|
| template<class SolverType , int NumberOfVolumesPerAxisInPatch, int HaloSize, int NumberOfUnknowns, int NumberOfAuxiliaryVariables, bool EvaluateFlux, bool EvaluateNonconservativeProduct, bool EvaluateSource, bool EvaluateMaximumEigenvalueAfterTimeStep, class TempDataEnumeratorType > |
| KeywordToAvoidDuplicateSymbolsForInlinedFunctions void | timeStepWithRusanovPatchwiseHeapStateless (int targetDevice, CellData< double, double > &patchData, tarch::timing::Measurement &measurement) InlineMethod |
| | The name patchwise is a little bit irritating in a GPU context.
|
| |
| template<class SolverType , int NumberOfVolumesPerAxisInPatch, int HaloSize, int NumberOfUnknowns, int NumberOfAuxiliaryVariables, bool EvaluateFlux, bool EvaluateNonconservativeProduct, bool EvaluateSource, bool EvaluateMaximumEigenvalueAfterTimeStep, class TempDataEnumeratorType > |
| KeywordToAvoidDuplicateSymbolsForInlinedFunctions void | timeStepWithRusanovPatchwiseHeapStateless (int targetDevice, CellData< double, double > &patchData) InlineMethod |
| |
| template<class SolverType , int NumberOfVolumesPerAxisInPatch, int HaloSize, int NumberOfUnknowns, int NumberOfAuxiliaryVariables, bool EvaluateFlux, bool EvaluateNonconservativeProduct, bool EvaluateSource, bool EvaluateMaximumEigenvalueAfterTimeStep, class TempDataEnumeratorType > |
| KeywordToAvoidDuplicateSymbolsForInlinedFunctions void | timeStepWithRusanovBatchedHeapStateless (int targetDevice, CellData< double, double > &patchData, tarch::timing::Measurement &measurement) InlineMethod |
| |
| template<class SolverType , int NumberOfVolumesPerAxisInPatch, int HaloSize, int NumberOfUnknowns, int NumberOfAuxiliaryVariables, bool EvaluateFlux, bool EvaluateNonconservativeProduct, bool EvaluateSource, bool EvaluateMaximumEigenvalueAfterTimeStep, class TempDataEnumeratorType > |
| KeywordToAvoidDuplicateSymbolsForInlinedFunctions void | timeStepWithRusanovBatchedHeapStateless (int targetDevice, CellData< double, double > &patchData) InlineMethod |
| |
| template<class SolverType , int NumberOfVolumesPerAxisInPatch, int HaloSize, int NumberOfUnknowns, int NumberOfAuxiliaryVariables, bool EvaluateFlux, bool EvaluateNonconservativeProduct, bool EvaluateSource, bool EvaluateMaximumEigenvalueAfterTimeStep, class TempDataEnumeratorType > |
| KeywordToAvoidDuplicateSymbolsForInlinedFunctions void | timeStepWithRusanovPatchwiseUSMStateless (int targetDevice, CellData< double, double > &patchData, tarch::timing::Measurement &measurement) InlineMethod |
| |
| template<class SolverType , int NumberOfVolumesPerAxisInPatch, int HaloSize, int NumberOfUnknowns, int NumberOfAuxiliaryVariables, bool EvaluateFlux, bool EvaluateNonconservativeProduct, bool EvaluateSource, bool EvaluateMaximumEigenvalueAfterTimeStep, class TempDataEnumeratorType > |
| KeywordToAvoidDuplicateSymbolsForInlinedFunctions void | timeStepWithRusanovPatchwiseUSMStateless (int targetDevice, CellData< double, double > &patchData) InlineMethod |
| |
| template<class SolverType , int NumberOfVolumesPerAxisInPatch, int HaloSize, int NumberOfUnknowns, int NumberOfAuxiliaryVariables, bool EvaluateFlux, bool EvaluateNonconservativeProduct, bool EvaluateSource, bool EvaluateMaximumEigenvalueAfterTimeStep, class TempDataEnumeratorType > |
| KeywordToAvoidDuplicateSymbolsForInlinedFunctions void | timeStepWithRusanovBatchedUSMStateless (int targetDevice, CellData< double, double > &patchData, tarch::timing::Measurement &measurement) InlineMethod |
| |
| template<class SolverType , int NumberOfVolumesPerAxisInPatch, int HaloSize, int NumberOfUnknowns, int NumberOfAuxiliaryVariables, bool EvaluateFlux, bool EvaluateNonconservativeProduct, bool EvaluateSource, bool EvaluateMaximumEigenvalueAfterTimeStep, class TempDataEnumeratorType > |
| KeywordToAvoidDuplicateSymbolsForInlinedFunctions void | timeStepWithRusanovBatchedUSMStateless (int targetDevice, CellData< double, double > &patchData) InlineMethod |
| |
| template<class SolverType , int NumberOfVolumesPerAxisInPatch, int HaloSize, int NumberOfUnknowns, int NumberOfAuxiliaryVariables, bool EvaluateFlux, bool EvaluateNonconservativeProduct, bool EvaluateSource, bool EvaluateMaximumEigenvalueAfterTimeStep, class TempDataEnumeratorType > |
| KeywordToAvoidDuplicateSymbolsForInlinedFunctions void | timeStepWithRusanovVolumewiseStateless (int targetDevice, CellData< double, double > &patchData, tarch::timing::Measurement &measurement) InlineMethod |
| |
| template<class SolverType , int NumberOfVolumesPerAxisInPatch, int HaloSize, int NumberOfUnknowns, int NumberOfAuxiliaryVariables, bool EvaluateFlux, bool EvaluateNonconservativeProduct, bool EvaluateSource, bool EvaluateMaximumEigenvalueAfterTimeStep, class TempDataEnumeratorType > |
| KeywordToAvoidDuplicateSymbolsForInlinedFunctions void | timeStepWithRusanovVolumewiseStateless (int targetDevice, CellData< double, double > &patchData) InlineMethod |
| |
template<class SolverType ,
int NumberOfVolumesPerAxisInPatch,
int HaloSize,
int NumberOfUnknowns,
int NumberOfAuxiliaryVariables,
bool EvaluateFlux,
bool EvaluateNonconservativeProduct,
bool EvaluateSource,
bool EvaluateMaximumEigenvalueAfterTimeStep, class TempDataEnumeratorType >
The name patchwise is a little bit irritating in a GPU context.
What it means is that the code runs logically through the patches one by one and issues all kernel on one patch before it continues with the next one. Obviously, the patches are all independent of each other. Therefore, we can deploy this sequence of patch updates to a distributed for over the team.
Memory management
If you run this routine without GPU offloading, we place all temporary data on the heap. If GPUs are to be used, then we create pointers to the heap, but we map the data explicitly to GPU memory. As we use explicit allocs and frees, the map clauses basically replicate the heap behaviour on the accelerator.
We assign the host pointers null, as the actual temporary data is to be created on the GPU. Such a 'create a pointer which is a nullptr' ensures that the code still compiles on the host even though it is semantically broken as the temporary data is not allocated.
Patch data transfer
All the data within the CellData object is trivially copyable besides the actual patch sizes and positions. Trivially copyable means that they are plain double arrays. That is, we can map them onto the device. The actual input and output data is stored as pointer to pointers. Here, we have to loop over the entries and copy each individual pointer over separately.
The positions and sizes of the patches are in principle trivially copyable as well, as they simply host arrays of known sizes. However, it is quite cumbersome to tell OpenMP that this stuff can just be copied. It also might become annoying the compiler decided to add padding bytes. So we copy those guys manually into one large buffer of double values, and then we reconstruct the objects on the device. In this particular context, the resulting overhead is acceptable, as this meta data is not extraordinary large.
If you copy the pointers to pointers from the host to the GPU, you have to be careful. It makes no sense to copy the pointers within the pointer array to the GPU, as these are host pointers. So first deploy the actual data to the GPU per patch. Then we grab the GPU pointer address to which the memory region has been mapped to. We store this data in mappedPointersToQIn. We transfer mappedPointersToQIn to the GPU and then reconstruct there the pointer-to-pointer map that we need within the kernels.
Referenced by runBenchmarks().
1.10.0 
