mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-03-13 18:37:27 +08:00
Code Issues Packages Projects Releases Wiki Activity

Remove old Clang compiler bug work-arounds. The two LLVM bugs referenced in the comments here have long been fixed. The workarounds were now detrimental because (1) they prevented using fused mul-add on Clang/ARM32 and (2) the unnecessary 'volatile' in 'asm volatile' prevented legitimate reordering by the compiler.

Browse Source
This commit is contained in:
Benoit Jacob 2020-09-15 11:07:57 -04:00
parent bb56a62582
commit cc0c38ace8
1 changed files with 2 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
Eigen/src/Core/arch/NEON/PacketMath.h
View File

@ -1010,17 +1010,8 @@ template<> EIGEN_STRONG_INLINE Packet2ul pdiv<Packet2ul>(const Packet2ul& /*a*/,
return pset1<Packet2ul>(0ULL);
}
// 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.
// 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.
#ifdef __ARM_FEATURE_FMA
template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c)
{ return vfmaq_f32(c,a,b); }
template<> EIGEN_STRONG_INLINE Packet2f pmadd(const Packet2f& a, const Packet2f& b, const Packet2f& c)
@ -1028,25 +1019,7 @@ template<> EIGEN_STRONG_INLINE Packet2f pmadd(const Packet2f& a, const Packet2f&
#else
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
}
template<> EIGEN_STRONG_INLINE Packet2f pmadd(const Packet2f& a, const Packet2f& b, const Packet2f& c)
{