get rid of OpenMP

This commit is contained in:
Matthias Hochsteger 2016-02-22 18:43:51 +01:00
parent da5e5dbdac
commit 1b4f596446
11 changed files with 71 additions and 42 deletions

View File

@ -144,12 +144,6 @@ if(APPLE)
# set(CMAKE_SHARED_LIBRARY_SUFFIX ".so")
endif(APPLE)
#######################################################################
if(NOT INTEL_MIC AND NOT WIN32 AND NOT APPLE)
find_package(OpenMP)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
endif(NOT INTEL_MIC AND NOT WIN32 AND NOT APPLE)
#######################################################################
find_package(ZLIB REQUIRED)
include_directories(${ZLIB_INCLUDE_DIRS})

View File

@ -14,6 +14,7 @@
namespace netgen
{
static mutex block_allocator_mutex;
BlockAllocator :: BlockAllocator (unsigned asize, unsigned ablocks)
: bablocks (0)
@ -36,8 +37,8 @@ namespace netgen
void * BlockAllocator :: Alloc ()
{
void * p;
#pragma omp critical (BlockAllocator)
{
lock_guard<mutex> guard(block_allocator_mutex);
// return new char[size];
if (!freelist)
{
@ -60,8 +61,8 @@ namespace netgen
void BlockAllocator :: Free (void * p)
{
#pragma omp critical (BlockAllocator)
{
lock_guard<mutex> guard(block_allocator_mutex);
if (bablocks.Size())
{
*(void**)p = freelist;

View File

@ -187,6 +187,27 @@ public:
#endif
// Simple ParallelFor function to replace OpenMP
template<typename TFunc>
void ParallelFor( int first, int next, const TFunc & f )
{
int nthreads = thread::hardware_concurrency();
thread * threads = new thread[nthreads];
for (int i=0; i<nthreads; i++)
{
threads[i] = std::thread( [&] ()
{
int myfirst = first + (next-first)*i/nthreads;
int mynext = first + (next-first)*(i+1)/nthreads;
f(myfirst, mynext);
});
}
for (int i=0; i<nthreads; i++)
threads[i].join();
delete [] threads;
}
}
#endif

View File

@ -21,6 +21,8 @@
#include <climits>
#include <algorithm>
#include <memory>
#include <thread>
#include <mutex>
#include <new>
@ -35,9 +37,6 @@
#include <unistd.h> // for usleep (only for parallel)
#endif
#ifdef _OPENMP
#include <omp.h>
#endif
/*

View File

@ -2363,8 +2363,6 @@ void LinkFunction ()
void Ng_TclCmd(string cmd)
{
#pragma omp critical(tcltodo)
{
*(multithread.tcl_todo) += cmd;
}
lock_guard<mutex> guard(tcl_todo_mutex);
*(multithread.tcl_todo) += cmd;
}

View File

@ -68,6 +68,8 @@ namespace netgen
Array<int> tets_in_qualclass;
mutex tcl_todo_mutex;
int h_argc = 0;
char ** h_argv = NULL;

View File

@ -29,6 +29,8 @@ namespace netgen
DLL_HEADER extern Array<int> tets_in_qualclass;
DLL_HEADER extern mutex tcl_todo_mutex;
class multithreadt
{
public:

View File

@ -4,6 +4,8 @@
namespace netgen
{
static mutex buildsearchtree_mutex;
Mesh :: Mesh ()
: surfarea(*this)
{
@ -4101,8 +4103,8 @@ namespace netgen
{
if (elementsearchtreets == GetTimeStamp()) return;
#pragma omp critical (buildsearchtree)
{
std::lock_guard<std::mutex> guard(buildsearchtree_mutex);
if (elementsearchtreets != GetTimeStamp())
{
NgLock lock(mutex);

View File

@ -2138,13 +2138,13 @@ int STLGeometry :: CheckGeometryOverlapping()
}
#pragma omp parallel
{
Array<int> inters;
mutex inters_mutex;
#pragma omp for
for (int i = 1; i <= GetNT(); i++)
{
ParallelFor( 1, GetNT()+1, [&] (int first, int next)
{
for (int i=first; i<next; i++) {
const STLTriangle & tri = GetTriangle(i);
Point<3> tpmin = tri.box.PMin();
@ -2176,7 +2176,7 @@ int STLGeometry :: CheckGeometryOverlapping()
if (IntersectTriangleTriangle (&trip1[0], &trip2[0]))
{
#pragma omp critical
lock_guard<mutex> guard(inters_mutex);
{
oltrigs++;
PrintMessage(5,"Intersecting Triangles: trig ",i," with ",inters.Get(j),"!");
@ -2186,6 +2186,7 @@ int STLGeometry :: CheckGeometryOverlapping()
}
}
}
});
}
PrintMessage(3,"Check overlapping geometry ... ", oltrigs, " triangles overlap");
return oltrigs;

View File

@ -2474,27 +2474,36 @@ namespace netgen
NgProfiler::RegionTimer reg1 (timer1);
int ne = mesh->GetNE();
double hminv = numeric_limits<double>::max();
double hmaxv = -numeric_limits<double>::max();
bool hhasit = false;
#if defined _OPENMP && _OPENMP >= 201107
#pragma omp parallel for reduction (max : hmaxv) reduction (min : hminv) reduction (|| : hhasit)
#endif
for (int i = 0; i < ne; i++)
{
double val;
bool considerElem = GetValue (sol, i, 0.333, 0.333, 0.333, comp, val);
if (considerElem)
{
if (val > hmaxv) hmaxv = val;
if (val < hminv) hminv = val;
hhasit = true;
}
}
minv = min(minv, hminv);
maxv = max(maxv, hmaxv);
hasit |= hhasit;
mutex min_mutex;
mutex max_mutex;
ParallelFor(0, ne, [&] (int first, int next)
{
double minv_local = numeric_limits<double>::max();
double maxv_local = -numeric_limits<double>::max();
for (int i=first; i<next; i++)
{
double val;
bool considerElem = GetValue (sol, i, 0.333, 0.333, 0.333, comp, val);
if (considerElem)
{
if (val > maxv_local) maxv_local = val;
if (val < minv_local) minv_local = val;
hasit = true;
}
}
if(minv_local < minv)
{
lock_guard<mutex> guard(min_mutex);
minv = minv_local;
}
if(maxv_local > maxv)
{
lock_guard<mutex> guard(max_mutex);
maxv = maxv_local;
}
});
}
if (sol->draw_surface)

View File

@ -710,8 +710,8 @@ namespace netgen
strcat (lstring, " 0");
Tcl_SetVar (interp, "::status_tetqualclasses", lstring, 0);
#pragma omp critical(tcltodo)
{
lock_guard<mutex> guard(tcl_todo_mutex);
if (multithread.tcl_todo->length())
{
Tcl_Eval (interp, multithread.tcl_todo->c_str());