Concurrent, page-based, dynamically-sized linear data structure with O(1) random access and STL-compliant iterators. It is primarily intended for applications that concurrently insert (a possibly unkown number of) elements into a dynamically growing linear array, and fast random access to said elements.
More...
|
| PagedArray () |
| Default constructor.
|
|
| ~PagedArray () |
| Destructor removed all allocated pages.
|
|
| PagedArray (const PagedArray &)=delete |
|
PagedArray & | operator= (const PagedArray &)=delete |
|
ValueBuffer | getBuffer () |
|
size_t | push_back_unsafe (const ValueType &value) |
|
void | shrink_to_fit () |
| Reduce the page table to fix the current size.
|
|
ValueType & | operator[] (size_t i) |
| Return a reference to the value at the specified offset.
|
|
const ValueType & | operator[] (size_t i) const |
| Return a const-reference to the value at the specified offset.
|
|
void | fill (const ValueType &v) |
| Set all elements in the page table to the specified value.
|
|
bool | copy (ValueType *p, size_t count) const |
| Copy the first count values in this PageArray into a raw c-style array, assuming it to be at least count elements long.
|
|
void | copy (ValueType *p) const |
|
void | resize (size_t size) |
| Resize this array to the specified size.
|
|
void | resize (size_t size, const ValueType &v) |
| Resize this array to the specified size and initialize all values to v.
|
|
size_t | size () const |
| Return the number of elements in this array.
|
|
size_t | capacity () const |
| Return the maximum number of elements that this array can contain without allocating more memory pages.
|
|
size_t | freeCount () const |
| Return the number of additional elements that can be added to this array without allocating more memory pages.
|
|
size_t | pageCount () const |
| Return the number of allocated memory pages.
|
|
size_t | memUsage () const |
| Return the memory footprint of this array in bytes.
|
|
bool | isEmpty () const |
| Return true if the container contains no elements.
|
|
bool | isPartiallyFull () const |
| Return true if the page table is partially full, i.e. the last non-empty page contains less than pageSize() elements.
|
|
void | clear () |
| Removes all elements from the array and delete all pages.
|
|
Iterator | begin () |
| Return a non-const iterator pointing to the first element.
|
|
Iterator | end () |
| Return a non-const iterator pointing to the past-the-last element.
|
|
ConstIterator | cbegin () const |
| Return a const iterator pointing to the first element.
|
|
ConstIterator | begin () const |
|
ConstIterator | cend () const |
| Return a const iterator pointing to the past-the-last element.
|
|
ConstIterator | end () const |
|
void | sort () |
| Parallel sort of all the elements in ascending order.
|
|
void | invSort () |
| Parallel sort of all the elements in descending order.
|
|
template<typename Functor > |
void | sort (Functor func) |
| Parallel sort of all the elements based on a custom functor with the api:
|
|
void | merge (PagedArray &other) |
| Transfer all the elements (and pages) from the other array to this array.
|
|
void | print (std::ostream &os=std::cout) const |
| Print information for debugging.
|
|
template<typename ValueT, size_t Log2PageSize = 10UL>
class openvdb::v11_0::util::PagedArray< ValueT, Log2PageSize >
Concurrent, page-based, dynamically-sized linear data structure with O(1) random access and STL-compliant iterators. It is primarily intended for applications that concurrently insert (a possibly unkown number of) elements into a dynamically growing linear array, and fast random access to said elements.
- Note
- Multiple threads can grow the page-table and push_back new elements concurrently. A ValueBuffer provides accelerated and threadsafe push_back at the cost of potentially re-ordering elements (when multiple instances are used).
This data structure employes contiguous pages of elements (a std::deque) which avoids moving data when the capacity is out-grown and new pages are allocated. The size of the pages can be controlled with the Log2PageSize template parameter (defaults to 1024 elements of type ValueT).
There are three fundamentally different ways to insert elements to this container - each with different advanteges and disadvanteges.
The simplest way to insert elements is to use PagedArray::push_back_unsafe which is not thread-safe:
Concurrent, page-based, dynamically-sized linear data structure with O(1) random access and STL-compl...
Definition PagedArray.h:144
size_t push_back_unsafe(const ValueType &value)
Definition PagedArray.h:193
The fastest way (by far) to insert elements is by means of a PagedArray::ValueBuffer:
for (
size_t i=0; i<100000; ++i) buffer.
push_back(i);
buffer.flush();
void push_back(const ValueT &v)
Add a value to the buffer and increment the size.
Definition PagedArray.h:558
ValueBuffer getBuffer()
Definition PagedArray.h:180
or
{
for (
size_t i=0; i<100000; ++i) buffer.
push_back(i);
}
or with TBB task-based multi-threading:
tbb::parallel_for(
tbb::blocked_range<size_t>(0, 10000, array.
pageSize()),
[&array](const tbb::blocked_range<size_t>& range) {
auto buffer = array.getBuffer();
for (size_t i=range.begin(); i!=range.end(); ++i) buffer.push_back(i);
}
);
static size_t pageSize()
Return the number of elements per memory page.
Definition PagedArray.h:334
or with TBB thread-local storage for even better performance (due to fewer concurrent instantiations of partially full ValueBuffers)
tbb::enumerable_thread_specific<PagedArray<size_t>::ValueBuffer>
pool(exemplar);
tbb::parallel_for(
tbb::blocked_range<size_t>(0, 10000, array.
pageSize()),
[&pool](const tbb::blocked_range<size_t>& range) {
auto &buffer = pool.local();
for (size_t i=range.begin(); i!=range.end(); ++i) buffer.push_back(i);
}
);
for (auto i=pool.begin(); i!=pool.end(); ++i) i->flush();
This technique generally outperforms PagedArray::push_back_unsafe, std::vector::push_back, std::deque::push_back and even tbb::concurrent_vector::push_back. Additionally it is thread-safe as long as each thread has it's own instance of a PagedArray::ValueBuffer. The only disadvantage is the ordering of the elements is undefined if multiple instance of a PagedArray::ValueBuffer are employed. This is typically the case in the context of multi-threading, where the ordering of inserts are undefined anyway. Note that a local scope can be used to guarentee that the ValueBuffer has inserted all its elements by the time the scope ends. Alternatively the ValueBuffer can be explicitly flushed by calling ValueBuffer::flush.
The third way to insert elements is to resize the container and use random access, e.g.
for (int i=0; i<100000; ++i) array[i] = i;
void resize(size_t size)
Resize this array to the specified size.
Definition PagedArray.h:288
or in terms of the random access iterator
for (
auto i=array.
begin(); i!=array.
end(); ++i) *i = i.pos();
Iterator begin()
Return a non-const iterator pointing to the first element.
Definition PagedArray.h:368
Iterator end()
Return a non-const iterator pointing to the past-the-last element.
Definition PagedArray.h:375
While this approach is both fast and thread-safe it suffers from the major disadvantage that the problem size, i.e. number of elements, needs to be known in advance. If that's the case you might as well consider using std::vector or a raw c-style array! In other words the PagedArray is most useful in the context of applications that involve multi-threading of dynamically growing linear arrays that require fast random access.