[src] Speed up VectorBase<Real>::Sum() by using BLAS (#2394)

src/cudamatrix/cu-vector-test.cc

@@ -345,7 +345,16 @@ template<typename Real> void CuVectorUnitTestSum() {
     A.SetRandn();
     ones.Set(1.0);
 
-    AssertEqual(VecVec(A, ones), A.Sum());
+    Real x = VecVec(A, ones);
+    Real y = A.Sum();
+    Real diff = std::abs(x - y);
+    // Note: CuVectorBase<> does not have an ApplyAbs() member
+    // function, so we copy back to a host vector for simplicity in
+    // this test case.
+    Vector<Real> A_host(A);
+    A_host.ApplyAbs();
+    Real s = A_host.Sum();
+    KALDI_ASSERT ( diff <= 1.0e-04 * s);
   }
 }
 

src/matrix/kaldi-vector.cc

@@ -691,9 +691,11 @@ void VectorBase<Real>::CopyDiagFromPacked(const PackedMatrix<Real> &M) {
 
 template<typename Real>
 Real VectorBase<Real>::Sum() const {
-  double sum = 0.0;
-  for (MatrixIndexT i = 0; i < dim_; i++) { sum += data_[i]; }
-  return sum;
+  // Do a dot-product with a size-1 array with a stride of 0 to
+  // implement sum. This allows us to access SIMD operations in a
+  // cross-platform way via your BLAS library.
+  Real one(1);
+  return cblas_Xdot(dim_, data_, 1, &one, 0);
 }
 
 template<typename Real>