Bug 1785: fix pround on x86 to use the same rounding mode as std::round.

This also adds pset1frombits helper to Packet[24]d.
Makes round ~45% slower for SSE: 1.65µs ± 1% before vs 2.45µs ± 2% after,
stil an order of magnitude faster than scalar version: 33.8µs ± 2%.

Eigen/src/Core/arch/AVX/PacketMath.h

diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h
index ed8acfc..00c3ae5 100644
--- a/Eigen/src/Core/arch/AVX/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -311,9 +311,6 @@
 #endif
 }
 
-template<> EIGEN_STRONG_INLINE Packet8f pround<Packet8f>(const Packet8f& a) { return _mm256_round_ps(a, _MM_FROUND_CUR_DIRECTION); }
-template<> EIGEN_STRONG_INLINE Packet4d pround<Packet4d>(const Packet4d& a) { return _mm256_round_pd(a, _MM_FROUND_CUR_DIRECTION); }
-
 template<> EIGEN_STRONG_INLINE Packet8f pceil<Packet8f>(const Packet8f& a) { return _mm256_ceil_ps(a); }
 template<> EIGEN_STRONG_INLINE Packet4d pceil<Packet4d>(const Packet4d& a) { return _mm256_ceil_pd(a); }
 
@@ -391,6 +388,19 @@
 #endif
 }
 
+template<> EIGEN_STRONG_INLINE Packet8f pround<Packet8f>(const Packet8f& a)
+{
+  const Packet8f mask = pset1frombits<Packet8f>(0x80000000u);
+  const Packet8f prev0dot5 = pset1frombits<Packet8f>(0x3EFFFFFFu);
+  return _mm256_round_ps(padd(por(pand(a, mask), prev0dot5), a), _MM_FROUND_TO_ZERO);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pround<Packet4d>(const Packet4d& a)
+{
+  const Packet4d mask = _mm256_castsi256_pd(_mm256_set_epi64x(0x8000000000000000ull, 0x8000000000000000ull, 0x8000000000000000ull, 0x8000000000000000ull));
+  const Packet4d prev0dot5 = _mm256_castsi256_pd(_mm256_set_epi64x(0x3FDFFFFFFFFFFFFFull, 0x3FDFFFFFFFFFFFFFull, 0x3FDFFFFFFFFFFFFFull, 0x3FDFFFFFFFFFFFFFull));
+  return _mm256_round_pd(padd(por(pand(a, mask), prev0dot5), a), _MM_FROUND_TO_ZERO);
+}
+
 template<> EIGEN_STRONG_INLINE Packet8f pselect<Packet8f>(const Packet8f& mask, const Packet8f& a, const Packet8f& b)
 { return _mm256_blendv_ps(b,a,mask); }
 template<> EIGEN_STRONG_INLINE Packet4d pselect<Packet4d>(const Packet4d& mask, const Packet4d& a, const Packet4d& b)

Eigen/src/Core/arch/SSE/PacketMath.h

diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index 4a772d0..c9aa391 100755
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -421,8 +421,20 @@
 template<int N> EIGEN_STRONG_INLINE Packet4i pshiftleft(Packet4i a) { return _mm_slli_epi32(a,N); }
 
 #ifdef EIGEN_VECTORIZE_SSE4_1
-template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return _mm_round_ps(a, 0); }
-template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return _mm_round_pd(a, 0); }
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a)
+{
+  // Unfortunatly _mm_round_ps doesn't have a rounding mode to implement numext::round.
+  const Packet4f mask = pset1frombits<Packet4f>(0x80000000u);
+  const Packet4f prev0dot5 = pset1frombits<Packet4f>(0x3EFFFFFFu);
+  return _mm_round_ps(padd(por(pand(a, mask), prev0dot5), a), _MM_FROUND_TO_ZERO);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a)
+{
+  const Packet2d mask = _mm_castsi128_pd(_mm_set_epi64x(0x8000000000000000ull, 0x8000000000000000ull));
+  const Packet2d prev0dot5 = _mm_castsi128_pd(_mm_set_epi64x(0x3FDFFFFFFFFFFFFFull, 0x3FDFFFFFFFFFFFFFull));
+  return _mm_round_pd(padd(por(pand(a, mask), prev0dot5), a), _MM_FROUND_TO_ZERO);
+}
 
 template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return _mm_ceil_ps(a); }
 template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return _mm_ceil_pd(a); }

test/packetmath.cpp

diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 00f5f04..ed0ec7e 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -518,6 +518,14 @@
   CHECK_CWISE1_IF(PacketTraits::HasCeil, numext::ceil, internal::pceil);
   CHECK_CWISE1_IF(PacketTraits::HasFloor, numext::floor, internal::pfloor);
 
+  // See bug 1785.
+  for (int i=0; i<size; ++i)
+   {
+     data1[i] = -1.5 + i;
+     data2[i] = -1.5 + i;
+   }
+  CHECK_CWISE1_IF(PacketTraits::HasRound, numext::round, internal::pround);
+
   for (int i=0; i<size; ++i)
   {
     data1[i] = internal::random<Scalar>(-1,1);