[squeak-dev] Input on class design for someone new to squeak

Hi!
Lately I have been playing with approaches to implement multi dimensional
arrays  in squeak with functionality similar to python's numpy library. The
structure I've come up with has a main array class with instance variables
pointing to a ByteArray and an array type object. Whenever a message is sent
to an array object, it is dispatched to the array type object which in turn
will then choose a concrete implementation from a plugin. The code in the
plugin is automatically generated from definitions contained in the various
array type classes. For an accessor like at:  for example I have a concrete
implementation as Slang plugin for each data type that casts the ByteArray
data to a pointer of the required type. So I have a structure looking like
this:

class ivars

SqArray buffer, arrayType

SqArrayType
- SqArrayInt8
- SqArrayInt16
- SqArrayFloat32
.....

SqArrayPlugin -> concrete implementations

Is this a sensible design for doing something like this in Smalltalk? Am I
abusing the Squeak plugin mechanism very badly? And finally, does something
better already exist?

Any input and constructive criticism is very much appreciated,
Danny

On Fri, Apr 18, 2008 at 10:39:27PM +0200, Danny Chan wrote:
> Hi!  Lately I have been playing with approaches to implement
> multi dimensional arrays  in squeak with functionality similar
> to python's numpy library.
>
> <snip>
>
> And finally, does something better already exist?

SmalLapack already kinda exists:
http://ncellier.ifrance.com/Smallapack/

--
Matthew Fulmer -- http://mtfulmer.wordpress.com/

Matthew Fulmer wrote:
Yes, that's a first attempt toward a scientific library.
Unfortunately, if all tests pass on Dolphin and VW (not 100% covergae
though), that's not the case of the Squeak version. Still bleding edge
with image blocking (win32) or crashing (linux).

I also have an unpublished link to FFTW in VW.

Nicolas

Am Samstag, 19. April 2008 12:20:39 schrieb nice:
Sounds interesting. There are two things I need in a matrix package, I would
be interested if this can be done with Smallapack:

I need the ability to generate views into a data set like it is possible in
python or Matlab. For example, in python I can do something like this to raw
image data in a 2 dimensional matrix:

# selects a sub matrix containing every second pixel in each dimension
redChannel = image[::2, ::2]              
# selects a square of 50 pixels in the top left of the image
topLeft = redChannel[0:50, 0:50]

Each of the sub matrices is a view into the data and does not copy values.

Another thing I need quite often is the ability to change the interpretaion of
the matrix data buffer on the fly. In python:

# charData is an array with n elements of type char
charData.dtype = dtype('int16')
# now charData is an array with n/2 elements of type short, interpreting the
sequence of two bytes as one short


If this is not possible, what would be needed to make Smallapack support this?

Danny

Danny Chan wrote:
It is possible to access an arbitrary sequence or rows.
Also possible to access an arbitrary sequence of columns.
So, you can do it in two stages by now. (Easy to enhance).
Also, message names are different depending on the type of argument.
        aMatrix atRow: anInteger.
        aMatrix atRows: aCollection.
I once tried generalized access messages to unify this, but used triple
dispatching at that time to handle generalizedAt:and:put:, something
quite horrible. I would tend to reify rows and columns accessors if i
had to implement it today.

Current implementation is to answer a copy like matlab.
However, it is easy to implement some kind of viewport object not
creating a copy if that is what you want.

There is no specific language in Smallapack, so you won't get these
short array notations, unless you implement your own language and
interpreter.

Smallapack supports only Floating point datatypes of Lapack (float,
double, float complex, double complex).
It was intended to perform linear algebra, not image processing.

As for the cast hack, I used such horrible thing in Smallapack to
extract real part of a complex matrix, pretending they were just made of
twice many reals...

It is possible to extend Smallapack to handle int, it is just a matter
of implementing a subclass.
But there is no fast primitive for int implemented in BLAS/LAPACK.
However, quite sure BitBlt is already addressing the problem.

Cheers

Nicolas

 > Lately I have been playing with approaches to implement multi
 > dimensional arrays...

