prathi3 · rarbore2 · prathi3
--- a/hercules_cg/src/gpu.rs 0 → 100644
+++ b/hercules_cg/src/gpu.rs 0 → 100644
+        }
+    }
+
+    /*
+     * This analysis determines the parallelization strategy within threadblocks.
+     * We run post-order traversal on the fork tree to get the thread quota per
+     * subtree. In particular, each fork starts with a base factor as the
+     * maximum over its descendants (leafs have base 1). We traverse up (details
+     * in helper) and pass the factor and a map from fork node to a tuple of
+     * (max quota of its siblings (including itself), its quota, its fork factor)
+     * from each node to its parents. The parent then compares
+     * - all three are needed for codegen. A node is in the map IFF it will be parallelized.
+     * If not, the fork will use the parent's quota and serialize over the Fork's
+     * ThreadIDs. Nodes may be removed from the map when traversing up the tree
+     * due to an ancestor having a larger factor that conflicts.
+     */
--- a/hercules_cg/src/fork_tree.rs 0 → 100644
+++ b/hercules_cg/src/fork_tree.rs 0 → 100644
+use std::collections::{HashMap, HashSet};
+
+use crate::*;
+
+/*
+ * Construct a map from fork node to all control nodes (including itself) satisfying:
+ * a) domination by F
+ * b) no domination by F's join
+ * c) no domination by any other fork that's also dominated by F, where we do count self-domination