diff --git a/PARALLEL_PROJECT_NOTES.md b/PARALLEL_PROJECT_NOTES.md new file mode 100644 index 000000000..b60263e4e --- /dev/null +++ b/PARALLEL_PROJECT_NOTES.md @@ -0,0 +1,218 @@ +# Parallel project notes + + + +We track notes for updates to +[smt/parallel.cpp](https://github.com/Z3Prover/z3/blob/master/src/smt/smt_parallel.cpp) +and possibly +[solver/parallel_tactic.cpp](https://github.com/Z3Prover/z3/blob/master/src/solver/parallel_tactical.cpp). + + + + + +## Variable selection heuristics + + + +* Lookahead solvers: + * lookahead in the smt directory performs a simplistic lookahead search using unit propagation. + * lookahead in the sat directory uses custom lookahead solver based on MARCH. March is described in Handbook of SAT and Knuth volumne 4. + * They both proxy on a cost model where the most useful variable to branch on is the one that _minimizes_ the set of new clauses maximally + through unit propagation. In other words, if a literal _p_ is set to true, and _p_ occurs in clause $\neg p \vee q \vee r$, then it results in + reducing the clause from size 3 to 2 (because $\neg p$ will be false after propagating _p_). + * Selected references: SAT handbook, Knuth Volumne 4, Marijn's March solver on github, [implementation of march in z3](https://github.com/Z3Prover/z3/blob/master/src/sat/sat_lookahead.cpp) +* VSIDS: + * As referenced in Matteo and Antti's solvers. + * Variable activity is a proxy for how useful it is to case split on a variable during search. Variables with a higher VSIDS are split first. + * VSIDS is updated dynamically during search. It was introduced in the paper with Moscovitz, Malik, et al in early 2000s. A good overview is in Armin's tutorial slides (also in my overview of SMT). + * VSIDS does not keep track of variable phases (if the variable was set to true or false). + * Selected refernces [DAC 2001](https://www.princeton.edu/~chaff/publication/DAC2001v56.pdf) and [Biere Tutorial, slide 64 on Variable Scoring Schemes](https://alexeyignatiev.github.io/ssa-school-2019/slides/ab-satsmtar19-slides.pdf) +* Proof prefix: + * Collect the literals that occur in learned clauses. Count their occurrences based on polarity. This gets tracked in a weighted score. + * The weight function can be formulated to take into account clause sizes. + * The score assignment may also decay similar to VSIDS. + * We could also use a doubly linked list for literals used in conflicts and keep reinsert literals into the list when they are used. This would be a "Variable move to front" (VMTF) variant. + * Selected references: [Battleman et al](https://www.cs.cmu.edu/~mheule/publications/proofix-SAT25.pdf) +* From local search: + * Note also that local search solvers can be used to assign variable branch priorities. + * We are not going to directly run a local search solver in the mix up front, but let us consider this heuristic for completeness. + * The heuristic is documented in Biere and Cai's journal paper on integrating local search for CDCL. + * Roughly, it considers clauses that move from the UNSAT set to the SAT set of clauses. It then keeps track of the literals involved. + * Selected references: [Cai et al](https://www.jair.org/index.php/jair/article/download/13666/26833/) +* Assignment trails: + * We could also consider the assignments to variables during search. + * Variables that are always assigned to the same truth value could be considered to be safe to assign that truth value. + * The cubes resulting from such variables might be a direction towards finding satisfying solutions. + * Selected references: [Alex and Vadim](https://link.springer.com/chapter/10.1007/978-3-319-94144-8_7) and most recently [Robin et al](https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.SAT.2024.9). + + +## Algorithms + +This section considers various possible algorithms. +In the following, $F$ refers to the original goal, $T$ is the number of CPU cores or CPU threads. + +### Base algorithm + +The existing algorithm in smt_parallel is as follows: + +1. Run a solver on $F$ with a bounded number of conflicts. +2. If the result is SAT/UNSAT, or UNKNOWN with an interrupt or timeout, return. If the maximal number of conflicts were reached continue. +3. Spawn $T$ solvers on $F$ with a bounded number of conflicts, wait until a thread returns UNSAT/SAT or all threads have reached a maximal number of conflicts. +4. Perform a similar check as in 2. +5. Share unit literals learned by each thread. +6. Compute unit cubes for each thread $T$. +7. Spawn $T$ solvers with $F \wedge \ell$, where $\ell$ is a unit literal determined by lookahead function in each thread. +8. Perform a similar check as in 2. But note that a thread can be UNSAT because the unit cube $\ell$ contradicted $F$. In this case learn the unit literal $\neg \ell$. +9. Shared unit literals learned by each thread, increase the maximal number of conflicts, go to 3. + +### Algorithm Variants + +* Instead of using lookahead solving to find unit cubes use the proof-prefix based scoring function. +* Instead of using independent unit cubes, perform a systematic (where systematic can mean many things) cube and conquer strategy. +* Spawn some threads to work in "SAT" mode, tuning to find models instead of short resolution proofs. +* Change the synchronization barrier discipline. +* [Future] Include in-processing + +### Cube and Conquer strategy + +We could maintain a global decomposition of the search space by maintaing a list of _cubes_. +Initially, the list of cubes has just one element, the cube with no literals $[ [] ]$. +By using a list of cubes instead of a _set_ of cubes we can refer to an ordering. +For example, cubes can be ordered by a suffix traversal of the _cube tree_ (the tree formed by +case splitting on the first literal, children of the _true_ branch are the cubes where the first +literal is true, children of the _false_ branch are the cubes where the first literal is false). + +The main question is going to be how the cube decomposition is created. + +#### Static cubing +We can aim for a static cube strategy that uses a few initial (concurrent) probes to find cube literals. +This strategy would be a parallel implementaiton of proof-prefix approach. The computed cubes are inserted +into the list of cubes and the list is consumed by a second round. + +#### Growing cubes on demand +Based on experiences with cubing so far, there is high variance in how easy cubes are to solve. +Some cubes will be harder than others to solve. For hard cubes, it is tempting to develop a recursive +cubing strategy. Ideally, a recursive cubing strategy is symmetric to top-level cubing. + +* The solver