Skip to content

factorizer

GalleyLogicFactorizer

Bases: LogicFactorizer

LogicLoader stage that runs Galley on each Query body (see optimizer), then forwards the Plan to the downstream loader ctx.

Default optimizer="bfs" is exact layered branch-and-bound; "dfs" uses the DFS kernel. Greedy k=1 bounds are used only on the layered exact path, not inside branch_and_bound_dfs. cost_model estimates the cost of each reduction, and defaults to FlopsCostModel.

Source code in src/finch/autoschedule/factorizer/galley_factorizer/galley_optimize.py 
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
class GalleyLogicFactorizer(LogicFactorizer):
    """
    LogicLoader stage that runs Galley on each ``Query`` body (see ``optimizer``),
    then forwards the Plan to the downstream loader ``ctx``.

    Default ``optimizer="bfs"`` is exact layered branch-and-bound; ``"dfs"`` uses the
    DFS kernel. Greedy ``k=1`` bounds are used only on the layered exact path, not
    inside ``branch_and_bound_dfs``. ``cost_model`` estimates the cost of each
    reduction, and defaults to ``FlopsCostModel``.
    """

    def __init__(
        self,
        ctx: LogicLoader,
        use_components: bool = True,
        *,
        optimizer: GalleyOptimizer = "bfs",
        cost_model: CostModel | None = None,
    ):
        self.ctx = ctx
        self.use_components = use_components
        self.optimizer = optimizer
        self.cost_model = cost_model

    def lower(
        self,
        term: LogicStatement,
        bindings: dict[Alias, TensorFType],
        stats: dict[Alias, TensorStats],
        stats_factory: StatsFactory,
    ):
        if not isinstance(term, Plan):
            raise ValueError(f"Unsupported program type: {type(term)}")
        logger.debug("Optimizing plan: %s", term)

        def transform(prgm, bindings):
            prgm = optimize_plan(
                prgm,
                stats_factory,
                stats,
                use_components=self.use_components,
                optimizer=self.optimizer,
                cost_model=self.cost_model,
            )
            return prgm, bindings

        # Galley only keeps the queries that compute produced aliases, so each
        # write to a bound tensor, including an in-place update, is renamed and
        # produced, then copied back to the tensor it updates. Galley only
        # reduces from literals, so aggregates which start from a tensor are
        # split first.
        term = split_aggregate_inits(desugar_query_into(term), bindings)
        term, bindings = with_unique_lhs(transform, term, bindings)
        assert isinstance(term, Plan)
        return self.ctx(flatten_plans(term), bindings, stats, stats_factory)

with_unique_lhs(f, root, bindings)

Ensures all left-hand sides (LHS) of queries are unique by inserting new tensors.

Source code in src/finch/autoschedule/factorizer/optimize.py 
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
def with_unique_lhs(
    f, root: LogicStatement, bindings: dict[Alias, TensorFType]
) -> tuple[LogicStatement, dict[Alias, TensorFType]]:
    """
    Ensures all left-hand sides (LHS) of queries are unique by inserting new
    tensors.
    """

    spc = Namespace(root)
    for var in bindings:
        spc.freshen(var.name)
    renames: dict[Alias, Alias] = {}
    bound = set(bindings.keys())
    writes: dict[Alias, Alias] = {}

    def rule_0(node):
        match node:
            case Query(Table(Alias() as lhs, idxs), rhs):
                if lhs in bound:
                    var = Alias(spc.freshen(lhs.name))
                    renames[lhs] = var
                    if lhs in bindings:
                        writes[lhs] = var
                    return Query(Table(var, idxs), rhs)
                bound.add(lhs)
                return None
            case Alias() as a if a in renames:
                return renames[a]
            case Produces(args):
                return Produces(args + tuple(writes.values()))

    root = Rewrite(PostWalk(rule_0))(root)
    root, bindings = f(root, bindings)

    unrenames = {v: k for k, v in renames.items()}

    # Some produces may be renamed during the optimization,
    # so we need to be a bit careful here.
    def rule_1(node):
        match node:
            case Produces(args):
                n_writes = len(writes)
                v_post_list = args[-n_writes:] if n_writes else ()

                bodies: list[LogicStatement] = []
                for (k, _v_pre), v_post in zip(
                    writes.items(), v_post_list, strict=True
                ):
                    idxs = tuple(
                        Field(spc.freshen("i")) for _ in range(bindings[k].ndim)
                    )
                    bodies.append(Query(Table(k, idxs), Table(v_post, idxs)))

                v_post_to_k = dict(zip(v_post_list, writes.keys(), strict=True))
                args_2 = tuple(
                    v_post_to_k.get(a, unrenames.get(a, a))
                    for a in args[: len(args) - n_writes]
                )
                return Plan(tuple(bodies) + (Produces(args_2),))

    return (Rewrite(PostWalk(rule_1))(root), bindings)