23.399999619
...

The text file contains a few hundred lines up to several thousands. I've tested five alternate solutions on R2016a. I attach the script and the text file. The results with tic/toc and profile are consistent.

&nbsp

The best code, "sscanf", is nearly twice as fast as the "standard".

The FEX-contribution, DateStr2Num by Jan Simon, is really fast. However, I failed to find a fast way to arrange the input data to fit the function. The line,

str = strcat( cac{1}(:,1), repmat({' '},[length(cac{1}(:,2)),1]), cac{1}(:,2) );

ruins the performance. There must be a better way!

Question: Which are the possibilities to increase the speed further?

&nbsp

ADDENDUM 2016-08-31

textsdn_2 (attached) is adapted to runperf. It contains two new cases. The summary result of runperf('textsdn_2.m') is

               Name               GroupCount    mean_MeasuredTime
    __________________________    __________    _________________
    text2sdn_2/Standard           4              1.0608          
    text2sdn_2/DateStr2Num        4              0.7491          
    text2sdn_2/sscanf             4              0.41834          
    text2sdn_2/fscanf             4              0.59519          
    text2sdn_2/datetime           4              1.2143          
    text2sdn_2/DateStr2Num_19c    4              0.3142          
    text2sdn_2/dtstr2dtnummx      4              1.0475

In production a text file is read once. In these tests the file is read the first time during warmup and is from that point in time available in the system cache.

text2sdn_2/DateStr2Num_19c is three times faster than text2sdn_2/Standard and more than twice as fast as text2sdn_2/DateStr2Num. One reason is that date and time are kept together by using the format '%19c%f'. DateStr2Num doesn't distinguish between tab and space.

cac = textscan( fid, '%19c%f', 'Headerlines',1, 'CollectOutput',true );

text2sdn_2/dtstr2dtnummx is only slightly faster than text2sdn_2/Standard, which is because this test is based on a column of 480 timestamps. With a single timestamp the relative difference is much larger.

Answer 1

[Yair Altman]

@Per - try to use dtstr2dtnummx(), as explained here: http://undocumentedmatlab.com/blog/datenum-performance

Yair Altman

http://UndocumentedMatlab.com