Re: [boost] Suggestions for TMP/function introspection problem?

Subject: Re: [boost] Suggestions for TMP/function introspection problem?
From: Hans Dembinski (hans.dembinski_at_[hidden])
Date: 2018-11-09 09:28:14

> On 8. Nov 2018, at 19:03, Steven Watanabe via Boost <boost_at_[hidden]> wrote:
>> 
>> It is for generic support of arbitrary accumulators per bin. Normally, a histogram gets some values equal to the number of axes, finds a corresponding bin and increments a counter for that bin. More generally, one can replace the counter by an accumulator that can accept an arbitrary number of values (the standard histogram is then the special case where the accumulator accepts zero arguments). Extra arguments passed to the histogram are forwarded to the accumulator. For example, an accumulator could compute a mean for values which end up in the same bin. This is what people in high energy physics call a "profile".
>> 
>> // bin in 'x', compute a mean of 'y' for each bin in 'x'
>> auto h = make_histogram_with(std::vector<mean_accumulator>(), axis::regular<>(10, 0, 1));
>> 
>> // first argument is used to find the bin, second argument is passed to the mean_accumulator instance of that bin
>> h(1, 2);
>> 
>> The operator() of the accumulator could accept more than one value. Maybe you want to bin in 'x' and compute means for 'y' and 'z' at the same time. This should work and then you would pass three values to the histogram in total. I want to use the introspection for the accumulator to detect the number of arguments.
> 
> Okay.  So the problem is that you're essentially
> packing two variadic argument lists together and
> you need to untangle them.  I don't think that's
> a good idea in the first place, as it makes the
> meaning of user code unclear.  I think it would
> be better to distinguish them at the call-site
> somehow.

Agreed, this solution is also more readable on the user-facing call-site. In the develop branch, it is now implemented like this (same histogram example as before)

// `sample` is a helper function which stores arguments passed to it as a std::tuple, to be forwarded to the accumulator
h(1, sample(2));

// sample may appear at the beginning or the end and can be combined in arbitrary ways with `weight`
// as long as the values passed to the histogram axis appear in one continuous block
h(weight(3.5), 1, sample(2));
h(1, weight(3.5), sample(2));
h(1, sample(2), weight(3.5));

>> I don't always know at compile-time how many arguments should be passed to the histogram axes (because one may use a dynamic histogram where the number of axis is only known at runtime), so I cannot just subtract the arguments that go to the axes and push the rest to the accumulator. But I could do it the other way round, subtract the arguments that need to go to the accumulator and pass the rest to the histogram axes.
>> 
> 
> In a dynamic histogram, why can't the accumulator
> also be dynamic?

Good point, that argument was the final nail in the coffin of my idea to make it work without the `sample` helper function.

Thank you for the valuable input!
Hans