would have to identify hard cubes vs. easy cubes. +* It would have to know when to stop working on a hard cube and replace it in the list of cubes by + a new list of sub-cubes. + +* Ideally, we don't need any static cubing and cubing is grown on demand while all threads are utilized. + * If we spawn $T$ threads to initially work with empty cubes, we could extract up to $T$ indepenent cubes + by examining the proof-prefix of their traces. This can form the basis for the first, up to $2^T$ cubes. + * After a round of solving with each thread churning on some cubes, we may obtain more proof-prefixes from + _hard_ cubes. It is not obvious that we want to share cubes from different proof prefixes at this point. + But a starting point is to split a hard cube into two by using the proof-prefix from attempting to solve it. + * Suppose we take the proof-prefix sampling algorithm at heart: It says to start with some initial cube prefix + and then sample for other cube literals. If we translate it to the case where multiple cubes are being processed + in parallel, then an analogy is to share candidates for new cube literals among cubes that are close to each-other. + For example, if thread $t_1$ processes cube $a, b, c$ and $t_2$ processes $a,b, \neg c$. They are close. They are only + separated by Hamming distance 1. If $t_1$ finds cube literal $d$ and $t_2$ finds cube literal $e$, we could consider the cubes + $a, b, c, d, e$, and $a, b, c, d, \neg e$, $\ldots$, $a, b, \neg c, \neg d, \neg e$. + +#### Representing cubes implicitly + +We can represent a list of cubes by using intervals and only represent start and end-points of the intervals. + +#### Batching +Threads can work on more than one cube in a batch. + +### Synchronization + +* The first thread to time out or finish could kill other threads instead of joining on all threads to finish. +* Instead of synchronization barriers have threads continue concurrently without terminating. They synchronize on signals and new units. This is trickier to implement, but in some guises accomplished in [sat/sat_parallel.cpp](https://github.com/Z3Prover/z3/blob/master/src/sat/sat_parallel.cpp) + + +## Parameter tuning + +The idea is to have parallel threads try out different parameter settings and search the parameter space of an optimal parameter setting. + +Let us assume that there is a set of tunable parameters $P$. The set comprises of a set of named parameters with initial values. +$P = \{ (p_1, v_1), \ldots, (p_n, v_n) \}$. +With each parameter associate a set of mutation functions $+=, -=, *=$, such as increment, decrement, scale a parameter by a non-negative multiplier (which can be less than 1). +We will initialize a search space of parameter settings by parameters, values and mutation functions that have assigned reward values. The reward value is incremented +if a parameter mutation step results in an improvement, and decremented if a mutation step degrades performance. +$P = \{ (p_1, v_1, \{ (r_{11}, m_{11}), \ldots, (r_{1k_1}, m_{1k_1}) \}), \ldots, (p_n, v_n, \{ (r_{n1}, m_{n1}), \ldots, (r_{nk_n}, m_{nk_n})\}) \}$. +The initial values of reward functions is fixed (to 1) and the initial values of parameters are the defaults. + +* The batch manager maintains a set of candidate parameters $CP = \{ (P_1, r_1), \ldots, (P_n, r_n) \}$. +* A worker thread picks up a parameter $P_i$ from $CP$ from the batch manager. +* It picks one or more parameter settings within $P_i$ whose mutation function have non-zero reward functions and applies a mutation. +* It then runs with a batch of cubes. +* It measures the reward for the new parameter setting based in number of cubes, cube depth, number of timeouts, and completions with number of conflicts. +* If the new reward is an improvement over $(P_i, r_i)$ it inserts the new parameter setting $(P_i', r_i')$ into the batch manager. +* The batch manager discards the worst parameter settings keeping the top $K$ ($K = 5$) parameter settings. + +When picking among mutation steps with reward functions use a weighted sampling algorithm. +Weighted sampling works as follows: You are given a set of items with weights $(i_1, w_1), \ldots, (i_k, w_k)$. +Add $w = \sum_j w_j$. Pick a random number $w_0$ in the range $0\ldots w$. +Then you pick item $i_n$ such that $n$ is the smallest index with $\sum_{j = 1}^n w_j \geq w_0$. + +SMT parameters that could be tuned: + +
+
+  arith.bprop_on_pivoted_rows (bool) (default: true)
+  arith.branch_cut_ratio (unsigned int) (default: 2)
+  arith.eager_eq_axioms (bool) (default: true)
+  arith.enable_hnf (bool) (default: true)
+  arith.greatest_error_pivot (bool) (default: false)
+  arith.int_eq_branch (bool) (default: false)
+  arith.min (bool) (default: false)
+  arith.nl.branching (bool) (default: true)
+  arith.nl.cross_nested (bool) (default: true)
+  arith.nl.delay (unsigned int) (default: 10)
+  arith.nl.expensive_patching (bool) (default: false)
+  arith.nl.expp (bool) (default: false)
+  arith.nl.gr_q (unsigned int) (default: 10)
+  arith.nl.grobner (bool) (default: true)
+  arith.nl.grobner_cnfl_to_report (unsigned int) (default: 1)
+  arith.nl.grobner_eqs_growth (unsigned int) (default: 10)
+  arith.nl.grobner_expr_degree_growth (unsigned int) (default: 2)
+  arith.nl.grobner_expr_size_growth (unsigned int) (default: 2)
+  arith.nl.grobner_frequency (unsigned int) (default: 4)
+  arith.nl.grobner_max_simplified (unsigned int) (default: 10000)
+  arith.nl.grobner_row_length_limit (unsigned int) (default: 10)
+  arith.nl.grobner_subs_fixed (unsigned int) (default: 1)
+  arith.nl.horner (bool) (default: true)
+  arith.nl.horner_frequency (unsigned int) (default: 4)
+  arith.nl.horner_row_length_limit (unsigned int) (default: 10)
+  arith.nl.horner_subs_fixed (unsigned int) (default: 2)
+  arith.nl.nra (bool) (default: true)
+  arith.nl.optimize_bounds (bool) (default: true)
+  arith.nl.order (bool) (default: true)
+  arith.nl.propagate_linear_monomials (bool) (default: true)
+  arith.nl.rounds (unsigned int) (default: 1024)
+  arith.nl.tangents (bool) (default: true)
+  arith.propagate_eqs (bool) (default: true)
+  arith.propagation_mode (unsigned int) (default: 1)
+  arith.random_initial_value (bool) (default: false)
+  arith.rep_freq (unsigned int) (default: 0)
+  arith.simplex_strategy (unsigned int) (default: 0)
+  dack (unsigned int) (default: 1)
+  dack.eq (bool) (default: false)
+  dack.factor (double) (default: 0.1)
+  dack.gc (unsigned int) (default: 2000)
+  dack.gc_inv_decay (double) (default: 0.8)
+  dack.threshold (unsigned int) (default: 10)
+  delay_units (bool) (default: false)
+  delay_units_threshold (unsigned int) (default: 32)
