From ea7f6c1e945284bddb219291a2f437ff740a563f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Joachim=20Sch=C3=B6berl?= <joachim.schoeberl@tuwien.ac.at>
Date: Tue, 22 Dec 2020 13:06:08 +0100
Subject: [PATCH] fnma intrinsic for avx512

---
 libsrc/core/simd_avx.hpp     |  8 ++++++++
 libsrc/core/simd_avx512.hpp  | 10 ++++++++++
 libsrc/core/simd_generic.hpp | 13 ++++++++++---
 3 files changed, 28 insertions(+), 3 deletions(-)

diff --git a/libsrc/core/simd_avx.hpp b/libsrc/core/simd_avx.hpp
index 281b10c8..37a0bab8 100644
--- a/libsrc/core/simd_avx.hpp
+++ b/libsrc/core/simd_avx.hpp
@@ -184,6 +184,14 @@ namespace ngcore
   {
     return _mm256_fmadd_pd (_mm256_set1_pd(a), b.Data(), c.Data());
   }
+  NETGEN_INLINE SIMD<double,4> FNMA (SIMD<double,4> a, SIMD<double,4> b, SIMD<double,4> c)
+  {
+    return _mm256_fnmadd_pd (a.Data(), b.Data(), c.Data());
+  }
+  NETGEN_INLINE SIMD<double,4> FNMA (const double & a, SIMD<double,4> b, SIMD<double,4> c)
+  {
+    return _mm256_fnmadd_pd (_mm256_set1_pd(a), b.Data(), c.Data());
+  }
 #endif
 
 #if defined(__FMA__) && !defined(__AVX512F__)
diff --git a/libsrc/core/simd_avx512.hpp b/libsrc/core/simd_avx512.hpp
index e453b7e4..bf57f4e1 100644
--- a/libsrc/core/simd_avx512.hpp
+++ b/libsrc/core/simd_avx512.hpp
@@ -234,6 +234,16 @@ namespace ngcore
   {
     return _mm512_fmadd_pd (_mm512_set1_pd(a), b.Data(), c.Data());
   }
+
+  NETGEN_INLINE SIMD<double,8> FNMA (SIMD<double,8> a, SIMD<double,8> b, SIMD<double,8> c)
+  {
+    return _mm512_fnmadd_pd (a.Data(), b.Data(), c.Data());
+  }
+  NETGEN_INLINE SIMD<double,8> FNMA (const double & a, SIMD<double,8> b, SIMD<double,8> c)
+  {
+    return _mm512_fnmadd_pd (_mm512_set1_pd(a), b.Data(), c.Data());
+  }
+
 }
 
 #endif // NETGEN_CORE_SIMD_AVX512_HPP
diff --git a/libsrc/core/simd_generic.hpp b/libsrc/core/simd_generic.hpp
index e479590d..1ad4ea99 100644
--- a/libsrc/core/simd_generic.hpp
+++ b/libsrc/core/simd_generic.hpp
@@ -513,11 +513,17 @@ namespace ngcore
 
     }
 
-  template <typename T1, typename T2, typename T3>
   // a*b+c
+  template <typename T1, typename T2, typename T3>
   NETGEN_INLINE auto FMA(T1 a, T2 b, T3 c)
   {
-    return a*b+c;
+    return c+a*b;
+  }
+
+  template <typename T1, typename T2, typename T3>
+  NETGEN_INLINE auto FNMA(T1 a, T2 b, T3 c)
+  {
+    return c-a*b;
   }
 
   // update form of fma
@@ -531,7 +537,8 @@ namespace ngcore
   template <int N>
   void FNMAasm (SIMD<double,N> a, SIMD<double,N> b, SIMD<double,N> & sum)
   {
-    sum -= a*b;
+    // sum -= a*b;
+    sum = FNMA(a,b,sum);
   }