1) Reduce the number of calls to malloc by grouping requests to larger blocks. To achieve this goal, the following ways are used:
- Containers of types sequence, list and map are initialized with an instance of NCollection_IncAllocator, at this taking care of the time of life of allocated objects, so that not to occupy huge amount of memory.
- Allocation of several arrays having the same and short life time is changed so that to allocate a buffer array of necessary size and to place arrays in this buffer.
2) In BRepMesh_FastDiscretFace, optimize the function filterParameters so that to avoid excess memory allocations.
3) In NCollection_CellFilter, change declaration of the method Reset to accept array by reference rather than by value.
4) Add Allocator() method in map, sequence and vector collection classes by analogy with list collection.
5) Correct the size of block for IncAllocator for x64 platform. In order free-ed block to be returned to the system its size should be at least 1024K on x64 and 512K on x86. This allows to retain free virtual space almost to the state before algorithm run.
6) Decrease amount of memory zeroed by calloc. For that, reduce theIncrement parameter of the embedded vectors of the classes NCollection_UBTreeFiller and BRepMesh_VertexInspector to default value 256.
7) Avoid computing bounding box when not necessary (if no relative deflection is used)
8) Cycles by wires of face using TopExp_Explorer are converted to use TopoDS_Iterator instead.
9) BRepMesh_FastDiscret::Add optimized to avoid storing sequences of faces and edges
10) The tests "mesh standard_* W7" are corrected to accept the new behavior. Earlier the following error took place:
Not connected mesh inside face 9
{12 13}
Now this error was replaced with another one:
free nodes (in pairs: face / node):
{9 12}
Actually it is not a regression, rather improvement, if we look at the snapshot.
11) Change other test cases to their actual state.
Simple primitives to parallelize loops type "for" and "foreach" were implemented. The primitives encapsulates complete logic for creating and managing parallel context of loops. Moreover the primitives may be a wrapper for some primitives from 3rd-party library - TBB.
To use it is necessary to implement TBB like interface which is based on functors. For example:
Class Functor
{
public:
void operator() ([proccesing instance]) const
{
//...
}
};
In the body of the operator () should be implemented thread-safe logic of computations that can be performed in parallel context. If parallelized loop iterates on the collections with direct access by index (such as Vector, Array), it is more efficient to use the primitive ParallelFor (because it has no critical section).
All parts of OCC code which are using tbb were changed on new primitives.
0024826: Wrapping of parallelisation algorithms
Small fix.
Fix compilation errors on Linux platform
Squeeze compilation warnings on Linux
Fix regressions
Back RemoveFaceAttribute for further reasons
Fix retrieving of polygon by index
Fix applying of location
Test case for issue CR23106
Fix memory leak regression 'test bugs vis bug79' occurred due to incorrect memory cleaning of inherited objects by MMgtRaw::Free through BRepMesh_IEdgeTool;
Replace BRepMesh_PDiscretRoot by pure pointer to BRepMesh_DiscretRoot;
Fix IVtkOCC_ShapeMesher.
Removed CDL declarations; Data collections are replaced by NCollections; Small code refactoring.
Remove definition of BRepMesh class. Code refactoring of BRepMesh_IncrementalMesh.
Function BRepMesh_Write storing BRepMesh_DataStructureOfDelaun to BRep file is added for debug needs.
Static method BRepMesh_GeomTool::IntLinLin has been added to eliminate code duplications in BRepMesh_Dealun and BRepMesh_CircleTool.
BRepMesh_CircleTool simplified method to find circumcircle.
Fix merging conflicts
Remove redundant function
Fix compilation warning on MacOS
Revert changes occurred during rebase
Resolved merging conflicts
Use parallel flag with BRepMesh_FastDiscret
Test cases for issue CR25039_2
BRepMesh_Classifier: Two-pass approach for intersection check with possibility to run it in parallel mode.
First pass - bounding boxes of segments are checked for overlapping;
Second pass - intersection point is calculated in case if overlapping is detected.
Make NCollection_UBTree::ChangeLastNode() exported due to compilation error on Linux platform.
Reason: method does not depend on template parameters, so it should be available.
Revert previous change and try to use another trick for Linux
Fix compilation warning on MacOS: remove redundant constant
Fix regressions: do not consider insignificant loops in case of self intersections on the same wire.
More sugar solution for compilation errors on NCollection_EBTree on Linux
Test cases for issue CR24968
