Optimize performance by using inline MatrixDotVectorInternal

This improves performace for the "best" models because it
avoids function calls.

The compiler also knows the passed values for the parameters
add_bias_fwd and skip_bias_back.

Signed-off-by: Stefan Weil <[email protected]>

Author: stweil

src/lstm/weightmatrix.cpp

@@ -38,6 +38,28 @@ const int kAdamCorrectionIterations = 200000;
 // Epsilon in Adam to prevent division by zero.
 const double kAdamEpsilon = 1e-8;
 
+// Computes matrix.vector v = Wu.
+// u is of size W.dim2() - add_bias_fwd and the output v is of size
+// W.dim1() - skip_bias_back.
+// If add_bias_fwd, u is imagined to have an extra element at the end with value
+// 1, to implement the bias, weight.
+// If skip_bias_back, we are actullay performing the backwards product on a
+// transposed matrix, so we need to drop the v output corresponding to the last
+// element in dim1.
+static inline void MatrixDotVectorInternal(const GENERIC_2D_ARRAY<double>& w,
+                                           bool add_bias_fwd,
+                                           bool skip_bias_back, const double* u,
+                                           double* v) {
+  int num_results = w.dim1() - skip_bias_back;
+  int extent = w.dim2() - add_bias_fwd;
+  for (int i = 0; i < num_results; ++i) {
+    const double* wi = w[i];
+    double total = WeightMatrix::DotProduct(wi, u, extent);
+    if (add_bias_fwd) total += wi[extent];  // The bias value.
+    v[i] = total;
+  }
+}
+
 // Copies the whole input transposed, converted to double, into *this.
 void TransposedArray::Transpose(const GENERIC_2D_ARRAY<double>& input) {
   int width = input.dim1();
@@ -401,26 +423,4 @@ void WeightMatrix::FloatToDouble(const GENERIC_2D_ARRAY<float>& wf,
   }
 }
 
-// Computes matrix.vector v = Wu.
-// u is of size W.dim2() - add_bias_fwd and the output v is of size
-// W.dim1() - skip_bias_back.
-// If add_bias_fwd, u is imagined to have an extra element at the end with value
-// 1, to implement the bias, weight.
-// If skip_bias_back, we are actullay performing the backwards product on a
-// transposed matrix, so we need to drop the v output corresponding to the last
-// element in dim1.
-void WeightMatrix::MatrixDotVectorInternal(const GENERIC_2D_ARRAY<double>& w,
-                                           bool add_bias_fwd,
-                                           bool skip_bias_back, const double* u,
-                                           double* v) {
-  int num_results = w.dim1() - skip_bias_back;
-  int extent = w.dim2() - add_bias_fwd;
-  for (int i = 0; i < num_results; ++i) {
-    const double* wi = w[i];
-    double total = DotProduct(wi, u, extent);
-    if (add_bias_fwd) total += wi[extent];  // The bias value.
-    v[i] = total;
-  }
-}
-
 }  // namespace tesseract.

src/lstm/weightmatrix.h

@@ -159,17 +159,6 @@ class WeightMatrix {
   static void FloatToDouble(const GENERIC_2D_ARRAY<float>& wf,
                             GENERIC_2D_ARRAY<double>* wd);
 
- private:
-  // Computes matrix.vector v = Wu.
-  // u is of size starts.back()+extents.back() and the output v is of size
-  // starts.size().
-  // The weight matrix w, is of size starts.size()xMAX(extents)+add_bias_fwd.
-  // If add_bias_fwd, an extra element at the end of w[i] is the bias weight
-  // and is added to v[i].
-  static void MatrixDotVectorInternal(const GENERIC_2D_ARRAY<double>& w,
-                                      bool add_bias_fwd, bool skip_bias_back,
-                                      const double* u, double* v);
-
  private:
   // Choice between float and 8 bit int implementations.
   GENERIC_2D_ARRAY<double> wf_;