FMM driver#
- boxtree.fmm.drive_fmm(wrangler: ExpansionWranglerInterface, src_weight_vecs, timing_data=None, global_src_idx_all_ranks=None, global_tgt_idx_all_ranks=None)[source]#
Top-level driver routine for a fast multipole calculation.
In part, this is intended as a template for custom FMMs, in the sense that you may copy and paste its source code as a starting point.
Nonetheless, many common applications (such as point-to-point FMMs) can be covered by supplying the right expansion_wrangler to this routine.
- Parameters:
expansion_wrangler â€“ An object exhibiting the
ExpansionWranglerInterface
. For distributed implementation, this wrangler should be a subclass ofboxtree.distributed.calculation.DistributedExpansionWrangler
.src_weight_vecs â€“ A sequence of source â€˜density/weights/chargesâ€™. Passed unmodified to expansion_wrangler. For distributed implementation, this argument is only significant on the root rank, but worker ranks still need to supply a dummy vector.
timing_data â€“ Either None, or a
dict
that is populated with timing information for the stages of the algorithm (in the form ofTimingResult
), if such information is available.global_src_idx_all_ranks â€“ Only used in the distributed implementation. A
list
of lengthnranks
, where the i-th entry is anumpy.ndarray
representing the global indices of sources in the local tree on rank i. Each entry can be returned from generate_local_tree. This argument is only significant on the root rank.global_tgt_idx_all_ranks â€“ Only used in the distributed implementation. A
list
of lengthnranks
, where the i-th entry is anumpy.ndarray
representing the global indices of targets in the local tree on rank i. Each entry can be returned from generate_local_tree. This argument is only significant on the root rank.
- Returns:
the potentials computed by expansion_wrangler. For the distributed implementation, the potentials are gathered and returned on the root rank; this function returns None on the worker ranks.
- class boxtree.fmm.TreeIndependentDataForWrangler[source]#
An object that can be used to store information for efficient wrangler execution that depends on the kernel but not the tree and/or the traversal.
Examples of such data include generated code for carrying out translations.
Note
Instances of this type should not hold a reference (and thereby be specific to) a
boxtree.Tree
instance. Their purpose is to host caches for generated translation code that is reusable across trees. It is OK for these instances to be specific to a given kernel (or set of kernels).
- class boxtree.fmm.ExpansionWranglerInterface(tree_indep: TreeIndependentDataForWrangler, traversal: FMMTraversalInfo)[source]#
Abstract expansion handling interface for use with
drive_fmm()
.See this test code for a very simple sample implementation.
Note
Wranglers may hold a reference (and thereby be specific to) a
boxtree.Tree
instance.TreeIndependentDataForWrangler
exists to hold data that is more broadly reusable.Functions that support returning timing data return a value supporting the
TimingFuture
interface.Changed in version 2018.1: Changed (a subset of) functions to return timing data.
- tree_indep#
An instance of (a typically wrangler-dependent subclass of)
TreeIndependentDataForWrangler
.
- traversal#
An instance of
FMMTraversalInfo
.
- tree#
Particle ordering
- abstract reorder_sources(source_array)[source]#
Return a copy of source_array in tree source order. source_array is in user source order.
- abstract reorder_potentials(potentials)[source]#
Return a copy of potentials in user target order. source_weights is in tree target order.
Views into arrays of expansions
Translations
- abstract form_multipoles(level_start_source_box_nrs, source_boxes, src_weight_vecs)[source]#
Return an expansions array containing multipole expansions in source_boxes due to sources with src_weight_vecs. All other expansions must be zero.
- Returns:
A pair (mpoles, timing_future).
- abstract coarsen_multipoles(level_start_source_parent_box_nrs, source_parent_boxes, mpoles)[source]#
For each box in source_parent_boxes, gather (and translate) the boxâ€™s childrenâ€™s multipole expansions in mpole and add the resulting expansion into the boxâ€™s multipole expansion in mpole.
- Returns:
A pair (mpoles, timing_future).
