Add masked pstoreu to AVX and AVX512 PacketMath

1 files changed, 26 insertions, 1 deletions

diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 200670b8c..e704a53ea 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -119,10 +119,16 @@ struct packet_helper
   inline Packet load(const T* from) const { return internal::pload<Packet>(from); }
 
   template<typename T>
+  inline Packet loadu(const T* from) const { return internal::ploadu<Packet>(from); }
+
+  template<typename T>
   inline Packet load(const T* from, unsigned long long umask) const { return internal::ploadu<Packet>(from, umask); }
 
   template<typename T>
   inline void store(T* to, const Packet& x) const { internal::pstore(to,x); }
+
+  template<typename T>
+  inline void store(T* to, const Packet& x, unsigned long long umask) const { internal::pstoreu(to, x, umask); }
 };
 
 template<typename Packet>
@@ -132,10 +138,16 @@ struct packet_helper<false,Packet>
   inline T load(const T* from) const { return *from; }
 
   template<typename T>
+  inline T loadu(const T* from) const { return *from; }
+
+  template<typename T>
   inline T load(const T* from, unsigned long long) const { return *from; }
 
   template<typename T>
   inline void store(T* to, const T& x) const { *to = x; }
+
+  template<typename T>
+  inline void store(T* to, const T& x, unsigned long long) const { *to = x; }
 };
 
 #define CHECK_CWISE1_IF(COND, REFOP, POP) if(COND) { \
@@ -203,18 +215,31 @@ template<typename Scalar,typename Packet> void packetmath()
 
   if (internal::unpacket_traits<Packet>::masked_load_available)
   {
+    packet_helper<internal::unpacket_traits<Packet>::masked_load_available, Packet> h;
     unsigned long long max_umask = (0x1ull << PacketSize);
+
     for (int offset=0; offset<PacketSize; ++offset)
     {
       for (unsigned long long umask=0; umask<max_umask; ++umask)
       {
-        packet_helper<internal::unpacket_traits<Packet>::masked_load_available, Packet> h;
         h.store(data2, h.load(data1+offset, umask));
         for (int k=0; k<PacketSize; ++k)
           data3[k] = ((umask & ( 0x1ull << k )) >> k) ? data1[k+offset] : Scalar(0);
         VERIFY(areApprox(data3, data2, PacketSize) && "internal::ploadu masked");
       }
     }
+
+    for (int offset=0; offset<PacketSize; ++offset)
+    {
+      for (unsigned long long umask=0; umask<max_umask; ++umask)
+      {
+        internal::pstore(data2, internal::pset1<Packet>(Scalar(0)));
+        h.store(data2, h.loadu(data1+offset), umask);
+        for (int k=0; k<PacketSize; ++k)
+          data3[k] = ((umask & ( 0x1ull << k )) >> k) ? data1[k+offset] : Scalar(0);
+        VERIFY(areApprox(data3, data2, PacketSize) && "internal::pstoreu masked");
+      }
+    }
   }
 
   for (int offset=0; offset<PacketSize; ++offset)