GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-12 19:59:05 +08:00
Code Issues Packages Projects Releases Wiki Activity

Get rid of EIGEN_HAS_AVX512_MATH workaround.

Browse Source
This commit is contained in:
Rasmus Munk Larsen 2023-02-23 23:16:41 +00:00
parent 00844e3865
commit 1c0a6cf228
5 changed files with 31 additions and 63 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
Eigen/src/Core/arch/AVX512/Complex.h
View File

@ -39,7 +39,7 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits
HasMul = 1, HasMul = 1,
HasDiv = 1, HasDiv = 1,
HasNegate = 1, HasNegate = 1,
HasSqrt = EIGEN_HAS_AVX512_MATH, HasSqrt = 1,
HasAbs = 0, HasAbs = 0,
HasAbs2 = 0, HasAbs2 = 0,
HasMin = 0, HasMin = 0,
@ -192,7 +192,7 @@ template<> struct packet_traits<std::complex<double> > : default_packet_traits
HasMul = 1, HasMul = 1,
HasDiv = 1, HasDiv = 1,
HasNegate = 1, HasNegate = 1,
HasSqrt = EIGEN_HAS_AVX512_MATH, HasSqrt = 1,
HasAbs = 0, HasAbs = 0,
HasAbs2 = 0, HasAbs2 = 0,
HasMin = 0, HasMin = 0,
@ -368,8 +368,6 @@ ptranspose(PacketBlock<Packet4cd,4>& kernel) {
kernel.packet[0] = Packet4cd(_mm512_shuffle_f64x2(T0, T2, (shuffle_mask<0,2,0,2>::mask))); // [a0 b0 c0 d0] kernel.packet[0] = Packet4cd(_mm512_shuffle_f64x2(T0, T2, (shuffle_mask<0,2,0,2>::mask))); // [a0 b0 c0 d0]
} }
#if EIGEN_HAS_AVX512_MATH
template<> EIGEN_STRONG_INLINE Packet4cd psqrt<Packet4cd>(const Packet4cd& a) { template<> EIGEN_STRONG_INLINE Packet4cd psqrt<Packet4cd>(const Packet4cd& a) {
return psqrt_complex<Packet4cd>(a); return psqrt_complex<Packet4cd>(a);
} }
@ -378,8 +376,6 @@ template<> EIGEN_STRONG_INLINE Packet8cf psqrt<Packet8cf>(const Packet8cf& a) {
return psqrt_complex<Packet8cf>(a); return psqrt_complex<Packet8cf>(a);
} }
#endif
} // end namespace internal } // end namespace internal
} // end namespace Eigen } // end namespace Eigen

13
Eigen/src/Core/arch/AVX512/MathFunctions.h
View File

@ -16,8 +16,6 @@ namespace Eigen {
namespace internal { namespace internal {
#if EIGEN_HAS_AVX512_MATH
#define EIGEN_DECLARE_CONST_Packet16f(NAME, X) \ #define EIGEN_DECLARE_CONST_Packet16f(NAME, X) \
const Packet16f p16f_##NAME = pset1<Packet16f>(X) const Packet16f p16f_##NAME = pset1<Packet16f>(X)
@ -165,8 +163,11 @@ psqrt<Packet16f>(const Packet16f& _x) {
template <> template <>
EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet8d EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet8d
psqrt<Packet8d>(const Packet8d& _x) { psqrt<Packet8d>(const Packet8d& _x) {
// Double requires 2 Newton-Raphson steps for convergence. #ifdef EIGEN_VECTORIZE_AVX512ER
return generic_sqrt_newton_step<Packet8d, /*Steps=*/1>::run(_x, _mm512_rsqrt28_pd(_x));
#else
return generic_sqrt_newton_step<Packet8d, /*Steps=*/2>::run(_x, _mm512_rsqrt14_pd(_x)); return generic_sqrt_newton_step<Packet8d, /*Steps=*/2>::run(_x, _mm512_rsqrt14_pd(_x));
#endif
} }
#else #else
template <> template <>
@ -185,7 +186,6 @@ BF16_PACKET_FUNCTION(Packet16f, Packet16bf, psqrt)
// prsqrt for float. // prsqrt for float.
#if defined(EIGEN_VECTORIZE_AVX512ER) #if defined(EIGEN_VECTORIZE_AVX512ER)
template <> template <>
EIGEN_STRONG_INLINE Packet16f prsqrt<Packet16f>(const Packet16f& x) { EIGEN_STRONG_INLINE Packet16f prsqrt<Packet16f>(const Packet16f& x) {
return _mm512_rsqrt28_ps(x); return _mm512_rsqrt28_ps(x);
@ -221,10 +221,10 @@ template<> EIGEN_STRONG_INLINE Packet16f preciprocal<Packet16f>(const Packet16f&
return generic_reciprocal_newton_step<Packet16f, /*Steps=*/1>::run(a, _mm512_rcp14_ps(a)); return generic_reciprocal_newton_step<Packet16f, /*Steps=*/1>::run(a, _mm512_rcp14_ps(a));
#endif #endif
} }
#endif
F16_PACKET_FUNCTION(Packet16f, Packet16h, preciprocal) F16_PACKET_FUNCTION(Packet16f, Packet16h, preciprocal)
BF16_PACKET_FUNCTION(Packet16f, Packet16bf, preciprocal) BF16_PACKET_FUNCTION(Packet16f, Packet16bf, preciprocal)
#endif
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
Packet16f plog1p<Packet16f>(const Packet16f& _x) { Packet16f plog1p<Packet16f>(const Packet16f& _x) {
@ -242,9 +242,6 @@ Packet16f pexpm1<Packet16f>(const Packet16f& _x) {
F16_PACKET_FUNCTION(Packet16f, Packet16h, pexpm1) F16_PACKET_FUNCTION(Packet16f, Packet16h, pexpm1)
BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pexpm1) BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pexpm1)
#endif // EIGEN_HAS_AVX512_MATH
template <> template <>
EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet16f EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet16f
psin<Packet16f>(const Packet16f& _x) { psin<Packet16f>(const Packet16f& _x) {

