diff --git a/src/cudamatrix/cu-vector-speed-test.cc b/src/cudamatrix/cu-vector-speed-test.cc
index 258e0840815..82c471db9f9 100644
--- a/src/cudamatrix/cu-vector-speed-test.cc
+++ b/src/cudamatrix/cu-vector-speed-test.cc
@@ -191,6 +191,31 @@ template<typename Real> void TestCuVectorAddDiagMatMat(int32 dim,
 }
 
 
+template<typename Real> void TestCuVectorAddDiagMat2OnVariousShapes(
+    int32 dim, MatrixTransposeType trans) {
+  BaseFloat time_in_secs = 0.02;
+  int32 size = 1024 * 32;
+  CuVector<Real> v(trans == kNoTrans ? size / dim : dim);
+  v.SetRandn();
+  CuMatrix<Real> N(size / dim, dim);
+  N.SetRandn();
+
+  Timer tim;
+  int32 iter = 0;
+
+  for (; tim.Elapsed() < time_in_secs; iter++) {
+    v.AddDiagMat2(1.0, N, trans, 0.0);
+  }
+
+  BaseFloat fdim = size;
+  BaseFloat gflops = (fdim * iter) / (tim.Elapsed() * 1.0e+09);
+  KALDI_LOG << "For CuVector::AddDiagMat2Shapes" << NameOf<Real>()
+            << (trans == kTrans ? "[trans]" : "[no-trans]") << ", for dim = ("
+            << size / dim << ", " << dim  << "), speed was " << gflops << " gigaflops.";
+}
+
+
+
 template<typename Real> void TestCuVectorAddDiagMat2(int32 dim, MatrixTransposeType trans) {
   BaseFloat time_in_secs = 0.02;
   CuVector<Real> v(dim);
@@ -343,7 +368,6 @@ template<typename Real> void TestCuVectorApplyCeilingNoCount(int32 dim) {
 
 template<typename Real> void CudaVectorSpeedTest() {
   std::vector<int32> sizes;
-  sizes.push_back(16);
   sizes.push_back(32);
   sizes.push_back(64);
   sizes.push_back(128);
@@ -369,6 +393,10 @@ template<typename Real> void CudaVectorSpeedTest() {
     TestCuVectorAddDiagMatMat<Real>(sizes[s], kTrans, kNoTrans);
     TestCuVectorAddDiagMatMat<Real>(sizes[s], kTrans, kTrans);
   }
+  for (int32 s = 0; s < ns; s++) {
+    TestCuVectorAddDiagMat2OnVariousShapes<Real>(sizes[s], kNoTrans);
+    TestCuVectorAddDiagMat2OnVariousShapes<Real>(sizes[s], kTrans);
+  }
   for (int32 s = 0; s < ns; s++) {
     TestCuVectorAddDiagMat2<Real>(sizes[s], kNoTrans);
     TestCuVectorAddDiagMat2<Real>(sizes[s], kTrans);
@@ -415,4 +443,3 @@ int main() {
 #endif
   KALDI_LOG << "Tests succeeded.";
 }
-
diff --git a/src/cudamatrix/cu-vector.cc b/src/cudamatrix/cu-vector.cc
index f85d20d37f1..0bb652808f2 100644
--- a/src/cudamatrix/cu-vector.cc
+++ b/src/cudamatrix/cu-vector.cc
@@ -569,8 +569,13 @@ void CuVectorBase<Real>::AddDiagMat2(Real alpha, const CuMatrixBase<Real> &M,
   if (CuDevice::Instantiate().Enabled()) {
     if (dim_ == 0) return;
     MatrixTransposeType other_trans = (trans == kTrans ? kNoTrans : kTrans);
-    this->AddDiagMatMat(alpha, M, trans,
-                        M, other_trans, beta);
+    KALDI_ASSERT(dim_ == (trans == kNoTrans ? M.NumRows() : M.NumCols()));
+    if (trans == kTrans && M.NumCols() < 512 && M.NumRows() > 8192) {
+      CuMatrix<Real> MT(M, kTrans);
+      this->AddDiagMatMat(alpha, MT, other_trans, MT, trans, beta);
+    } else {
+      this->AddDiagMatMat(alpha, M, trans, M, other_trans, beta);
+    }
   } else
 #endif
   {