Re: TempAlloc (Was: Library Development: What to finish/flesh out?)
On 3/28/2011 7:38 PM, Masahiro Nakagawa wrote: > On Sun, 27 Mar 2011 12:09:03 +0900, dsimcha wrote: > >> On 3/18/2011 2:56 AM, Don wrote: 3. TempAlloc: A memory allocator based on a thread-local segmented stack, useful for allocating large temporary buffers in things like numerics code. Also comes with a hash table, hash set and AVL tree optimized for this allocation scheme. The advantages over plain old stack allocation are that it's independent of function calls (meaning you can return pointers to TempAlloc-allocated memory from a function, etc.) and it's segmented, meaning you can allocate huge buffers w/o risking stack overflow. Its main weakness is that this stack is not scanned by the GC, meaning that you can't store the only reference to a GC-allocated piece of memory here. However, in practice large arrays of primitives are an extremely common case in performance-critical code. I find this module immensely useful in dstats and Lars Kyllingstad uses it in SciD. Getting it into Phobos would make it easy for other scientific/numerics code to use it. Completion state: Working and used. Needs a litte cleanup and documentation. (Phobos candidate) >>> >>> This is #1. Far and away. Belongs in druntime. >>> I would use it instantly in BigInt. >> >> This is not up for "official" review yet because std.net.isemail and >> std.parallelism (and possibly more stuff depending on how we manage >> the review queue) are ahead of it. If you want a sneak preview: >> >> Code: >> >> https://github.com/dsimcha/TempAlloc >> >> Docs: >> >> http://cis.jhu.edu/~dsimcha/d/phobos/core_tempalloc.html >> >> I put in a few small, tangentially related memory allocation functions >> that I've wanted in druntime/Phobos for awhile in this proposal, too. >> alignedMalloc() and alignedFree() are used by TempAlloc but can be >> made private if people don't like them. newVoid() is not used by >> TempAlloc and might fit better in std.array, but I think it belongs >> somewhere in Phobos or druntime. > > Now, I am thinking of light-weight process for concurrency. > Your TempAlloc(segmented stack) seems to be a good parts for > implementing such feature :) > > > Masahiro I'm not sure what you're suggesting. I'd definitely need more details to help you in any way. If it involves fibers, remember that thread-local storage only works with OS threads. TempAlloc stores its state in thread-local storage. Other than that, I'm interested in hearing about this. If there are any enhancements I can add to make what you want to do easier, let me know.
Re: TempAlloc (Was: Library Development: What to finish/flesh out?)
On Sun, 27 Mar 2011 12:09:03 +0900, dsimcha wrote: On 3/18/2011 2:56 AM, Don wrote: 3. TempAlloc: A memory allocator based on a thread-local segmented stack, useful for allocating large temporary buffers in things like numerics code. Also comes with a hash table, hash set and AVL tree optimized for this allocation scheme. The advantages over plain old stack allocation are that it's independent of function calls (meaning you can return pointers to TempAlloc-allocated memory from a function, etc.) and it's segmented, meaning you can allocate huge buffers w/o risking stack overflow. Its main weakness is that this stack is not scanned by the GC, meaning that you can't store the only reference to a GC-allocated piece of memory here. However, in practice large arrays of primitives are an extremely common case in performance-critical code. I find this module immensely useful in dstats and Lars Kyllingstad uses it in SciD. Getting it into Phobos would make it easy for other scientific/numerics code to use it. Completion state: Working and used. Needs a litte cleanup and documentation. (Phobos candidate) This is #1. Far and away. Belongs in druntime. I would use it instantly in BigInt. This is not up for "official" review yet because std.net.isemail and std.parallelism (and possibly more stuff depending on how we manage the review queue) are ahead of it. If you want a sneak preview: Code: https://github.com/dsimcha/TempAlloc Docs: http://cis.jhu.edu/~dsimcha/d/phobos/core_tempalloc.html I put in a few small, tangentially related memory allocation functions that I've wanted in druntime/Phobos for awhile in this proposal, too. alignedMalloc() and alignedFree() are used by TempAlloc but can be made private if people don't like them. newVoid() is not used by TempAlloc and might fit better in std.array, but I think it belongs somewhere in Phobos or druntime. Now, I am thinking of light-weight process for concurrency. Your TempAlloc(segmented stack) seems to be a good parts for implementing such feature :) Masahiro
TempAlloc (Was: Library Development: What to finish/flesh out?)
On 3/18/2011 2:56 AM, Don wrote: 3. TempAlloc: A memory allocator based on a thread-local segmented stack, useful for allocating large temporary buffers in things like numerics code. Also comes with a hash table, hash set and AVL tree optimized for this allocation scheme. The advantages over plain old stack allocation are that it's independent of function calls (meaning you can return pointers to TempAlloc-allocated memory from a function, etc.) and it's segmented, meaning you can allocate huge buffers w/o risking stack overflow. Its main weakness is that this stack is not scanned by the GC, meaning that you can't store the only reference to a GC-allocated piece of memory here. However, in practice large arrays of primitives are an extremely common case in performance-critical code. I find this module immensely useful in dstats and Lars Kyllingstad uses it in SciD. Getting it into Phobos would make it easy for other scientific/numerics code to use it. Completion state: Working and used. Needs a litte cleanup and documentation. (Phobos candidate) This is #1. Far and away. Belongs in druntime. I would use it instantly in BigInt. This is not up for "official" review yet because std.net.isemail and std.parallelism (and possibly more stuff depending on how we manage the review queue) are ahead of it. If you want a sneak preview: Code: https://github.com/dsimcha/TempAlloc Docs: http://cis.jhu.edu/~dsimcha/d/phobos/core_tempalloc.html I put in a few small, tangentially related memory allocation functions that I've wanted in druntime/Phobos for awhile in this proposal, too. alignedMalloc() and alignedFree() are used by TempAlloc but can be made private if people don't like them. newVoid() is not used by TempAlloc and might fit better in std.array, but I think it belongs somewhere in Phobos or druntime.
Re: Library Development: What to finish/flesh out?
On 3/26/2011 4:25 AM, Johan Granberg wrote: Definitely some would love it, me among them. This to me feels like the most useful. Does your (dsimcha) implementation support spare matrices as well, and if not are the api general engough to support them if someone has time to implement them. When you say looks like normal matlab does that include b=A\x? Regardless any standardization of linear algebra libraries would improve on the current state in most languages. No support for sparse matrices, but the API should have no problem supporting them eventually. It's all templates. I'd like to support b=A\x, but D doesn't have a \ operator. At any rate, it's not much more verbose to do b = solve(A, x).
Re: Library Development: What to finish/flesh out?
On 03/26/2011 04:43 AM, Jonathan M Davis wrote: I find the responses to this list to be rather interesting. Most of it, I find to be of mild interest at best (certainly for the sort of stuff that _I_ do anyway), but others find some of them to be very desirable. > List in order from most to least finished: > > 1. Rational: A library for handling rational numbers exactly. Templated > on integer type, can use BigInts for guaranteed accuracy, or fixed-width > integers for more speed where the denominator and numerator will be small. >Completion state: Mostly finished. Just need to fix a litte bit rot and > submit for review. (Phobos candidate) Potentially interesting, but I don't know if I'd ever use it. That's it: never ever listen to people to know what they /really/ need, really find useful, really appreciate. Many are able to push like crazy for thingies that just look cool (esp at design time), but none of them will actually use for three years. (experienced ;-) Denis -- _ vita es estrany spir.wikidot.com
Re: Library Development: What to finish/flesh out?
Jonathan M Davis wrote: > On 2011-03-17 08:33, dsimcha wrote: >> I've accumulated a bunch of little libraries via various evening and >> weekend hacking projects over the past year or so, in various states of >> completion. Most are things I'm at least half-considering for Phobos, >> though some belong as third-party libs. I definitely don't have time to >> finish/flesh out all of them anytime soon, so I've decided to ask the >> community what to prioritize. Below is a summary of everything I've been >> working on, with its current level of completion. Please let me know the >> following: >> >> 1. A relative ordering of how useful you think these libraries would be >> to the community. >> >> 2. In absolute terms, would you find this useful? >> >> 3. For the Phobos candidates, whether they're general enough to belong >> in the **standard** library. > > I find the responses to this list to be rather interesting. Most of it, I > find to be of mild interest at best (certainly for the sort of stuff that > _I_ do anyway), but others find some of them to be very desirable. > >> List in order from most to least finished: >> >> 1. Rational: A library for handling rational numbers exactly. >> Templated on integer type, can use BigInts for guaranteed accuracy, or >> fixed-width integers for more speed where the denominator and numerator >> will be small. >> Completion state: Mostly finished. Just need to fix a litte bit rot >> and >> submit for review. (Phobos candidate) > > Potentially interesting, but I don't know if I'd ever use it. > >> 2. RandAA: A hash table implementation with deterministic memory >> management, based on randomized probing. Main advantage over builtin AAs >> is that it plays much nicer with the GC and multithreaded programs. >> Lookup times are also expected O(1) no matter how many collisions exist >> in modulus hash space, as long as there are few collisions in full 32- or >> 64-bit hash space. Completion state: Mostly finished. Just needs a >> little doc improvement, a few benchmarks and submission for review. >> (Phobos candidate) > > I'm afraid that I don't understand how this is better than the current > hash table, and if it really is better, perhaps it should replace > implementation of the built in one? > >> 3. TempAlloc: A memory allocator based on a thread-local segmented >> stack, useful for allocating large temporary buffers in things like >> numerics code. Also comes with a hash table, hash set and AVL tree >> optimized for this >> allocation scheme. The advantages over plain old stack allocation are >> that it's independent of function calls (meaning you can return pointers >> to TempAlloc-allocated memory from a function, etc.) and it's segmented, >> meaning you can allocate huge buffers w/o risking stack overflow. Its >> main weakness is that this stack is not scanned by the GC, meaning that >> you can't store the only reference to a GC-allocated piece of memory >> here. >> However, in practice large arrays of primitives are an extremely common >> case in >> performance-critical code. I find this module immensely useful in dstats >> and Lars Kyllingstad uses it in SciD. Getting it into Phobos would make >> it easy for other scientific/numerics code to use it. Completion state: >> Working and used. Needs a litte cleanup and documentation. (Phobos >> candidate) > > Personally, I see zero use for this in my stuff, but obviously others find > it very compelling. > >> 4. Streaming CSV Parser: Parses CSV files as they're read in, a few >> convenience functions for extracting columns into structs. If Phobos >> every gets SQLite support I'll probably add sugar for turning a CSV file >> into an >> SQLite database, too. Completion state: Prototype working, needs >> testing, >> cleanup and documentation. (Phobos candidate) > > This definitely sounds useful. I've had to deal with CSV parsing in Java > before, and I'd love to see a solid CSV parser in Phobos, but I don't know > how much I'd actually end up using it. In the few cases where I'd be > looking to deal with CSV files though, it could be invaluable. > >> 5. Matrix operations: SciD improvements that allow you to write matrix >> operations that look like normal math/MATLAB and optimizes them via >> expression templates so that a minimal number of temporary matrices are >> created. Uses/will use BLAS for multiplication. Completion state: >> Addition implemented. Multiplication not. > > I don't expect that I'd _ever_ need this, but I'd fully expect that some > folks would love it. Definitely some would love it, me among them. This to me feels like the most useful. Does your (dsimcha) implementation support spare matrices as well, and if not are the api general engough to support them if someone has time to implement them. When you say looks like normal matlab does that include b=A\x? Regardless any standardization of linear algebra libraries would improve on the current stat
Re: Library Development: What to finish/flesh out?
On 2011-03-17 08:33, dsimcha wrote: > I've accumulated a bunch of little libraries via various evening and > weekend hacking projects over the past year or so, in various states of > completion. Most are things I'm at least half-considering for Phobos, > though some belong as third-party libs. I definitely don't have time to > finish/flesh out all of them anytime soon, so I've decided to ask the > community what to prioritize. Below is a summary of everything I've been > working on, with its current level of completion. Please let me know the > following: > > 1. A relative ordering of how useful you think these libraries would be to > the community. > > 2. In absolute terms, would you find this useful? > > 3. For the Phobos candidates, whether they're general enough to belong in > the **standard** library. I find the responses to this list to be rather interesting. Most of it, I find to be of mild interest at best (certainly for the sort of stuff that _I_ do anyway), but others find some of them to be very desirable. > List in order from most to least finished: > > 1. Rational: A library for handling rational numbers exactly. Templated > on integer type, can use BigInts for guaranteed accuracy, or fixed-width > integers for more speed where the denominator and numerator will be small. > Completion state: Mostly finished. Just need to fix a litte bit rot and > submit for review. (Phobos candidate) Potentially interesting, but I don't know if I'd ever use it. > 2. RandAA: A hash table implementation with deterministic memory > management, based on randomized probing. Main advantage over builtin AAs > is that it plays much nicer with the GC and multithreaded programs. > Lookup times are also expected O(1) no matter how many collisions exist in > modulus hash space, as long as there are few collisions in full 32- or > 64-bit hash space. Completion state: Mostly finished. Just needs a > little doc improvement, a few benchmarks and submission for review. > (Phobos candidate) I'm afraid that I don't understand how this is better than the current hash table, and if it really is better, perhaps it should replace implementation of the built in one? > 3. TempAlloc: A memory allocator based on a thread-local segmented stack, > useful for allocating large temporary buffers in things like numerics code. > Also comes with a hash table, hash set and AVL tree optimized for this > allocation scheme. The advantages over plain old stack allocation are that > it's independent of function calls (meaning you can return pointers to > TempAlloc-allocated memory from a function, etc.) and it's segmented, > meaning you can allocate huge buffers w/o risking stack overflow. Its > main weakness is that this stack is not scanned by the GC, meaning that > you can't store the only reference to a GC-allocated piece of memory here. > However, in practice large arrays of primitives are an extremely common > case in > performance-critical code. I find this module immensely useful in dstats > and Lars Kyllingstad uses it in SciD. Getting it into Phobos would make > it easy for other scientific/numerics code to use it. Completion state: > Working and used. Needs a litte cleanup and documentation. (Phobos > candidate) Personally, I see zero use for this in my stuff, but obviously others find it very compelling. > 4. Streaming CSV Parser: Parses CSV files as they're read in, a few > convenience functions for extracting columns into structs. If Phobos every > gets SQLite support I'll probably add sugar for turning a CSV file into an > SQLite database, too. Completion state: Prototype working, needs testing, > cleanup and documentation. (Phobos candidate) This definitely sounds useful. I've had to deal with CSV parsing in Java before, and I'd love to see a solid CSV parser in Phobos, but I don't know how much I'd actually end up using it. In the few cases where I'd be looking to deal with CSV files though, it could be invaluable. > 5. Matrix operations: SciD improvements that allow you to write matrix > operations that look like normal math/MATLAB and optimizes them via > expression templates so that a minimal number of temporary matrices are > created. Uses/will use BLAS for multiplication. Completion state: > Addition implemented. Multiplication not. I don't expect that I'd _ever_ need this, but I'd fully expect that some folks would love it. > 6. Machine learning: Decision trees, KNN, Random Forest, Logistic > Regression, SVM, Naive Bayes, etc. This would be a dstats module. > Completion state: Decision trees prototyped, logistic regression working. I'd have to see the actual library to know whether I'd find much use in it. Probably not though. > 7. std.mixins: Mixins for commonly needed boilerplate code. I stopped > working on this when Andrei suggested that making a collection of mixins > into a module is a bad idea. I've thought about it some more and I > respectfu
Re: Library Development: What to finish/flesh out?
On 3/25/2011 5:59 PM, Denis Koroskin wrote: On Sat, 26 Mar 2011 00:26:40 +0300, dsimcha wrote: On 3/25/2011 3:50 PM, Sean Kelly wrote: On Mar 24, 2011, at 1:00 PM, dsimcha wrote: BTW, the TempAlloc module also includes a hash table, hash set and AVL tree that are specifically optimized for TempAlloc. Should these be included in the submission? The disadvantages I see here is that they are less generally useful (possibly too high level for druntime) and that they will make the review take a heck of a lot longer. Are they necessary for TempAlloc to function? If so, I'd add them but hidden, as I imagine there's more code than you'd want to simply drop in a private block in core.memory. It may be time for core to get a core.internal package for this kind of stuff. No, they are just data structures built on top of TempAlloc and optimized for it. I'd love to see them, in a separate module probably. This suggests two separate proposals. The more I think about it, the more I think this is the way to go. TempAlloc per se is much more self-evidently useful than the extra data structures and doesn't need the extra data structures to work. The extras shouldn't hold up its inclusion. The extra data structures only use (or only should use; I don't remember whether I bend this rule) TempAlloc's public API. Furthermore, I'm not sure they're generally useful enough to belong in Phobos. I'd like feedback from others when/if TempAlloc is in Phobos and more people are familiar with it.
Re: Library Development: What to finish/flesh out?
On Sat, 26 Mar 2011 00:26:40 +0300, dsimcha wrote: On 3/25/2011 3:50 PM, Sean Kelly wrote: On Mar 24, 2011, at 1:00 PM, dsimcha wrote: BTW, the TempAlloc module also includes a hash table, hash set and AVL tree that are specifically optimized for TempAlloc. Should these be included in the submission? The disadvantages I see here is that they are less generally useful (possibly too high level for druntime) and that they will make the review take a heck of a lot longer. Are they necessary for TempAlloc to function? If so, I'd add them but hidden, as I imagine there's more code than you'd want to simply drop in a private block in core.memory. It may be time for core to get a core.internal package for this kind of stuff. No, they are just data structures built on top of TempAlloc and optimized for it. I'd love to see them, in a separate module probably.
Re: Library Development: What to finish/flesh out?
On 3/25/2011 3:50 PM, Sean Kelly wrote: On Mar 24, 2011, at 1:00 PM, dsimcha wrote: BTW, the TempAlloc module also includes a hash table, hash set and AVL tree that are specifically optimized for TempAlloc. Should these be included in the submission? The disadvantages I see here is that they are less generally useful (possibly too high level for druntime) and that they will make the review take a heck of a lot longer. Are they necessary for TempAlloc to function? If so, I'd add them but hidden, as I imagine there's more code than you'd want to simply drop in a private block in core.memory. It may be time for core to get a core.internal package for this kind of stuff. No, they are just data structures built on top of TempAlloc and optimized for it.
Re: Library Development: What to finish/flesh out?
On Mar 24, 2011, at 1:00 PM, dsimcha wrote: > > BTW, the TempAlloc module also includes a hash table, hash set and AVL tree > that > are specifically optimized for TempAlloc. Should these be included in the > submission? The disadvantages I see here is that they are less generally > useful > (possibly too high level for druntime) and that they will make the review > take a > heck of a lot longer. Are they necessary for TempAlloc to function? If so, I'd add them but hidden, as I imagine there's more code than you'd want to simply drop in a private block in core.memory. It may be time for core to get a core.internal package for this kind of stuff.
Re: Library Development: What to finish/flesh out?
== Quote from Sean Kelly (s...@invisibleduck.org)'s article > On Mar 17, 2011, at 11:56 PM, Don wrote: > > dsimcha wrote: > >> I've accumulated a bunch of little libraries via various evening and > weekend > >> hacking projects over the past year or so, in various states of > completion. > >> Most are things I'm at least half-considering for Phobos, though some > belong > >> as third-party libs. I definitely don't have time to finish/flesh > out all of > >> them anytime soon, so I've decided to ask the community what to > prioritize. > >> Below is a summary of everything I've been working on, with its > current level > >> of completion. Please let me know the following: > > > >> 3. TempAlloc: A memory allocator based on a thread-local segmented > stack, > >> useful for allocating large temporary buffers in things like numerics > code. > >> Also comes with a hash table, hash set and AVL tree optimized for > this > >> allocation scheme. The advantages over plain old stack allocation > are that > >> it's independent of function calls (meaning you can return pointers > to > >> TempAlloc-allocated memory from a function, etc.) and it's segmented, > meaning > >> you can allocate huge buffers w/o risking stack overflow. Its main > weakness > >> is that this stack is not scanned by the GC, meaning that you can't > store the > >> only reference to a GC-allocated piece of memory here. However, in > practice > >> large arrays of primitives are an extremely common case in > >> performance-critical code. I find this module immensely useful in > dstats and > >> Lars Kyllingstad uses it in SciD. Getting it into Phobos would make > it easy > >> for other scientific/numerics code to use it. Completion state: > Working and > >> used. Needs a litte cleanup and documentation. (Phobos candidate) > > > > This is #1. Far and away. Belongs in druntime. > Stuff like this is why core.memory isn't called core.gc. BTW, the TempAlloc module also includes a hash table, hash set and AVL tree that are specifically optimized for TempAlloc. Should these be included in the submission? The disadvantages I see here is that they are less generally useful (possibly too high level for druntime) and that they will make the review take a heck of a lot longer.
Re: Library Development: What to finish/flesh out?
== Quote from Sean Kelly (s...@invisibleduck.org)'s article > On Mar 17, 2011, at 11:56 PM, Don wrote: > > dsimcha wrote: > >> I've accumulated a bunch of little libraries via various evening and > weekend > >> hacking projects over the past year or so, in various states of > completion. > >> Most are things I'm at least half-considering for Phobos, though some > belong > >> as third-party libs. I definitely don't have time to finish/flesh > out all of > >> them anytime soon, so I've decided to ask the community what to > prioritize. > >> Below is a summary of everything I've been working on, with its > current level > >> of completion. Please let me know the following: > > > >> 3. TempAlloc: A memory allocator based on a thread-local segmented > stack, > >> useful for allocating large temporary buffers in things like numerics > code. > >> Also comes with a hash table, hash set and AVL tree optimized for > this > >> allocation scheme. The advantages over plain old stack allocation > are that > >> it's independent of function calls (meaning you can return pointers > to > >> TempAlloc-allocated memory from a function, etc.) and it's segmented, > meaning > >> you can allocate huge buffers w/o risking stack overflow. Its main > weakness > >> is that this stack is not scanned by the GC, meaning that you can't > store the > >> only reference to a GC-allocated piece of memory here. However, in > practice > >> large arrays of primitives are an extremely common case in > >> performance-critical code. I find this module immensely useful in > dstats and > >> Lars Kyllingstad uses it in SciD. Getting it into Phobos would make > it easy > >> for other scientific/numerics code to use it. Completion state: > Working and > >> used. Needs a litte cleanup and documentation. (Phobos candidate) > > > > This is #1. Far and away. Belongs in druntime. > Stuff like this is why core.memory isn't called core.gc. Ok, this seems like a popular choice. I should have time after my )#*#$ thesis proposal and after std.parallelism is done to clean this up and submit for review. I'm sure the documentation will need to be improved, but as with std.parallelism I'm not sure **how** it will be judged deficient by people not as intimately familiar with the library as I am.
Re: Library Development: What to finish/flesh out?
On Mar 17, 2011, at 11:56 PM, Don wrote: > dsimcha wrote: >> I've accumulated a bunch of little libraries via various evening and weekend >> hacking projects over the past year or so, in various states of completion. >> Most are things I'm at least half-considering for Phobos, though some belong >> as third-party libs. I definitely don't have time to finish/flesh out all of >> them anytime soon, so I've decided to ask the community what to prioritize. >> Below is a summary of everything I've been working on, with its current level >> of completion. Please let me know the following: > >> 3. TempAlloc: A memory allocator based on a thread-local segmented stack, >> useful for allocating large temporary buffers in things like numerics code. >> Also comes with a hash table, hash set and AVL tree optimized for this >> allocation scheme. The advantages over plain old stack allocation are that >> it's independent of function calls (meaning you can return pointers to >> TempAlloc-allocated memory from a function, etc.) and it's segmented, meaning >> you can allocate huge buffers w/o risking stack overflow. Its main weakness >> is that this stack is not scanned by the GC, meaning that you can't store the >> only reference to a GC-allocated piece of memory here. However, in practice >> large arrays of primitives are an extremely common case in >> performance-critical code. I find this module immensely useful in dstats and >> Lars Kyllingstad uses it in SciD. Getting it into Phobos would make it easy >> for other scientific/numerics code to use it. Completion state: Working and >> used. Needs a litte cleanup and documentation. (Phobos candidate) > > This is #1. Far and away. Belongs in druntime. Stuff like this is why core.memory isn't called core.gc.
Re: Library Development: What to finish/flesh out?
On Fri, Mar 18, 2011 at 2:19 PM, filgood wrote: > Hi Lars, > > I agree on your orderbut would like to see Matrix ops in Phobos over > time (my understanding was that it can work without BLAS (just slower), > people can always in BLAS when they need to extra performance, no?). > > David, thanks a lot for your hard work... > > > How about Eigen? http://eigen.tuxfamily.org/index.php?title=Benchmark Those benchmarks are old, but still. Eigen v3 performs even better and they've added multi-threading and tons of other features and enhancements.
Re: Library Development: What to finish/flesh out?
Hi Lars, I agree on your orderbut would like to see Matrix ops in Phobos over time (my understanding was that it can work without BLAS (just slower), people can always in BLAS when they need to extra performance, no?). David, thanks a lot for your hard work... On 18/03/2011 09:26, Lars T. Kyllingstad wrote: On Thu, 17 Mar 2011 15:33:10 +, dsimcha wrote: I've accumulated a bunch of little libraries via various evening and weekend hacking projects over the past year or so, in various states of completion. Most are things I'm at least half-considering for Phobos, though some belong as third-party libs. I definitely don't have time to finish/flesh out all of them anytime soon, so I've decided to ask the community what to prioritize. Below is a summary of everything I've been working on, with its current level of completion. Please let me know the following: 1. A relative ordering of how useful you think these libraries would be to the community. In order: 1. TempAlloc 2. Matrix ops (I'm biased here, of course...) 3. RandAA 4. CSV parser 5. Rational That said, it does make sense to start with the things which require the least amount of work. If it would take you half an hour to complete the rationals lib, for instance, that may be a good starting point. Regarding std.mixins, I have to agree with Andrei and the others that code should be organised by functionality and not implementation method. Having thought some more about it, I also think std.file is not the right place for GZip support. Phobos needs an std.compression package/module, and a method for bulk reading/writing of gzip files may well be a good start. Finally, I know next to nothing about machine learning, so I won't express any opinion about it. 2. In absolute terms, would you find this useful? Absolutely! 3. For the Phobos candidates, whether they're general enough to belong in the **standard** library. I agree with Don that TempAlloc is a candidate for druntime. Other than that, yes. I note that others have suggested that the matrix stuff go into Phobos. As long as it depends on BLAS, I would say that's out of the question. -Lars
Re: Library Development: What to finish/flesh out?
On 03/17/2011 11:25 PM, dsimcha wrote: On 3/17/2011 6:18 PM, spir wrote: I'd have much use for both below. On 03/17/2011 04:33 PM, dsimcha wrote: 1. Rational: A library for handling rational numbers exactly. Templated on integer type, can use BigInts for guaranteed accuracy, or fixed-width integers for more speed where the denominator and numerator will be small. Completion state: Mostly finished. Just need to fix a litte bit rot and submit for review. (Phobos candidate) For decimal exactitude, what about plain fixed point (with decimal factor and binary mantissa)? I wouldn't mind having this, but I see it as completely orthogonal to rational numbers and don't have any near-term intentions of implementing it. Right. 2. RandAA: A hash table implementation with deterministic memory management, based on randomized probing. Main advantage over builtin AAs is that it plays much nicer with the GC and multithreaded programs. Lookup times are also expected O(1) no matter how many collisions exist in modulus hash space, as long as there are few collisions in full 32- or 64-bit hash space. Completion state: Mostly finished. Just needs a little doc improvement, a few benchmarks and submission for review. (Phobos candidate) How complicated would it be to add (optional) support for keeping insertion order (for iteration only)? Thought at a // array with pointers to the cells holding key/value pairs. It's a good idea, but IMHO something like this should be templated on the type of the associative array and work with builtin AAs, etc., too. It should be a decorator or something: /** Wraps any type that conforms to the duck interface of an associative array to preserve ordering for iteration. */ struct OrderedAA(AA) { alias typeof(AA.init.keys.front) K; alias typeof(AA.init.values.front) V; K[] order; AA aa; void opIndexAssign(V val, K key) { if(!(key in aa)) { order ~= key; } aa[key] = val; } // opApply, etc. } Right too. But the reason I have not implemented it yet is I don't want to keep a // array of keys, which require AA lookup for each key. Instead, I thought at an array of pointers to where the (key:value) cells are stored (cell "buckets"), to avoid AA key lookups. This, I guess, requires tweaking the implementation of the actual data structure. Reason why I asked you as you are dfevelopping a new one (so, you may have such a use case in mind before design is frozen). I have not yet had a look at the implementation of builtin AAs to see whether this would be easy. I guess not: it just require catching the very moment where a new pair is placed into a given bucket corresponding to its hash value. (And possibly the same thing at re-hash time.) Denis -- _ vita es estrany spir.wikidot.com
Re: Library Development: What to finish/flesh out?
On Thu, 17 Mar 2011 15:33:10 +, dsimcha wrote: > I've accumulated a bunch of little libraries via various evening and > weekend hacking projects over the past year or so, in various states of > completion. Most are things I'm at least half-considering for Phobos, > though some belong as third-party libs. I definitely don't have time to > finish/flesh out all of them anytime soon, so I've decided to ask the > community what to prioritize. Below is a summary of everything I've been > working on, with its current level of completion. Please let me know > the following: > > 1. A relative ordering of how useful you think these libraries would be > to the community. In order: 1. TempAlloc 2. Matrix ops (I'm biased here, of course...) 3. RandAA 4. CSV parser 5. Rational That said, it does make sense to start with the things which require the least amount of work. If it would take you half an hour to complete the rationals lib, for instance, that may be a good starting point. Regarding std.mixins, I have to agree with Andrei and the others that code should be organised by functionality and not implementation method. Having thought some more about it, I also think std.file is not the right place for GZip support. Phobos needs an std.compression package/module, and a method for bulk reading/writing of gzip files may well be a good start. Finally, I know next to nothing about machine learning, so I won't express any opinion about it. > 2. In absolute terms, would you find this useful? Absolutely! > 3. For the Phobos candidates, whether they're general enough to belong > in the **standard** library. I agree with Don that TempAlloc is a candidate for druntime. Other than that, yes. I note that others have suggested that the matrix stuff go into Phobos. As long as it depends on BLAS, I would say that's out of the question. -Lars
Re: Library Development: What to finish/flesh out?
dsimcha wrote: I've accumulated a bunch of little libraries via various evening and weekend hacking projects over the past year or so, in various states of completion. Most are things I'm at least half-considering for Phobos, though some belong as third-party libs. I definitely don't have time to finish/flesh out all of them anytime soon, so I've decided to ask the community what to prioritize. Below is a summary of everything I've been working on, with its current level of completion. Please let me know the following: 3. TempAlloc: A memory allocator based on a thread-local segmented stack, useful for allocating large temporary buffers in things like numerics code. Also comes with a hash table, hash set and AVL tree optimized for this allocation scheme. The advantages over plain old stack allocation are that it's independent of function calls (meaning you can return pointers to TempAlloc-allocated memory from a function, etc.) and it's segmented, meaning you can allocate huge buffers w/o risking stack overflow. Its main weakness is that this stack is not scanned by the GC, meaning that you can't store the only reference to a GC-allocated piece of memory here. However, in practice large arrays of primitives are an extremely common case in performance-critical code. I find this module immensely useful in dstats and Lars Kyllingstad uses it in SciD. Getting it into Phobos would make it easy for other scientific/numerics code to use it. Completion state: Working and used. Needs a litte cleanup and documentation. (Phobos candidate) This is #1. Far and away. Belongs in druntime. I would use it instantly in BigInt.
Re: Library Development: What to finish/flesh out?
On 3/17/2011 8:33 AM, dsimcha wrote: 8. GZip support in std.file: I'll leave the stream stuff for someone else, but just simple stuff like read(), write(), append() IMHO belongs in std.file. Completion state: Not started, but this is the easiest of the bunch to implement. (Phobos candidate) I'd definitely like to see gzip support in Phobos. But it shouldn't be in std.file, as there are many compression schemes in use besides gzip, and such should be composable with the file interface using ranges. In fact, gzip has nothing necessarily to do with files. It compresses/decompresses a stream (a range in D). Whether that is a file or something else is quite irrelevant to gzip. DMD, for example, uses a simple compressor for long symbol names. I suggest making a package for compressors, and then have gzip be a module within that, as in: std.compressor.gzip
Re: Library Development: What to finish/flesh out?
