search

boostorg / histogram

Fast multi-dimensional generalized histogram with convenient interface for C++14
Boost Software License 1.0
313 stars 74 forks source link

Feature request: strided array support for fill? #331

Open henryiii opened 2 years ago

henryiii commented 2 years ago

In #327, the bug seems to be that single values are not supported correctly. In https://github.com/scikit-hep/hist/issues/303, scalars are currently not allowed for samples - in working on this, I found that h.fill(0.3, sample=[1, 2], weight=1) is not supported in boost-histogram, due to the spans not matching (may be solvable on the boost-histogram side).

A way to support this easily could be to support strided arrays, like NumPy, Eigen, Boost.MultiArray, etc. all provide. If an array is described by a stride and an offset, then setting that stride to 0 would allow a scalar value without having a scalar datatype. Then there would be half as many items in the fill variant, just arrays, no scalars anymore.

A related issue is that boost-histogram has to copy the input array if it's not already a dense 1D array. If describing an array with a starting pointer, stride, and offset was possible, then this would no longer be necessary, only the datatype would have to match to avoid a copy.

Thoughts?

HDembinski commented 2 years ago

h.fill(0.3, sample=[1, 2], weight=1) should be supported in C++, but currently it is not. What is already supported is h.fill(0.3, [1, 2]) for a multidimensional histogram. This support just does not currently extend to sample and weight, but it should.

Supporting strided arrays in general is much more involved than this special case, which is easy to support without supporting strided input.

HDembinski commented 2 years ago

What I am currently doing internally is smarter than just setting the stride to 0 if the user passes a value instead of an iterable. A naive implementation would call axis::index(value) N times, but I am calling it exactly once in this case.