+  dt_lazy_splits (unsigned int) (default: 1)
+  lemma_gc_strategy (unsigned int) (default: 0)
+  phase_caching_off (unsigned int) (default: 100)
+  phase_caching_on (unsigned int) (default: 400)
+  phase_selection (unsigned int) (default: 3)
+  qi.eager_threshold (double) (default: 10.0)
+  qi.lazy_threshold (double) (default: 20.0)
+  qi.quick_checker (unsigned int) (default: 0)
+  relevancy (unsigned int) (default: 2)
+  restart_factor (double) (default: 1.1)
+  restart_strategy (unsigned int) (default: 1)
+  seq.max_unfolding (unsigned int) (default: 1000000000)
+  seq.min_unfolding (unsigned int) (default: 1)
+  seq.split_w_len (bool) (default: true)
+
+ + diff --git a/run_local_tests.sh b/run_local_tests.sh new file mode 100755 index 000000000..e9bd45bad --- /dev/null +++ b/run_local_tests.sh @@ -0,0 +1,32 @@ +#!/bin/bash +# run from inside ./z3/build + +Z3=./z3 +OPTIONS="-v:0 -st smt.threads=4" +OUT_FILE="../z3_results.txt" +BASE_PATH="../../z3-poly-testing/inputs/" + +# List of relative test files (relative to BASE_PATH) +REL_TEST_FILES=( + "QF_NIA_small/Ton_Chanh_15__Singapore_v1_false-termination.c__p27381_terminationG_0.smt2" + "QF_UFDTLIA_SAT/52759_bec3a2272267494faeecb6bfaf253e3b_10_QF_UFDTLIA.smt2" +) + +# Clear output file +> "$OUT_FILE" + +# Loop through and run Z3 on each file +for rel_path in "${REL_TEST_FILES[@]}"; do + full_path="$BASE_PATH$rel_path" + test_name="$rel_path" + + echo "Running: $test_name" + echo "===== $test_name =====" | tee -a "$OUT_FILE" + + # Run Z3 and pipe output to both screen and file + $Z3 "$full_path" $OPTIONS 2>&1 | tee -a "$OUT_FILE" + + echo "" | tee -a "$OUT_FILE" +done + +echo "Results written to $OUT_FILE" diff --git a/src/math/polynomial/polynomial.cpp b/src/math/polynomial/polynomial.cpp index 9a0f572dd..0ad9639f2 100644 --- a/src/math/polynomial/polynomial.cpp +++ b/src/math/polynomial/polynomial.cpp @@ -5153,6 +5153,8 @@ namespace polynomial { // unsigned sz = R->size(); for (unsigned i = 0; i < sz; i++) { + if (sz > 100 && i % 100 == 0) + checkpoint(); monomial * m = R->m(i); numeral const & a = R->a(i); if (m->degree_of(x) == deg_R) { @@ -5571,6 +5573,7 @@ namespace polynomial { h = mk_one(); while (true) { + checkpoint(); TRACE(resultant, tout << "A: " << A << "\nB: " << B << "\n";); degA = degree(A, x); degB = degree(B, x); diff --git a/src/smt/priority_queue.h b/src/smt/priority_queue.h new file mode 100644 index 000000000..39deab9bb --- /dev/null +++ b/src/smt/priority_queue.h @@ -0,0 +1,191 @@ +// SOURCE: https://github.com/Ten0/updatable_priority_queue/blob/master/updatable_priority_queue.h + +#include +#include + +namespace updatable_priority_queue { + template + struct priority_queue_node { + Priority priority; + Key key; + priority_queue_node(const Key& key, const Priority& priority) : priority(priority), key(key) {} + friend bool operator<(const priority_queue_node& pqn1, const priority_queue_node& pqn2) { + return pqn1.priority > pqn2.priority; + } + friend bool operator>(const priority_queue_node& pqn1, const priority_queue_node& pqn2) { + return pqn1.priority < pqn2.priority; + } + }; + + /** Key has to be an uint value (convertible to size_t) + * This is a max heap (max is on top), to match stl's pQ */ + template + class priority_queue { + protected: + std::vector id_to_heappos; + std::vector> heap; + std::size_t max_size = 4; // std::numeric_limits::max(); // Create a variable max_size that defaults to the largest size_t value possible + + public: + // priority_queue() {} + priority_queue(std::size_t max_size = std::numeric_limits::max()): max_size(max_size) {} + + // Returns a const reference to the internal heap storage + const std::vector>& get_heap() const { + return heap; + } + + bool empty() const { return heap.empty(); } + std::size_t size() const { return heap.size(); } + + /** first is priority, second is key */ + const priority_queue_node& top() const { return heap.front(); } + + void pop(bool remember_key=false) { + if(size() == 0) return; + id_to_heappos[heap.front().key] = -1-remember_key; + if(size() > 1) { + *heap.begin() = std::move(*(heap.end()-1)); + id_to_heappos[heap.front().key] = 0; + } + heap.pop_back(); + sift_down(0); + } + + priority_queue_node pop_value(bool remember_key=true) { + if(size() == 0) return priority_queue_node(-1, Priority()); + priority_queue_node ret = std::move(*heap.begin()); + id_to_heappos[ret.key] = -1-remember_key; + if(size() > 1) { + *heap.begin() = std::move(*(heap.end()-1)); + id_to_heappos[heap.front().key] = 0; + } + heap.pop_back(); + sift_down(0); + return ret; + } + + /** Sets the priority for the given key. If not present, it will be added, otherwise it will be updated + * Returns true if the priority was changed. + * */ + bool set(const Key& key, const Priority& priority, bool only_if_higher=false) { + if(key < id_to_heappos.size() && id_to_heappos[key] < ((size_t)-2)) // This key is already in the pQ + return update(key, priority, only_if_higher); + else + return push(key, priority, only_if_higher); + } + + std::pair get_priority(const Key& key) { + if(key < id_to_heappos.size()) { + size_t pos = id_to_heappos[key]; + if(pos < ((size_t)-2)) { + return {true, heap[pos].priority}; + } + } + return {false, 0}; + } + + /** Returns true if the key was not inside and was added, otherwise does nothing and returns false + * If the key was remembered and only_if_unknown is true, does nothing and returns false + * */ + bool push(const Key& key, const Priority& priority, bool only_if_unknown = false) { + extend_ids(key); + if (id_to_heappos[key] < ((size_t)-2)) return false; // already inside + if (only_if_unknown && id_to_heappos[key] == ((size_t)-2)) return false; // was evicted and only_if_unknown prevents re-adding + + if (heap.size() < max_size) { + // We have room: just add new element + size_t n = heap.size(); + id_to_heappos[key] = n; + heap.emplace_back(key, priority); + sift_up(n); + return true; + } else { + // Heap full: heap[0] is the smallest priority in the top-k (min-heap) + if (priority <= heap[0].priority) { + // New element priority too small or equal, discard it + return false; + } + // Evict smallest element at heap[0] + Key evicted_key = heap[0].key; + id_to_heappos[evicted_key] = -2; // Mark evicted + + heap[0] = priority_queue_node(key, priority); + id_to_heappos[key] = 0; + sift_down(0); // restore min-heap property + return true; + } + } + + + + /** Returns true if the key was already inside and was updated, otherwise does nothing and returns false */ + bool update(const Key& key, const Priority& new_priority, bool only_if_higher=false) { + if(key >= id_to_heappos.size()) return false; + size_t heappos = id_to_heappos[key]; + if(heappos >= ((size_t)-2)) return false; + Priority& priority = heap[heappos].priority; + if(new_priority > priority) { + priority = new_priority; + sift_up(heappos); + return true; + } + else if(!only_if_higher && new_priority < priority) { + priority = new_priority; + sift_down(heappos); + return true; + } + return false; + } + + void clear() { + heap.clear(); + id_to_heappos.clear(); + } + + + private: + void extend_ids(Key k) { + size_t new_size = k+1; + if(id_to_heappos.size() < new_size) + id_to_heappos.resize(new_size, -1); + } + + void sift_down(size_t heappos) { + size_t len = heap.size(); + size_t child = heappos*2+1; + if(len < 2 || child >= len) return; + if(child+1 < len && heap[child+1] > heap[child]) ++child; // Check whether second child is higher + if(!