RSS News Feed

Re fast gaussian filter

Message-ID:<6l9g6aFbe59hU1@mid.individual.net>
Subject:

Re: fast gaussian filter

Date:Fri, 10 Oct 2008 18:03:07 +0100


finecur wrote:
> Do you know any fast implementation of Gaussian smoothing? I mean
> beyond making it seperatable...
> 
> I found a recursive implementation from the paper "Recursive
> implementation of the Gaussian filter", If I use integer but not float
> point, however, the result is not accurate.
> 
> Any other method? Photoshop's Gaussian blur is very fast, I am
> wondering how they did it.

You need to exploit integer (fixed point) SIMD.

Paul

Message-ID:
Subject:

Re: fast gaussian filter

Date:Fri, 10 Oct 2008 18:44:03 +0100


On Oct 10, 10:03=A0am, Paul Russell  wrote:
> finecur wrote:
> > Do you know any fast implementation of Gaussian smoothing? I mean
> > beyond making it seperatable...
>
> > I found a recursive implementation from the paper "Recursive
> > implementation of the Gaussian filter", If I use integer but not float
> > point, however, the result is not accurate.
>
> > Any other method? Photoshop's Gaussian blur is very fast, I am
> > wondering how they did it.
>
> You need to exploit integer (fixed point) SIMD.
>
> Paul

SIMD is not available for the platform I am working on.

Message-ID:<48f23ab8$0$18148$4fafbaef@reader3.news.tin.it>
Subject:

Re: fast gaussian filter

Date:Sun, 12 Oct 2008 18:59:13 +0100



"aruzinsky"  ha scritto nel messaggio 
news:bd16d583-f324-42c4-9a7c-46e14ef018de@v15g2000hsa.googlegroups.com...
On Oct 10, 11:44 am, finecur  wrote:

> How do you know Photoshop doesn't use SIMD?

maybe with "SIMD is not available for the platform I am working on" he means 
he uses an old PC not supporting SIMD instructions, therefore they can't be 
used by Photoshop.

Message-ID:<7985894e-b733-47be-a852-d2b31b8a7396@u65g2000hsc.googlegroups.com>
Subject:

Re: fast gaussian filter

Date:Sun, 12 Oct 2008 19:46:57 +0100


On Oct 10, 12:33=A0pm, finecur  wrote:
> Do you know any fast implementation of Gaussian smoothing? I mean
> beyond making it seperatable...
>
> I found a recursive implementation from the paper "Recursive
> implementation of the Gaussian filter", If I use integer but not float
> point, however, the result is not accurate.
>
> Any other method? Photoshop's Gaussian blur is very fast, I am
> wondering how they did it.

Judging from this, http://www.stereopsis.com/kpt/kptblur.html
KPT blurs were even more efficient than Photoshop's.

Message-ID:
Subject:

Re: fast gaussian filter

Date:Mon, 13 Oct 2008 17:29:43 +0100


aruzinsky wrote:
> On Oct 10, 10:33 am, finecur  wrote:
>> Do you know any fast implementation of Gaussian smoothing? I mean
>> beyond making it seperatable...
>>
>> I found a recursive implementation from the paper "Recursive
>> implementation of the Gaussian filter", If I use integer but not float
>> point, however, the result is not accurate.
>>
>> Any other method? Photoshop's Gaussian blur is very fast, I am
>> wondering how they did it.
> 
> How fast is Photoshop's Gaussian with a Std.Dev. (which I assume is
> radius) of 100?
> 
> Using a separable kernel with a support of 6 Std.Dev. and X87 floating
> point, I get 8.7 seconds/megapixel on my computer.  That seems
> adequate for most intents and purposes.  I

On gimp I just applied a 100 pixel gaussian to an RGB image
of 3264x2448 pixels on a fujitsu-seimens laptop.

It took 10.38 seconds.

