change to unaligned load

dmlc · Oct 17, 2015 · aa77e19 · aa77e19
1 parent de24af8
commit aa77e19
Show file tree

Hide file tree

Showing 3 changed files with 12 additions and 1 deletion.
diff --git a/mshadow/extension/implicit_gemm.h b/mshadow/extension/implicit_gemm.h
@@ -77,7 +77,7 @@ struct Plan<ImplicitGEMMExp<LhsExp, RhsExp, DType>, DType> {
       for (index_t j = 0; j < Packet::kSize; ++j) {
         rhs_temp[j] = rhs_.Eval(i + j, x);
       }
-      sum = sum + Packet::Load(lhs_temp) * Packet::Load(rhs_temp);
+      sum = sum + Packet::LoadUnAligned(lhs_temp) * Packet::LoadUnAligned(rhs_temp);
     }
     DType ret_result = sum.Sum();
 

diff --git a/mshadow/packet/plain-inl.h b/mshadow/packet/plain-inl.h
@@ -30,6 +30,10 @@ struct Packet<DType, kPlain> {
   MSHADOW_CINLINE static Packet<DType, kPlain> Load(const DType* src) {
     return Packet<DType, kPlain>(*src);
   }
+  // load from address
+  MSHADOW_CINLINE static Packet<DType, kPlain> LoadUnAligned(const DType* src) {
+    return Packet<DType, kPlain>(*src);
+  }
   // fill it with value s
   MSHADOW_CINLINE Packet<DType, kPlain>& operator=(DType s) {
     data_ = s;

diff --git a/mshadow/packet/sse-inl.h b/mshadow/packet/sse-inl.h
@@ -32,6 +32,10 @@ struct Packet<float, kSSE2> {
   MSHADOW_CINLINE static Packet<float, kSSE2> Load(const float* src) {
     return Packet<float, kSSE2>(_mm_load_ps(src));
   }
+  // load from address
+  MSHADOW_CINLINE static Packet<float, kSSE2> LoadUnAligned(const float* src) {
+    return Packet<float, kSSE2>(_mm_loadu_ps(src));
+  }
   // fill it with value s
   MSHADOW_CINLINE Packet<float, kSSE2>& operator=(float s) {
     data_ = _mm_set1_ps(s);
@@ -73,6 +77,9 @@ struct Packet<double, kSSE2> {
   MSHADOW_CINLINE static Packet<double, kSSE2> Load(const double* src) {
     return Packet<double, kSSE2>(_mm_load_pd(src));
   }
+  MSHADOW_CINLINE static Packet<double, kSSE2> LoadUnAligned(const double* src) {
+    return Packet<double, kSSE2>(_mm_loadu_pd(src));
+  }
   // fill it with value s
   MSHADOW_CINLINE Packet<double, kSSE2>& operator=(double s) {
     data_ = _mm_set1_pd(s);