From: Maximilian Bergbauer <maximilian.bergbauer@tum.de>
Date: Mon, 22 May 2023 13:28:41 +0000 (+0200)
Subject: Use intrinsics for vertical addition
X-Git-Tag: v9.5.0-rc1~196^2~4
X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=3a2b35e47f8a99ba155526cbd4d364c37316e026;p=dealii.git

Use intrinsics for vertical addition
---

diff --git a/include/deal.II/base/vectorization.h b/include/deal.II/base/vectorization.h
index 06a29e0d96..e81ab787e7 100644
--- a/include/deal.II/base/vectorization.h
+++ b/include/deal.II/base/vectorization.h
@@ -4971,7 +4971,7 @@ inline double
 VectorizedArray<double, 4>::horizontal_add()
 {
   VectorizedArray<double, 2> t1;
-  t1.data = this->get_low() + this->get_high();
+  t1.data = _mm_add_pd(this->get_low(), this->get_high());
   return t1.horizontal_add();
 }
 
@@ -4981,7 +4981,7 @@ inline float
 VectorizedArray<float, 8>::horizontal_add()
 {
   VectorizedArray<float, 4> t1;
-  t1.data = this->get_low() + this->get_high();
+  t1.data = _mm_add_ps(this->get_low(), this->get_high());
   return t1.horizontal_add();
 }
 #endif
@@ -4993,7 +4993,7 @@ inline double
 VectorizedArray<double, 8>::horizontal_add()
 {
   VectorizedArray<double, 4> t1;
-  t1.data = this->get_low() + this->get_high();
+  t1.data = _mm256_add_pd(this->get_low(), this->get_high());
   return t1.horizontal_add();
 }
 
@@ -5003,7 +5003,7 @@ inline float
 VectorizedArray<float, 16>::horizontal_add()
 {
   VectorizedArray<float, 8> t1;
-  t1.data = this->get_low() + this->get_high();
+  t1.data = _mm256_add_ps(this->get_low(), this->get_high());
   return t1.horizontal_add();
 }
 #endif