From 74a6dc8f336867864cbef9857fc73f68fa6722b2 Mon Sep 17 00:00:00 2001
From: "The gemma.cpp Authors" <no-reply@google.com>
Date: Mon, 22 Jul 2024 10:08:31 -0700
Subject: [PATCH] Use all CPU sockets when pinning threads to cores

PiperOrigin-RevId: 654800375
---
 util/app.h | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/util/app.h b/util/app.h
index d8b7405..bbd2047 100644
--- a/util/app.h
+++ b/util/app.h
@@ -83,9 +83,10 @@ static inline std::vector<size_t> AssignCpusFromTopology(
   // This would result in the least amount of sharing of the last-level
   // cache slices. All assignments are made from Package 0.
   std::vector<std::vector<size_t>> clusters;
-  clusters.reserve(topology.packages[0].clusters.size());
-  for (auto& cluster : topology.packages[0].clusters) {
-    clusters.push_back(LpsToCpus(cluster.lps));
+  for (auto& package : topology.packages) {
+    for (auto& cluster : package.clusters) {
+      clusters.push_back(LpsToCpus(cluster.lps));
+    }
   }
   std::vector<size_t> assigned_cpus;
   assigned_cpus.reserve(num_workers);