While some channel types are classes, a built-in type such as ‘unsigned char’ is a valid channel. We have to be careful so we don’t slow down operations on common 8-bit images. I think it is safe to replace “Channel” with “const Channel” but “const Channel&” has a size of a pointer and this is far bigger than unsigned char.

Perhaps to do this right one needs to create a metafunction that returns the type by value or by const reference, whichever is faster.

Lubomir

From: Eloi Du Bois [mailto:eloi.du.bois@gmail.com]
Sent: Friday, November 27, 2009 3:37 PM
To: boost-users@lists.boost.org
Cc: Lubomir Bourdev
Subject: Re: [Boost-users] [GIL] Performance warning...

I am not an expert but I think on primitive types, there is no need to put a reference (and it may slows down the process).
But the line I mentioned before concern a channel object, and I think this object stores many datas (maybe a byte array ?).

2009/11/28 Christian Henning <chhenning@gmail.com>

Lubomir, looking at the default_channel_converter the channels already
arrive as a const reference. All I did was to make sure they are still
const references in the channel_converter functions. In all cases the
source channels should at least be const. Which allows for more
compiler optimizations.

I did run the performance test in gil_test suite and the changes were
marginal. Some tests were faster some were slower after the change.

Christian

On Fri, Nov 27, 2009 at 6:13 PM, Lubomir Bourdev <lbourdev@adobe.com> wrote:
> This was done on purpose. I don't know if it still holds, but it used to be
> that a const reference to a simple type like an int or char is slower than
> passing the type by value. This is the same reason most STL algorithms take
> iterators by value. If you want to make this change it is important to
> verify that performance of simple built-in channel types is not adversely
> affected.
>
> Lubomir

>
> On 11/27/09 12:17 PM, "Christian Henning" <chhenning@gmail.com> wrote:
>
>> Hi Eloi, you're correct and I'll fix that. I can even make it const
>> reference. Running the test shows no real difference but that might be
>> only because no expensive channels are being used.
>>
>> I'll commit the changes asap.
>>
>> Lubomir, please review the changes if your time allows. Affected files
>> are channel_algorithm.hpp and gil_concept.hpp.
>>
>> Thanks,
>> Christian
>>
>
>

_______________________________________________
Boost-users mailing list
Boost-users@lists.boost.org
http://lists.boost.org/mailman/listinfo.cgi/boost-users