From c11a6fe10189009220bed168707123af34376de2 Mon Sep 17 00:00:00 2001
From: Dominic Kempf <dominic.kempf@iwr.uni-heidelberg.de>
Date: Wed, 24 Jul 2019 14:17:14 +0200
Subject: [PATCH] Provide fallback if no arch flag is given
---
dune/codegen/sumfact/horizontaladd.hh | 35 ++++++++++++++++++---------
1 file changed, 23 insertions(+), 12 deletions(-)
diff --git a/dune/codegen/sumfact/horizontaladd.hh b/dune/codegen/sumfact/horizontaladd.hh
index 90dbf038..7dd122af 100644
--- a/dune/codegen/sumfact/horizontaladd.hh
+++ b/dune/codegen/sumfact/horizontaladd.hh
@@ -1,21 +1,11 @@
#ifndef DUNE_CODEGEN_SUMFACT_HORIZONTALADD_HH
#define DUNE_CODEGEN_SUMFACT_HORIZONTALADD_HH
-#include<immintrin.h>
#include<dune/codegen/common/vectorclass.hh>
-template<class V>
-typename base_floatingpoint<V>::value horizontal_add_lower(const V& x)
-{
- return horizontal_add(x.get_low());
-}
-
-template<class V>
-typename base_floatingpoint<V>::value horizontal_add_upper(const V& x)
-{
- return horizontal_add(x.get_high());
-}
+// Only use our custom implementations if we have AVX2 or later!
+#if INSTRSET >= 8
/** Implement a variant of horizontal_add(Vec2d) that avoids the haddpd
* instruction and instead uses the shuffle port.
@@ -49,6 +39,27 @@ static inline double permuting_horizontal_add(const Vec8d& a)
#endif
+#else
+template<typename V>
+static inline double permuting_horizontal_add (const V& a)
+{
+ return horizontal_add(a);
+}
+
+#endif
+
+template<class V>
+typename base_floatingpoint<V>::value horizontal_add_lower(const V& x)
+{
+ return horizontal_add(x.get_low());
+}
+
+template<class V>
+typename base_floatingpoint<V>::value horizontal_add_upper(const V& x)
+{
+ return horizontal_add(x.get_high());
+}
+
template<class V>
typename base_floatingpoint<V>::value permuting_horizontal_add_lower(const V& x)
{
--
GitLab