Explicit copy memset nodes#2380

Open

ThrudPrimrose wants to merge 75 commits into

mainfrom

explicit-copy-memset-nodes

Commits on May 15, 2026

Commits on May 16, 2026

Unify libnode stream connector on __dace_current_stream (valid in both legacy and experimental codegen)
ThrudPrimrose
committed
Remove libnode stream-input plumbing; rename STREAM_CONN -> CURRENT_STREAM_NAME
ThrudPrimrose
committed

Commits on May 17, 2026

Commits on May 20, 2026

Merge branch 'main' into explicit-copy-memset-nodes
ThrudPrimrose
committed
Move connector names to class constants; use dtypes.{CPU,GPU}_RESIDENT_STORAGES
ThrudPrimrose
committed
Apply connector-contract + inner-literal consolidation to memset and copy_node
ThrudPrimrose
committed
helpers: use imported nodes module for MapEntry/Tasklet/LibraryNode isinstance checks
ThrudPrimrose
committed
Accept Fortran-packed layouts in CopyNDTemplate; route mixed C/F to MappedTasklet
ThrudPrimrose
committed
MappedTasklet handles rank-mismatch via 1-D walker + int_floor/% delinearization; selector bypasses CopyND
ThrudPrimrose
committed
Revert MappedTasklet rank-mismatch; CopyNDTemplate is the only supported reshape impl
ThrudPrimrose
committed
Refactor multi-dim copy tests onto shared helper + add unsupported-case feature-regression tests
ThrudPrimrose
committed
Extract auto_dispatch + merge memset test helpers + type hints on new helpers
ThrudPrimrose
committed
DRY/YAGNI: drop dead CopyExpansion fields, reuse collapse_shape_and_strides in memset, Range.num_elements over reduce
ThrudPrimrose
committed
Drop ExpandCopyNDTemplate; MappedTasklet handles rank-mismatch (CopyND only for shared memory)
ThrudPrimrose
committed
Simplify _coarse_pick, _cuda2d_strides_are_supported, merge cudaMemcpy2D 1D+2D branches
ThrudPrimrose
committed
Tailor _build_copynd_call to its sole shared-memory caller
ThrudPrimrose
committed
Drop dynamic-input connectors from copy/memset libnodes; subset symbols must already be in scope
ThrudPrimrose
committed
Prune dead code: drop length-one<->scalar pass (unused), TODOs, require_contiguous param, _make_memset_skeleton unused returns
ThrudPrimrose
committed
Soften _is_cross_cpu_gpu docstring on Register handling assumption
ThrudPrimrose
committed
Merge remote-tracking branch 'origin/main' into explicit-copy-memset-nodes
ThrudPrimrose
committed
Inline _require_contiguous_subset as combined check at both call sites
ThrudPrimrose
committed
Add type hints (StorageType, Range, SymExpr, LibraryNode forward refs) on copy/memset libnode helpers
ThrudPrimrose
committed
Improve _refine_cuda_impl_for_subsets docstring (accurate routing table)
ThrudPrimrose
committed
Drop _memcpy_connector_typing: DaCe handles single-element pointer connectors correctly
ThrudPrimrose
committed
Add 5 Auto-routed copy edge-case tests: 4D/1D flatten, 1D/4D inflate, 3D/2D collapse, F-order collapse, step-2 stride
ThrudPrimrose
committed
Trim inner-tasklet connector comments; unify memset inner connector to '_out'; add transpose-pattern + dst-non-contig rejection tests
ThrudPrimrose
committed
Rename memset tests: <expansion>_<rank>_<storage>; reject-test names the failure mode
ThrudPrimrose
committed
Drop new_gpu_codegen_only markers; remove stale connector-types test
ThrudPrimrose
committed
Unify SDFG-construction helpers in copy_node_test.py
ThrudPrimrose
committed
Reject transpose pattern upfront in CopyLibraryNode + trim test docstrings
ThrudPrimrose
committed
Pin contract: same-rank copy needs matching per-dim subset sizes, not array shapes
ThrudPrimrose
committed
Normalize storage-type references to the full dace.dtypes.StorageType.X form
ThrudPrimrose
committed
MemsetLibraryNode: Auto falls back to 'pure' for non-contiguous subsets
ThrudPrimrose
committed
Prune dead helpers, args, and stale docstring
ThrudPrimrose
committed
InsertExplicitCopies: lift stage-in / stage-out copies to libnodes inside the map scope
ThrudPrimrose
committed
IEC cleanup: unify staging-lift methods + drop dead param
ThrudPrimrose
committed
copy_node_test: assert no dace::CopyND in generated C++ at every compile
ThrudPrimrose
committed
Inline single-callsite helpers, drop dead docs; simplify __main__ blocks; factor staging-test scaffold
ThrudPrimrose
committed
Trim verbose WHY comment on symbolic-< fallback in _is_consecutive_reshape
ThrudPrimrose
committed

Commits on May 21, 2026

Merge branch 'main' into explicit-copy-memset-nodes
ThrudPrimrose
committed

Commits on May 26, 2026

Merge origin/main into explicit-copy-memset-nodes (prefer theirs)
ThrudPrimrose
committed

Commits on May 29, 2026

Commits on May 30, 2026

Commits on Jun 4, 2026

Improve register location detection
ThrudPrimrose
committed

Commits on Jun 8, 2026

Updated how `get_gpu_{backend, runtime}()` worked.

philip-paul-mueller
authored and
ThrudPrimrose
committed
The expansion of the copy library node always assumed CUDA and not Hip. Now they use the configured backend.

philip-paul-mueller
authored and
ThrudPrimrose
committed
Removed a stray `print()`.

philip-paul-mueller
authored and
ThrudPrimrose
committed

Commits on Jun 9, 2026

Commits on Jun 29, 2026