Performance degradation due to default initialisation of elements in standard containers

Question

(Yes, I know there is a question with almost the same title, but the answer was not satisfactory, see below)

EDIT Sorry, the original question didn't use compiler optimization. This is now fixed, but to avoid trivial optimization and to come closer to my actual use case, the test has been split into two compilation units.

The fact that the constructor of std::vector<> has linear complexity is a nuisance when it comes to performance-critical applications. Consider this simple code

// compilation unit 1:
void set_v0(type*x, size_t n)
{
  for(size_t i=0; i<n; ++i)
    x[i] = simple_function(i);
}

// compilation unit 2:
std::vector<type> x(n);                     // default initialisation is wasteful
set_v0(x.data(),n);                         // over-writes initial values

when a significant amount of time is wasted by constructing x. The conventional way around this, as explored by this question, seems to be to merely reserve the storage and use push_back() to fill in the data:

// compilation unit 1:
void set_v1(std::vector<type>&x, size_t n)
{
  x.reserve(n);
  for(size_t i=0; i<n; ++i)
    x.push_back(simple_function(i));
}

// compilation unit 2:
std::vector<type> x(); x.reserve(n);        // no initialisation
set_v1(x,n);                                // using push_back()

However, as indicated by my comment, the push_back() is inherently slow, making this second approach actually slower than the first one for sufficiently simply constructible objects, such as size_ts, when for

simple_function = [](size_t i) { return i; };

I get the following timings (using gcc 4.8 with -O3; clang 3.2 produced ~10% slower code)

timing vector::vector(n) + set_v0();
n=10000 time: 3.9e-05 sec
n=100000 time: 0.00037 sec
n=1000000 time: 0.003678 sec
n=10000000 time: 0.03565 sec
n=100000000 time: 0.373275 sec

timing vector::vector() + vector::reserve(n) + set_v1();
n=10000 time: 1.9e-05 sec
n=100000 time: 0.00018 sec
n=1000000 time: 0.00177 sec
n=10000000 time: 0.020829 sec
n=100000000 time: 0.435393 sec

The speed-up actually possible if one could elide the default construction of elements can be estimated by the following cheating version

// compilation unit 2
std::vector<type> x; x.reserve(n);          // no initialisation
set_v0(x,n);                                // error: write beyond end of vector
                                            // note: vector::size() == 0

when we get

timing vector::vector + vector::reserve(n) + set_v0();          (CHEATING)
n=10000 time: 8e-06 sec
n=100000 time: 7.2e-05 sec
n=1000000 time: 0.000776 sec
n=10000000 time: 0.01119 sec
n=100000000 time: 0.298024 sec

So, my first question: Is there any legal way to use a standard library container which would give these latter timings? Or do I have to resort to manage the memory myself?

Now, what I really want, is to use multi-threading to fill in the container. The naive code (using openMP in this example for simplicity, which excludes clang for the moment)

// compilation unit 1
void set_v0(type*x, size_t n)
{
#pragma omp for                       // only difference to set_v0() from above 
  for(size_t i=0; i<n; ++i)
    x[i] = simple_function(i);
}

// compilation unit 2:
std::vector<type> x(n);               // default initialisation not mutli-threaded
#pragma omp parallel
set_v0(x,n);                          // over-writes initial values in parallel

now suffers from the fact that the default initialization of all elements is not multi-threaded, resulting in an potentially serious performance degradation. Here are the timings for set_omp_v0() and a equivalent cheating method (using my macbook's intel i7 chip with 4 cores, 8 hyperthreads):

timing std::vector::vector(n) + omp parallel set_v0()
n=10000 time: 0.000389 sec
n=100000 time: 0.000226 sec
n=1000000 time: 0.001406 sec
n=10000000 time: 0.019833 sec
n=100000000 time: 0.35531 sec

timing vector::vector + vector::reserve(n) + omp parallel set_v0(); (CHEATING)
n=10000 time: 0.000222 sec
n=100000 time: 0.000243 sec
n=1000000 time: 0.000793 sec
n=10000000 time: 0.008952 sec
n=100000000 time: 0.089619 sec

Note that the cheat version is ~3.3 times faster than the serial cheat version, roughly as expected, but the standard version is not.

So, my second question: Is there any legal way to use a standard library container which would give these latter timings in multi-threaded situations?

PS. I found this question, where std::vector is tricked into avoiding the default initialization by providing it with a uninitialized_allocator. This is no longer standard compliant, but works very well for my test case (see my own answer below and this question for details).

