Unified sse pldexp_double api
diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index 1d640cf..25705e7 100755
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -45,6 +45,7 @@
typedef eigen_packet_wrapper<__m128i, 0> Packet4i;
typedef eigen_packet_wrapper<__m128i, 1> Packet16b;
+typedef eigen_packet_wrapper<__m128i, 2> Packet2l;
template<> struct is_arithmetic<__m128> { enum { value = true }; };
template<> struct is_arithmetic<__m128i> { enum { value = true }; };
@@ -192,6 +193,7 @@
template<> struct unpacket_traits<Packet2d> {
typedef double type;
typedef Packet2d half;
+ typedef Packet2l integer_packet;
enum {size=2, alignment=Aligned16, vectorizable=true, masked_load_available=false, masked_store_available=false};
};
template<> struct unpacket_traits<Packet4i> {
@@ -483,6 +485,9 @@
template<int N> EIGEN_STRONG_INLINE Packet4i parithmetic_shift_right(Packet4i a) { return _mm_srai_epi32(a,N); }
template<int N> EIGEN_STRONG_INLINE Packet4i plogical_shift_right(Packet4i a) { return _mm_srli_epi32(a,N); }
template<int N> EIGEN_STRONG_INLINE Packet4i plogical_shift_left(Packet4i a) { return _mm_slli_epi32(a,N); }
+template<int N> EIGEN_STRONG_INLINE Packet2l plogical_shift_right(Packet2l a) { return _mm_srli_epi64(a,N); }
+template<int N> EIGEN_STRONG_INLINE Packet2l plogical_shift_left(Packet2l a) { return _mm_slli_epi64(a,N); }
+
#ifdef EIGEN_VECTORIZE_SSE4_1
template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a)
@@ -753,12 +758,7 @@
}
template<> EIGEN_STRONG_INLINE Packet2d pldexp<Packet2d>(const Packet2d& a, const Packet2d& exponent) {
- const Packet4i cst_1023_0 = _mm_setr_epi32(1023, 1023, 0, 0);
- Packet4i emm0 = _mm_cvttpd_epi32(exponent);
- emm0 = padd(emm0, cst_1023_0);
- emm0 = _mm_slli_epi32(emm0, 20);
- emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(1,2,0,3));
- return pmul(a, Packet2d(_mm_castsi128_pd(emm0)));
+ return pldexp_double(a,exponent);
}
// with AVX, the default implementations based on pload1 are faster
diff --git a/Eigen/src/Core/arch/SSE/TypeCasting.h b/Eigen/src/Core/arch/SSE/TypeCasting.h
index 1b8e9a5..09a6966 100644
--- a/Eigen/src/Core/arch/SSE/TypeCasting.h
+++ b/Eigen/src/Core/arch/SSE/TypeCasting.h
@@ -69,6 +69,14 @@
return _mm_cvtps_pd(a);
}
+template<> EIGEN_STRONG_INLINE Packet2l pcast<Packet2d, Packet2l>(const Packet2d& a) {
+ return _mm_set_epi64x(int64_t(a[1]), int64_t(a[0]));
+ }
+
+template<> EIGEN_STRONG_INLINE Packet2d pcast<Packet2l, Packet2d>(const Packet2l& a) {
+ return _mm_set_pd(double(_mm_cvtsi128_si64(_mm_unpackhi_epi64(a, a))), double(_mm_cvtsi128_si64(a)));
+}
+
template<> EIGEN_STRONG_INLINE Packet4i preinterpret<Packet4i,Packet4f>(const Packet4f& a) {
return _mm_castps_si128(a);
}
@@ -77,6 +85,13 @@
return _mm_castsi128_ps(a);
}
+template<> EIGEN_STRONG_INLINE Packet2l preinterpret<Packet2l,Packet2d>(const Packet2d& a) {
+ return _mm_castpd_si128(a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d preinterpret<Packet2d, Packet2l>(const Packet2l& a) {
+ return _mm_castsi128_pd(a);
+}
// Disable the following code since it's broken on too many platforms / compilers.
//#elif defined(EIGEN_VECTORIZE_SSE) && (!EIGEN_ARCH_x86_64) && (!EIGEN_COMP_MSVC)
