From 7b6d0ff1f6afdb7f6137c92a76d20e7497e2fa2a Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Fri, 7 Dec 2018 15:14:50 +0100
Subject: Enable FMA with MSVC (through /arch:AVX2). To make this possible, I
 also has to turn the #warning regarding AVX512-FMA to a #error.

---
 CMakeLists.txt                               | 15 ++++++++++++++-
 Eigen/src/Core/arch/AVX/PacketMath.h         |  2 +-
 Eigen/src/Core/arch/SSE/PacketMath.h         |  4 ++--
 Eigen/src/Core/util/ConfigureVectorization.h |  8 +++++---
 4 files changed, 22 insertions(+), 7 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 45cbc75ee..e18428e50 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -250,7 +250,7 @@ if(NOT MSVC)
 
   option(EIGEN_TEST_AVX512 "Enable/Disable AVX512 in tests/examples" OFF)
   if(EIGEN_TEST_AVX512)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512f -DEIGEN_ENABLE_AVX512")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512f -mfma -DEIGEN_ENABLE_AVX512")
     if (NOT "${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
       set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fabi-version=6")
     endif()
@@ -350,6 +350,19 @@ else(NOT MSVC)
     endif(NOT CMAKE_CL_64)
     message(STATUS "Enabling SSE2 in tests/examples")
   endif(EIGEN_TEST_SSE2)
+
+  option(EIGEN_TEST_AVX "Enable/Disable AVX in tests/examples" OFF)
+  if(EIGEN_TEST_AVX)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
+    message(STATUS "Enabling AVX in tests/examples")
+  endif()
+
+  option(EIGEN_TEST_FMA "Enable/Disable FMA/AVX2 in tests/examples" OFF)
+  if(EIGEN_TEST_FMA AND NOT EIGEN_TEST_NEON)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
+    message(STATUS "Enabling FMA/AVX2 in tests/examples")
+  endif()
+
 endif(NOT MSVC)
 
 option(EIGEN_TEST_NO_EXPLICIT_VECTORIZATION "Disable explicit vectorization in tests/examples" OFF)
diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h
index 35c821472..9bc986575 100644
--- a/Eigen/src/Core/arch/AVX/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -181,7 +181,7 @@ template<> EIGEN_STRONG_INLINE Packet8i pdiv<Packet8i>(const Packet8i& /*a*/, co
   return pset1<Packet8i>(0);
 }
 
-#ifdef __FMA__
+#ifdef EIGEN_VECTORIZE_FMA
 template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) {
 #if ( (EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC<80) || (EIGEN_COMP_CLANG) )
   // Clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers,
diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index 971a31a4d..b80b7f46a 100755
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -266,7 +266,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const
 
 // for some weird raisons, it has to be overloaded for packet of integers
 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
-#ifdef __FMA__
+#ifdef EIGEN_VECTORIZE_FMA
 template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return _mm_fmadd_ps(a,b,c); }
 template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return _mm_fmadd_pd(a,b,c); }
 #endif
@@ -1013,7 +1013,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pinsertlast(const Packet2d& a, double b)
 }
 
 // Scalar path for pmadd with FMA to ensure consistency with vectorized path.
-#ifdef __FMA__
+#ifdef EIGEN_VECTORIZE_FMA
 template<> EIGEN_STRONG_INLINE float pmadd(const float& a, const float& b, const float& c) {
   return ::fmaf(a,b,c);
 }
diff --git a/Eigen/src/Core/util/ConfigureVectorization.h b/Eigen/src/Core/util/ConfigureVectorization.h
index 2584717b5..121476394 100644
--- a/Eigen/src/Core/util/ConfigureVectorization.h
+++ b/Eigen/src/Core/util/ConfigureVectorization.h
@@ -250,15 +250,17 @@
       #define EIGEN_VECTORIZE_SSE4_1
       #define EIGEN_VECTORIZE_SSE4_2
     #endif
-    #ifdef __FMA__
+    #if defined(__FMA__) || (EIGEN_COMP_MSVC && defined(__AVX2__))
+      // MSVC does not expose a switch dedicated for FMA
+      // For MSVC, AVX2 => FMA
       #define EIGEN_VECTORIZE_FMA
     #endif
     #if defined(__AVX512F__)
       #ifndef __FMA__
       #if EIGEN_COMP_GNUC
-      #warning Please add -mfma to your compiler flags: compiling with -mavx512f alone without SSE/AVX FMA is not supported (bug 1638).
+      #error Please add -mfma to your compiler flags: compiling with -mavx512f alone without SSE/AVX FMA is not supported (bug 1638).
       #else
-      #error   Please enable FMA in your compiler flags (e.g. -mfma): compiling with AVX512 alone without SSE/AVX FMA is not supported (bug 1638).
+      #error Please enable FMA in your compiler flags (e.g. -mfma): compiling with AVX512 alone without SSE/AVX FMA is not supported (bug 1638).
       #endif
       #endif
       #define EIGEN_VECTORIZE_AVX512
-- 
cgit v1.2.3