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

Add more missing vectorized casts for int on x86, and remove redundant unit tests

Browse Source
This commit is contained in:
Rasmus Munk Larsen
2023-03-24 16:02:00 +00:00
parent 33e206f714
commit b8b8a26145
5 changed files with 41 additions and 110 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -64,7 +64,6 @@ struct type_casting_traits<float, bool> {
};
#endif // EIGEN_VECTORIZE_AVX512
template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) {
return _mm256_cvttps_epi32(a);
}
@@ -77,6 +76,10 @@ template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet4d, Packet8f>(const Packet4d
return _mm256_set_m128(_mm256_cvtpd_ps(b), _mm256_cvtpd_ps(a));
}
template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet4d, Packet8i>(const Packet4d& a, const Packet4d& b) {
return _mm256_set_m128i(_mm256_cvtpd_epi32(b), _mm256_cvtpd_epi32(a));
}
template <>
EIGEN_STRONG_INLINE Packet16b pcast<Packet8f, Packet16b>(const Packet8f& a,
const Packet8f& b) {

5
Eigen/src/Core/arch/AVX512/PacketMath.h
View File

@@ -544,6 +544,8 @@ EIGEN_STRONG_INLINE Packet8d pmax<PropagateNaN, Packet8d>(const Packet8d& a, con
template<int I_> EIGEN_STRONG_INLINE Packet8f extract256(Packet16f x) { return _mm512_extractf32x8_ps(x,I_); }
template<int I_> EIGEN_STRONG_INLINE Packet2d extract128(Packet8d x) { return _mm512_extractf64x2_pd(x,I_); }
EIGEN_STRONG_INLINE Packet16f cat256(Packet8f a, Packet8f b) { return _mm512_insertf32x8(_mm512_castps256_ps512(a),b,1); }
EIGEN_STRONG_INLINE Packet16i cat256i(Packet8i a, Packet8i b) {
return _mm512_insertf32x8(_mm512_castsi256_si512(a),b,1); }
#else
// AVX512F does not define _mm512_extractf32x8_ps to extract _m256 from _m512
template<int I_> EIGEN_STRONG_INLINE Packet8f extract256(Packet16f x) {
@@ -559,6 +561,9 @@ EIGEN_STRONG_INLINE Packet16f cat256(Packet8f a, Packet8f b) {
return _mm512_castsi512_ps(_mm512_inserti64x4(_mm512_castsi256_si512(_mm256_castps_si256(a)),
_mm256_castps_si256(b),1));
}
EIGEN_STRONG_INLINE Packet16i cat256i(Packet8i a, Packet8i b) {
return _mm512_inserti64x4(_mm512_castsi256_si512(a), b, 1);
}
#endif
// Helper function for bit packing snippet of low precision comparison.

4
Eigen/src/Core/arch/AVX512/TypeCasting.h
View File

@@ -55,6 +55,10 @@ template<> EIGEN_STRONG_INLINE Packet16f pcast<Packet8d, Packet16f>(const Packet
return cat256(_mm512_cvtpd_ps(a), _mm512_cvtpd_ps(b));
}
template<> EIGEN_STRONG_INLINE Packet16i pcast<Packet8d, Packet16i>(const Packet8d& a, const Packet8d& b) {
return cat256i(_mm512_cvtpd_epi32(a), _mm512_cvtpd_epi32(b));
}
template<> EIGEN_STRONG_INLINE Packet16i preinterpret<Packet16i, Packet16f>(const Packet16f& a) {
return _mm512_castps_si512(a);
}

25
Eigen/src/Core/arch/SSE/TypeCasting.h
View File

@@ -27,13 +27,14 @@ struct type_casting_traits<float, bool> {
};
template <>
struct type_casting_traits<float, int> {
struct type_casting_traits<float, double> {
enum {
VectorizedCast = 1,
SrcCoeffRatio = 1,
TgtCoeffRatio = 1
TgtCoeffRatio = 2
};
};
#endif
template <>
struct type_casting_traits<int, float> {
@@ -45,14 +46,22 @@ struct type_casting_traits<int, float> {
};
template <>
struct type_casting_traits<float, double> {
struct type_casting_traits<float, int> {
enum {
VectorizedCast = 1,
SrcCoeffRatio = 1,
TgtCoeffRatio = 2
TgtCoeffRatio = 1
};
};
template <>
struct type_casting_traits<double, int> {
enum {
VectorizedCast = 1,
SrcCoeffRatio = 2,
TgtCoeffRatio = 1
};
};
#endif
template <>
struct type_casting_traits<double, float> {
@@ -91,6 +100,12 @@ template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet2d, Packet4f>(const Packet2d
return _mm_shuffle_ps(_mm_cvtpd_ps(a), _mm_cvtpd_ps(b), (1 << 2) | (1 << 6));
}
template<> EIGEN_STRONG_INLINE Packet4i pcast<Packet2d, Packet4i>(const Packet2d& a, const Packet2d& b) {
return _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(_mm_cvtpd_epi32(a)),
_mm_castsi128_ps(_mm_cvtpd_epi32(b)),
(1 << 2) | (1 << 6)));
}
template<> EIGEN_STRONG_INLINE Packet2d pcast<Packet4f, Packet2d>(const Packet4f& a) {
// Simply discard the second half of the input
return _mm_cvtps_pd(a);