I've appended /proc/cpuinfo to the post, but the highlights are:
Intel(R) Celeron(R) M processor         1500MHz

    BugBear

Message-ID:
Subject:

Re: fast gaussian filter

Date:Tue, 14 Oct 2008 09:47:37 +0100


aruzinsky wrote:
> On Oct 13, 10:29 am, bugbear 
> wrote:
>> aruzinsky wrote:
>>> On Oct 10, 10:33 am, finecur  wrote:
>>>> Do you know any fast implementation of Gaussian smoothing? I mean
>>>> beyond making it seperatable...
>>>> I found a recursive implementation from the paper "Recursive
>>>> implementation of the Gaussian filter", If I use integer but not float
>>>> point, however, the result is not accurate.
>>>> Any other method? Photoshop's Gaussian blur is very fast, I am
>>>> wondering how they did it.
>>> How fast is Photoshop's Gaussian with a Std.Dev. (which I assume is
>>> radius) of 100?
>>> Using a separable kernel with a support of 6 Std.Dev. and X87 floating
>>> point, I get 8.7 seconds/megapixel on my computer.  That seems
>>> adequate for most intents and purposes.  I
>> On gimp I just applied a 100 pixel gaussian to an RGB image
>> of 3264x2448 pixels on a fujitsu-seimens laptop.
>>
>> It took 10.38 seconds.
>>
>> I've appended /proc/cpuinfo to the post, but the highlights are:
>> Intel(R) Celeron(R) M processor         1500MHz
>>
>>     BugBear- Hide quoted text -
>>
>> - Show quoted text -
> 
> I didn't say "100 pixel."  The size of my kernel was 601 pixels.

You said:
"How fast is Photoshop's Gaussian with a Std.Dev. (which I assume is
radius) of 100?"

So I did radius 100 in Gimp (for comparison with the request
you made of Photoshop)

    BugBear

Message-ID:<3164d8cd-291b-4ba8-95a1-f56d0cc7c9c6@s9g2000prg.googlegroups.com>
Subject:

Re: fast gaussian filter

Date:Tue, 14 Oct 2008 18:45:33 +0100


On Oct 14, 2:47=A0am, bugbear  wrote:
> aruzinsky wrote:
> > On Oct 13, 10:29 am, bugbear 
> > wrote:
> >> aruzinsky wrote:
> >>> On Oct 10, 10:33 am, finecur  wrote:
> >>>> Do you know any fast implementation of Gaussian smoothing? I mean
> >>>> beyond making it seperatable...
> >>>> I found a recursive implementation from the paper "Recursive
> >>>> implementation of the Gaussian filter", If I use integer but not flo=
at
> >>>> point, however, the result is not accurate.
> >>>> Any other method? Photoshop's Gaussian blur is very fast, I am
> >>>> wondering how they did it.
> >>> How fast is Photoshop's Gaussian with a Std.Dev. (which I assume is
> >>> radius) of 100?
> >>> Using a separable kernel with a support of 6 Std.Dev. and X87 floatin=
g
> >>> point, I get 8.7 seconds/megapixel on my computer. =A0That seems
> >>> adequate for most intents and purposes. =A0I
> >> On gimp I just applied a 100 pixel gaussian to an RGB image
> >> of 3264x2448 pixels on a fujitsu-seimens laptop.
>
> >> It took 10.38 seconds.
>
> >> I've appended /proc/cpuinfo to the post, but the highlights are:
> >> Intel(R) Celeron(R) M processor =A0 =A0 =A0 =A0 1500MHz
>
> >> =A0 =A0 BugBear- Hide quoted text -
>
> >> - Show quoted text -
>
> > I didn't say "100 pixel." =A0The size of my kernel was 601 pixels.
>
> You said:
> "How fast is Photoshop's Gaussian with a Std.Dev. (which I assume is
> radius) of 100?"
>
> So I did radius 100 in Gimp (for comparison with the request
> you made of Photoshop)
>
> =A0 =A0 BugBear- Hide quoted text -
>
> - Show quoted text -

Sorry.  You should compare impulse responses to see whether there is a
difference, otherwise we may be comparing apples and oranges.