Add packet up "pones". Write pnot(a) as pxor(pones(a), a).

1 files changed, 16 insertions, 0 deletions

diff --git a/Eigen/src/Core/arch/GPU/PacketMathHalf.h b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
index 3e35f96cc..c4dfedcf8 100644
--- a/Eigen/src/Core/arch/GPU/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
@@ -143,6 +143,10 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pabs<half2>(const half2&
   return result;
 }
 
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pones<half2>(const half2& a) {
+  half2 result;
+  *(reinterpret_cast<unsigned*>(&(result))) = 0xffffffffu;
+}  
 
 EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void
 ptranspose(PacketBlock<half2,2>& kernel) {
@@ -640,6 +644,14 @@ EIGEN_STRONG_INLINE Packet16h float2half(const Packet16f& a) {
 #endif
 }
 
+template<> EIGEN_STRONG_INLINE Packet16h pnot(const Packet16h& a) {
+  Packet16h r; r.x = _mm256_xor_si256(a.x, pcmp_eq(a.x, a.x)); return r;
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pones(const Packet16h& a) {
+  Packet16h r; r.x = Packet8i(pones(a.x)); return r;
+}
+
 template<> EIGEN_STRONG_INLINE Packet16h por(const Packet16h& a,const Packet16h& b) {
   // in some cases Packet8i is a wrapper around __m256i, so we need to 
   // cast to Packet8i to call the correct overload.
@@ -1085,6 +1097,10 @@ EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) {
 #endif
 }
 
+template<> EIGEN_STRONG_INLINE Packet8h pones(const Packet8h& a) {
+  Packet8h r; r.x = _mm_cmpeq_epi32(a.x, a.x); return r;
+}
+
 template<> EIGEN_STRONG_INLINE Packet8h por(const Packet8h& a,const Packet8h& b) {
   // in some cases Packet4i is a wrapper around __m128i, so we either need to 
   // cast to Packet4i to directly call the intrinsics as below: