refactor(search): Answer quiescence and the root through one fail soft value - #374
Merged
Merged
Conversation
…t value A search's fail soft answer (the window as it stands, the best score and the move that scored it, the taint, the count of moves searched, and whether alpha rose) was written three times. `Node` held it with `absorb` and `raised_alpha`. Quiescence kept six locals for it and repeated `absorb` by hand in its capture loop. `search_root` kept seven and did the same in its move loop, with `alpha != opening_alpha` as whether an aborted iteration may answer. The answer half of `Node` is now its own value, `FailSoft`, which `Node` holds as `answer`. It has `open`, `absorb` (which still moves alpha and the root bounds together at a rise) and `raised_alpha`, and a `stand_pat` step for quiescence that sets the best score and alpha from the static score with no move and no count. Quiescence opens one before its stand pat test, absorbs each capture, reads a side in check with no legal move off a searched count of zero, and stores as before (its standing evaluation with the entry) with `raised_alpha` from the answer. The root opens one at the aspiration window with both root bounds, absorbs each move, stops at beta, answers with the best move and score, and reads whether an aborted iteration may answer off `raised_alpha`. The late move rules, the census, the effort instrument and the ledger read the bounds and the count through `node.answer`, a field path that costs nothing over the old one. `Node::open` takes the opened answer, and its `too_many_arguments` allowance goes. The name keeps clear of the protocol's `Answer` and the effort instrument's `Answered`, which are finished answers rather than one being built. `absorb` asks beta only of a score above alpha, as quiescence and the root did, where `Node` asked beta first. With alpha under beta the two give the same answer, and a move that fails low (most of them) asks one question. A debug assertion holds the window open. Asking beta first measured 206,734 more instructions over the depth 5 bench (0.1%), most of it in quiescence. The root's store keeps its exemption from the taint policy, now as `store_root_answer` with the reason beside it: the reported line is read back from the root's slot, so the root stores whatever its taint, and under the skip policy the rare tainted cutoff that slot offers is refused as under refuse. No test held the exemption, so one now searches a root a half move short of the fifty move draw under the skip policy and reads that its one tainted store landed. The answer's tests move to the new value, and two more read it with no board: the stand pat sets the best score and alpha and counts no move, and a fresh answer counts nothing searched until a move is absorbed. The bench and a fixed depth search open the root at the full window, so no test reached the root's ceiling or its abort below alpha. Two more call the root at a window no move beats. A depth five search of the opening at 500 to 560 answers a ceiling, and the same search stopped one node short answers nothing rather than the closest move. A root whose every move scores exactly alpha answers a ceiling too, since meeting alpha does not raise it. The tree is unchanged. The bench counts 5,965,973, and its per position rows print as on the parent under each of the four taint policies, with the time masked. Every instrument's rows at depth 5 print as on the parent: cutoffs, reductions, effort with no switch off and with null_move off, residuals, and forced with every kind and with null_move, reverse_futility, and skip and trusted_scout from depth 2 alone, at their default sampling and at every 1. The info lines of `go depth 12` from the start position through the protocol match the parent's, with the time masked. Callgrind over the depth 5 bench counts 194,139,961 instructions on the parent and 194,078,291 here, -61,670 (-0.03%). By symbol, summing each function's two copies, alpha_beta is 180,733 fewer and search_root 21,250, windowed is unchanged, and quiescence is 140,252 more (1.2% of its own). Quiescence's own lines in engine.rs lost 228,276 and the integer comparisons inlined into it gained 267,714, much of that the same comparisons credited to the operator's line rather than the caller's. Its mate test asks the searched count before the check flag. Written the other way round, the compiler kept the flag on the stack rather than in a register, every test of it read memory, and quiescence ran 45,834 more. Bench: 5965973 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Speed against
|
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
1 participant
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.
A search's fail soft answer (the window as it stands, the best score and its
move, the taint, the count searched, whether alpha rose) was written three
times:
Nodeheld it, quiescence kept six locals and repeatedabsorbbyhand, and
search_rootkept seven and did the same. It is now one value,FailSoft, whichNodeholds and quiescence and the root absorb through.Quiescence's stand pat is a named step on it, and the root's store is a named
store_root_answerwith its documented exemption from the taint policystated where it happens.
What a reviewer should know
arche benchcounts 5,965,973 on the base and thebranch, and its per position rows are identical under each of the four taint
policies. Every instrument's rows at depth 5 are identical at the default
sampling and at
every 1: cutoffs, reductions, effort with and withoutnull_moveoff, residuals, and the new forced instrument with every kindand three selections of kinds. Every
infoline ofgo depth 12from thestart position through the protocol matches. A second review's own builds
also held
bench 7, the reference configuration at depths 7 and 8, thegames suite under three policies, seven switches off, and node limit
sessions including an aborted iteration reported as a floor.
bench 5against the parent: -61,670 instructions (-0.03%).alpha_betaand the root fell; quiescence rose 1.2% of its own cost, whichis integer comparisons credited to the operator's line rather than the
caller's. Two shapes were measured on the way:
absorbasking beta firstcost 0.1% more, almost all in quiescence, so it asks beta only of a score
above alpha, as quiescence and the root did before; and quiescence's mate
test written flag first made the compiler spill the check flag to the
stack, seen in the disassembly, so the count is asked first. A speed round
on an idle machine read -0.7% with a 95% interval from -2.9% to +0.6%, not
resolved, as a change this size should read. A first round under load was
set aside.
case (the stand pat's baseline, the order of the beta test, the best move
and the store, the abort rule, the bound, the game over read, the taint)
and found no behaviour defect. It found the root's abort rule had no test
off the full window, since the bench and a fixed depth search open the root
at the full window. Two tests now drive the root at an aspiration window no
move beats: the search answers a ceiling, the same search stopped one node
short answers nothing, and a root whose every move scores exactly alpha
answers a ceiling.
documented on
TaintPolicy::Skip; a test now holds it.Tests go from 784 release and 786 debug in the workspace to 789 and 791.
🤖 Generated with Claude Code