43
Eigen/src/Core/arch/AVX512/PacketMath.h
View File

@ -30,13 +30,6 @@ namespace internal {
#endif #endif
#endif #endif
// Disable the code for older versions of gcc that don't support many of the required avx512 math instrinsics.
#if EIGEN_GNUC_STRICT_AT_LEAST(5,3,0) || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC >= 1923 || EIGEN_COMP_ICC >= 1900
#define EIGEN_HAS_AVX512_MATH 1
#else
#define EIGEN_HAS_AVX512_MATH 0
#endif
typedef __m512 Packet16f; typedef __m512 Packet16f;
typedef __m512i Packet16i; typedef __m512i Packet16i;
typedef __m512d Packet8d; typedef __m512d Packet8d;
@ -84,14 +77,14 @@ struct packet_traits<half> : default_packet_traits {
HasMax = 1, HasMax = 1,
HasConj = 1, HasConj = 1,
HasSetLinear = 0, HasSetLinear = 0,
HasLog = EIGEN_HAS_AVX512_MATH, HasSqrt = 1,
HasLog1p = EIGEN_HAS_AVX512_MATH, HasRsqrt = 1,
HasExp = EIGEN_HAS_AVX512_MATH, HasLog = 1,
HasExpm1 = EIGEN_HAS_AVX512_MATH, HasLog1p = 1,
HasSqrt = EIGEN_HAS_AVX512_MATH, HasExp = 1,
HasRsqrt = EIGEN_HAS_AVX512_MATH, HasExpm1 = 1,
HasBessel = EIGEN_HAS_AVX512_MATH, HasBessel = 1,
HasNdtri = EIGEN_HAS_AVX512_MATH, HasNdtri = 1,
HasSin = EIGEN_FAST_MATH, HasSin = EIGEN_FAST_MATH,
HasCos = EIGEN_FAST_MATH, HasCos = EIGEN_FAST_MATH,
HasTanh = EIGEN_FAST_MATH, HasTanh = EIGEN_FAST_MATH,
@ -126,19 +119,17 @@ template<> struct packet_traits<float> : default_packet_traits
HasASin = 1, HasASin = 1,
HasATan = 1, HasATan = 1,
HasATanh = 1, HasATanh = 1,
#if EIGEN_HAS_AVX512_MATH HasSqrt = 1,
HasRsqrt = 1,
HasLog = 1, HasLog = 1,
HasLog1p = 1, HasLog1p = 1,
HasExpm1 = 1, HasExpm1 = 1,
HasNdtri = 1, HasNdtri = 1,
HasBessel = 1, HasBessel = 1,
HasExp = 1, HasExp = 1,
HasSqrt = EIGEN_FAST_MATH,
HasRsqrt = EIGEN_FAST_MATH,
HasReciprocal = EIGEN_FAST_MATH, HasReciprocal = EIGEN_FAST_MATH,
HasTanh = EIGEN_FAST_MATH, HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH, HasErf = EIGEN_FAST_MATH,
#endif
HasCmp = 1, HasCmp = 1,
HasDiv = 1, HasDiv = 1,
HasRound = 1, HasRound = 1,
@ -156,12 +147,10 @@ template<> struct packet_traits<double> : default_packet_traits
AlignedOnScalar = 1, AlignedOnScalar = 1,
size = 8, size = 8,
HasHalfPacket = 1, HasHalfPacket = 1,
#if EIGEN_HAS_AVX512_MATH HasSqrt = 1,
HasRsqrt = 1,
HasLog = 1, HasLog = 1,
HasExp = 1, HasExp = 1,
HasSqrt = EIGEN_FAST_MATH,
HasRsqrt = EIGEN_FAST_MATH,
#endif
HasATan = 1, HasATan = 1,
HasCmp = 1, HasCmp = 1,
HasDiv = 1, HasDiv = 1,
@ -2294,20 +2283,18 @@ struct packet_traits<bfloat16> : default_packet_traits {
HasInsert = 1, HasInsert = 1,
HasSin = EIGEN_FAST_MATH, HasSin = EIGEN_FAST_MATH,
HasCos = EIGEN_FAST_MATH, HasCos = EIGEN_FAST_MATH,
#if EIGEN_HAS_AVX512_MATH HasSqrt = 1,
HasRsqrt = 1,
#ifdef EIGEN_VECTORIZE_AVX512DQ #ifdef EIGEN_VECTORIZE_AVX512DQ
HasLog = 1, // Currently fails test with bad accuracy. HasLog = 1, // Currently fails test with bad accuracy.
HasLog1p = 1, HasLog1p = 1,
HasExpm1 = 1, HasExpm1 = 1,
HasNdtri = 1, HasNdtri = 1,
HasBessel = 1, HasBessel = 1,
#endif #endif
HasExp = 1, HasExp = 1,
HasSqrt = EIGEN_FAST_MATH,
HasRsqrt = EIGEN_FAST_MATH,
HasTanh = EIGEN_FAST_MATH, HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH, HasErf = EIGEN_FAST_MATH,
#endif
HasCmp = 1, HasCmp = 1,
HasDiv = 1 HasDiv = 1
}; };

