Simplify and improve AtomicQuadTree #5
Conversation
src/kernels.h
Outdated
| tree.next_nodes[0] = is_leaf<Index_t>; | ||
| tree.node_status[0].store(NodeStatus::EmptyLeaf, memory_order_relaxed); | ||
| } | ||
|
|
||
| template<typename T, typename Index_t> | ||
| auto clear_tree(System<T>& system, AtomicQuadTreeContainer<T, Index_t> tree) { | ||
| auto clear_tree(System<T>& system, AtomicQuadTree<T, Index_t> tree) { | ||
| // clear the tree, ready for next iteration | ||
| auto r = system.body_indices(); | ||
| std::for_each_n( | ||
| std::execution::par_unseq, | ||
| r.begin(), tree.bump_allocator->load(memory_order_acquire), |
There was a problem hiding this comment.
The tree is not properly initialised in the first iteration and the first force step does not calculate acceleration properly. Applying the for loop to the entire tree I think is fine.
| r.begin(), tree.bump_allocator->load(memory_order_acquire), | |
| r.begin(), tree.capacity, |
Furthermore, I am concerned that this can run over the end of the iterator? I'm not sure how this works with std::views::iota, but e.g. there could be 1000 bodies, so r counts up to 1000, but there may be 4000 tree nodes.
There was a problem hiding this comment.
Regarding the std::views::iota question, std::views::iota(v) (with one argument, not two) generates an integer range that starts at the value and spans until numeric_limits<T>::max(). There are no values of type T greater than numeric_limits<T>::max(), so the number of leafs can't be larger (it'd have already overflown the integer type somewhere else).
There was a problem hiding this comment.
The tree is not properly initialised in the first iteration and the first force step does not calculate acceleration properly.
I've tried to fix it in this commit (505dcbc) but I'm not 100% sure I did that correctly.
Also, i'm not sure if it is better to always clean the tree up to the capacity, or up to the last allocated node, but we can explore tweaking that in a sub-sequent PR so I've left things "as is".
gonzalobg
force-pushed
the
atom_quad_tree
branch
from
5bd7e65 to
27a83a3
Compare
Start one thread per tree node, and filter out threads that do not start at a leaf node, instead of preparing and updating a leaf node count.
gonzalobg
force-pushed
the
atom_quad_tree
branch
from
505dcbc to
8970037
Compare
|
Rebased on top of |
This PR simplifies the owning AtomicQuadTree and its Const/Non-Const reference types into a single "reference" type that can be easily passed to parallel algorithms by value.
It also performs some initial optimizations and simplifications on AtomicQuadTree and its algorithms:
atomic_calc_massalgorithm is optimized to process one tree node per thread (threads assigned to empty nodes just exit). The last thread to process a node advances to the next one. This allow us to remove theleaf_count, which simplifies the quadtree,atomic_calc_mass, andatomic_insert. This improves performance by 1.9x on GPUs.