fix collision particle invalid memory

Fix changes introduced in #4745 to avoid invalid memory access.
Avoid using void pointer as place holder for the later used frame list.

include/picongpu/particles/collision/InterCollision.hpp

@@ -127,8 +127,8 @@ namespace picongpu::particles::collision
 
  PMACC_SMEM(worker, nppc, memory::Array<uint32_t, numCellsPerSuperCell>);
 
- PMACC_SMEM(worker, parCellList0, detail::ListEntry<numCellsPerSuperCell>);
- PMACC_SMEM(worker, parCellList1, detail::ListEntry<numCellsPerSuperCell>);
+ PMACC_SMEM(worker, parCellList0, detail::ListEntry<T_ParBox0, numCellsPerSuperCell>);
+ PMACC_SMEM(worker, parCellList1, detail::ListEntry<T_ParBox1, numCellsPerSuperCell>);
  PMACC_SMEM(worker, densityArray0, memory::Array<float_X, numCellsPerSuperCell>);
  PMACC_SMEM(worker, densityArray1, memory::Array<float_X, numCellsPerSuperCell>);
 
@@ -187,8 +187,8 @@ namespace picongpu::particles::collision
  superCellIdx,
  densityArray0[linearIdx],
  densityArray1[linearIdx],
- parCellList0.template getParticlesAccessor<FramePtr0>(linearIdx),
- parCellList1.template getParticlesAccessor<FramePtr1>(linearIdx),
+ parCellList0.getParticlesAccessor(linearIdx),
+ parCellList1.getParticlesAccessor(linearIdx),
  linearIdx);
  });
 

include/picongpu/particles/collision/IntraCollision.hpp

@@ -105,7 +105,7 @@ namespace picongpu::particles::collision
  constexpr uint32_t numCellsPerSuperCell = pmacc::math::CT::volume<SuperCellSize>::type::value;
 
  PMACC_SMEM(worker, nppc, memory::Array<uint32_t, numCellsPerSuperCell>);
- PMACC_SMEM(worker, parCellList, detail::ListEntry<numCellsPerSuperCell>);
+ PMACC_SMEM(worker, parCellList, detail::ListEntry<T_ParBox, numCellsPerSuperCell>);
  PMACC_SMEM(worker, densityArray, memory::Array<float_X, numCellsPerSuperCell>);
 
  constexpr bool ifAverageLog = !std::is_same<T_SumCoulombLogBox, std::nullptr_t>::value;
@@ -148,7 +148,7 @@ namespace picongpu::particles::collision
  auto collisionFunctorCtx = forEachCell(
  [&](int32_t const idx)
  {
- auto parAccess = parCellList.template getParticlesAccessor<FramePtr>(idx);
+ auto parAccess = parCellList.getParticlesAccessor(idx);
  uint32_t const sizeAll = parAccess.size();
  uint32_t potentialPartners = sizeAll - 1u + sizeAll % 2u;
  auto collisionFunctor = srcCollisionFunctor(

include/picongpu/particles/collision/detail/ListEntry.hpp

@@ -32,52 +32,12 @@ namespace picongpu::particles::collision
 {
  namespace detail
  {
- /** Access all particles in a cell.
- *
- * An accessor provides access to all particles within a cell (NOT supercell).
- *
- * @tparam T_FramePtrType frame pointer type
- */
- template<typename T_FramePtrType>
- struct ParticleAccessor
- {
- T_FramePtrType* m_framePtrList = nullptr;
- uint32_t* m_parIdxList = nullptr;
- uint32_t m_numPars = 0u;
-
- static constexpr uint32_t frameSize = T_FramePtrType::type::frameSize;
-
- DINLINE ParticleAccessor(uint32_t* parIdxList, uint32_t const numParticles, T_FramePtrType* framePtrList)
- : m_framePtrList(framePtrList)
- , m_parIdxList(parIdxList)
- , m_numPars(numParticles)
- {
- }
-
- /** Number of particles within the cell
- */
- DINLINE uint32_t size() const
- {
- return m_numPars;
- }
-
- /** get particle
- *
- * @param idx index of the particle, range [0;size())
- */
- DINLINE auto operator[](uint32_t idx) const
- {
- const uint32_t inSuperCellIdx = m_parIdxList[idx];
- return m_framePtrList[inSuperCellIdx / frameSize][inSuperCellIdx % frameSize];
- }
- };
-
  /* Storage for particle.
  *
  * Storage for particles of collision eligible macro particles. It comes with a simple
  * shuffling algorithm.
  */
- template<uint32_t T_numElem>
+ template<typename T_ParticlesBox, uint32_t T_numElem>
  struct ListEntry
  {
  private:
@@ -87,24 +47,58 @@ namespace picongpu::particles::collision
  //! number of particles per cell
  memory::Array<uint32_t, T_numElem> numParticles;
 
