Repository navigation
Add GETAWAY 3D descriptors - #358
Merged
Merged
Conversation
groupAllReduceSum shuffled with __activemask() as its mask, which does not guarantee the group's lanes are converged; it now reduces with cooperative_groups::reduce over an explicit thread_block_tile, so every lane of the tile must participate. USR's closest/farthest-atom search uses the same reduction with its tie-breaking comparison. The pairwise kernel's combine across a warp's four groups (lanes with equal lane % kGroupSize), which no tile covers, moves into one helper, sumAcrossGroups. All 3D properties are bit-identical to before in both precisions on 484 ChEMBL + planar conformers, compute-sanitizer synccheck reports no errors, and kernel times are unchanged.
RDKit's 273-value GETAWAY vector (CalcGETAWAY order) in a new Getaway family, one warp per conformer: - Leverages from H = X (X^T X)^+ X^T with RDKit's 1e-3 pseudo-inverse cutoff, in float64 (they are rounded and clustered for ITH/ISH). - Pair loop over the atom-pair triangle with a bond-distance search capped at 8 (shared searchBondDepths, now templated on its cap): H, HATS, R and R+ per lag 1-8 and 7 channels; farther or disconnected pairs enter the totals only, as in RDKit. - REIG by power iteration on the nonnegative symmetric R (its Perron eigenvalue); the first product gives row sums for RARS and RCON. - HIC uses the PBF of the molecule's default conformer, HGM a sequential leverage product, and ITH/ISH a literal port of RDKit's %.*g rounding and digit-string clustering. GetawayOptions.precision (default 2) is RDKit's precision argument, limited to 1-6. FULL matches RDKit exactly on all 120,000+ values of 439 single-fragment ChEMBL + planar conformers; SINGLE is within one 0.001 rounding unit. RDKit's lag loop runs to its 1e8 distance for disconnected atoms, so it does not finish on salts; the benchmark's RDKit GETAWAY reference skips multi-fragment molecules and reports how many. 24,200 conformers, SINGLE: all properties 333 -> 531 ms (GETAWAY kernel 194 ms); requests without GETAWAY are unchanged. Serial RDKit GETAWAY is ~26 ms per conformer.
The group maximum reduces with cooperative_groups::reduce over the group's tile instead of __activemask() shuffles; the cross-group sums and maximum use sumAcrossGroups and a new maxAcrossGroups beside it. GETAWAY output is unchanged (FULL still matches RDKit exactly) and synccheck is clean.
ITH/ISH ran on one lane per conformer, rebuilding decimal text for every IsClose2 comparison during an insertion sort and greedy scan. The lanes now round each heavy-atom leverage and reduce its printed text to the integers IsClose2 reads (leading zeros, length, leading integer, whether it is all digits), rank the heavy atoms in parallel, and lane 0 runs only the greedy scan on those keys. GETAWAY kernel on 24,200 conformers, SINGLE: 192 -> 150 ms. FULL still matches RDKit exactly, including ITH and ISH.
Bond distances depend only on the molecule, but both kernels searched them per conformer. A per-molecule table (one byte per atom pair, depth 10) is now built once per batch and read by both; batches with molecules above 3072 atoms, or past a 256 MB budget, keep the per-row searches. Outputs are bit-identical. SINGLE, all properties, 24,200 conformers: 487 -> 409 ms end to end (getaway3DKernel 152 -> 111 ms, pairwise3DKernel 122 -> 62 ms, table build 13 ms). FULL is unchanged within noise.
Compute the lag-0 sums after the pair loop and write the H and R blocks before REIG, so neither set of accumulators is live through the other's loop. float32: 168 -> 128 registers (3 -> 4 blocks per SM); float64: 218 -> 186. Outputs are bit-identical. GETAWAY only, 24,200 conformers: SINGLE kernel 106.7 -> 93.1 ms (end to end 178.8 -> 164.3 ms); FULL unchanged (2276 -> 2273 ms).
Contributor
 |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
evasnow1992
reviewed
Oct 7, 2026
Collaborator
There was a problem hiding this comment.
The docstring here currently says the is3D flag is preserved for PBF, but GETAWAY's HIC now uses it as well. It may also be worth mentioning that, while the RDKit-molecule path follows RDKit in using the first conformer's flatness for HIC, device-coordinate inputs use each conformer's own geometry.
evasnow1992
approved these changes
Oct 7, 2026
evasnow1992
left a comment
evasnow1992
left a comment
Collaborator
There was a problem hiding this comment.
Changes look good to me.
- ITH and ISH are NaN when a heavy-atom leverage is not finite, instead of clustering unwritten sort keys; non-finite values skip significant-digit rounding. - GETAWAY precision is validated and passed only when GETAWAY is requested, so unused options are ignored as documented. - The bond-distance table is sized by the batch's actual atom pairs (Property3DDeviceInputs::moleculeAtomPairs) rather than the largest molecule times the molecule count. - Document that GETAWAY's HIC reads is3D and takes its flatness from each molecule's first conformer, or each row's own with device coordinates.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.