MADNESS
0.10.1
|
Declares the macrotaskq
and MacroTaskBase classes.
More...
#include <madness/world/cloud.h>
#include <madness/world/world.h>
#include <madness/mra/macrotaskpartitioner.h>
Go to the source code of this file.
Namespaces | |
madness | |
File holds all helper structures necessary for the CC_Operator and CC2 class. | |
Functions | |
template<typename T > | |
void | madness::gaxpy (const double a, ScalarResult< T > &left, const double b, const T &right, const bool fence=true) |
the result type of a macrotask must implement gaxpy More... | |
template<typename T > | |
std::vector< std::shared_ptr< ScalarResult< T > > > | madness::scalar_result_shared_ptr_vector (World &world, std::size_t n) |
helper function to create a vector of ScalarResult, circumventing problems with the constructors More... | |
Declares the macrotaskq
and MacroTaskBase classes.
A MacroTaskq executes tasks on World objects, e.g. differentiation of a function or other arithmetic. Complex algorithms can be implemented.
The universe world is split into subworlds, each of them executing macrotasks of the task queue. This improves locality and speedups for large number of compute nodes, by reducing communications within worlds.
The user defines a macrotask (an example is found in test_vectormacrotask.cc), the tasks are lightweight and carry only bookkeeping information, actual input and output are stored in a cloud (see cloud.h)
The user-defined macrotask is derived from MacroTaskIntermediate and must implement the run() method. A heterogeneous task queue is possible.
TODO: priority q TODO: task submission from inside task (serialize task instead of replicate) TODO: update documentation TODO: consider serializing task member variables