23
Eigen/src/Core/arch/AVX512/PacketMathFP16.h
View File

@ -16,13 +16,6 @@ namespace Eigen {
namespace internal { namespace internal {
// Disable the code for older versions of gcc that don't support many of the required avx512 math instrinsics.
#if EIGEN_GNUC_STRICT_AT_LEAST(5,3,0) || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC >= 1923 || EIGEN_COMP_ICC >= 1900
#define EIGEN_HAS_AVX512_MATH 1
#else
#define EIGEN_HAS_AVX512_MATH 0
#endif
typedef __m512h Packet32h; typedef __m512h Packet32h;
typedef eigen_packet_wrapper<__m256i, 1> Packet16h; typedef eigen_packet_wrapper<__m256i, 1> Packet16h;
typedef eigen_packet_wrapper<__m128i, 2> Packet8h; typedef eigen_packet_wrapper<__m128i, 2> Packet8h;
@ -54,15 +47,15 @@ struct packet_traits<half> : default_packet_traits {
HasMax = 1, HasMax = 1,
HasConj = 1, HasConj = 1,
HasSetLinear = 0, HasSetLinear = 0,
HasLog = 1,
HasLog1p = 1,
HasExp = 1,
HasExpm1 = 1,
HasSqrt = 1,
HasRsqrt = 1,
// These ones should be implemented in future // These ones should be implemented in future
HasLog = EIGEN_HAS_AVX512_MATH, HasBessel = 0,
HasLog1p = EIGEN_HAS_AVX512_MATH, HasNdtri = 0,
HasExp = EIGEN_HAS_AVX512_MATH,
HasExpm1 = EIGEN_HAS_AVX512_MATH,
HasSqrt = EIGEN_HAS_AVX512_MATH,
HasRsqrt = EIGEN_HAS_AVX512_MATH,
HasBessel = 0, // EIGEN_HAS_AVX512_MATH,
HasNdtri = 0, // EIGEN_HAS_AVX512_MATH,
HasSin = EIGEN_FAST_MATH, HasSin = EIGEN_FAST_MATH,
HasCos = EIGEN_FAST_MATH, HasCos = EIGEN_FAST_MATH,
HasTanh = EIGEN_FAST_MATH, HasTanh = EIGEN_FAST_MATH,

5
unsupported/Eigen/src/SpecialFunctions/arch/AVX512/BesselFunctions.h
View File

@ -4,9 +4,6 @@
namespace Eigen { namespace Eigen {
namespace internal { namespace internal {
// Bessel functions only available for some compilers.
#if EIGEN_HAS_AVX512_MATH
F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_i0) F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_i0)
BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_i0) BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_i0)
@ -43,8 +40,6 @@ BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_y0)
F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_y1) F16_PACKET_FUNCTION(Packet16f, Packet16h, pbessel_y1)
BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_y1) BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pbessel_y1)
#endif
} // namespace internal } // namespace internal
} // namespace Eigen } // namespace Eigen