(heap[child] > heap[heappos])) return; // Already in heap order + + priority_queue_node val = std::move(heap[heappos]); + do { + heap[heappos] = std::move(heap[child]); + id_to_heappos[heap[heappos].key] = heappos; + heappos = child; + child = 2*child+1; + if(child >= len) break; + if(child+1 < len && heap[child+1] > heap[child]) ++child; + } while(heap[child] > val); + heap[heappos] = std::move(val); + id_to_heappos[heap[heappos].key] = heappos; + } + + void sift_up(size_t heappos) { + size_t len = heap.size(); + if(len < 2 || heappos <= 0) return; + size_t parent = (heappos-1)/2; + if(!(heap[heappos] > heap[parent])) return; + priority_queue_node val = std::move(heap[heappos]); + do { + heap[heappos] = std::move(heap[parent]); + id_to_heappos[heap[heappos].key] = heappos; + heappos = parent; + if(heappos <= 0) break; + parent = (parent-1)/2; + } while(val > heap[parent]); + heap[heappos] = std::move(val); + id_to_heappos[heap[heappos].key] = heappos; + } + }; +} \ No newline at end of file diff --git a/src/smt/smt_context.h b/src/smt/smt_context.h index 2fbc1d705..63316a331 100644 --- a/src/smt/smt_context.h +++ b/src/smt/smt_context.h @@ -50,6 +50,8 @@ Revision History: #include "model/model.h" #include "solver/progress_callback.h" #include "solver/assertions/asserted_formulas.h" +#include "smt/priority_queue.h" +#include "util/dlist.h" #include // there is a significant space overhead with allocating 1000+ contexts in @@ -189,6 +191,17 @@ namespace smt { unsigned_vector m_lit_occs; //!< occurrence count of literals svector m_bdata; //!< mapping bool_var -> data svector m_activity; + updatable_priority_queue::priority_queue m_pq_scores; + + struct lit_node : dll_base { + literal lit; + lit_node(literal l) : lit(l) { init(this); } + }; + lit_node* m_dll_lits; + + // svector> m_lit_scores; + svector m_lit_scores[2]; + clause_vector m_aux_clauses; clause_vector m_lemmas; vector m_clauses_to_reinit; @@ -933,6 +946,17 @@ namespace smt { ast_pp_util m_lemma_visitor; void dump_lemma(unsigned n, literal const* lits); void dump_axiom(unsigned n, literal const* lits); + void add_scores(unsigned n, literal const* lits); + void reset_scores() { + for (auto& e : m_lit_scores[0]) + e = 0; + for (auto& e : m_lit_scores[1]) + e = 0; + m_pq_scores.clear(); // Clear the priority queue heap as well + } + double get_score(literal l) const { + return m_lit_scores[l.sign()][l.var()]; + } public: void ensure_internalized(expr* e); diff --git a/src/smt/smt_internalizer.cpp b/src/smt/smt_internalizer.cpp index 9aa6d68f4..c7e257fac 100644 --- a/src/smt/smt_internalizer.cpp +++ b/src/smt/smt_internalizer.cpp @@ -931,6 +931,10 @@ namespace smt { set_bool_var(id, v); m_bdata.reserve(v+1); m_activity.reserve(v+1); + m_lit_scores[0].reserve(v + 1); + m_lit_scores[1].reserve(v + 1); + + m_lit_scores[0][v] = m_lit_scores[1][v] = 0.0; m_bool_var2expr.reserve(v+1); m_bool_var2expr[v] = n; literal l(v, false); @@ -1419,6 +1423,7 @@ namespace smt { break; case CLS_LEARNED: dump_lemma(num_lits, lits); + add_scores(num_lits, lits); break; default: break; @@ -1527,6 +1532,27 @@ namespace smt { }} } + // void context::add_scores(unsigned n, literal const* lits) { + // for (unsigned i = 0; i < n; ++i) { + // auto lit = lits[i]; + // unsigned v = lit.var(); + // m_lit_scores[v][lit.sign()] += 1.0 / n; + // } + // } + + void context::add_scores(unsigned n, literal const* lits) { + for (unsigned i = 0; i < n; ++i) { + auto lit = lits[i]; + unsigned v = lit.var(); // unique key per literal + + m_lit_scores[lit.sign()][v] += 1.0 / n; + + auto new_score = m_lit_scores[0][v] * m_lit_scores[1][v]; + m_pq_scores.set(v, new_score); + + } + } + void context::dump_axiom(unsigned n, literal const* lits) { if (m_fparams.m_axioms2files) { literal_buffer tmp; diff --git a/src/smt/smt_lookahead.h b/src/smt/smt_lookahead.h index 5deccad2c..d53af58e4 100644 --- a/src/smt/smt_lookahead.h +++ b/src/smt/smt_lookahead.h @@ -30,11 +30,13 @@ namespace smt { struct compare; - double get_score(); + // double get_score(); void choose_rec(expr_ref_vector& trail, expr_ref_vector& result, unsigned depth, unsigned budget); public: + double get_score(); + lookahead(context& ctx); expr_ref choose(unsigned budget = 2000); diff --git a/src/smt/smt_parallel.cpp b/src/smt/smt_parallel.cpp index 4941e4df9..d08a7c045 100644 --- a/src/smt/smt_parallel.cpp +++ b/src/smt/smt_parallel.cpp @@ -36,237 +36,463 @@ namespace smt { #else #include +#include namespace smt { + + void parallel::worker::run() { + ast_translation g2l(p.ctx.m, m); // global to local context -- MUST USE p.ctx.m, not ctx->m, AS GLOBAL MANAGER!!! + ast_translation l2g(m, p.ctx.m); // local to global context + while (m.inc()) { // inc: increase the limit and check if it is canceled, vs m.limit().is_canceled() is readonly. the .limit() is also not necessary (m.inc() etc provides a convenience wrapper) + vector cubes; + b.get_cubes(g2l, cubes); + if (cubes.empty()) + return; + collect_shared_clauses(g2l); + for (auto& cube : cubes) { + if (!m.inc()) { + b.set_exception("context cancelled"); + return; + } + IF_VERBOSE(1, verbose_stream() << "Worker " << id << " cube: " << mk_bounded_pp(mk_and(cube), m, 3) << "\n"); + lbool r = check_cube(cube); + if (m.limit().is_canceled()) { + IF_VERBOSE(1, verbose_stream() << "Worker " << id << " context cancelled\n"); + return; + } + switch (r) { + case l_undef: { + IF_VERBOSE(1, verbose_stream() << "Worker " << id << " found undef cube\n"); + // return unprocessed cubes to the batch manager + // add a split literal to the batch manager. + // optionally process other cubes and delay sending back unprocessed cubes to batch manager. + vector returned_cubes; + returned_cubes.push_back(cube); + auto split_atoms = get_split_atoms(); + b.return_cubes(l2g, returned_cubes, split_atoms); + update_max_thread_conflicts(); + break; + } + case