X-Git-Url: http://git.rrze.uni-erlangen.de/gitweb/?p=LbmBenchmarkKernelsPublic.git;a=blobdiff_plain;f=src%2FBenchKernelD3Q19ListPullSplitNt.c;fp=src%2FBenchKernelD3Q19ListPullSplitNt.c;h=0132dc9058f26d7f18f153e8a0c11f37d38efa74;hp=dfab54afbdc574c3936d146e28e9925e479a1088;hb=e3f82424829ebb623343ce0092238f83b4a1b8c2;hpb=ecf590ae9bb13ba2b2f01c3bf7a53056a8b1467b diff --git a/src/BenchKernelD3Q19ListPullSplitNt.c b/src/BenchKernelD3Q19ListPullSplitNt.c index dfab54a..0132dc9 100644 --- a/src/BenchKernelD3Q19ListPullSplitNt.c +++ b/src/BenchKernelD3Q19ListPullSplitNt.c @@ -29,6 +29,7 @@ #include "Memory.h" #include "Vtk.h" #include "Vector.h" +#include "LikwidIf.h" #include #include @@ -118,6 +119,8 @@ void FNAME(KernelPullSplitNt1S)(LatticeDesc * ld, KernelData * kernelData, CaseD KernelStatistics(kd, ld, cd, 0); #endif + + X_LIKWID_START("list-pull-split-nt-1s"); #ifdef _OPENMP #pragma omp parallel default(none) \ shared(nFluid, nCells, kd, kdl, adjList, src, dst, \ @@ -184,6 +187,7 @@ void FNAME(KernelPullSplitNt1S)(LatticeDesc * ld, KernelData * kernelData, CaseD for(int iter = 0; iter < maxIterations; ++iter) { + #if 1 #define INDEX_START blIndexStart #define INDEX_STOP blIndexVec @@ -201,6 +205,8 @@ void FNAME(KernelPullSplitNt1S)(LatticeDesc * ld, KernelData * kernelData, CaseD #define INDEX_STOP blIndexStop #include "BenchKernelD3Q19ListPullSplitNt1SScalar.h" #endif + + #pragma omp barrier #pragma omp single @@ -235,6 +241,9 @@ void FNAME(KernelPullSplitNt1S)(LatticeDesc * ld, KernelData * kernelData, CaseD MemFree((void **)&tmpArray); } + + X_LIKWID_STOP("list-pull-split-nt-1s"); + #ifdef VTK_OUTPUT if (cd->VtkOutput) { kd->PdfsActive = src; @@ -321,6 +330,10 @@ void FNAME(KernelPullSplitNt2S)(LatticeDesc * ld, KernelData * kernelData, CaseD KernelStatistics(kd, ld, cd, 0); #endif + + X_LIKWID_START("list-pull-split-nt-2s"); + + #ifdef _OPENMP #pragma omp parallel default(none) \ shared(nFluid, nCells, kd, kdl, adjList, src, dst, \ @@ -406,6 +419,7 @@ void FNAME(KernelPullSplitNt2S)(LatticeDesc * ld, KernelData * kernelData, CaseD #endif #pragma omp barrier + #pragma omp single { #ifdef VERIFICATION @@ -438,6 +452,8 @@ void FNAME(KernelPullSplitNt2S)(LatticeDesc * ld, KernelData * kernelData, CaseD MemFree((void **)&tmpArray); } + X_LIKWID_STOP("list-pull-split-nt-2s"); + #ifdef VTK_OUTPUT if (cd->VtkOutput) { kd->PdfsActive = src;