July 2020 - Numeric - Haskell.org

Fwd: HMatrix priorities
by dominic＠steinitz.org 10 Jul '20

10 Jul '20

Hi Huw, Thanks very much for the questions and working on hmatrix. Apologies for not having more informative answers. Dominic Steinitz dominic(a)steinitz.org http://idontgetoutmuch.org Twitter: @idontgetoutmuch > On 8 Jul 2020, at 08:18, Huw Campbell <huw.campbell(a)gmail.com> wrote: > > Hi all, > > I've put a bit of time into hmatrix over the past week. My main > priorities at the moment are fixing bugs, segfaults, memory leaks, and > other infelicities. But after that's done I'm planning on tackling > some broader API and design issues. > > I have a few questions: > > - Why are CSR and CSC 1 indexed? The early literature uses 0 indexing > AFAICS, and it means that Ben Gamari's libsvd wrapper has to do extra > leg work reindexing before calling its C code. I didn’t even know it did this. I don’t think the author (Alberto Ruiz) is subscribed to this list. We could try asking him directly. > - Why do we have our own mutable vector type? I would have thought the > version in > the vector package would be sufficient. My guess would be that when the code was written mutable vectors were not available in vector. > - Should we try to move more of the library towards the general matrix > `GMatrix`? At the moment there are not very many functions which > actually use it (and it has partial record labels), and one can't > project out back to any of the underlying matrix forms. I have no view on this other than to be careful about performance which reminds me: we should really do some performance measurements. > - Can we bring the test suites and benchmarks back into the libraries > proper? The test suite is currently lagging the major version, and on > master doesn't compile with GHC 8.6+ (I have a PR for that). Yes please. > - CI looks like it's only testing one version of GHC, do we have a > policy on how many versions we support? No policy but I guess we should aim for a few more versions. > - What are all of your priorities? Dominic has mentioned a few use cases around > sparse matrices people have been working on. Some time ago, I created hmatrix-sundials but gave it its own repo. The idea was to have an identical interface to ODE solving so that you could switch between SUNDIALS and GSL. Roman Cheplyaka has vastly improved hmatrix-sundials and it now uses suitesparse for sparse matrices. In terms of priorities, my own are focussed on solving PDEs so I really need sparse matrices. And SUNDIALS is really a lot better for differential equations than GSL. I really like the statically typed matrices (the rows and columns being typed by their size). Are you thinking just about the base package or the other packages like glpk? > We've also spoken about github etiquette. I agreed with Dominic that > branches and pull requests are probably the best approach at the > moment. I won't merge my own PRs without input from another maintainer > at this stage, so please do have a look. I am not totally convinced having a monorepo is the best way forward but I don’t feel strongly. > Regards, > Huw > _______________________________________________ > Numeric mailing list > Numeric(a)haskell.org > http://mail.haskell.org/cgi-bin/mailman/listinfo/numeric

1 0

HMatrix priorities
by Huw Campbell 08 Jul '20

08 Jul '20

Hi all, I've put a bit of time into hmatrix over the past week. My main priorities at the moment are fixing bugs, segfaults, memory leaks, and other infelicities. But after that's done I'm planning on tackling some broader API and design issues. I have a few questions: - Why are CSR and CSC 1 indexed? The early literature uses 0 indexing AFAICS, and it means that Ben Gamari's libsvd wrapper has to do extra leg work reindexing before calling its C code. - Why do we have our own mutable vector type? I would have thought the version in the vector package would be sufficient. - Should we try to move more of the library towards the general matrix `GMatrix`? At the moment there are not very many functions which actually use it (and it has partial record labels), and one can't project out back to any of the underlying matrix forms. - Can we bring the test suites and benchmarks back into the libraries proper? The test suite is currently lagging the major version, and on master doesn't compile with GHC 8.6+ (I have a PR for that). - CI looks like it's only testing one version of GHC, do we have a policy on how many versions we support? - What are all of your priorities? Dominic has mentioned a few use cases around sparse matrices people have been working on. We've also spoken about github etiquette. I agreed with Dominic that branches and pull requests are probably the best approach at the moment. I won't merge my own PRs without input from another maintainer at this stage, so please do have a look. Regards, Huw

2 1

Re: [Numeric] HMatrix collaboration
by dominic＠steinitz.org 07 Jul '20

07 Jul '20

Hi Huw, You are right. I have given you access to the hmatrix repo. We don’t currently have any modus operandi for the organisation in which this repo lives. I assume that we will all create branches for review before merging? I’ve cc’d the somewhat unused mailing list. If that doesn’t work for e.g. Q&A then we can try something else. Dominic Steinitz dominic(a)steinitz.org http://idontgetoutmuch.org Twitter: @idontgetoutmuch > On 6 Jul 2020, at 08:09, Huw Campbell <huw.campbell(a)gmail.com> wrote: > > Hi Dominic, > > Thanks very much. I still can't seem to push to the hmatrix repository, I think the permissions are separate to the haskell-numerics organisation. > > Is there a mailing list or unofficial list of email addresses we can use for questions and change proposals? Some questions (like, why is CSR 1 indexed?) feel a bit strange for an issue tracker. > > Regards, > Huw > > On Sat, Jul 4, 2020 at 5:56 PM <dominic(a)steinitz.org <mailto:dominic@steinitz.org>> wrote: > Hi Huw > > Great 😁 > > I am just about to go out but some thoughts > > Roman Cheplyaka has updated hmatrix-sundials to use sparse matrices - I have an application for this solving some PDEs which I have yet to try. > > I think Henning Thielemann has also done some work on sparse matrices / bindings to BLAS / LAPACK. > > The collaborative effort to improve random number generation worked really well - hopefully we can repeat the experience in this important area. > > Dominic Steinitz > dominic(a)steinitz.org <mailto:dominic@steinitz.org> > http://idontgetoutmuch.org <http://idontgetoutmuch.org/> > Twitter: @idontgetoutmuch > >> On 4 Jul 2020, at 05:05, Huw Campbell <huw.campbell(a)gmail.com <mailto:huw.campbell@gmail.com>> wrote: >> >> Hi Dominic, >> >> I've been thinking about putting some more work into hmatrix for a few >> years now, ever since I used it effectively in my neural network >> library Grenade. >> >> I currently have a few PRs up to fix and speed up sparse matrix creation. >> >> I would like to really tighten up its benchmark and test suites (bring >> them into the packages themselves instead of hosting them in an >> external package; remove the poorly conceived `Testable` type class; >> using criterion a lot more; and possibly porting tests to hedgehog). >> >> Other items on the cards are using im2col to speed up convolutions (I >> have this working in Grenade already) and make them easier to use with >> user defined padding and stride. >> >> It's quite a bit to do, so I was wondering if I could become a collaborator? >> >> In terms of Haskell projects I maintain, the main open source one >> these days is optparse-applicative, which I've been incrementally >> improving over the past 5 years. >> >> Kind regards, >> Huw >

2 1