      What's the intended use?


-C

Am Dienstag, 22. April 2008 20:13:14 schrieb Craig Latta:
>  > Lately I have been playing with approaches to implement multi
>  > dimensional arrays...
>
>       What's the intended use?
>
>
> -C

At the moment, mostly for image processing, both of offline images as well as
from live image streams captured from hardware. Therefore I need flexible
indexing schemes, different data types and some support for basic statistics,
later on support for some basic least squares fitting. Probably more things
in time. And I must be able to do most things inplace without a lot of
copying.

Basically I would like to do the things I can do in my favourite environment
for scientific computing (python + numpy) in my favourite development
environment, and it seems right now, that Smalltalk is starting to become
this environment.

Danny

I'm not certain but I suspect Tim is refering to Diego Gomez Deck

http://wiki.squeak.org/squeak/1603
http://wiki.squeak.org/squeak/3765
http://wiki.squeak.org/squeak/2411

Ken

On Thu, 2008-04-24 at 13:55 -0700, tim Rowledge wrote:

On 24-Apr-08, at 2:08 PM, Ken Causey wrote:

> I'm not certain but I suspect Tim is refering to Diego Gomez Deck
>
> http://wiki.squeak.org/squeak/1603
> http://wiki.squeak.org/squeak/3765
> http://wiki.squeak.org/squeak/2411
That's the stuff - thanks Ken!

I was impressed with it, though I'm no image processing expert and my  
opinion wouldn't really count for much.

tim
--
tim Rowledge; [hidden email]; http://www.rowledge.org/tim
Logic:   The art of being wrong with confidence...

May be you are talking about my old PhotoSqueak project, at
http://www.jvuletich.org/Squeak/PhotoSqueak/PhotoSqueak1/ImageProcessingEng.html 
,
http://www.jvuletich.org/Squeak/PhotoSqueak/PhotoSqueak2/DownloadEng.html 
and
http://www.jvuletich.org/Squeak/PhotoSqueak/PhotoSqueakGallery/PhotoSqueakGallery.html

Cheers,
Juan Vuletich

tim Rowledge wrote:

On 24-Apr-08, at 6:57 PM, Juan Vuletich wrote:

> May be you are talking about my old PhotoSqueak project,
Hmm, yes maybe that was what I was thinking about. Interesting that we  
have two interesting image processor packages.

tim
--
tim Rowledge; [hidden email]; http://www.rowledge.org/tim
Useful random insult:- Suffers from Clue Deficit Disorder.

Hi!
Both seem to be very interesting, I will definitely have a look. But maybe I
should explain a little bit more what I need to do. I am not interested so
much in the image chain that is applied to a raw image from a camera source.
I am working with camera modules that have raw output, there is almost no
image processing applied within the module. I have to analyze the images for
defects and problems of the sensor, the optics and the interaction of both.
Typical tasks are for example:
- Find the optical center from a homogeneously illuminated image
- Find specific target structures in the image and try to extract a sharpness
measure
- Develop an algorithm that allows to quantify dust and flare
- Analyze typical optical problems like distortion, chromatic aberration,
vignetting
- Analyze color shift problems caused by mismatch between sensor and optics /
sensor settings
- Retrieve different metrics for several regions of interest on the image like
color offsets, statistical measures, contrast...

All the processing performed on the image later on is quite interesting to me,
because I need to take the influence of the image pipe on the final result
into account. But my focus is really on the analysis of large matrix data
sets. Typically I need 2D matrices for analysis of luminance values only, 3D
matrices for analysis of RGB data. I also need 4D matrices for analysis of
color data from raw images (in most image sensors, the color information is
encoded by having a different color filter on each pixel, and later on
retrieving a full RGB value for a pixel by interpolating the color
information from neighbouring pixels).

So it is really rather data analysis of very differerent data sets , it only
happens to be mostly imaging related at the moment.



Am Freitag, 25. April 2008 04:04:41 schrieb tim Rowledge: