On Sat, 28 Aug 2010, Dave Abrahams wrote:

On Aug 28, 2010, at 3:58 PM, Jeremiah Willcock wrote:

...

There are several container types (such as typical deque implementations, if you have access to their internals) that can have for_each() implemented for them in a way that is faster than using iterators and a generic for_each() function. Is there a way to take advantage of that in Boost.Range, such as by overloading boost::for_each()? Is there some kind of Iterable concept that doesn't require actual iterators or allows them to be slow?

I presume you're familiar with http://lafstern.org/matt/segmented.pdf?

I had heard of segmented iterators but hadn't looked at them in detail, so thanks for the link. I'm not sure I want something that's just two-level and still based on iterators, though; one case I was thinking of is compressed data that truly has a linear sequence but it might be faster to have the decompressor control the program's control flow rather than having the algorithm do it (of course, that isn't as general as iterators, but works for some algorithms). -- Jeremiah Willcock

On 29/08/2010 00:53, Dave Abrahams wrote:

...

...

I presume you're familiar with http://lafstern.org/matt/segmented.pdf?

I had heard of segmented iterators but hadn't looked at them in detail, so thanks for the link. I'm not sure I want something that's just two-level

Read the paper. It works for as many levels as you want. The localiterator can itself be segmented.

I see some potential for some interesting discussion here, sorry in advance for hijacking this thread. I think the whole segmented iterators/ranges is a bad idea, as it has quite a few disadvantages and there are alternative, simpler solutions that provide the same benefits. segmented ranges have the following disadvantages: - while segmented iterators are easier to write than flat iterators in certain cases, they're still not as easy to write as they could be - any code or algorithm that deals with regular ranges has to be rewritten, not just in order to gain the performance gains, but to get basic functionality working as well. That's about as bad as it gets from an integration point of view. - only trivial operations that manipulate element per element (for_each, copy, transform, accumulate, erase_if) can really exploit the performance gains of segmented ranges - making non-trivial operations (any operation that needs to manipulate several adjacent elements [1]) deal with segmented ranges can be so complicated that they can realistically only be written with a flattened non-segmented view of the range. [1] examples include character encoding conversion, base64 encoding/decoding, gzip encoding/decoding, etc.

...

and still based on iterators, though; one case I was thinking of is compressed data that truly has a linear sequence but it might be faster to have the decompressor control the program's control flow rather than having the algorithm do it (of course, that isn't as general as iterators, but works for some algorithms).

Ah, push or pull—the eternal struggle.

Push and pull are the opposite ends of the same thing. When you want to generate data, you'd rather push rather than be pulled, when you want to read it, you'd rather pull than be pushed. So what's the solution? Generators. An by that, I mean a function that writes data to an output iterator. A generator is a pure-push primitive, and is trivial to write efficiently. Trivial operations can be implemented with no cost on top of a generator, by using something such as boost::function_output_iterator. There are generic techniques to build a (normal, non-segmented) range that you can then easily pull from on top of a generator, but those are more complicated and introduce some overhead; overhead you would need anyway to do any non-trivial operation. There are two techniques to achieve this. The first one, the magical silver bullet, involves using a coroutine. Every time you generate a value, yield. This gives control back to the other end, that can do whatever it wants with the element, and when it pulls again it goes back in the generator exactly where it was, and can keep on pushing. The code keeps swapping between a pushing and a pulling context and achieves exactly the push->pull integration we wanted. It however raises a few concerns: context size, allocation and management, and potentially excessive context swapping. The other technique is to generate data step by step, each step having a fixed maximum size. Then you can run a step on a temporary buffer, read off that temporary buffer, then run a new step once everything has been read. An iterator that does this transparently on-top of a generator can be written. Albeit it does introduce some overhead, it could be less than that of the coroutine solution. The biggest concern is that it makes the iterator fat, since the temporary buffer is part of the iterator; and iterators are typically passed by value. The latter approach is the one I used to implement my generic Converter system as part of my Unicode library. A converter is a step-generator, except it takes input to convert as well. I've got an iterator adaptor that can then run a converter as the data is being iterated, and thus lazily perform any arbitrary conversion. Note that the end result, in the case of Unicode, in not very different from the Unicode iterator adaptors hand-written by John Maddock, except the implementation is much simpler, and can be easily used for efficient pushing as well. So, at the end of this long message, what do I propose to solve the original problem? Have your container provide a step-generator. You can then implement your iterator on top of it, or not, your choice (but I'd recommend using it to avoid unnecessary code duplication and maintenance). A type trait could then be introduced to tell whether a sequence provides a generator or not. Boost.Range could make use of that type trait to specialize trivial operations to use the generator instead of the iterators. We end up with: - an easy to write solution - compatibility will all code that takes ranges as input - only the operations that want to can be specialized to exploit the performance gains - no restriction imposed on operations that cannot be implemented in terms of the generator directly

On 29/08/2010 17:53, Dave Abrahams wrote:

On Sun, Aug 29, 2010 at 4:34 AM, Mathias Gaunard wrote:

...

So, at the end of this long message, what do I propose to solve the original problem?

Have your container provide a step-generator. You can then implement your iterator on top of it, or not, your choice (but I'd recommend using it to avoid unnecessary code duplication and maintenance).

I don't understand your suggestion . To make your proposal understandable, could you please illustrate, in code, what a step-generator would look like for std::deque, and how it could be used?

I'm not really familiar with the implementation of std::deque, so I'll just assume it's a vector of statically-sized arrays. (I'm also assuming each statically-sized array is full for the sake of simplicity, but in truth the ones on both ends wouldn't be) I don't know yet what a generator that just runs a step should look like, here's a first try: template<typename T> struct deque_step_generator { typedef T output_type; static const size_t max_output = whatever; deque_step_generator(deque& deq_) : deq(deq_), it(deq.begin()) {} template<typename Out> Out operator()(Out out) { return copy(*it++, out); } bool at_end() { return it == deq.end(); } private: typedef vector< array > deque; deque &deq; typename deque::const_iterator it; }; One solution for the caller to detect it has reached the end is to have an at_end member function. Another solution could be to generate nothing at the end -- in a similar fashion to the 'read' POSIX function. What's the best design remains to be evaluated. It might also be needed to have a way to integrate right-to-left generation to allow for bidirectional ranges. Of course, for certain data structures, providing a step-generator, as I call it, can still be too complicated, and they could only have a 'global' generator, which, in the case or our deque, would be: template<typename T> struct deque_generator { typedef T output_type; deque_generator(deque& deq_) : deq(deq_) {} template<typename Out> Out operator()(Out out) { for(typename deque::const_iterator it = deq.begin(); it != deq.end(); ++it) out = copy(*it++, out); return out; } private: typedef vector< array > deque; deque &deq; }; In the case of a global generator, implementing algorithms such as std::accumulate should be basically the same thing as the string_appender example here: http://www.boost.org/doc/libs/1_44_0/libs/iterator/doc/function_output_itera... In the case of the step one, it's the same except you need to do it in a loop. In order to be converted to a flat iterable range, a global generator would require a coroutine, while a step one could be converted using either the coroutine or the buffered technique. Writing the code that does that generic conversion could be a bit harder to write than that, so I'm not really up to the task at this late hour (to the point where I even cheated a fair bit on the deque generators...).

On 30/08/2010 13:05, David Abrahams wrote:

Hi Mathias,

Thanks, but you only answered the first part of my question. I don't understand how you'd use such a thing.

I don't really understand your question, so I've put together a practical example, that is available on the vault: http://www.boostpro.com/vault/index.php?action=downloadfile&filename=generator_iterator.zip&directory=& To demonstrate the idea, I've implemented a simplistic binary tree in binary_tree.hpp, and provided for it a forward iterator and a global generator. As can be clearly seen, the global generator has a trivial, (and possibly more efficient) implementation, while the iterator is a bit more complicated and shouldn't fare as well. I'm not even sure the iterator implementation is correct, which shows writing iterators is quite more complicated. I've implemented a range-based 'copy' function, that behaves just like boost::copy from Boost.Range, except that if the type has a generator, it can use it instead. The same thing could be done for all standard algorithms: for_each, accumulate, etc. I've also implemented a facility to build an iterator from a generator using coroutines (uses the latest version of boost.coroutine I could find). I tried to do some simple timings, but I must be doing it wrong, since the results can be extremely variable.