l_true: { + IF_VERBOSE(1, verbose_stream() << "Worker " << id << " found sat cube\n"); + model_ref mdl; + ctx->get_model(mdl); + b.set_sat(l2g, *mdl); + return; + } + case l_false: { + // if unsat core only contains (external) assumptions (i.e. all the unsat core are asms), then unsat and return as this does NOT depend on cubes + // otherwise, extract lemmas that can be shared (units (and unsat core?)). + // share with batch manager. + // process next cube. + expr_ref_vector const& unsat_core = ctx->unsat_core(); + IF_VERBOSE(1, verbose_stream() << "unsat core: " << unsat_core << "\n"); + // If the unsat core only contains assumptions, + // unsatisfiability does not depend on the current cube and the entire problem is unsat. + if (all_of(unsat_core, [&](expr* e) { return asms.contains(e); })) { + IF_VERBOSE(1, verbose_stream() << "Worker " << id << " determined formula unsat\n"); + b.set_unsat(l2g, unsat_core); + return; + } + for (expr* e : unsat_core) + if (asms.contains(e)) + b.report_assumption_used(l2g, e); // report assumptions used in unsat core, so they can be used in final core + + IF_VERBOSE(1, verbose_stream() << "Worker " << id << " found unsat cube\n"); + b.collect_clause(l2g, id, mk_not(mk_and(unsat_core))); + break; + } + } + } + share_units(l2g); + } + } + + parallel::worker::worker(unsigned id, parallel& p, expr_ref_vector const& _asms): id(id), p(p), b(p.m_batch_manager), m_smt_params(p.ctx.get_fparams()), asms(m) { + ast_translation g2l(p.ctx.m, m); + for (auto e : _asms) + asms.push_back(g2l(e)); + IF_VERBOSE(1, verbose_stream() << "Worker " << id << " created with " << asms.size() << " assumptions\n"); + m_smt_params.m_preprocess = false; + ctx = alloc(context, m, m_smt_params, p.ctx.get_params()); + context::copy(p.ctx, *ctx, true); + ctx->set_random_seed(id + m_smt_params.m_random_seed); + + m_max_thread_conflicts = ctx->get_fparams().m_threads_max_conflicts; + m_max_conflicts = ctx->get_fparams().m_max_conflicts; + } + + void parallel::worker::share_units(ast_translation& l2g) { + // Collect new units learned locally by this worker and send to batch manager + unsigned sz = ctx->assigned_literals().size(); + for (unsigned j = m_num_shared_units; j < sz; ++j) { // iterate only over new literals since last sync + literal lit = ctx->assigned_literals()[j]; + expr_ref e(ctx->bool_var2expr(lit.var()), ctx->m); // turn literal into a Boolean expression + if (lit.sign()) + e = m.mk_not(e); // negate if literal is negative + b.collect_clause(l2g, id, e); + } + m_num_shared_units = sz; + } + + void parallel::batch_manager::collect_clause(ast_translation& l2g, unsigned source_worker_id, expr* clause) { + std::scoped_lock lock(mux); + expr* g_clause = l2g(clause); + if (!shared_clause_set.contains(g_clause)) { + shared_clause_set.insert(g_clause); + shared_clause sc{source_worker_id, expr_ref(g_clause, m)}; + shared_clause_trail.push_back(sc); + } + } + + void parallel::worker::collect_shared_clauses(ast_translation& g2l) { + expr_ref_vector new_clauses = b.return_shared_clauses(g2l, m_shared_clause_limit, id); // get new clauses from the batch manager + // iterate over new clauses and assert them in the local context + for (expr* e : new_clauses) { + expr_ref local_clause(e, g2l.to()); // e was already translated to the local context in the batch manager!! + ctx->assert_expr(local_clause); // assert the clause in the local context + IF_VERBOSE(1, verbose_stream() << "Worker " << id << " asserting shared clause: " << mk_bounded_pp(local_clause, m, 3) << "\n"); + } + } + + // get new clauses from the batch manager and assert them in the local context + expr_ref_vector parallel::batch_manager::return_shared_clauses(ast_translation& g2l, unsigned& worker_limit, unsigned worker_id) { + std::scoped_lock lock(mux); + expr_ref_vector result(g2l.to()); + for (unsigned i = worker_limit; i < shared_clause_trail.size(); ++i) { + if (shared_clause_trail[i].source_worker_id == worker_id) + continue; // skip clauses from the requesting worker + result.push_back(g2l(shared_clause_trail[i].clause.get())); + } + worker_limit = shared_clause_trail.size(); // update the worker limit to the end of the current trail + return result; + } + + lbool parallel::worker::check_cube(expr_ref_vector const& cube) { + IF_VERBOSE(1, verbose_stream() << "Worker " << id << " checking cube\n";); + for (auto& atom : cube) + asms.push_back(atom); + lbool r = l_undef; + + ctx->get_fparams().m_max_conflicts = std::min(m_max_thread_conflicts, m_max_conflicts); + try { + r = ctx->check(asms.size(), asms.data()); + } + catch (z3_error& err) { + b.set_exception(err.error_code()); + } + catch (z3_exception& ex) { + b.set_exception(ex.what()); + } + catch (...) { + b.set_exception("unknown exception"); + } + asms.shrink(asms.size() - cube.size()); + IF_VERBOSE(1, verbose_stream() << "Worker " << id << " DONE checking cube " << r << "\n";); + return r; + } + + void parallel::batch_manager::get_cubes(ast_translation& g2l, vector& cubes) { + std::scoped_lock lock(mux); + if (m_cubes.size() == 1 && m_cubes[0].size() == 0) { + // special initialization: the first cube is emtpy, have the worker work on an empty cube. + cubes.push_back(expr_ref_vector(g2l.to())); + return; + } + + for (unsigned i = 0; i < std::min(m_max_batch_size / p.num_threads, (unsigned)m_cubes.size()) && !m_cubes.empty(); ++i) { + auto& cube = m_cubes.back(); + expr_ref_vector l_cube(g2l.to()); + for (auto& e : cube) { + l_cube.push_back(g2l(e)); + } + cubes.push_back(l_cube); + m_cubes.pop_back(); + } + } + + void parallel::batch_manager::set_sat(ast_translation& l2g, model& m) { + std::scoped_lock lock(mux); + if (m_state != state::is_running) + return; + m_state = state::is_sat; + p.ctx.set_model(m.translate(l2g)); + cancel_workers(); + } + + void parallel::batch_manager::set_unsat(ast_translation& l2g, expr_ref_vector const& unsat_core) { + std::scoped_lock lock(mux); + if (m_state != state::is_running) + return; + m_state = state::is_unsat; + + // every time we do a check_sat call, don't want to have old info coming from a prev check_sat call + // the unsat core gets reset internally in the context after each check_sat, so we assert this property here + // takeaway: each call to check_sat needs to have a fresh unsat core + SASSERT(p.ctx.m_unsat_core.empty()); + for (expr* e : unsat_core) + p.ctx.m_unsat_core.push_back(l2g(e)); + cancel_workers(); + } + + void parallel::batch_manager::set_exception(unsigned error_code) { + std::scoped_lock lock(mux); + if (m_state != state::is_running) + return; + m_state = state::is_exception_code; + m_exception_code = error_code; + cancel_workers(); + } + + void parallel::batch_manager::set_exception(std::string const& msg) { + std::scoped_lock lock(mux); + if (m_state != state::is_running || m.limit().is_canceled()) + return; + m_state = state::is_exception_msg; + m_exception_msg = msg; + cancel_workers(); + } + + void parallel::batch_manager::report_assumption_used(ast_translation& l2g, expr* assumption) { + std::scoped_lock lock(mux); + p.m_assumptions_used.insert(l2g(assumption)); + } + + lbool parallel::batch_manager::get_result() const { + if (m.limit().is_canceled()) + return l_undef; // the main context was cancelled, so we return undef. + switch (m_state) { + case state::is_running: // batch manager is still running, but all threads have processed their cubes, which means all cubes were unsat + if (!m_cubes.empty()) + throw default_exception("inconsistent end state"); + if (!p.m_assumptions_used.empty()) { + // collect unsat core from assumptions used, if any --> case when all cubes were unsat, but depend on nonempty asms, so we need to add these asms to final unsat core + SASSERT(p.ctx.m_unsat_core.empty()); + for (auto a : p.m_assumptions_used) + p.ctx.m_unsat_core.push_back(a); + } + return l_false; + case state::is_unsat: + return l_false; + case state::is_sat: + return l_true; + case state::is_exception_msg: + throw default_exception(m_exception_msg.c_str()); + case state::is_exception_code: + throw z3_error(m_exception_code); + default: + UNREACHABLE(); + return l_undef; + } + } + + /* + Batch manager maintains C_batch, A_batch. + C_batch - set of cubes + A_batch - set of split atoms. + return_cubes is called with C_batch A_batch C A. + C_worker - one or more cubes + A_worker - split atoms form the worker thread. - lbool parallel::operator()(expr_ref_vector const& asms) { + Assumption: A_worker does not occur in C_worker. + + ------------------------------------------------------------------------------------------------------------------------------------------------------------ + Greedy strategy: + For each returned cube c from the worker, you split it on all split atoms not in it (i.e., A_batch \ atoms(c)), plus any new atoms from A_worker. + For each existing cube in the batch, you also split it on the new atoms from A_worker. + + return_cubes C_batch A_batch C_worker A_worker: + C_batch <- { cube * 2^(A_worker u (A_batch \ atoms(cube)) | cube in C_worker } u + { cube * 2^(A_worker \ A_batch) | cube in C_batch } + = + let C_batch' = C_batch u { cube * 2^(A_batch \ atoms(cube)) | cube in C_worker } + in { cube * 2^(A_worker \ A_batch) | cube in C_batch' } + A_batch <- A_batch u A_worker + + ------------------------------------------------------------------------------------------------------------------------------------------------------------ + Frugal strategy: only split on worker cubes + + case 1: thread returns no cubes, just atoms: just create 2^k cubes from all combinations of atoms so far. + return_cubes C_batch A_batch [[]] A_worker: + C_batch <- C_batch u 2^(A_worker u A_batch), + A_batch <- A_batch u A_worker + + case 2: thread returns both cubes and atoms + Only the returned cubes get split by the newly discovered atoms (A_worker). Existing cubes are not touched. + return_cubes C_batch A_batch C_worker A_worker: + C_batch <- C_batch u { cube * 2^A_worker | cube in C_worker }. + A_batch <- A_batch u A_worker + + This means: + Only the returned cubes get split by the newly discovered atoms (A_worker). + Existing cubes are not touched. + + ------------------------------------------------------------------------------------------------------------------------------------------------------------ + Hybrid: Between Frugal and Greedy: (generalizes the first case of empty cube returned by worker) -- don't focus on this approach + i.e. Expand only the returned cubes, but allow them to be split on both new and old atoms not already in them. + + C_batch <- C_batch u { cube * 2^(A_worker u (A_batch \ atoms(cube)) | cube in C_worker } + A_batch <- A_batch u A_worker + + ------------------------------------------------------------------------------------------------------------------------------------------------------------ + Final thought (do this!): use greedy strategy by a policy when C_batch, A_batch, A_worker are "small". -- want to do this. switch to frugal strategy after reaching size limit + */ - lbool result = l_undef; - unsigned num_threads = std::min((unsigned) std::thread::hardware_concurrency(), ctx.get_fparams().m_threads); - flet _nt(ctx.m_fparams.m_threads, 1); - unsigned thread_max_conflicts = ctx.get_fparams().m_threads_max_conflicts; - unsigned max_conflicts = ctx.get_fparams().m_max_conflicts; - - // try first sequential with a low conflict budget to make super easy problems cheap - unsigned max_c = std::min(thread_max_conflicts, 40u); - flet _mc(ctx.get_fparams().m_max_conflicts, max_c); - result = ctx.check(asms.size(), asms.data()); - if (result != l_undef || ctx.m_num_conflicts < max_c) { - return result; - } - - enum par_exception_kind { - DEFAULT_EX, - ERROR_EX + // currenly, the code just implements the greedy strategy + void parallel::batch_manager::return_cubes(ast_translation& l2g, vectorconst& C_worker, expr_ref_vector const& A_worker) { + auto atom_in_cube = [&](expr_ref_vector const& cube, expr* atom) { + return any_of(cube, [&](expr* e) { return e == atom || (m.is_not(e, e) && e == atom); }); }; - vector smt_params; - scoped_ptr_vector pms; - scoped_ptr_vector pctxs; - vector pasms; + auto add_split_atom = [&](expr* atom, unsigned start) { + unsigned stop = m_cubes.size(); + for (unsigned i = start; i < stop; ++i) { + m_cubes.push_back(m_cubes[i]); + m_cubes.back().push_back(m.mk_not(atom)); + m_cubes[i].push_back(atom); + } + }; + std::scoped_lock lock(mux); + unsigned max_cubes = 1000; + bool greedy_mode = (m_cubes.size() <= max_cubes); + unsigned a_worker_start_idx = 0; + + // + // --- Phase 1: Greedy split of *existing* cubes on new A_worker atoms (greedy) --- + // + if (greedy_mode) { + for (; a_worker_start_idx < A_worker.size(); ++a_worker_start_idx) { + expr_ref g_atom(l2g(A_worker[a_worker_start_idx]), l2g.to()); + if (m_split_atoms.contains(g_atom)) + continue; + m_split_atoms.push_back(g_atom); + + add_split_atom(g_atom, 0); // split all *existing* cubes + if (m_cubes.size() > max_cubes) { + greedy_mode = false; + ++a_worker_start_idx; // start frugal from here + break; + } + } + } + + unsigned initial_m_cubes_size = m_cubes.size(); // where to start processing the worker cubes after splitting the EXISTING cubes on the new worker atoms + + // --- Phase 2: Process worker cubes (greedy) --- + for (auto& c : C_worker) { + expr_ref_vector g_cube(l2g.to()); + for (auto& atom : c) + g_cube.push_back(l2g(atom)); + + unsigned start = m_cubes.size(); // update start after adding each cube so we only process the current cube being added + m_cubes.push_back(g_cube); + + if (greedy_mode) { + // Split new cube on all existing m_split_atoms not in it + for (auto g_atom : m_split_atoms) { + if (!atom_in_cube(g_cube, g_atom)) { + add_split_atom(g_atom, start); + if (m_cubes.size() > max_cubes) { + greedy_mode = false; + break; + } + } + } + } + } + + // --- Phase 3: Frugal fallback: only process NEW worker cubes with NEW atoms --- + if (!greedy_mode) { + for (unsigned i = a_worker_start_idx; i < A_worker.size(); ++i) { + expr_ref g_atom(l2g(A_worker[i]), l2g.to()); + if (!m_split_atoms.contains(g_atom)) + m_split_atoms.push_back(g_atom); + add_split_atom(g_atom, initial_m_cubes_size); + } + } + } + + expr_ref_vector parallel::worker::get_split_atoms() { + unsigned k = 2; + + auto candidates = ctx->m_pq_scores.get_heap(); + + std::sort(candidates.begin(), candidates.end(), + [](const auto& a, const auto& b) { return a.priority > b.priority; }); + + expr_ref_vector top_lits(m); + for (const auto& node: candidates) { + if (ctx->get_assignment(node.key) != l_undef) + continue; + + expr* e = ctx->bool_var2expr(node.key); + if (!e) + continue; + + top_lits.push_back(expr_ref(e, m)); + if (top_lits.size() >= k) + break; + } + IF_VERBOSE(1, verbose_stream() << "top literals " << top_lits << " head size " << ctx->m_pq_scores.size() << " num vars " << ctx->get_num_bool_vars() << "\n"); + return top_lits; + } + + void parallel::batch_manager::initialize() { + m_state = state::is_running; + m_cubes.reset(); + m_cubes.push_back(expr_ref_vector(m)); // push empty cube + m_split_atoms.reset(); + } + + lbool parallel::operator()(expr_ref_vector const& asms) { ast_manager& m = ctx.m; - scoped_limits sl(m.limit()); - unsigned finished_id = UINT_MAX; - std::string ex_msg; - par_exception_kind ex_kind = DEFAULT_EX; - unsigned error_code = 0; - bool done = false; - unsigned num_rounds = 0; + if (m.has_trace_stream()) throw default_exception("trace streams have to be off in parallel mode"); - - params_ref params = ctx.get_params(); - for (unsigned i = 0; i < num_threads; ++i) { - smt_params.push_back(ctx.get_fparams()); - smt_params.back().m_preprocess = false; - } - - for (unsigned i = 0; i < num_threads; ++i) { - ast_manager* new_m = alloc(ast_manager, m, true); - pms.push_back(new_m); - pctxs.push_back(alloc(context, *new_m, smt_params[i], params)); - context& new_ctx = *pctxs.back(); - context::copy(ctx, new_ctx, true); - new_ctx.set_random_seed(i + ctx.get_fparams().m_random_seed); - ast_translation tr(m, *new_m); - pasms.push_back(tr(asms)); - sl.push_child(&(new_m->limit())); - } - - auto cube = [](context& ctx, expr_ref_vector& lasms, expr_ref& c) { - lookahead lh(ctx); - c = lh.choose(); - if (c) { - if ((ctx.get_random_value() % 2) == 0) - c = c.get_manager().mk_not(c); - lasms.push_back(c); - } + struct scoped_clear_table { + obj_hashtable& ht; + scoped_clear_table(obj_hashtable& ht) : ht(ht) {} // Constructor: Takes a reference to a hash table when the object is created and saves it. + ~scoped_clear_table() { ht.reset(); } // Destructor: When the scoped_clear_table object goes out of scope, it automatically calls reset() on that hash table, clearing it }; + scoped_clear_table clear(m_assumptions_used); // creates a scoped_clear_table named clear, bound to m_assumptions_used - obj_hashtable unit_set; - expr_ref_vector unit_trail(ctx.m); - unsigned_vector unit_lim; - for (unsigned i = 0; i < num_threads; ++i) unit_lim.push_back(0); - - std::function collect_units = [&,this]() { - //return; -- has overhead - for (unsigned i = 0; i < num_threads; ++i) { - context& pctx = *pctxs[i]; - pctx.pop_to_base_lvl(); - ast_translation tr(pctx.m, ctx.m); - unsigned sz = pctx.assigned_literals().size(); - for (unsigned j = unit_lim[i]; j < sz; ++j) { - literal lit = pctx.assigned_literals()[j]; - //IF_VERBOSE(0, verbose_stream() << "(smt.thread " << i << " :unit " << lit << " " << pctx.is_relevant(lit.var()) << ")\n";); - if (!pctx.is_relevant(lit.var())) - continue; - expr_ref e(pctx.bool_var2expr(lit.var()), pctx.m); - if (lit.sign()) e = pctx.m.mk_not(e); - expr_ref ce(tr(e.get()), ctx.m); - if (!unit_set.contains(ce)) { - unit_set.insert(ce); - unit_trail.push_back(ce); - } - } - } - - unsigned sz = unit_trail.size(); - for (unsigned i = 0; i < num_threads; ++i) { - context& pctx = *pctxs[i]; - ast_translation tr(ctx.m, pctx.m); - for (unsigned j = unit_lim[i]; j < sz; ++j) { - expr_ref src(ctx.m), dst(pctx.m); - dst = tr(unit_trail.get(j)); - pctx.assert_expr(dst); - } - unit_lim[i] = pctx.assigned_literals().size(); - } - IF_VERBOSE(1, verbose_stream() << "(smt.thread :units " << sz << ")\n"); - }; - - std::mutex mux; - - auto worker_thread = [&](int i) { - try { - context& pctx = *pctxs[i]; - ast_manager& pm = *pms[i]; - expr_ref_vector lasms(pasms[i]); - expr_ref c(pm); - - pctx.get_fparams().m_max_conflicts = std::min(thread_max_conflicts, max_conflicts); - if (num_rounds > 0 && (num_rounds % pctx.get_fparams().m_threads_cube_frequency) == 0) - cube(pctx, lasms, c); - IF_VERBOSE(1, verbose_stream() << "(smt.thread " << i; - if (num_rounds > 0) verbose_stream() << " :round " << num_rounds; - if (c) verbose_stream() << " :cube " << mk_bounded_pp(c, pm, 3); - verbose_stream() << ")\n";); - lbool r = pctx.check(lasms.size(), lasms.data()); - - if (r == l_undef && pctx.m_num_conflicts >= max_conflicts) - ; // no-op - else if (r == l_undef && pctx.m_num_conflicts >= thread_max_conflicts) - return; - else if (r == l_false && pctx.unsat_core().contains(c)) { - IF_VERBOSE(1, verbose_stream() << "(smt.thread " << i << " :learn " << mk_bounded_pp(c, pm, 3) << ")"); - pctx.assert_expr(mk_not(mk_and(pctx.unsat_core()))); - return; - } + { + m_batch_manager.initialize(); + m_workers.reset(); + scoped_limits sl(m.limit()); + flet _nt(ctx.m_fparams.m_threads, 1); + SASSERT(num_threads > 1); + for (unsigned i = 0; i < num_threads; ++i) + m_workers.push_back(alloc(worker, i, *this, asms)); // i.e. "new worker(i, *this, asms)" + // THIS WILL ALLOW YOU TO CANCEL ALL THE CHILD THREADS + // within the lexical scope of the code block, creates a data structure that allows you to push children + // objects to the limit object, so if someone cancels the parent object, the cancellation propagates to the children + // and that cancellation has the lifetime of the scope + // even if this code doesn't expliclty kill the main thread, still applies bc if you e.g. Ctrl+C the main thread, the children threads