Answer 1

With g++ 4.5 I was able to realize an approximate 20% reduction in runtime from v0 (1.0s to 0.8s) and slightly less from 0.95s to 0.8s for v1 by using a generator to construct directly:

struct Generator : public std::iterator<std::forward_iterator_tag, int>
{
    explicit Generator(int start) : value_(start) { }
    void operator++() { ++value_; }
    int operator*() const { return value_; }

    bool operator!=(Generator other) const { return value_ != other.value_; }

    int value_;
};

int main()
{
    const int n = 100000000;
    std::vector<int> v(Generator(0), Generator(n));

    return 0;
}

Answer 2

Okay, here is what I've learned since asking this question.

Q1 (Is there any legal way to use a standard library container which would give these latter timings?) Yes to some degree, as shown in the answers by Mark and Evgeny. The method of providing a generator to the constructor of std::vector elides the default construction.

Q2 (Is there any legal way to use a standard library container which would give these latter timings in multi-threaded situations?) No, I don't think so. The reason is that on construction any standard-compliant container must initialise its elements, already to ensure that the call to the element destructors (upon destruction or resizing of the container) is well-formed. As the std library containers do not support the usage of multi-threading in constructing their elements, the trick of Q1 cannot be replicated here, so we cannot elide the default construction.

Thus, if we want to use C++ for high-performance computing our options are somewhat limited when it comes to managing large amounts of data. We can

1 declare a container object and, in the same compilation unit, immediately fill it (concurrently), when the compiler hopefully optimizes the initialization at construction away;

2 resort to new[] and delete[] or even malloc() and free(), when all the memory management and, in the latter case, construction of elements is our responsibility and our potential usage of the C++ standard library very limited.

3 trick a std::vector to not initialise its elements using a custom unitialised_allocator that elides the default construction. Following the ideas of Jared Hoberock such an allocator could look like this (see also here):

// based on a design by Jared Hoberock
// edited (Walter) 10-May-2013, 23-Apr-2014
template<typename T, typename base_allocator = std::allocator<T> >
struct uninitialised_allocator
  : base_allocator
{
  static_assert(std::is_same<T,typename base_allocator::value_type>::value,
                "allocator::value_type mismatch");

  template<typename U>
  using base_t =
    typename std::allocator_traits<base_allocator>::template rebind_alloc<U>;

  // rebind to base_t<U> for all U!=T: we won't leave other types uninitialised!
  template<typename U>
  struct rebind
  {
    typedef typename
    std::conditional<std::is_same<T,U>::value,
                     uninitialised_allocator, base_t<U> >::type other; 
  }

  // elide trivial default construction of objects of type T only
  template<typename U>
  typename std::enable_if<std::is_same<T,U>::value && 
                          std::is_trivially_default_constructible<U>::value>::type
  construct(U*) {}

  // elide trivial default destruction of objects of type T only
  template<typename U>
  typename std::enable_if<std::is_same<T,U>::value && 
                          std::is_trivially_destructible<U>::value>::type
  destroy(U*) {}

  // forward everything else to the base
  using base_allocator::construct;
  using base_allocator::destroy;
};

Then an unitialised_vector<> template could be defined like this:

template<typename T, typename base_allocator = std::allocator<T>>
using uninitialised_vector = std::vector<T,uninitialised_allocator<T,base_allocator>>;

and we can still use almost all of the standard library's functionality. Though it must be said that the uninitialised_allocator, and hence by implication the unitialised_vector are not standard compliant, because its elements are not default constructed (e.g. a vector<int> will not have all 0 after construction).

When using this tool for my little test problem, I get excellent results:

timing vector::vector(n) + set_v0();
n=10000 time: 3.7e-05 sec
n=100000 time: 0.000334 sec
n=1000000 time: 0.002926 sec
n=10000000 time: 0.028649 sec
n=100000000 time: 0.293433 sec

timing vector::vector() + vector::reserve() + set_v1();
n=10000 time: 2e-05 sec
n=100000 time: 0.000178 sec
n=1000000 time: 0.001781 sec
n=10000000 time: 0.020922 sec
n=100000000 time: 0.428243 sec

timing vector::vector() + vector::reserve() + set_v0();
n=10000 time: 9e-06 sec
n=100000 time: 7.3e-05 sec
n=1000000 time: 0.000821 sec
n=10000000 time: 0.011685 sec
n=100000000 time: 0.291055 sec

timing vector::vector(n) + omp parllel set_v0();
n=10000 time: 0.00044 sec
n=100000 time: 0.000183 sec
n=1000000 time: 0.000793 sec
n=10000000 time: 0.00892 sec
n=100000000 time: 0.088051 sec

timing vector::vector() + vector::reserve() + omp parallel set_v0();
n=10000 time: 0.000192 sec
n=100000 time: 0.000202 sec
n=1000000 time: 0.00067 sec
n=10000000 time: 0.008596 sec
n=100000000 time: 0.088045 sec

when there is no difference any more between the cheating and "legal" versions.