Scattered container store each type's data member sequentially for all objects. That is, the first data member of all objects is stored contiguosly in memory, then the second data member, and so on as shown in the following figure:
This improves cache line utilization when iterating sequentially over a container without accessing all object's data members. They also allow asynchronous processing of object's data member without false sharing.
The following containers are available:
scattered::vector<T>
(analogous tostd::vector<T>
).
Scattered is a Boost Software License'd header only C++1y library and is tested with Boost 1.54 (1.55 not supported yet, see issue tracker) and trunk clang/libc++. It depends on Boost.MPL and Boost.Fusion.
#include <algorithm>
#include <boost/range/algorithm.hpp>
#include <boost/fusion/adapted/struct/adapt_assoc_struct.hpp>
#include "scattered/vector.hpp"
/// T is a struct; k contains keys to access the struct elements:
struct T {
float x; double y; int i; bool b;
struct k { struct x {}; struct y {}; struct i {}; struct b {}; };
};
/// This adapts the class as an associative fusion sequence
BOOST_FUSION_ADAPT_ASSOC_STRUCT(
T, (float, x, T::k::x)(double, y, T::k::y)
(int , i, T::k::i)(bool , b, T::k::b))
int main() {
using scattered::get;
using k = T::k;
scattered::vector<T> vec(10);
/// To modify the elements of the vector in place
/// the keys are used on a reference proxy:
int count = 0;
for (auto i : vec) {
get<k::x>(i) = static_cast<float>(count);
get<k::y>(i) = static_cast<double>(count);
get<k::i>(i) = count;
get<k::b>(i) = count % 2 == 0;
++count;
}
/// Cache lines contain only "y" data-members:
for (auto i : vec) { get<k::y>(i) += get<k::y>(i); }
/// Boost and STL algorithms work out of the box
boost::stable_sort(vec, [](auto i, auto j) {
return get<k::x>(i) > get<k::x>(j);
});
/// Reference proxy implicitly converts to T
boost::transform(vec, std::begin(vec), [](T i) { i.y *= i.y; return i; });
vec.push_back(T{4.0, 3.0, 2, false});
}
./configure.sh
is used to provide libc++'s path and select the compilation mode (debug/release/sanitizers are provided, see./configure,sh -h
).make
compiles the tests,ctest
launches all tests.make docs
builds the documentation.make update-format
reformats the code with clang-format.make bench
runs all benchmarks.
Scattered containers use proxy reference types and, as a consequence, have the
following caveats similar to those of std::vector<bool>
:
// i's type = scattered::vector<T>::reference, not scattered::vector<T>::value_type
auto i = *scatteredVector.begin();
// j's type = scattered::vector<T>::reference&
auto& j = *scatteredVector.begin();
// this modifies scatteredVector:
i = T{};
// to get a value you need to use value_type (or T)
T value = *scatteredVector.begin();
// This fails because T& cannot bind to scattered::vector<T>::reference&
T& ref = *catteredVector.begin(); // Compilation Error
Scattered relies on
BOOST_FUSION_ADAPT_ASSOC_X
to adapt classes as associative Boost.Fusion sequences. The
scattered::get<key>(reference_proxy)
function returns a reference to the
object's data member associated with the key
.
The aim of the library is to maximize memory bandwidth usage for algorithms that
iterate sequentially over the container. The following benchmarks are located
in the benchmarks/
directory and can be run with make bench
.
See the roadmaps page in the issue list.
- Finish:
scattered::vector
will be provided. - Other containers:
scattered::flat_set
andscattered::unordered_map
.
The Scattered library resulted from efforts to improve the cache performance of numerical fluid mechanics codes at the Institute of Aerodynamics, Aachen. I'd like to thank Georg Geiser for introducing me to What Every Programmer Should Know About Memory. Furthermore, I want to thank the guests of the LoungeC++ for their discussions, company, and help. In particular, to Evgeny Panasyuk who motivated me to write this library.