Revision control
++ // Copyright (c) 2001-2004 NovodeX AG. All rights reserved.
++ // Copyright (c) 2004-2008 AGEIA Technologies, Inc. All rights reserved.
++ // SPDX-FileCopyrightText: Copyright (c) 2008-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
++ // SPDX-License-Identifier: Apache-2.0
++
++ #ifndef PXS_PARTITION_EDGE_H
++ #define PXS_PARTITION_EDGE_H
++
++ // PT: this is a temporary place for code related to PartitionEdge. This seems to be a GPU-only class
++ // but for some reason some CPU bits do need it. Ideally it would be fully contained inside the GPU DLL.
++
++ #include "PxsIslandSim.h"
++ #include "PxcNpWorkUnit.h"
++
++ namespace physx
++ {
++ // PT: TODO: mNextPatch is almost always null so it would make sense to store that cold data somewhere else, e.g:
++ // - use one bit to mark the general case where mNextPatch is null
++ // - store non-null mNextPatch in a hashmap indexed by mUniqueIndex (no need to reserve the full memory for it)
++ //
++ // The only annoying bit is that the mechanism needed to actually walk the linked list (i.e. the hashmap) needs to be
++ // available in processPartitionEdges below, so that's more GPU stuff exposed to the CPU code. But I guess we crossed
++ // that line a while ago when the heads of the LLs moved to the low-level island manager anyway. And in fact we could
++ // put the two in the same structure eventually, like a PartitionEdgeManager.
++ //
++ // In any case the benefit of the change would be a smaller PartitionEdge. Going further, the nodes can be retrieved from
++ // the edge index, so if mNextPatch also disappears then only the unique ID remains, which.... can be derived from
++ // the PartitionEdge address?! So it would be just the "edge index" with some bits encoded in it, just 4 bytes.
++ //
++ // This is per-edge data so we could also merge this with CPUExternalData, which already contains the node indices, and
++ // has an implicit unique index as the index into mEdgeNodeIndices. But maybe we cannot because there can be multiple
++ // PartitionEdge for the same source edge (hence the linked list).
++ //
++ // Another idea would be to store the edge index instead. You would need access to the edge manager in CPU code but no
++ // hashmap or new structure is needed.
++
++ struct PartitionEdge
++ {
++ enum Enum
++ {
++ HAS_INFINITE_MASS0 = (1<<0),
++ HAS_INFINITE_MASS1 = (1<<1),
++ HAS_THRESHOLD = (1<<2),
++ IS_CONTACT = (1<<3),
++ SPECIAL_HANDLED = (1<<4),
++
++ NB_BITS = 5
++ };
++
++ PxNodeIndex mNode0; //! The node index for node 0. Can be obtained from the edge index alternatively
++ PxNodeIndex mNode1; //! The node index for node 1. Can be obtained from the edge index alternatively
++ PartitionEdge* mNextPatch; //! for the contact manager has more than 1 patch, we have next patch's edge and previous patch's edge to connect to this edge
++ private:
++ IG::EdgeIndex mEdgeIndex; //! The edge index into the island manager. Used to identify the contact manager/constraint
++ public:
++ PxU32 mUniqueIndex; //! a unique ID for this edge
++
++ PX_FORCE_INLINE IG::EdgeIndex getEdgeIndex() const { return mEdgeIndex >> NB_BITS; }
++
++ PX_FORCE_INLINE PxU32 isArticulation0() const { return mNode0.isArticulation(); }
++ PX_FORCE_INLINE PxU32 isArticulation1() const { return mNode1.isArticulation(); }
++
++ PX_FORCE_INLINE PxU32 hasInfiniteMass0() const { return mEdgeIndex & HAS_INFINITE_MASS0; }
++ PX_FORCE_INLINE PxU32 hasInfiniteMass1() const { return mEdgeIndex & HAS_INFINITE_MASS1; }
++
++ PX_FORCE_INLINE void setInfiniteMass0() { mEdgeIndex |= HAS_INFINITE_MASS0; }
++ PX_FORCE_INLINE void setInfiniteMass1() { mEdgeIndex |= HAS_INFINITE_MASS1; }
++
++ PX_FORCE_INLINE void setHasThreshold() { mEdgeIndex |= HAS_THRESHOLD; }
++ PX_FORCE_INLINE PxU32 hasThreshold() const { return mEdgeIndex & HAS_THRESHOLD; }
++
++ PX_FORCE_INLINE void setIsContact() { mEdgeIndex |= IS_CONTACT; }
++ PX_FORCE_INLINE PxU32 isContact() const { return mEdgeIndex & IS_CONTACT; }
++
++ PX_FORCE_INLINE void setSpecialHandled() { mEdgeIndex |= SPECIAL_HANDLED; }
++ PX_FORCE_INLINE void clearSpecialHandled() { mEdgeIndex &= ~SPECIAL_HANDLED; }
++ PX_FORCE_INLINE PxU32 isSpecialHandled() const { return mEdgeIndex & SPECIAL_HANDLED; }
++
++ //KS - This constructor explicitly does not set mUniqueIndex. It is filled in by the pool allocator and this constructor
++ //is called afterwards. We do not want to stomp the uniqueIndex value
++ PartitionEdge(IG::EdgeIndex index) :
++ mNextPatch(NULL),
++ mEdgeIndex(index << NB_BITS)
++ {
++ PX_ASSERT(!(index & 0xf8000000)); // PT: reserve 5 bits for internal flags
++ }
++ };
++ PX_COMPILE_TIME_ASSERT(sizeof(PartitionEdge)<=32); // PT: 2 of them per cache-line
++
++ static PX_FORCE_INLINE void processPartitionEdges(const IG::GPUExternalData* gpuData, const PxcNpWorkUnit& unit)
++ {
++ if(gpuData && !(unit.mFlags & PxcNpWorkUnitFlag::eDISABLE_RESPONSE))
++ {
++ PxU32* edgeNodeIndices = gpuData->getEdgeNodeIndexPtr();
++ if(edgeNodeIndices) // PT: only non-null for GPU version
++ {
++ const PartitionEdge* partitionEdge = gpuData->getFirstPartitionEdge(unit.mEdgeIndex);
++ while(partitionEdge)
++ {
++ edgeNodeIndices[partitionEdge->mUniqueIndex] = unit.mNpIndex;
++ partitionEdge = partitionEdge->mNextPatch;
++ }
++ }
++ }
++ }
++ }
++
++ #endif