Perfomance Considerations

[Alexander Adam]

Hi,

What I am having is this -- I got a string start (type is wchar_t) and
the end of a string to compare. Now I've got around 500 words to
compare to and if matched, I want to get an unique integer for the
word. Currently this is done by filling up a std::map<wstrin g, int>
with all words and doing an: iterator it = my_map.find(wch ar_t* value)
and return it->second if found, otherwise return 0. Now this routine
takes pretty much time and I was wondering whether doing a manual list
of

if (strncmp(...) == 0)
...
else if (strncmp(...) == 0)
...

would be faster / more efficient? I was also thinking about doing
something like

switch (*str)
{
case 'a':
if (strncmp(...))
}

i.e. first compaing the first char and then do some deeper evaluation
to avoid calling the strncmp subroutine but I doubt that this even
makes sense.

Any idea so far or any closer description required? Basically as said,
I am having a wchar_t* str_start and an wchar_t* str_end which I want
to compare to predefined words and match an unique integer code for
it.

Regards
Alexander

[red floyd]

Alexander Adam wrote:

Hi,

What I am having is this -- I got a string start (type is wchar_t) and
the end of a string to compare. Now I've got around 500 words to
compare to and if matched, I want to get an unique integer for the
word. Currently this is done by filling up a std::map<wstrin g, int>
with all words and doing an: iterator it = my_map.find(wch ar_t* value)
and return it->second if found, otherwise return 0. Now this routine
takes pretty much time and I was wondering whether doing a manual list
of

if (strncmp(...) == 0)
..
else if (strncmp(...) == 0)
..

would be faster / more efficient? I was also thinking about doing
something like

switch (*str)
{
case 'a':
if (strncmp(...))
}

i.e. first compaing the first char and then do some deeper evaluation
to avoid calling the strncmp subroutine but I doubt that this even
makes sense.

Any idea so far or any closer description required? Basically as said,
I am having a wchar_t* str_start and an wchar_t* str_end which I want
to compare to predefined words and match an unique integer code for
it.

This looks like a clear case of Hoare's Dictum. Have you even
determined if this is your bottleneck? Have you looked into alternate
datastructures and algorithms? Is there a reason your're using strncmp
instead of wcsncmp (since you're dealing with wchar_t)? Why don't you
convert your wchar_t array to a wstring?

There's lots of stuff to be done up front before you worry about this
sort of micro-optimization.

[Marco Manfredini]

Alexander Adam wrote:

Hi,

What I am having is this -- I got a string start (type is wchar_t) and
the end of a string to compare. Now I've got around 500 words to
compare to and if matched, I want to get an unique integer for the
word. Currently this is done by filling up a std::map<wstrin g, int>
with all words and doing an: iterator it = my_map.find(wch ar_t* value)
and return it->second if found, otherwise return 0. Now this routine
takes pretty much time and I was wondering whether doing a manual list
of

if (strncmp(...) == 0)
..
else if (strncmp(...) == 0)
..

No, because on average you have to make 250 compares per word, while
std::map does 8 or 9. If some of your search words show up more often
than others then you might sort the compares, but that's not really
helping you.

>
would be faster / more efficient? I was also thinking about doing
something like

switch (*str)
{
case 'a':
if (strncmp(...))
}

Yes that's better. I don't want to bother you with further hints on how
to improve this further, or if you should use hash_maps instead. If
(what I guess from your suggestion to strcmp everything) you have a
*fixed* number of words that you would like to recognize, then a very
good bet is a code generator which turns your wordlist into a piece of
code to recognize these words as fast a possible.
I'm quite happy with re2c (http://re2c.org/), which, with its default
settings, basically does the switch-thing, alas to the bitter end. And
it supports wide-characters.

--
IYesNo yes=YesNoFactor y.getFactoryIns tance().YES;
yes.getDescript ion().equals(ar ray[0].toUpperCase()) ;