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

Move eigen_packet_wrapper to GenericPacketMath.h and use it for SSE/AVX/AVX512 as it is already used for NEON.

Browse Source 
This will allow us to define multiple packet types backed by the same vector type, e.g., __m128i.
Use this machanism to define packets for half and clean up the packet op implementations.
This commit is contained in:
Rasmus Munk Larsen
2020-04-15 18:17:19 +00:00
parent 4aae8ac693
commit 5ab87d8aba
5 changed files with 134 additions and 205 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
Eigen/src/Core/arch/SSE/PacketMath.h
View File

@@ -36,29 +36,15 @@ namespace internal {
// One solution is to increase ABI version using -fabi-version=4 (or greater).
// Otherwise, we workaround this inconvenience by wrapping 128bit types into the following helper
// structure:
template<typename T>
struct eigen_packet_wrapper
{
EIGEN_ALWAYS_INLINE operator T&() { return m_val; }
EIGEN_ALWAYS_INLINE operator const T&() const { return m_val; }
EIGEN_ALWAYS_INLINE eigen_packet_wrapper() {}
EIGEN_ALWAYS_INLINE eigen_packet_wrapper(const T &v) : m_val(v) {}
EIGEN_ALWAYS_INLINE eigen_packet_wrapper& operator=(const T &v) {
m_val = v;
return *this;
}
T m_val;
};
typedef eigen_packet_wrapper<__m128> Packet4f;
typedef eigen_packet_wrapper<__m128i> Packet4i;
typedef eigen_packet_wrapper<__m128d> Packet2d;
#else
typedef __m128 Packet4f;
typedef __m128i Packet4i;
typedef __m128d Packet2d;
#endif
typedef eigen_packet_wrapper<__m128i, 0> Packet4i;
template<> struct is_arithmetic<__m128> { enum { value = true }; };
template<> struct is_arithmetic<__m128i> { enum { value = true }; };
template<> struct is_arithmetic<__m128d> { enum { value = true }; };
@@ -400,7 +386,6 @@ template<> EIGEN_STRONG_INLINE Packet2d pcmp_eq(const Packet2d& a, const Packet2
template<> EIGEN_STRONG_INLINE Packet4i pcmp_lt(const Packet4i& a, const Packet4i& b) { return _mm_cmplt_epi32(a,b); }
template<> EIGEN_STRONG_INLINE Packet4i pcmp_eq(const Packet4i& a, const Packet4i& b) { return _mm_cmpeq_epi32(a,b); }
template<> EIGEN_STRONG_INLINE Packet4i ptrue<Packet4i>(const Packet4i& a) { return _mm_cmpeq_epi32(a, a); }
template<> EIGEN_STRONG_INLINE Packet4f
ptrue<Packet4f>(const Packet4f& a) {
@@ -521,7 +506,6 @@ template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
}
template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
{
return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd(reinterpret_cast<const double*>(from))), 0, 0, 1, 1);