From ab310943d62982a45734f52fed782ec57c9aa6a0 Mon Sep 17 00:00:00 2001
From: Rasmus Munk Larsen <rmlarsen@google.com>
Date: Fri, 9 Aug 2024 10:48:24 -0700
Subject: [PATCH] Add a yield instruction in the two spinloops of the threaded
 matmul implementation.

---
 Eigen/src/Core/products/GeneralMatrixMatrix.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/Eigen/src/Core/products/GeneralMatrixMatrix.h b/Eigen/src/Core/products/GeneralMatrixMatrix.h
index e9d0cae8d..ebfac0146 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrix.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrix.h
@@ -97,6 +97,7 @@ struct general_matrix_matrix_product<Index, LhsScalar, LhsStorageOrder, Conjugat
         // Then, we set info->task_info[tid].users to the number of threads to mark that all other threads are going to
         // use it.
         while (info->task_info[tid].users != 0) {
+          std::this_thread::yield();
         }
         info->task_info[tid].users = threads;
 
@@ -115,6 +116,7 @@ struct general_matrix_matrix_product<Index, LhsScalar, LhsStorageOrder, Conjugat
           // However, no need to wait for the B' part which has been updated by the current thread!
           if (shift > 0) {
             while (info->task_info[i].sync != k) {
+              std::this_thread::yield();
             }
           }