Several improvements regarding packet-bitwise operations:

- add unit tests
- optimize their AVX512f implementation
- add missing implementations (half, Packet4f, ...)

1 files changed, 26 insertions, 0 deletions

diff --git a/Eigen/src/Core/arch/GPU/PacketMathHalf.h b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
index 8787adcde..cdd2b001b 100644
--- a/Eigen/src/Core/arch/GPU/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
@@ -640,6 +640,19 @@ EIGEN_STRONG_INLINE Packet16h float2half(const Packet16f& a) {
 #endif
 }
 
+template<> EIGEN_STRONG_INLINE Packet16h por(const Packet16h& a,const Packet16h& b) {
+  Packet16h r; r.x = por(a.x,b.x); return r;
+}
+template<> EIGEN_STRONG_INLINE Packet16h pxor(const Packet16h& a,const Packet16h& b) {
+  Packet16h r; r.x = pxor(a.x,b.x); return r;
+}
+template<> EIGEN_STRONG_INLINE Packet16h pand(const Packet16h& a,const Packet16h& b) {
+  Packet16h r; r.x = pand(a.x,b.x); return r;
+}
+template<> EIGEN_STRONG_INLINE Packet16h pandnot(const Packet16h& a,const Packet16h& b) {
+  Packet16h r; r.x = pandnot(a.x,b.x); return r;
+}
+
 template<> EIGEN_STRONG_INLINE Packet16h pnegate(const Packet16h& a) {
   // FIXME we could do that with bit manipulation
   Packet16f af = half2float(a);
@@ -1063,6 +1076,19 @@ EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) {
 #endif
 }
 
+template<> EIGEN_STRONG_INLINE Packet8h por(const Packet8h& a,const Packet8h& b) {
+  Packet8h r; r.x = por(a.x,b.x); return r;
+}
+template<> EIGEN_STRONG_INLINE Packet8h pxor(const Packet8h& a,const Packet8h& b) {
+  Packet8h r; r.x = pxor(a.x,b.x); return r;
+}
+template<> EIGEN_STRONG_INLINE Packet8h pand(const Packet8h& a,const Packet8h& b) {
+  Packet8h r; r.x = pand(a.x,b.x); return r;
+}
+template<> EIGEN_STRONG_INLINE Packet8h pandnot(const Packet8h& a,const Packet8h& b) {
+  Packet8h r; r.x = pandnot(a.x,b.x); return r;
+}
+
 template<> EIGEN_STRONG_INLINE Packet8h pconj(const Packet8h& a) { return a; }
 
 template<> EIGEN_STRONG_INLINE Packet8h pnegate(const Packet8h& a) {