- /** Frame pointer array.
+ using FramePtrType = typename T_ParticlesBox::FramePtr;
+
+ /**! pointer to an array of frames for the selected supercell index
  *
- * A Frame pointer contains only a pointer therefore storing data as void* is allowed (but not
- * nice). This keeps the ListEntry signature equal for all species.
+ * Since ListEntry is designed to be located in shared memory the list located in global memory will be
+ * shared by all workers.
  */
- void** framePtr = nullptr;
+ FramePtrType* framePtr = nullptr;
 
  public:
- DINLINE uint32_t& size(uint32_t cellIdx)
+ /** Access all particles in a cell.
+ *
+ * An accessor provides access to all particles within a cell (NOT supercell).
+ *
+ */
+
+ struct ParticleAccessor
  {
- return numParticles[cellIdx];
- }
+ FramePtrType* m_framePtrList = nullptr;
+ uint32_t* m_parIdxList = nullptr;
+ uint32_t m_numPars = 0u;
+
+ static constexpr uint32_t frameSize = FramePtrType::type::frameSize;
+
+ DINLINE ParticleAccessor(uint32_t* parIdxList, uint32_t const numParticles, FramePtrType* framePtrList)
+ : m_framePtrList(framePtrList)
+ , m_parIdxList(parIdxList)
+ , m_numPars(numParticles)
+ {
+ }
+
+ /** Number of particles within the cell
+ */
+ DINLINE uint32_t size() const
+ {
+ return m_numPars;
+ }
+
+ /** get particle
+ *
+ * @param idx index of the particle, range [0;size())
+ */
+ DINLINE auto operator[](uint32_t idx) const
+ {
+ const uint32_t inSuperCellIdx = m_parIdxList[idx];
+ return m_framePtrList[inSuperCellIdx / frameSize][inSuperCellIdx % frameSize];
+ }
+ };
 
- template<typename T_FramePtrType>
- DINLINE T_FramePtrType& frameData(uint32_t frameIdx)
+ DINLINE uint32_t& size(uint32_t cellIdx)
  {
- static_assert(sizeof(void*) == sizeof(T_FramePtrType));
- return reinterpret_cast<T_FramePtrType*>(framePtr)[frameIdx];
+ return numParticles[cellIdx];
  }
 
  /** Get particle index array.
@@ -127,15 +121,11 @@ namespace picongpu::particles::collision
  * @param superCellIdx supercell index, relative to the guard origin
  * @param numParArray array with number of particles for each cell
  */
- template<
- typename T_Worker,
- typename T_DeviceHeapHandle,
- typename T_ParticlesBox,
- typename T_NumParticlesArray>
+ template<typename T_Worker, typename T_DeviceHeapHandle, typename T_NumParticlesArray>
  DINLINE void init(
  T_Worker const& worker,
  T_DeviceHeapHandle& deviceHeapHandle,
- T_ParticlesBox pb,
+ T_ParticlesBox& pb,
  DataSpace<simDim> superCellIdx,
  T_NumParticlesArray& numParArray)
  {
@@ -152,14 +142,14 @@ namespace picongpu::particles::collision
 
  // Chunk size in bytes based on the typical initial number of frames within a supercell.
  constexpr uint32_t frameListChunkSize = cellListChunkSize * framePtrBytes;
- framePtr = (void**)
+ framePtr = (FramePtrType*)
  allocMem<frameListChunkSize>(worker, numFrames * framePtrBytes, deviceHeapHandle);
 
  auto frame = pb.getFirstFrame(superCellIdx);
  uint32_t frameId = 0u;
  while(frame.isValid())
  {
- frameData<decltype(frame)>(frameId) = frame;
+ framePtr[frameId] = frame;
  frame = pb.getNextFrame(frame);
  ++frameId;
  }
@@ -265,13 +255,9 @@ namespace picongpu::particles::collision
  * @param cellIdx cell index within the supercell, range [0, number of cells in supercell)
  * @return accessor to access particles via index
  */
- template<typename T_FramePtrType>
  DINLINE auto getParticlesAccessor(uint32_t cellIdx)
  {
- return ParticleAccessor<T_FramePtrType>(
- particleIds(cellIdx),
- size(cellIdx),
- reinterpret_cast<T_FramePtrType*>(framePtr));
+ return ParticleAccessor(particleIds(cellIdx), size(cellIdx), framePtr);
  }
 
  private:
@@ -375,20 +361,19 @@ namespace picongpu::particles::collision
  template<
  typename T_Worker,
  typename T_ForEachCell,
- typename T_ParticlesBox,
  typename T_DeviceHeapHandle,
  typename T_Array,
  typename T_Filter,
+ typename T_ParticlesBox,
  uint32_t T_numListEntryCells>
  DINLINE void prepareList(
  T_Worker const& worker,
  T_ForEachCell forEachCell,
- T_ParticlesBox pb,
+ T_ParticlesBox& pb,
  DataSpace<simDim> superCellIdx,
  T_DeviceHeapHandle deviceHeapHandle,
- ListEntry<T_numListEntryCells>& listEntrys,
+ ListEntry<T_ParticlesBox, T_numListEntryCells>& listEntrys,
  T_Array& nppc,
-
  T_Filter filter)
  {
  // Initialize nppc with zeros.

include/picongpu/particles/collision/detail/cellDensity.hpp

@@ -42,7 +42,7 @@ namespace picongpu::particles::collision::detail
  forEachCell(
  [&](uint32_t const linearIdx)
  {
- auto parAccess = parCellList.template getParticlesAccessor<T_FramePtr>(linearIdx);
+ auto parAccess = parCellList.getParticlesAccessor(linearIdx);
  uint32_t const numParInCell = parAccess.size();
  float_X density(0.0);
  for(uint32_t partIdx = 0; partIdx < numParInCell; partIdx++)