Re: [Lcms-user] lcms2-2.8 release candidate 1 available for testing

[Marti M. <mar...@li...>]

 

Hi,

 

Looks great!

 

I have added your changes to the git repository. Many thanks for your
contribution.

 

Best regards,

Marti Maria

The LittleCMS project

http://www.littlecms.com 

 

From: Noel Carboni [mailto:NCa...@Pr...] 
Sent: martes, 7 de junio de 2016 21:34
To: 'Marti Maria' <mar...@li...>;
lcm...@li...
Subject: RE: [Lcms-user] lcms2-2.8 release candidate 1 available for testing

 

Hi Marti,

 

More info related to the cmsFLAGS_COPY_ALPHA flag...

 

Please allow me to make some suggestions to improve the speed of in-place
color conversions (i.e., conversions of colors in a single image, where no
copy is occurring from one buffer to another):

 

 

1.  Add these statements to the _cmsHandleExtraChannels() function:

 

               // Exit early if in-place color-management is occurring - no
need to copy extra channels to themselves.

               if (p->InputFormat == p->OutputFormat && in == out)

                              return;

 

I've tested the above and it does make a difference in speed for the
common/typical operation of in-place color-management.

 

 

2.  I've also tested hand-optimizing copying the extra info for one extra
channel each for input/output - another very typical scenario - and by
reducing the array variables to simple variables the entire innermost loop
can be run only from registers.  It does make a difference in the speed when
one extra channel is copied (e.g., for RGBA).  That actually brought the
LittleCMS time to do the task on 8 bit to 8 bit transforms in line with the
optimized code we had already written.

 

Here is a copy of the optimized c file.  It's all yours to include in the
package if you'd like to use it.  All of the changes are in the
_cmsHandleExtraChannels() function.

 

http://Noel.ProDigitalSoftware.com/temp/cmsalpha.c

 

Using Visual Studio 2015, the assembly instructions for the innermost
per-pixel loop for the optimized "one extra field" function are reduced by
the optimizer to this quite efficient set:

 

$LL7@cmsHandleE:

 

; 466  :                                                copyValueFn(DestPtr,
SourcePtr);

 

            mov    rdx, rbx

            mov    rcx, rdi

            call      r15

 

; 468  :                                                SourcePtr +=
SourceIncrements[0];

 

            add      rbx, r14

 

; 469  :                                                DestPtr   +=
DestIncrements[0];

 

            add      rdi, r12

            sub      rsi, 1

            jne       SHORT $LL7@cmsHandleE

 

-Noel

 

From: Marti Maria [mailto:mar...@li...] 
Sent: Mon, June 6, 2016 3:36 PM
To: lcm...@li... <mailto:lcm...@li...>

Subject: [Lcms-user] lcms2-2.8 release candidate 1 available for testing

 

 

Hi,

 

After some time of inactivity I've taken again the development of lcms, so
here is the first release candidate for lcms 2.8

 

Most of code is sponsored by Alien Skin Software, thank you so much for this
great contribution!

 

Aside of bugfixes, 2.8 does support alpha channel transportation and a new
transform stride function and plug-in which can be used to increase
significantly the processing speed.

 

See here the release candidate http://www.littlecms.com/lcms2-2.8rc1.tar.gz

 

Best regards

 

Marti Maria

The LittleCMS project

http://www.littlecms.com