Probably. But I suspect we need a smart data structure that has part of its state stored on disk. Some pictures weight over 1Gb. You cannot have all this in memory. Making sure that only part of the matrix that interest you is in memory while the remaining is left on disk is crucial.
Alexandre On May 8, 2014, at 5:01 PM, Ronie Salgado <[email protected]> wrote: > Alex, > > We need a DSL with the following features: > - Lazy evaluation. > - Matrix/vector operations > - Multiples backends, for fallback if OpenCL is not available in some > platform. > > The trickiest part is the one of Lazy evaluation. This is important, > specially because the number of computations required for (AB)C or A(BC) can > be orders of magnitude different. > > Greetings, > Ronie > > > 2014-05-08 16:53 GMT-04:00 kilon alios <[email protected]>: > is this project going to be open source, or you will keep it private ? > > > On Thu, May 8, 2014 at 11:22 PM, Alexandre Bergel <[email protected]> > wrote: > We have matrices of size 2000 x 2000. Soon we would like to work with > matrices 2000 x 2000 x 10 000. > Your work on OpenCL is __highly__ important to us > > Alexandre > > > On May 8, 2014, at 4:02 PM, Ronie Salgado <[email protected]> wrote: > > > Alex, > > > > What is the size of those matrices? > > > > > > 2014-05-08 15:56 GMT-04:00 Alexandre Bergel <[email protected]>: > > My first thought would be to have some particular C routines to perform a > > transformation on a matrix. > > This would not be that difficult I guess. Well.. We have to see. We will > > likely enter the optimization phase in a couple of months from now. > > > > Alexandre > > > > > > On May 8, 2014, at 2:15 PM, Chris Cunningham <[email protected]> > > wrote: > > > > > So, Hannes may have been referring to optimizing the slow parts of the > > > process in Native Boost directly instead of using an external library (if > > > not, may I suggest it?). It is possible that just a few tweaks using > > > Assembler in the process would get the speed up to an acceptable level, > > > which would be determined by profiling the performance. Of course, that > > > path requires someone on your project that can understand Assembler to > > > get familiar with NativeBoost. > > > -cbc > > > > > > On Thu, May 8, 2014 at 8:05 AM, Alexandre Bergel > > > <[email protected]> wrote: > > > We will surely review all the different options for interacting with > > > external libraries. Ronie is our local expert :-) > > > > > > Currently, we have no preview of the Roassal book to share. Give us a > > > couple of months :-) > > > > > > Cheers, > > > Alexandre > > > > > > > > > On May 8, 2014, at 9:23 AM, H. Hirzel <[email protected]> wrote: > > > > > > > Interesting project, Alexandre! > > > > > > > > On 5/6/14, Alexandre Bergel <[email protected]> wrote: > > > >> Thanks Kilon for your nice words. > > > >> Currently, AstroCloud is a bit slow. For example, loading this image > > > >> takes > > > >> much more than it should. For now, speed is not our priority. > > > >> AstroCloud has > > > >> to be fast and we will make it fast. Probably relying on external > > > >> library is > > > >> the way to go. > > > > > > > > I assume you consider or have considered NativeBoost as well? > > > > > > > >> > > > >> Just to give a bit more context about why we are doing AstroCloud. > > > >> ALMA is > > > >> the largest observatory on Earth, making it the largest astronomical > > > >> project > > > >> in existence (http://www.almaobservatory.org). This is what humanity > > > >> has > > > >> best to detect life outside Earth. > > > >> ALMA is made of 66 high-precision antenas making it possible to see > > > >> deep in > > > >> the universe. Unfortunately, there is no good software to process these > > > >> immensely complex pictures. Some tool suite naturally exist, but they > > > >> were > > > >> not made for ALMA, making these existing tools largely suboptimal to > > > >> process > > > >> complex picture of ALMA. > > > >> Faviola’s dream (and mine) is to produce a tool ready to be consumed by > > > >> astronomers to process ALMA pictures. Faviola and I are currently > > > >> seeking > > > >> for additional resources on making this happen. > > > > > > > > > > > >> Regarding your help about documenting Roassal. It is very welcome! In > > > >> one of > > > >> my previous post, I said that 2014 will be the year of documenting > > > >> Roassal. > > > >> A book is on its way. > > > > > > > > > > > >> We will need people to review it. So thanks! > > > > > > > > Preview of part of it on line? > > > > > > > > > > > > Regards > > > > Hannes > > > > > > > > _______________________________________________ > > > > Moose-dev mailing list > > > > [email protected] > > > > https://www.iam.unibe.ch/mailman/listinfo/moose-dev > > > > > > -- > > > _,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;: > > > Alexandre Bergel http://www.bergel.eu > > > ^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;. > > > > > > > > > > > > > > > > > > > -- > > _,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;: > > Alexandre Bergel http://www.bergel.eu > > ^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;. > > > > > > > > > > > > -- > _,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;: > Alexandre Bergel http://www.bergel.eu > ^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;. > > > > > > -- _,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;: Alexandre Bergel http://www.bergel.eu ^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;._,.;:~^~:;.
