template<class ArrayT>
| static void Impala::Core::Array::Pattern::PxGatherArrayOFT | ( | ArrayT * | glob, | |
| ArrayT * | loc, | |||
| int | root, | |||
| bool | bdata | |||
| ) | [inline, static] |
Definition at line 833 of file PxDistribution.h.
References _myCPU, _nrCPUs, ArrayBD(), ArrayBH(), ArrayBW(), ArrayCPB(), ArrayH(), ArrayPB(), ArrayW(), GATH_TAG, PxLclArrayCopy(), PxLclDepth(), PxLclHeight(), PxLclStart(), PxLclWidth(), and TO.
Referenced by PxGatherArray().
00834 { 00835 /*** Gather array data using One-level Flat Tree ***/ 00836 00837 int eSize = ArrayT::ElemSize(); 00838 int sSize = eSize*sizeof(typename ArrayT::StorType); 00839 MPI_Datatype elem, blk2d, blk3d; 00840 MPI_Status stat; 00841 MPI_Type_contiguous(sSize, MPI_BYTE, &elem); 00842 00843 if (_myCPU == root) { // receive array data from all other CPUs 00844 PxLclArrayCopy(loc, TO, glob, root, bdata); 00845 for (int partner=0; partner<_nrCPUs; partner++) { 00846 if (partner != _myCPU) { 00847 int xSize = PxLclWidth(partner); 00848 int ySize = PxLclHeight(partner); 00849 int zSize = PxLclDepth(partner); 00850 int offset = PxLclStart(partner,ArrayBW(glob), 00851 ArrayBH(glob), ArrayBD(glob)); 00852 if (bdata) { 00853 xSize += 2*ArrayBW(glob); 00854 ySize += 2*ArrayBH(glob); 00855 zSize += 2*ArrayBD(glob); 00856 offset -= 00857 (ArrayW(glob)*ArrayH(glob)*ArrayBD(glob) + 00858 ArrayW(glob)*ArrayBH(glob) + ArrayBW(glob)); 00859 } 00860 if (xSize != 0 && ySize != 0 && zSize != 0) { 00861 MPI_Type_vector(ySize, xSize, 00862 ArrayW(glob), elem, &blk2d); 00863 MPI_Type_hvector(zSize, 1, 00864 ArrayW(glob)*ArrayH(glob)*sSize, 00865 blk2d, &blk3d); 00866 MPI_Type_commit(&blk3d); 00867 MPI_Recv(ArrayPB(glob) + offset*eSize, 1, blk3d, 00868 partner, GATH_TAG, MPI_COMM_WORLD, &stat); 00869 MPI_Type_free(&blk3d); 00870 MPI_Type_free(&blk2d); 00871 } 00872 } 00873 } 00874 } else { // send array data to root 00875 int xSize = PxLclWidth(_myCPU); 00876 int ySize = PxLclHeight(_myCPU); 00877 int zSize = PxLclDepth(_myCPU); 00878 if (bdata) { 00879 xSize += 2*ArrayBW(glob); 00880 ySize += 2*ArrayBH(glob); 00881 zSize += 2*ArrayBD(glob); 00882 } 00883 if (xSize != 0 && ySize != 0 && zSize != 0) { 00884 MPI_Type_vector(ySize, xSize, ArrayW(loc), elem, &blk2d); 00885 MPI_Type_hvector(zSize, 1, 00886 ArrayW(loc)*ArrayH(loc)*sSize, blk2d, &blk3d); 00887 MPI_Type_commit(&blk3d); 00888 MPI_Send((bdata) ? ArrayPB(loc) : ArrayCPB(loc), 00889 1, blk3d, root, GATH_TAG, MPI_COMM_WORLD); 00890 MPI_Type_free(&blk3d); 00891 MPI_Type_free(&blk2d); 00892 } 00893 } 00894 MPI_Type_free(&elem); 00895 }
Here is the call graph for this function:
