Fix a few issues for AVX512. This change enables vectorized versions of log, exp, log1p, expm1 when AVX512DQ is not available.

2 files changed, 7 insertions, 10 deletions

diff --git a/Eigen/src/Core/arch/AVX512/MathFunctions.h b/Eigen/src/Core/arch/AVX512/MathFunctions.h
index 2c34868a7..c04da6795 100644
--- a/Eigen/src/Core/arch/AVX512/MathFunctions.h
+++ b/Eigen/src/Core/arch/AVX512/MathFunctions.h
@@ -35,7 +35,6 @@ namespace internal {
 #define _EIGEN_DECLARE_CONST_Packet16bf_FROM_INT(NAME, X) \
   const Packet16bf p16bf_##NAME =  preinterpret<Packet16bf,Packet16i>(pset1<Packet16i>(X))
 
-#if defined(EIGEN_VECTORIZE_AVX512DQ)
 template <>
 EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
 plog<Packet16f>(const Packet16f& _x) {
@@ -50,7 +49,6 @@ plog<Packet8d>(const Packet8d& _x) {
 
 F16_PACKET_FUNCTION(Packet16f, Packet16h, plog)
 BF16_PACKET_FUNCTION(Packet16f, Packet16bf, plog)
-#endif
 
 // Exponential function. Works by writing "x = m*log(2) + r" where
 // "m = floor(x/log(2)+1/2)" and "r" is the remainder. The result is then
@@ -334,7 +332,6 @@ EIGEN_STRONG_INLINE Packet8d prsqrt<Packet8d>(const Packet8d& x) {
 }
 #endif
 
-#if defined(EIGEN_VECTORIZE_AVX512DQ)
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet16f plog1p<Packet16f>(const Packet16f& _x) {
   return generic_plog1p(_x);
@@ -350,7 +347,6 @@ Packet16f pexpm1<Packet16f>(const Packet16f& _x) {
 
 F16_PACKET_FUNCTION(Packet16f, Packet16h, pexpm1)
 BF16_PACKET_FUNCTION(Packet16f, Packet16bf, pexpm1)
-#endif
 
 #endif
 
diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h
index 9acec3439..2bbcd802c 100644
--- a/Eigen/src/Core/arch/AVX512/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -108,13 +108,11 @@ template<> struct packet_traits<float>  : default_packet_traits
     HasSin = EIGEN_FAST_MATH,
     HasCos = EIGEN_FAST_MATH,
 #if EIGEN_GNUC_AT_LEAST(5, 3) || (!EIGEN_COMP_GNUC_STRICT)
-#ifdef EIGEN_VECTORIZE_AVX512DQ
     HasLog = 1,
     HasLog1p  = 1,
     HasExpm1  = 1,
     HasNdtri = 1,
     HasBessel  = 1,
-#endif
     HasExp = 1,
     HasSqrt = EIGEN_FAST_MATH,
     HasRsqrt = EIGEN_FAST_MATH,
@@ -139,9 +137,7 @@ template<> struct packet_traits<double> : default_packet_traits
     size = 8,
     HasHalfPacket = 1,
 #if EIGEN_GNUC_AT_LEAST(5, 3) || (!EIGEN_COMP_GNUC_STRICT)
-#ifdef EIGEN_VECTORIZE_AVX512DQ
     HasLog  = 1,
-#endif
     HasSqrt = EIGEN_FAST_MATH,
     HasRsqrt = EIGEN_FAST_MATH,
 #endif
@@ -888,11 +884,16 @@ EIGEN_STRONG_INLINE Packet16f pfrexp<Packet16f>(const Packet16f& a, Packet16f& e
 }
 
 template<>
-EIGEN_STRONG_INLINE Packet8d pfrexp<Packet8d>(const Packet8d& a, Packet8d& exponent){
+EIGEN_STRONG_INLINE Packet8d pfrexp<Packet8d>(const Packet8d& a, Packet8d& exponent) {
   const Packet8d cst_1022d = pset1<Packet8d>(1022.0);
+#ifdef EIGEN_TEST_AVX512DQ
+  exponent = psub(_mm512_cvtepi64_pd(_mm512_srli_epi64(_mm512_castpd_si512(a), 52)), cst_1022d);
+#else
+  exponent = psub(_mm512_cvtepi32_pd(_mm512_cvtepi64_epi32(_mm512_srli_epi64(_mm512_castpd_si512(a), 52))),
+                                     cst_1022d);
+#endif
   const Packet8d cst_half = pset1<Packet8d>(0.5);
   const Packet8d cst_inv_mant_mask  = pset1frombits<Packet8d>(static_cast<uint64_t>(~0x7ff0000000000000ull));
-  exponent = psub(_mm512_cvtepi64_pd(_mm512_srli_epi64(_mm512_castpd_si512(a), 52)), cst_1022d);
   return por(pand(a, cst_inv_mant_mask), cst_half);
 }