- abstract eval_direct(target_boxes, neighbor_sources_starts, neighbor_sources_lists, src_weight_vecs)[source]#
For each box in target_boxes, evaluate the influence of the neighbor sources due to src_weight_vecs, which use CSR-like interaction list storage and are indexed like target_boxes.
- Returns:
A pair (pot, timing_future), where pot is a a new potential array.
- abstract multipole_to_local(level_start_target_or_target_parent_box_nrs, target_or_target_parent_boxes, starts, lists, mpole_exps)[source]#
For each box in target_or_target_parent_boxes, translate and add the influence of the multipole expansion in mpole_exps into a new array of local expansions. starts and lists use CSR-like interaction list storage, and starts is indexed like target_or_target_parent_boxes.
- Returns:
A pair (pot, timing_future) where pot is a new (local) expansion array.
- abstract eval_multipoles(target_boxes_by_source_level, from_sep_smaller_by_level, mpole_exps)[source]#
For a level i, each box in target_boxes_by_source_level[i], evaluate the multipole expansion in mpole_exps in the nearby boxes given in from_sep_smaller_by_level, and return a new potential array. starts and lists in from_sep_smaller_by_level[i] use CSR-like interaction list storage and starts is indexed like target_boxes_by_source_level[i].
- Returns:
A pair (pot, timing_future) where pot is a new potential array.
- abstract form_locals(level_start_target_or_target_parent_box_nrs, target_or_target_parent_boxes, starts, lists, src_weight_vecs)[source]#
For each box in target_or_target_parent_boxes, form local expansions due to the sources in the nearby boxes given in starts and lists, and return a new local expansion array. starts and lists use CSR-like interaction list storage and starts is indexed like target_or_target_parent_boxes.
- Returns:
A pair (pot, timing_future) where pot is a new local expansion array.
- abstract refine_locals(level_start_target_or_target_parent_box_nrs, target_or_target_parent_boxes, local_exps)[source]#
For each box in child_boxes, translate the boxâ€™s parentâ€™s local expansion in local_exps and add the resulting expansion into the boxâ€™s local expansion in local_exps.
- Returns:
A pair (local_exps, timing_future).
- abstract eval_locals(level_start_target_box_nrs, target_boxes, local_exps)[source]#
For each box in target_boxes, evaluate the local expansion in local_exps and return a new potential array.
- Returns:
A pair (pot, timing_future) where pot is a new potential array.
- abstract finalize_potentials(potentials, template_ary)[source]#
Postprocess the reordered potentials. This is where global scaling factors could be applied. This is distinct from
reorder_potentials()
because some derived FMMs (notably the QBX FMM) do their own reordering.- Parameters:
template_ary â€“ If the array type used inside of the FMM is different from the array type used by the user (e.g.
boxtree.pyfmmlib_integration.FMMLibExpansionWrangler
usesnumpy.ndarray
internally, this array can be used to help convert the output back to the userâ€™s array type (typicallypyopencl.array.Array
).
Integrates boxtree
with
pyfmmlib.
- class boxtree.pyfmmlib_integration.FMMLibTreeIndependentDataForWrangler(dim, kernel, ifgrad=False)[source]#
- class boxtree.pyfmmlib_integration.FMMLibExpansionWrangler(tree_indep, traversal, *, helmholtz_k=None, fmm_level_to_order=None, dipole_vec=None, dipoles_already_reordered=False, order=None, optimized_m2l_precomputation_memory_cutoff_bytes=100000000, rotation_data=None)[source]#
Implements the
boxtree.fmm.ExpansionWranglerInterface
by using pyfmmlib.Timing results returned by this wrangler contains the values wall_elapsed and (optionally, if supported) process_elapsed, which measure wall time and process time in seconds, respectively.
Internal bits#
- class boxtree.pyfmmlib_integration.FMMLibRotationDataInterface[source]#
Abstract interface for additional, optional data for precomputation of rotation matrices passed to the expansion wrangler.
- class boxtree.pyfmmlib_integration.FMMLibRotationData(queue, trav)[source]#
An implementation of the
FMMLibRotationDataInterface
.