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

Clean up packetmath tests and fix various bugs to make bfloat16 pass (almost) all packetmath tests with SSE, AVX, and AVX512.

Browse Source
This commit is contained in:
Rasmus Munk Larsen
2020-10-09 20:05:49 +00:00
parent 7a8d3d5b81
commit 4e4d3f32d1
8 changed files with 114 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
Eigen/src/Core/arch/AVX/PacketMath.h
View File

@@ -1205,6 +1205,11 @@ EIGEN_STRONG_INLINE Packet8bf pmax<Packet8bf>(const Packet8bf& a,
return F32ToBf16(pmax<Packet8f>(Bf16ToF32(a), Bf16ToF32(b)));
}
template <>
EIGEN_STRONG_INLINE Packet8bf plset<Packet8bf>(const bfloat16& a) {
return F32ToBf16(plset<Packet8f>(static_cast<float>(a)));
}
template<> EIGEN_STRONG_INLINE Packet8bf por(const Packet8bf& a,const Packet8bf& b) {
return _mm_or_si128(a,b);
}

48
Eigen/src/Core/arch/AVX/TypeCasting.h
View File

@@ -35,23 +35,6 @@ struct type_casting_traits<int, float> {
};
template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) {
return _mm256_cvttps_epi32(a);
}
template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8i, Packet8f>(const Packet8i& a) {
return _mm256_cvtepi32_ps(a);
}
template<> EIGEN_STRONG_INLINE Packet8i preinterpret<Packet8i,Packet8f>(const Packet8f& a) {
return _mm256_castps_si256(a);
}
template<> EIGEN_STRONG_INLINE Packet8f preinterpret<Packet8f,Packet8i>(const Packet8i& a) {
return _mm256_castsi256_ps(a);
}
#ifndef EIGEN_VECTORIZE_AVX512
template <>
@@ -63,9 +46,6 @@ struct type_casting_traits<Eigen::half, float> {
};
};
template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8h, Packet8f>(const Packet8h& a) {
return half2float(a);
}
template <>
struct type_casting_traits<float, Eigen::half> {
@@ -85,10 +65,6 @@ struct type_casting_traits<bfloat16, float> {
};
};
template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8bf, Packet8f>(const Packet8bf& a) {
return Bf16ToF32(a);
}
template <>
struct type_casting_traits<float, bfloat16> {
enum {
@@ -100,6 +76,30 @@ struct type_casting_traits<float, bfloat16> {
#endif // EIGEN_VECTORIZE_AVX512
template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) {
return _mm256_cvttps_epi32(a);
}
template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8i, Packet8f>(const Packet8i& a) {
return _mm256_cvtepi32_ps(a);
}
template<> EIGEN_STRONG_INLINE Packet8i preinterpret<Packet8i,Packet8f>(const Packet8f& a) {
return _mm256_castps_si256(a);
}
template<> EIGEN_STRONG_INLINE Packet8f preinterpret<Packet8f,Packet8i>(const Packet8i& a) {
return _mm256_castsi256_ps(a);
}
template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8h, Packet8f>(const Packet8h& a) {
return half2float(a);
}
template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8bf, Packet8f>(const Packet8bf& a) {
return Bf16ToF32(a);
}
template<> EIGEN_STRONG_INLINE Packet8h pcast<Packet8f, Packet8h>(const Packet8f& a) {
return float2half(a);
}