Avoid compiling gbdt models with many conditions

The average path length grows slower with larger trees (you need to
traverse a smaller part of the total model). This makes it
harder(slower) for llvm to compile and run the code compared to the
VMForest option.

1 files changed, 3 insertions, 0 deletions

diff --git a/eval/src/vespa/eval/eval/gbdt.cpp b/eval/src/vespa/eval/eval/gbdt.cpp
index 3422228b03c..7ab4c4ae822 100644
--- a/eval/src/vespa/eval/eval/gbdt.cpp
+++ b/eval/src/vespa/eval/eval/gbdt.cpp
@@ -154,6 +154,9 @@ Optimize::select_best(const ForestStats &stats,
     if ((stats.tree_sizes.back().size > 12) && (path_len > 2500.0)) {
         return apply_chain(VMForest::optimize_chain, stats, trees);
     }
+    if (stats.total_size > 25000) {
+        return apply_chain(VMForest::optimize_chain, stats, trees);
+    }
     return Optimize::Result();
 }