On 3/17/2011 6:18 PM, spir wrote: I'd have much use for both below. On 03/17/2011 04:33 PM, dsimcha wrote: 1. Rational: A library for handling rational numbers exactly. Templated on integer type, can use BigInts for guaranteed accuracy, or fixed-width integers for more speed where the denominator and numerator will be small. Completion state: Mostly finished. Just need to fix a litte bit rot and submit for review. (Phobos candidate) For decimal exactitude, what about plain fixed point (with decimal factor and binary mantissa)? I wouldn't mind having this, but I see it as completely orthogonal to rational numbers and don't have any near-term intentions of implementing it. 2. RandAA: A hash table implementation with deterministic memory management, based on randomized probing. Main advantage over builtin AAs is that it plays much nicer with the GC and multithreaded programs. Lookup times are also expected O(1) no matter how many collisions exist in modulus hash space, as long as there are few collisions in full 32- or 64-bit hash space. Completion state: Mostly finished. Just needs a little doc improvement, a few benchmarks and submission for review. (Phobos candidate) How complicated would it be to add (optional) support for keeping insertion order (for iteration only)? Thought at a // array with pointers to the cells holding key/value pairs. It's a good idea, but IMHO something like this should be templated on the type of the associative array and work with builtin AAs, etc., too. It should be a decorator or something: /** Wraps any type that conforms to the duck interface of an associative array to preserve ordering for iteration. */ struct OrderedAA(AA) { alias typeof(AA.init.keys.front) K; alias typeof(AA.init.values.front) V; K[] order; AA aa; void opIndexAssign(V val, K key) { if(!(key in aa)) { order ~= key; } aa[key] = val; } // opApply, etc. }
Re: Library Development: What to finish/flesh out?
I'd have much use for both below. On 03/17/2011 04:33 PM, dsimcha wrote: 1. Rational: A library for handling rational numbers exactly. Templated on integer type, can use BigInts for guaranteed accuracy, or fixed-width integers for more speed where the denominator and numerator will be small. Completion state: Mostly finished. Just need to fix a litte bit rot and submit for review. (Phobos candidate) For decimal exactitude, what about plain fixed point (with decimal factor and binary mantissa)? 2. RandAA: A hash table implementation with deterministic memory management, based on randomized probing. Main advantage over builtin AAs is that it plays much nicer with the GC and multithreaded programs. Lookup times are also expected O(1) no matter how many collisions exist in modulus hash space, as long as there are few collisions in full 32- or 64-bit hash space. Completion state: Mostly finished. Just needs a little doc improvement, a few benchmarks and submission for review. (Phobos candidate) How complicated would it be to add (optional) support for keeping insertion order (for iteration only)? Thought at a // array with pointers to the cells holding key/value pairs. Denis -- _ vita es estrany spir.wikidot.com
Re: Library Development: What to finish/flesh out?
On Thu, 17 Mar 2011 15:33:10 +, dsimcha wrote: > 5. Matrix operations: SciD improvements that allow you to write matrix > operations that look like normal math/MATLAB and optimizes them via > expression templates so that a minimal number of temporary matrices are > created. Uses/will use BLAS for multiplication. Completion state: > Addition implemented. Multiplication not. Nice matrix ops get my vote for what I would find most useful. Having matrices with m.rownames and m.colnames (similar to R's rownames(m) and colnames(m) for a matrix m), would be great too.
Re: Library Development: What to finish/flesh out?
dsimcha napisał: > I've accumulated a bunch of little libraries via various evening and weekend > hacking projects over the past year or so, in various states of completion. > Most are things I'm at least half-considering for Phobos, though some belong > as third-party libs. I definitely don't have time to finish/flesh out all of > them anytime soon, so I've decided to ask the community what to prioritize. > Below is a summary of everything I've been working on, with its current level > of completion. Please let me know the following: > > 1. A relative ordering of how useful you think these libraries would be to > the community. > > 2. In absolute terms, would you find this useful? > > 3. For the Phobos candidates, whether they're general enough to belong in the > **standard** library. > > List in order from most to least finished: > > 1. Rational: A library for handling rational numbers exactly. Templated on > integer type, can use BigInts for guaranteed accuracy, or fixed-width integers > for more speed where the denominator and numerator will be small. Completion > state: Mostly finished. Just need to fix a litte bit rot and submit for > review. (Phobos candidate) I'd find it useful. As for its presence in Phobos, I'm uncertain if it's in enough demand. > 2. RandAA: A hash table implementation with deterministic memory management, > based on randomized probing. Main advantage over builtin AAs is that it plays > much nicer with the GC and multithreaded programs. Lookup times are also > expected O(1) no matter how many collisions exist in modulus hash space, as > long as there are few collisions in full 32- or 64-bit hash space. Completion > state: Mostly finished. Just needs a little doc improvement, a few > benchmarks and submission for review. (Phobos candidate) Useful for me and in Phobos. > 3. TempAlloc: A memory allocator based on a thread-local segmented stack, > useful for allocating large temporary buffers in things like numerics code. > Also comes with a hash table, hash set and AVL tree optimized for this > allocation scheme. The advantages over plain old stack allocation are that > it's independent of function calls (meaning you can return pointers to > TempAlloc-allocated memory from a function, etc.) and it's segmented, meaning > you can allocate huge buffers w/o risking stack overflow. Its main weakness > is that this stack is not scanned by the GC, meaning that you can't store the > only reference to a GC-allocated piece of memory here. However, in practice > large arrays of primitives are an extremely common case in > performance-critical code. I find this module immensely useful in dstats and > Lars Kyllingstad uses it in SciD. Getting it into