inline void AllocatedChunk::removeReference( WD const &wd ) {
//ensure(_refs > 0, "invalid removeReference, chunk has 0 references!");
if ( _refs == 0 ) {
*myThread->_file << " removeReference ON A CHUNK WITH 0 REFS!!!" << std::endl;
}
_refs--;
_refWdId[&wd]--;
if ( _refWdId[&wd] == 0 ) {
_refLoc[wd.getId()].clear();
}
//std::cerr << "del ref to chunk "<< (void*)this << " " << _refs.value() << std::endl;
//if ( _refs == (unsigned int) -1 ) {
// std::cerr << "overflow at references chunk "<< (void*)this << std::endl; sys.printBt();
//} else if ( _refs == 0 ) {
// std::cerr << "zeroed at references chunk "<< (void*)this << std::endl;
//}
}
void MemCacheCopy::generateInOps( BaseAddressSpaceInOps &ops, bool input, bool output, WD const &wd, unsigned int copyIdx ) {
//NANOS_INSTRUMENT( InstrumentState inst4(NANOS_CC_CDIN_OP_GEN); );
_reg.key->lockObject();
if ( input && output ) {
//re read version, in case of this being a commutative or concurrent access
if ( _reg.getVersion() > _version ) {
*myThread->_file << "[!!!] WARNING: concurrent or commutative detected, wd " << wd.getId() << " " << (wd.getDescription()!=NULL?wd.getDescription():"[no desc]") << " index " << copyIdx << " _reg.getVersion() " << _reg.getVersion() << " _version " << _version << std::endl;
_version = _reg.getVersion();
}
}
if ( ops.getPE()->getMemorySpaceId() != 0 ) {
/* CACHE ACCESS */
if ( input ) {
if ( _policy == RegionCache::FPGA ) {
_chunk->copyRegionFromHost( ops, _reg.id, _version, wd, copyIdx );
} else {
_chunk->NEWaddReadRegion2( ops, _reg.id, _version, _locations, wd, copyIdx );
}
} else if ( output ) {
_chunk->NEWaddWriteRegion( _reg.id, _version, &wd, copyIdx );
} else {
fatal("invalid path");
}
} else {
/* HOST ACCESS */
if ( input ) {
ops.copyInputData( *this, wd, copyIdx );
}
}
//NANOS_INSTRUMENT( inst4.close(); );
_reg.key->unlockObject();
}
bool GPUThread::inlineWorkDependent ( WD &wd )
{
GPUProcessor &myGPU = * ( GPUProcessor * ) myThread->runningOn();
if ( GPUConfig::isOverlappingInputsDefined() ) {
// Wait for the input transfer stream to finish
NANOS_GPU_CREATE_IN_CUDA_RUNTIME_EVENT( NANOS_GPU_CUDA_INPUT_STREAM_SYNC_EVENT );
cudaStreamSynchronize( myGPU.getGPUProcessorInfo()->getInTransferStream() );
NANOS_GPU_CLOSE_IN_CUDA_RUNTIME_EVENT;
// Erase the wait input list and synchronize it with cache
myGPU.getInTransferList()->clearMemoryTransfers();
myGPU.freeInputPinnedMemory();
}
// Check if someone is waiting for our data
myGPU.getOutTransferList()->clearRequestedMemoryTransfers();
// We wait for wd inputs, but as we have just waited for them, we could skip this step
wd.start( WD::IsNotAUserLevelThread );
GPUDD &dd = ( GPUDD & ) wd.getActiveDevice();
NANOS_INSTRUMENT ( InstrumentStateAndBurst inst1( "user-code", wd.getId(), NANOS_RUNNING ) );
( dd.getWorkFct() )( wd.getData() );
if ( !GPUConfig::isOverlappingOutputsDefined() && !GPUConfig::isOverlappingInputsDefined() ) {
// Wait for the GPU kernel to finish
NANOS_GPU_CREATE_IN_CUDA_RUNTIME_EVENT( NANOS_GPU_CUDA_DEVICE_SYNC_EVENT );
#ifdef NANOS_GPU_USE_CUDA32
cudaThreadSynchronize();
#else
cudaDeviceSynchronize();
#endif
NANOS_GPU_CLOSE_IN_CUDA_RUNTIME_EVENT;
// Normally this instrumentation code is inserted by the compiler in the task outline.
// But because the kernel call is asynchronous for GPUs we need to raise them manually here
// when we know the kernel has really finished
NANOS_INSTRUMENT ( raiseWDClosingEvents() );
// Copy out results from tasks executed previously
// Do it always, as another GPU may be waiting for results
myGPU.getOutTransferList()->executeMemoryTransfers();
}
else {
myGPU.getOutTransferList()->executeMemoryTransfers();
}
if ( GPUConfig::isPrefetchingDefined() ) {
WD * last = &wd;
while ( canPrefetch() ) {
// Get next task in order to prefetch data to device memory
WD *next = Scheduler::prefetch( ( nanos::BaseThread * ) this, *last );
if ( next != NULL ) {
next->init();
addNextWD( next );
last = next;
} else {
break;
}
}
}
if ( GPUConfig::isOverlappingOutputsDefined() ) {
NANOS_GPU_CREATE_IN_CUDA_RUNTIME_EVENT( NANOS_GPU_CUDA_OUTPUT_STREAM_SYNC_EVENT );
cudaStreamSynchronize( myGPU.getGPUProcessorInfo()->getOutTransferStream() );
NANOS_GPU_CLOSE_IN_CUDA_RUNTIME_EVENT;
myGPU.freeOutputPinnedMemory();
}
if ( GPUConfig::isOverlappingOutputsDefined() || GPUConfig::isOverlappingInputsDefined() ) {
// Wait for the GPU kernel to finish, if we have not waited before
//cudaThreadSynchronize();
NANOS_GPU_CREATE_IN_CUDA_RUNTIME_EVENT( NANOS_GPU_CUDA_KERNEL_STREAM_SYNC_EVENT );
cudaStreamSynchronize( myGPU.getGPUProcessorInfo()->getKernelExecStream() );
NANOS_GPU_CLOSE_IN_CUDA_RUNTIME_EVENT;
// Normally this instrumentation code is inserted by the compiler in the task outline.
// But because the kernel call is asynchronous for GPUs we need to raise them manually here
// when we know the kernel has really finished
NANOS_INSTRUMENT ( raiseWDClosingEvents() );
}
return true;
}
inline void AllocatedChunk::addReference( WD const &wd, unsigned int loc ) {
_refs++;
_refWdId[&wd]++;
_refLoc[wd.getId()].insert(loc);
//std::cerr << "add ref to chunk "<< (void*)this << " " << _refs.value() << std::endl;
}