diff options
Diffstat (limited to 'Eigen/src/Core/arch/NEON/PacketMath.h')
| -rw-r--r-- | Eigen/src/Core/arch/NEON/PacketMath.h | 28 |
1 files changed, 26 insertions, 2 deletions
diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index fead02916..3224c36bd 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -177,7 +177,11 @@ template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, co return pset1<Packet4i>(0); } -#ifdef __ARM_FEATURE_FMA +// Clang/ARM wrongly advertises __ARM_FEATURE_FMA even when it's not available, +// then implements a slow software scalar fallback calling fmaf()! +// Filed LLVM bug: +// https://llvm.org/bugs/show_bug.cgi?id=27216 +#if (defined __ARM_FEATURE_FMA) && !(EIGEN_COMP_CLANG && EIGEN_ARCH_ARM) // 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. @@ -186,7 +190,27 @@ template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, co // 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); } +template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { +#if EIGEN_COMP_CLANG && EIGEN_ARCH_ARM + // Clang/ARM will replace VMLA by VMUL+VADD at least for some values of -mcpu, + // at least -mcpu=cortex-a8 and -mcpu=cortex-a7. Since the former is the default on + // -march=armv7-a, that is a very common case. + // See e.g. this thread: + // http://lists.llvm.org/pipermail/llvm-dev/2013-December/068806.html + // Filed LLVM bug: + // https://llvm.org/bugs/show_bug.cgi?id=27219 + Packet4f r = c; + asm volatile( + "vmla.f32 %q[r], %q[a], %q[b]" + : [r] "+w" (r) + : [a] "w" (a), + [b] "w" (b) + : ); + return r; +#else + return vmlaq_f32(c,a,b); +#endif +} #endif // No FMA instruction for int, so use MLA unconditionally. |