diff options
author | 2015-01-30 17:45:03 -0500 | |
---|---|---|
committer | 2015-01-30 17:45:03 -0500 | |
commit | 5ef95fabee5e9a9357c082cd32ae3b4affb2eff6 (patch) | |
tree | 7eccbf0fa491942ffbffacbef0966503b627ecfc /Eigen/src/Core/arch/NEON/PacketMath.h | |
parent | 0f216136980503c3792a90e382b4d6bbdbb870c0 (diff) |
bug #936, patch 3/3: Properly detect FMA support on ARM (requires VFPv4)
and use it instead of MLA when available, because it's both more accurate,
and faster.
Diffstat (limited to 'Eigen/src/Core/arch/NEON/PacketMath.h')
-rw-r--r-- | Eigen/src/Core/arch/NEON/PacketMath.h | 19 |
1 files changed, 17 insertions, 2 deletions
diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index 71255ac85..9afd86bec 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -177,8 +177,19 @@ template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, co return pset1<Packet4i>(0); } -// for some weird raisons, it has to be overloaded for packet of integers +#ifdef __ARM_FEATURE_FMA +// See bug 936. +// FMA is available on VFPv4 i.e. when compiling with -mfpu=neon-vfpv4. +// FMA is a true fused multiply-add i.e. only 1 rounding at the end, no intermediate rounding. +// MLA is not fused i.e. does 2 roundings. +// In addition to giving better accuracy, FMA also gives better performance here on a Krait (Nexus 4): +// MLA: 10 GFlop/s ; FMA: 12 GFlops/s. +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vfmaq_f32(c,a,b); } +#else template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vmlaq_f32(c,a,b); } +#endif + +// No FMA instruction for int, so use MLA unconditionally. template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return vmlaq_s32(c,a,b); } template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vminq_f32(a,b); } @@ -551,8 +562,12 @@ template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vdivq_f64(a,b); } -// for some weird raisons, it has to be overloaded for packet of integers +#ifdef __ARM_FEATURE_FMA +// See bug 936. See above comment about FMA for float. +template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vfmaq_f64(c,a,b); } +#else template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vmlaq_f64(c,a,b); } +#endif template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vminq_f64(a,b); } |