![]() |
Reference documentation for deal.II version 9.6.0
|
Namespaces | |
namespace | distributed |
namespace | fullydistributed |
namespace | internal |
namespace | shared |
Classes | |
class | CellWeights |
class | DistributedTriangulationBase |
struct | ParallelForInteger |
class | TriangulationBase |
Functions | |
template<typename InputIterator , typename OutputIterator , typename Function > | |
void | transform (const InputIterator &begin_in, const InputIterator &end_in, OutputIterator out, const Function &function, const unsigned int grainsize) |
template<typename InputIterator1 , typename InputIterator2 , typename OutputIterator , typename Function > | |
void | transform (const InputIterator1 &begin_in1, const InputIterator1 &end_in1, InputIterator2 in2, OutputIterator out, const Function &function, const unsigned int grainsize) |
template<typename InputIterator1 , typename InputIterator2 , typename InputIterator3 , typename OutputIterator , typename Function > | |
void | transform (const InputIterator1 &begin_in1, const InputIterator1 &end_in1, InputIterator2 in2, InputIterator3 in3, OutputIterator out, const Function &function, const unsigned int grainsize) |
template<typename Iterator , typename Function > | |
void | apply_to_subranges (const Iterator &begin, const std_cxx20::type_identity_t< Iterator > &end, const Function &f, const unsigned int grainsize) |
template<typename ResultType , typename Iterator , typename Function > | |
ResultType | accumulate_from_subranges (const Function &f, const Iterator &begin, const std_cxx20::type_identity_t< Iterator > &end, const unsigned int grainsize) |
A namespace in which we define classes and algorithms that deal with running in parallel on shared memory machines when deal.II is configured to use multiple threads (see Parallel computing with multiple processors accessing shared memory), as well as running things in parallel on distributed memory machines (see Parallel computing with multiple processors using distributed memory).
void parallel::transform | ( | const InputIterator & | begin_in, |
const InputIterator & | end_in, | ||
OutputIterator | out, | ||
const Function & | function, | ||
const unsigned int | grainsize ) |
An algorithm that performs the action <code>*out++ = function(*in++)</code> where the <code>in</code> iterator ranges over the given input range. This algorithm does pretty much what std::transform does. The difference is that the function can run in parallel when deal.II is configured to use multiple threads. If running in parallel, the iterator range is split into several chunks that are each packaged up as a task and given to the Threading Building Blocks scheduler to work on as compute resources are available. The function returns once all chunks have been worked on. The last argument denotes the minimum number of elements of the iterator range per task; the number must be large enough to amortize the startup cost of new tasks, and small enough to ensure that tasks can be reasonably load balanced. For a discussion of the kind of problems to which this function is applicable, see the @ref threads "Parallel computing with multiple processors" topic. @note This class, function, or variable is a template, and it can only be instantiated if the following condition is true: @code (std::invocable<Function,
decltype(*std::declval<InputIterator>())> && std::assignable_from<decltype(*std::declval<OutputIterator>()), std::invoke_result_t<Function, decltype(*std::declval<InputIterator>())>>) If your compiler supports the C++20 standard, then this constraint will be enforced by a C++20 requires clause.
Definition at line 165 of file parallel.h.
void parallel::transform | ( | const InputIterator1 & | begin_in1, |
const InputIterator1 & | end_in1, | ||
InputIterator2 | in2, | ||
OutputIterator | out, | ||
const Function & | function, | ||
const unsigned int | grainsize ) |
An algorithm that performs the action <code>*out++ = function(*in1++, *in2++)</code> where the <code>in1</code> iterator ranges over the given input range, using the parallel for operator of tbb. This algorithm does pretty much what std::transform does. The difference is that the function can run in parallel when deal.II is configured to use multiple threads. If running in parallel, the iterator range is split into several chunks that are each packaged up as a task and given to the Threading Building Blocks scheduler to work on as compute resources are available. The function returns once all chunks have been worked on. The last argument denotes the minimum number of elements of the iterator range per task; the number must be large enough to amortize the startup cost of new tasks, and small enough to ensure that tasks can be reasonably load balanced. For a discussion of the kind of problems to which this function is applicable, see the @ref threads "Parallel computing with multiple processors" topic. @note This class, function, or variable is a template, and it can only be instantiated if the following condition is true: @code (std::invocable<Function,
decltype(*std::declval<InputIterator1>()), decltype(*std::declval<InputIterator2>())> && std::assignable_from<decltype(*std::declval<OutputIterator>()), std::invoke_result_t<Function, decltype(*std::declval<InputIterator1>()), decltype(*std::declval<InputIterator2>())>>) If your compiler supports the C++20 standard, then this constraint will be enforced by a C++20 requires clause.
