xref: /llvm-project/mlir/lib/Conversion/PDLToPDLInterp/RootOrdering.h (revision 9eb8e7b176e9fc38c8df86bd927663c6409ac262)
1 //===- RootOrdering.h - Optimal root ordering  ------------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains definition for a cost graph over candidate roots and
10 // an implementation of an algorithm to determine the optimal ordering over
11 // these roots. Each edge in this graph indicates that the target root can be
12 // connected (via a chain of positions) to the source root, and their cost
13 // indicates the estimated cost of such traversal. The optimal root ordering
14 // is then formulated as that of finding a spanning arborescence (i.e., a
15 // directed spanning tree) of minimal weight.
16 //
17 //===----------------------------------------------------------------------===//
18 
19 #ifndef MLIR_LIB_CONVERSION_PDLTOPDLINTERP_ROOTORDERING_H_
20 #define MLIR_LIB_CONVERSION_PDLTOPDLINTERP_ROOTORDERING_H_
21 
22 #include "mlir/IR/Value.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/SmallVector.h"
25 #include <functional>
26 #include <vector>
27 
28 namespace mlir {
29 namespace pdl_to_pdl_interp {
30 
31 /// The information associated with an edge in the cost graph. Each node in
32 /// the cost graph corresponds to a candidate root detected in the pdl.pattern,
33 /// and each edge in the cost graph corresponds to connecting the two candidate
34 /// roots via a chain of operations. The cost of an edge is the smallest number
35 /// of upward traversals required to go from the source to the target root, and
36 /// the connector is a `Value` in the intersection of the two subtrees rooted at
37 /// the source and target root that results in that smallest number of upward
38 /// traversals. Consider the following pattern with 3 roots op3, op4, and op5:
39 ///
40 ///                 argA ---> op1 ---> op2 ---> op3 ---> res3
41 ///                            ^        ^
42 ///                            |        |
43 ///                           argB     argC
44 ///                            |        |
45 ///                            v        v
46 ///                 res4 <--- op4      op5 ---> res5
47 ///                            ^        ^
48 ///                            |        |
49 ///                           op6      op7
50 ///
51 /// The cost of the edge op3 -> op4 is 1 (the upward traversal argB -> op4),
52 /// with argB being the connector `Value` and similarly for op3 -> op5 (cost 1,
53 /// connector argC). The cost of the edge op4 -> op3 is 3 (upward traversals
54 /// argB -> op1 -> op2 -> op3, connector argB), while the cost of edge op5 ->
55 /// op3 is 2 (uwpard traversals argC -> op2 -> op3). There are no edges between
56 /// op4 and op5 in the cost graph, because the subtrees rooted at these two
57 /// roots do not intersect. It is easy to see that the optimal root for this
58 /// pattern is op3, resulting in the spanning arborescence op3 -> {op4, op5}.
59 struct RootOrderingEntry {
60   /// The depth of the connector `Value` w.r.t. the target root.
61   ///
62   /// This is a pair where the first value is the additive cost (the depth of
63   /// the connector), and the second value is a priority for breaking ties
64   /// (with 0 being the highest). Typically, the priority is a unique edge ID.
65   std::pair<unsigned, unsigned> cost;
66 
67   /// The connector value in the intersection of the two subtrees rooted at
68   /// the source and target root that results in that smallest depth w.r.t.
69   /// the target root.
70   Value connector;
71 };
72 
73 /// A directed graph representing the cost of ordering the roots in the
74 /// predicate tree. It is represented as an adjacency map, where the outer map
75 /// is indexed by the target node, and the inner map is indexed by the source
76 /// node. Each edge is associated with a cost and the underlying connector
77 /// value.
78 using RootOrderingGraph = DenseMap<Value, DenseMap<Value, RootOrderingEntry>>;
79 
80 /// The optimal branching algorithm solver. This solver accepts a graph and the
81 /// root in its constructor, and is invoked via the solve() member function.
82 /// This is a direct implementation of the Edmonds' algorithm, see
83 /// https://en.wikipedia.org/wiki/Edmonds%27_algorithm. The worst-case
84 /// computational complexity of this algorithm is O(N^3), for a single root.
85 /// The PDL-to-PDLInterp lowering calls this N times (once for each candidate
86 /// root), so the overall complexity root ordering is O(N^4). If needed, this
87 /// could be reduced to O(N^3) with a more efficient algorithm. However, note
88 /// that the underlying implementation is very efficient, and N in our
89 /// instances tends to be very small (<10).
90 class OptimalBranching {
91 public:
92   /// A list of edges (child, parent).
93   using EdgeList = std::vector<std::pair<Value, Value>>;
94 
95   /// Constructs the solver for the given graph and root value.
96   OptimalBranching(RootOrderingGraph graph, Value root);
97 
98   /// Runs the Edmonds' algorithm for the current `graph`, returning the total
99   /// cost of the minimum-weight spanning arborescence (sum of the edge costs).
100   /// This function first determines the optimal local choice of the parents
101   /// and stores this choice in the `parents` mapping. If this choice results
102   /// in an acyclic graph, the function returns immediately. Otherwise, it
103   /// takes an arbitrary cycle, contracts it, and recurses on the new graph
104   /// (which is guaranteed to have fewer nodes than we began with). After we
105   /// return from recursion, we redirect the edges to/from the contracted node,
106   /// so the `parents` map contains a valid solution for the current graph.
107   unsigned solve();
108 
109   /// Returns the computed parent map. This is the unique predecessor for each
110   /// node (root) in the optimal branching.
getRootOrderingParents()111   const DenseMap<Value, Value> &getRootOrderingParents() const {
112     return parents;
113   }
114 
115   /// Returns the computed edges as visited in the preorder traversal.
116   /// The specified array determines the order for breaking any ties.
117   EdgeList preOrderTraversal(ArrayRef<Value> nodes) const;
118 
119 private:
120   /// The graph whose optimal branching we wish to determine.
121   RootOrderingGraph graph;
122 
123   /// The root of the optimal branching.
124   Value root;
125 
126   /// The computed parent mapping. This is the unique predecessor for each node
127   /// in the optimal branching. The keys of this map correspond to the keys of
128   /// the outer map of the input graph, and each value is one of the keys of
129   /// the inner map for this node. Also used as an intermediate (possibly
130   /// cyclical) result in the optimal branching algorithm.
131   DenseMap<Value, Value> parents;
132 };
133 
134 } // namespace pdl_to_pdl_interp
135 } // namespace mlir
136 
137 #endif // MLIR_CONVERSION_PDLTOPDLINTERP_ROOTORDERING_H_
138