diff options
author | Gael Guennebaud <g.gael@free.fr> | 2018-11-28 17:14:20 +0100 |
---|---|---|
committer | Gael Guennebaud <g.gael@free.fr> | 2018-11-28 17:14:20 +0100 |
commit | 3e95e398b6cee0218c0a01880fecd6215a7f0d28 (patch) | |
tree | 474587609f1ad126e88b5fa590e148a289a26512 /Eigen/src/Core/arch/SSE | |
parent | aa6097395bc9147b771b33d4cbcb55e55ed409a7 (diff) |
pmin/pmax o SSE: make sure to use AVX instruction with AVX enabled, and disable gcc workaround for fixed gcc versions
Diffstat (limited to 'Eigen/src/Core/arch/SSE')
-rwxr-xr-x | Eigen/src/Core/arch/SSE/PacketMath.h | 14 |
1 files changed, 12 insertions, 2 deletions
diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h index 1a2710a3d..004c09f5a 100755 --- a/Eigen/src/Core/arch/SSE/PacketMath.h +++ b/Eigen/src/Core/arch/SSE/PacketMath.h @@ -270,13 +270,18 @@ template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& #endif template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { -#if EIGEN_COMP_GNUC +#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63 // There appears to be a bug in GCC, by which the optimizer may // flip the argument order in calls to _mm_min_ps, so we have to // resort to inline ASM here. This is supposed to be fixed in gcc6.3, // see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867 + #ifdef EIGEN_VECTORIZE_AVX + Packet4f res; + asm("vminps %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b)); + #else Packet4f res = b; asm("minps %[a], %[res]" : [res] "+x" (res) : [a] "x" (a)); + #endif return res; #else // Arguments are reversed to match NaN propagation behavior of std::min. @@ -284,13 +289,18 @@ template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const #endif } template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { -#if EIGEN_COMP_GNUC +#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63 // There appears to be a bug in GCC, by which the optimizer may // flip the argument order in calls to _mm_min_pd, so we have to // resort to inline ASM here. This is supposed to be fixed in gcc6.3, // see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867 + #ifdef EIGEN_VECTORIZE_AVX + Packet2d res; + asm("vminpd %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b)); + #else Packet2d res = b; asm("minpd %[a], %[res]" : [res] "+x" (res) : [a] "x" (a)); + #endif return res; #else // Arguments are reversed to match NaN propagation behavior of std::min. |