On Mon, Aug 30, 2010 at 3:22 PM, Dave Abrahams wrote:

Sent from my iPhone On Aug 30, 2010, at 3:26 PM, Mathias Gaunard wrote:

I don't really understand your question, so I've put together a practical example, that is available on the vault: http://www.boostpro.com/vault/index.php?action=downloadfile&filename=generator_iterator.zip&directory=&

And if you'd put that in a github repo I'd have been able to peruse it with my coveted but awkward mobile device ;-)

My coveted, yet awkward mobile device[0] can handle that link just fine, extract the zip, and peruse the source code fine. ;-) [0] HTC Touch Pro, running a dual-boot between Windows Mobile 6.5 and Android 2.2.

On 8/29/2010 1:03 PM, Jeremiah Willcock wrote:

I think just overloading for_each() and writing algorithms in terms of that when possible would get the same benefits, though.

This way lies madness. The separation of containers and algorithms in the STL is what keeps it manageable. You don't want to implement N algorithms over M container types. You want to define a concept on which you can "overload" for_each. -- Eric Niebler BoostPro Computing http://www.boostpro.com

On 8/29/2010 2:27 PM, Jeremiah Willcock wrote:

On Sun, 29 Aug 2010, Eric Niebler wrote:

...

On 8/29/2010 1:03 PM, Jeremiah Willcock wrote:

...

I think just overloading for_each() and writing algorithms in terms of that when possible would get the same benefits, though.

This way lies madness. The separation of containers and algorithms in the STL is what keeps it manageable. You don't want to implement N algorithms over M container types. You want to define a concept on which you can "overload" for_each.

Yes, that's what I meant. Have a default implementation, but specifically allow it to be overloaded, then make other algorithms use for_each() so that they can use overloaded versions. Maybe the base algorithm should be some kind of fold like it is in a functional language, though.

No. There is no one-algorithm-to-rule-them-all, with which all other algorithms can be implemented. There are different algorithms and different sequence types. The trick is matching the best algorithm implementation with the specific capabilities supported by the sequence. Hence the iterator categories. And segmented iterators are an extra category for which the algorithms can be specialized. But I'm pretty sure you know all this, so I feel a bit strange saying it. Have I missed something? -- Eric Niebler BoostPro Computing http://www.boostpro.com

On Sun, Aug 29, 2010 at 5:34 AM, Mathias Gaunard wrote:

On 29/08/2010 00:53, Dave Abrahams wrote:

...

...

...

I presume you're familiar with http://lafstern.org/matt/segmented.pdf?

I had heard of segmented iterators but hadn't looked at them in detail, so thanks for the link.  I'm not sure I want something that's just two-level

Read the paper. It works for as many levels as you want. The localiterator can itself be segmented.

I see some potential for some interesting discussion here, sorry in advance for hijacking this thread.

I think the whole segmented iterators/ranges is a bad idea, as it has quite a few disadvantages and there are alternative, simpler solutions that provide the same benefits.

segmented ranges have the following disadvantages:

- while segmented iterators are easier to write than flat iterators in certain cases, they're still not as easy to write as they could be

I don't think anyone is suggesting replacing flat iterators with segmented iterators -- they would coexist. I don't see anything wrong with letting generic algorithms (like for_each) transparently take advantage of the underlying storage of a sequence.

- making non-trivial operations (any operation that needs to manipulate several adjacent elements [1]) deal with segmented ranges can be so complicated that they can realistically only be written with a flattened non-segmented view of the range.

[1] examples include character encoding conversion, base64 encoding/decoding, gzip encoding/decoding, etc.

While developing a HTTP parser, I wanted to be able to parse between two or more buffers (segments), but that was quite slow. I wanted full performance for the 99% of the time that a valid sequence would exist in a single buffer. I wanted to avoid complicated code to move between segments, and didn't want to copy buffers around. I ended up letting the parser keep taking flat iterators as input -- as you say, clearly these are the simplest kind to work with. On top of that I built another parser that first uses a parser as far as it can go on the first segment, and if that doesn't have enough data to return something, it uses a parser on as little of the full sequence as possible. Worked wonders, required very little code to do, and kept the core parsing logic simple. Granted HTTP is not all things, and base64/character encoding conversion/gzip are not all things, but I think this way would work fine for at least those. I don't think it's right to say "the whole segmented iterators/ranges is a bad idea" -- I can see plenty of applications where writing for segmented iterators *would* get overly complicated, but I don't see any reason why we'd be forced to use them if something better was available. -- Cory Nelson http://int64.org