From e19ece822de2a516f66712068a2e5fa4fe18150b Mon Sep 17 00:00:00 2001 From: Gael Guennebaud Date: Wed, 28 Nov 2018 17:56:24 +0100 Subject: [PATCH] Disable fma gcc's workaround for gcc >= 8 (based on GEMM benchmarks) --- Eigen/src/Core/arch/AVX/PacketMath.h | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h index 4c1abe43f..abfeb4338 100644 --- a/Eigen/src/Core/arch/AVX/PacketMath.h +++ b/Eigen/src/Core/arch/AVX/PacketMath.h @@ -174,7 +174,7 @@ template<> EIGEN_STRONG_INLINE Packet8i pdiv(const Packet8i& /*a*/, co #ifdef __FMA__ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) { -#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) ) +#if ( (EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC<80) || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) ) // clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers, // and gcc stupidly generates a vfmadd132ps instruction, // so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate @@ -187,7 +187,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& #endif } template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) { -#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) ) +#if ( (EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC<80) || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) ) // see above Packet4d res = c; __asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));