Definition at line 240 of file parallel.h.
void parallel::transform | ( | const InputIterator1 & | begin_in1, |
const InputIterator1 & | end_in1, | ||
InputIterator2 | in2, | ||
InputIterator3 | in3, | ||
OutputIterator | out, | ||
const Function & | function, | ||
const unsigned int | grainsize ) |
An algorithm that performs the action <code>*out++ = function(*in1++, *in2++, *in3++)</code> where the <code>in1</code> iterator ranges over the given input range. This algorithm does pretty much what std::transform does. The difference is that the function can run in parallel when deal.II is configured to use multiple threads. If running in parallel, the iterator range is split into several chunks that are each packaged up as a task and given to the Threading Building Blocks scheduler to work on as compute resources are available. The function returns once all chunks have been worked on. The last argument denotes the minimum number of elements of the iterator range per task; the number must be large enough to amortize the startup cost of new tasks, and small enough to ensure that tasks can be reasonably load balanced. For a discussion of the kind of problems to which this function is applicable, see the @ref threads "Parallel computing with multiple processors" topic. @note This class, function, or variable is a template, and it can only be instantiated if the following condition is true: @code (std::invocable<Function,
decltype(*std::declval<InputIterator1>()), decltype(*std::declval<InputIterator2>()), decltype(*std::declval<InputIterator3>())> && std::assignable_from<decltype(*std::declval<OutputIterator>()), std::invoke_result_t<Function, decltype(*std::declval<InputIterator1>()), decltype(*std::declval<InputIterator2>()), decltype(*std::declval<InputIterator3>())>>) If your compiler supports the C++20 standard, then this constraint will be enforced by a C++20 requires clause.
Definition at line 322 of file parallel.h.
void parallel::apply_to_subranges | ( | const Iterator & | begin, |
const std_cxx20::type_identity_t< Iterator > & | end, | ||
const Function & | f, | ||
const unsigned int | grainsize ) |
This function applies the given function argument f
to all elements in the range [begin,end)
and may do so in parallel. An example of its use is given in step-69.
However, in many cases it is not efficient to call a function on each element, so this function calls the given function object on sub-ranges. In other words: if the given range [begin,end)
is smaller than grainsize or if multithreading is not enabled, then we call f(begin,end)
; otherwise, we may execute, possibly in parallel, a sequence of calls f(b,e)
where [b,e)
are subintervals of [begin,end)
and the collection of calls we do to f(.,.)
will happen on disjoint subintervals that collectively cover the original interval [begin,end)
.
Oftentimes, the called function will of course have to get additional information, such as the object to work on for a given value of the iterator argument. This can be achieved by binding certain arguments. For example, here is an implementation of a matrix-vector multiplication \(y=Ax\) for a full matrix \(A\) and vectors \(x,y\):
Note how we use the lambda function to convert mat_vec_on_subranges
from a function that takes 5 arguments to one taking 2 by binding the remaining arguments. The resulting function object requires only two arguments, begin_row
and end_row
, with all other arguments fixed.
The code, if in single-thread mode, will call mat_vec_on_subranges
on the entire range [0,n_rows)
exactly once. In multi-threaded mode, however, it may be called multiple times on subranges of this interval, possibly allowing more than one CPU core to take care of part of the work.
The grainsize
argument (50 in the example above) makes sure that subranges do not become too small, to avoid spending more time on scheduling subranges to CPU resources than on doing actual work.
For a discussion of the kind of problems to which this function is applicable, see also the Parallel computing with multiple processors topic.
Definition at line 452 of file parallel.h.
ResultType parallel::accumulate_from_subranges | ( | const Function & | f, |
const Iterator & | begin, | ||
const std_cxx20::type_identity_t< Iterator > & | end, | ||
const unsigned int | grainsize ) |
This function works a lot like the apply_to_subranges() function, but it allows to accumulate numerical results computed on each subrange into one number. The type of this number is given by the ResultType
template argument that needs to be explicitly specified, and results are added up (i.e., the reduction of results from subranges happens by adding up these results).
An example of use of this function is to compute the value of the expression \(x^T A x\) for a square matrix \(A\) and a vector \(x\). The sum over rows can be parallelized and the whole code might look like this:
Here, mat_norm_sqr_on_subranges
is called on the entire range [0,A.n_rows())
if this range is less than the minimum grainsize (above chosen as 50) or if deal.II is configured to not use multithreading. Otherwise, it may be called on subsets of the given range, with results from the individual subranges accumulated internally.
For a discussion of the kind of problems to which this function is applicable, see also the Parallel computing with multiple processors topic.
Definition at line 608 of file parallel.h.