need to be cancelled + for (auto w : m_workers) + sl.push_child(&(w->limit())); - bool first = false; - { - std::lock_guard lock(mux); - if (finished_id == UINT_MAX) { - finished_id = i; - first = true; - result = r; - done = true; - } - if (!first && r != l_undef && result == l_undef) { - finished_id = i; - result = r; - } - else if (!first) return; - } - - for (ast_manager* m : pms) { - if (m != &pm) m->limit().cancel(); - } - - } - catch (z3_error & err) { - if (finished_id == UINT_MAX) { - error_code = err.error_code(); - ex_kind = ERROR_EX; - done = true; - } - } - catch (z3_exception & ex) { - if (finished_id == UINT_MAX) { - ex_msg = ex.what(); - ex_kind = DEFAULT_EX; - done = true; - } - } - catch (...) { - if (finished_id == UINT_MAX) { - ex_msg = "unknown exception"; - ex_kind = ERROR_EX; - done = true; - } - } - }; - - // for debugging: num_threads = 1; - - while (true) { + // Launch threads vector threads(num_threads); for (unsigned i = 0; i < num_threads; ++i) { - threads[i] = std::thread([&, i]() { worker_thread(i); }); + threads[i] = std::thread([&, i]() { + m_workers[i]->run(); + }); } - for (auto & th : threads) { + + // Wait for all threads to finish + for (auto& th : threads) th.join(); - } - if (done) break; - collect_units(); - ++num_rounds; - max_conflicts = (max_conflicts < thread_max_conflicts) ? 0 : (max_conflicts - thread_max_conflicts); - thread_max_conflicts *= 2; + for (auto w : m_workers) + w->collect_statistics(ctx.m_aux_stats); } - for (context* c : pctxs) { - c->collect_statistics(ctx.m_aux_stats); - } - - if (finished_id == UINT_MAX) { - switch (ex_kind) { - case ERROR_EX: throw z3_error(error_code); - default: throw default_exception(std::move(ex_msg)); - } - } - - model_ref mdl; - context& pctx = *pctxs[finished_id]; - ast_translation tr(*pms[finished_id], m); - switch (result) { - case l_true: - pctx.get_model(mdl); - if (mdl) - ctx.set_model(mdl->translate(tr)); - break; - case l_false: - ctx.m_unsat_core.reset(); - for (expr* e : pctx.unsat_core()) - ctx.m_unsat_core.push_back(tr(e)); - break; - default: - break; - } - - return result; + m_workers.clear(); + return m_batch_manager.get_result(); // i.e. all threads have finished all of their cubes -- so if state::is_running is still true, means the entire formula is unsat (otherwise a thread would have returned l_undef) } } diff --git a/src/smt/smt_parallel.h b/src/smt/smt_parallel.h index 07b04019d..b337d5e45 100644 --- a/src/smt/smt_parallel.h +++ b/src/smt/smt_parallel.h @@ -19,16 +19,124 @@ Revision History: #pragma once #include "smt/smt_context.h" +#include namespace smt { class parallel { context& ctx; + unsigned num_threads; + + struct shared_clause { + unsigned source_worker_id; + expr_ref clause; + }; + + class batch_manager { + enum state { + is_running, + is_sat, + is_unsat, + is_exception_msg, + is_exception_code + }; + + ast_manager& m; + parallel& p; + std::mutex mux; + state m_state = state::is_running; + expr_ref_vector m_split_atoms; // atoms to split on + vector m_cubes; + unsigned m_max_batch_size = 10; + unsigned m_exception_code = 0; + std::string m_exception_msg; + vector shared_clause_trail; // store all shared clauses with worker IDs + obj_hashtable shared_clause_set; // for duplicate filtering on per-thread clause expressions + + // called from batch manager to cancel other workers if we've reached a verdict + void cancel_workers() { + IF_VERBOSE(1, verbose_stream() << "Canceling workers\n"); + for (auto& w : p.m_workers) + w->cancel(); + } + + public: + batch_manager(ast_manager& m, parallel& p) : m(m), p(p), m_split_atoms(m) { } + + void initialize(); + + void set_unsat(ast_translation& l2g, expr_ref_vector const& unsat_core); + void set_sat(ast_translation& l2g, model& m); + void set_exception(std::string const& msg); + void set_exception(unsigned error_code); + + // + // worker threads ask the batch manager for a supply of cubes to check. + // they pass in a translation function from the global context to local context (ast-manager). It is called g2l. + // The batch manager returns a list of cubes to solve. + // + void get_cubes(ast_translation& g2l, vector& cubes); + + // + // worker threads return unprocessed cubes to the batch manager together with split literal candidates. + // the batch manager re-enqueues unprocessed cubes and optionally splits them using the split_atoms returned by this and workers. + // + void return_cubes(ast_translation& l2g, vectorconst& cubes, expr_ref_vector const& split_atoms); + void report_assumption_used(ast_translation& l2g, expr* assumption); + void collect_clause(ast_translation& l2g, unsigned source_worker_id, expr* e); + expr_ref_vector return_shared_clauses(ast_translation& g2l, unsigned& worker_limit, unsigned worker_id); + lbool get_result() const; + }; + + class worker { + unsigned id; // unique identifier for the worker + parallel& p; + batch_manager& b; + ast_manager m; + expr_ref_vector asms; + smt_params m_smt_params; + scoped_ptr ctx; + unsigned m_max_conflicts = 800; // the global budget for all work this worker can do across cubes in the current run. + unsigned m_max_thread_conflicts = 100; // the per-cube limit for how many conflicts the worker can spend on a single cube before timing out on it and moving on + unsigned m_num_shared_units = 0; + unsigned m_shared_clause_limit = 0; // remembers the index into shared_clause_trail marking the boundary between "old" and "new" clauses to share + void share_units(ast_translation& l2g); + lbool check_cube(expr_ref_vector const& cube); + void update_max_thread_conflicts() { + m_max_thread_conflicts *= 2; + } // allow for backoff scheme of conflicts within the thread for cube timeouts. + public: + worker(unsigned id, parallel& p, expr_ref_vector const& _asms); + void run(); + expr_ref_vector get_split_atoms(); + void collect_shared_clauses(ast_translation& g2l); + + void cancel() { + IF_VERBOSE(1, verbose_stream() << "Worker " << id << " canceling\n"); + m.limit().cancel(); + } + void collect_statistics(::statistics& st) const { + IF_VERBOSE(1, verbose_stream() << "Collecting statistics for worker " << id << "\n"); + ctx->collect_statistics(st); + } + reslimit& limit() { + return m.limit(); + } + }; + + obj_hashtable m_assumptions_used; // assumptions used in unsat cores, to be used in final core + batch_manager m_batch_manager; + ptr_vector m_workers; + public: - parallel(context& ctx): ctx(ctx) {} + parallel(context& ctx) : + ctx(ctx), + num_threads(std::min( + (unsigned)std::thread::hardware_concurrency(), + ctx.get_fparams().m_threads)), + m_batch_manager(ctx.m, *this) {} lbool operator()(expr_ref_vector const& asms); - }; }