devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

bug #1631: fix compilation with ARM NEON and clang, and cleanup the weird pshiftright_and_cast and pcast_and_shiftleft functions.

Browse Source
This commit is contained in:
Gael Guennebaud
2018-11-27 23:45:00 +01:00
parent a1a5fbbd21
commit b131a4db24
8 changed files with 124 additions and 90 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
Eigen/src/Core/arch/AVX/PacketMath.h
View File

@@ -256,7 +256,17 @@ template<> EIGEN_STRONG_INLINE Packet8f pselect<Packet8f>(const Packet8f& mask,
template<> EIGEN_STRONG_INLINE Packet4d pselect<Packet4d>(const Packet4d& mask, const Packet4d& a, const Packet4d& b)
{ return _mm256_blendv_pd(b,a,mask); }
template<int N> EIGEN_STRONG_INLINE Packet8i pshiftleft(const Packet8i& a) {
template<int N> EIGEN_STRONG_INLINE Packet8i pshiftright(Packet8i a) {
#ifdef EIGEN_VECTORIZE_AVX2
return _mm256_srli_epi32(a, N);
#else
__m128i lo = _mm_srli_epi32(_mm256_extractf128_si256(a, 0), N);
__m128i hi = _mm_srli_epi32(_mm256_extractf128_si256(a, 1), N);
return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1);
#endif
}
template<int N> EIGEN_STRONG_INLINE Packet8i pshiftleft(Packet8i a) {
#ifdef EIGEN_VECTORIZE_AVX2
return _mm256_slli_epi32(a, N);
#else
@@ -409,33 +419,10 @@ template<> EIGEN_STRONG_INLINE Packet4d pabs(const Packet4d& a)
return _mm256_and_pd(a,mask);
}
template<> EIGEN_STRONG_INLINE Packet8f pshiftright_and_cast<Packet8f>(Packet8f v, int n)
{
#ifdef EIGEN_VECTORIZE_AVX2
return _mm256_cvtepi32_ps(_mm256_srli_epi32(_mm256_castps_si256(v), n));
#else
__m128i lo = _mm_srli_epi32(_mm256_extractf128_si256(_mm256_castps_si256(v), 0), n);
__m128i hi = _mm_srli_epi32(_mm256_extractf128_si256(_mm256_castps_si256(v), 1), n);
return _mm256_cvtepi32_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1));
#endif
}
template<> EIGEN_STRONG_INLINE Packet8f pfrexp<Packet8f>(const Packet8f& a, Packet8f& exponent) {
return pfrexp_float(a,exponent);
}
template<> EIGEN_STRONG_INLINE Packet8f pcast_and_shiftleft<Packet8f>(Packet8f v, int n)
{
Packet8i vi = _mm256_cvttps_epi32(v);
#ifdef EIGEN_VECTORIZE_AVX2
return _mm256_castsi256_ps(_mm256_slli_epi32(vi, n));
#else
__m128i lo = _mm_slli_epi32(_mm256_extractf128_si256(vi, 0), n);
__m128i hi = _mm_slli_epi32(_mm256_extractf128_si256(vi, 1), n);
return _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1));
#endif
}
template<> EIGEN_STRONG_INLINE Packet8f pldexp<Packet8f>(const Packet8f& a, const Packet8f& exponent) {
return pldexp_float(a,exponent);
}