Phobos would make it easy > for other scientific/numerics code to use it. Completion state: Working and > used. Needs a litte cleanup and documentation. (Phobos candidate) Useful for me, don't know if for everyone else. > 4. Streaming CSV Parser: Parses CSV files as they're read in, a few > convenience functions for extracting columns into structs. If Phobos every > gets SQLite support I'll probably add sugar for turning a CSV file into an > SQLite database, too. Completion state: Prototype working, needs testing, > cleanup and documentation. (Phobos candidate) You mean a lazy slurp? It'd be useful for everyone. > 5. Matrix operations: SciD improvements that allow you to write matrix > operations that look like normal math/MATLAB and optimizes them via expression > templates so that a minimal number of temporary matrices are created. > Uses/will use BLAS for multiplication. Completion state: Addition > implemented. Multiplication not. It is worth considering standardizing at least matrix expressions in Phobos. The motivation is analogous to ranges -- to run an algorithm from lib A on a matrix container from lib B. C++ would be green with envy. I'd be glad to be part of the effort once I'm done with xml. > 6. Machine learning: Decision trees, KNN, Random Forest, Logistic > Regression, SVM, Naive Bayes, etc. This would be a dstats module. Completion > state: Decision trees prototyped, logistic regression working. I'd find it useful, I think anyone who's into this would too. > 7. std.mixins: Mixins for commonly needed boilerplate code. I stopped > working on this when Andrei suggested that making a collection of mixins into > a module is a bad idea. I've thought about it some more and I respectfully > disagree. std.mixins would be a one-stop shop for pretty much any boilerplate > you need to inject, and most of this code doesn't fit in any other obvious > place. Completion state: A few things (struct comparison, simple class > constructors, Singleton pattern) prototyped. (Phobos candidate) I'm afraid I also think functionality should be categorized by the purpose it serves rather than implementation technique. > 8. GZip support in std.file: I'll leave the stream stuff fo
Library Development: What to finish/flesh out?
I've accumulated a bunch of little libraries via various evening and weekend hacking projects over the past year or so, in various states of completion. Most are things I'm at least half-considering for Phobos, though some belong as third-party libs. I definitely don't have time to finish/flesh out all of them anytime soon, so I've decided to ask the community what to prioritize. Below is a summary of everything I've been working on, with its current level of completion. Please let me know the following: 1. A relative ordering of how useful you think these libraries would be to the community. 2. In absolute terms, would you find this useful? 3. For the Phobos candidates, whether they're general enough to belong in the **standard** library. List in order from most to least finished: 1. Rational: A library for handling rational numbers exactly. Templated on integer type, can use BigInts for guaranteed accuracy, or fixed-width integers for more speed where the denominator and numerator will be small. Completion state: Mostly finished. Just need to fix a litte bit rot and submit for review. (Phobos candidate) 2. RandAA: A hash table implementation with deterministic memory management, based on randomized probing. Main advantage over builtin AAs is that it plays much nicer with the GC and multithreaded programs. Lookup times are also expected O(1) no matter how many collisions exist in modulus hash space, as long as there are few collisions in full 32- or 64-bit hash space. Completion state: Mostly finished. Just needs a little doc improvement, a few benchmarks and submission for review. (Phobos candidate) 3. TempAlloc: A memory allocator based on a thread-local segmented stack, useful for allocating large temporary buffers in things like numerics code. Also comes with a hash table, hash set and AVL tree optimized for this allocation scheme. The advantages over plain old stack allocation are that it's independent of function calls (meaning you can return pointers to TempAlloc-allocated memory from a function, etc.) and it's segmented, meaning you can allocate huge buffers w/o risking stack overflow. Its main weakness is that this stack is not scanned by the GC, meaning that you can't store the only reference to a GC-allocated piece of memory here. However, in practice large arrays of primitives are an extremely common case in performance-critical code. I find this module immensely useful in dstats and Lars Kyllingstad uses it in SciD. Getting it into Phobos would make it easy for other scientific/numerics code to use it. Completion state: Working and used. Needs a litte cleanup and documentation. (Phobos candidate) 4. Streaming CSV Parser: Parses CSV files as they're read in, a few convenience functions for extracting columns into structs. If Phobos every gets SQLite support I'll probably add sugar for turning a CSV file into an SQLite database, too. Completion state: Prototype working, needs testing, cleanup and documentation. (Phobos candidate) 5. Matrix operations: SciD improvements that allow you to write matrix operations that look like normal math/MATLAB and optimizes them via expression templates so that a minimal number of temporary matrices are created. Uses/will use BLAS for multiplication. Completion state: Addition implemented. Multiplication not. 6. Machine learning: Decision trees, KNN, Random Forest, Logistic Regression, SVM, Naive Bayes, etc. This would be a dstats module. Completion state: Decision trees prototyped, logistic regression working. 7. std.mixins: Mixins for commonly needed boilerplate code. I stopped working on this when Andrei suggested that making a collection of mixins into a module is a bad idea. I've thought about it some more and I respectfully disagree. std.mixins would be a one-stop shop for pretty much any boilerplate you need to inject, and most of this code doesn't fit in any other obvious place. Completion state: A few things (struct comparison, simple class constructors, Singleton pattern) prototyped. (Phobos candidate) 8. GZip support in std.file: I'll leave the stream stuff for someone else, but just simple stuff like read(), write(), append() IMHO belongs in std.file. Completion state: Not started, but this is the easiest of the bunch to implement. (Phobos candidate)