Xが2次元、Yが1次元の場合の処理を追加

python/test/function/test_broadcast_to.py

@@ -22,17 +22,25 @@
     function_tester,
     list_ctx_and_func_name)
 
+
 def copying_to_leaf(x, y, axis):
     return (len(x.shape) - len(y.shape) - axis) == 0
 
+
 def ref_broadcast_to(x, y, axis):
     if axis < 0 or copying_to_leaf(x, y, axis):
+        # Copy data to leaf
         return np.ones(x.shape) * y
     else:
-        return np.ones(x.shape)
+        # Copy data from specified axis
+        if len(x.shape) == 2:
+            t = y[:,np.newaxis]
+            t.transpose()
+            return np.tile(t, (1, x.shape[1]))
 
 
 PARAMS = [
+    ((2, 3), (2), 0),
     ((2, 3), (3), 1),
     ((2, 3, 4), (4), 2),
     ((2, 3, 4), (3, 4), 1),

src/nbla/function/broadcast_to.cpp

@@ -111,8 +111,17 @@ void BroadcastTo<T>::forward_impl(const Variables &inputs, const Variables &outp
 				// Y dimension size is always 1
 				switch(axis_) {
 					case 0:
-						// X: (2,3) Y: (2) axis=0
-						// copy Y values vertically
+						{
+							// X: (2,3) Y: (2) axis=0
+							// copy Y values vertically
+							Size_t height = xs[0];
+							Size_t width = xs[1];
+							for (Size_t v=0; v<height; v++) {
+								T val = y[v];
+								T* zrow = &z[v*width];
+								std::fill(zrow, zrow+width, val);
+							}
+						}
 						break;
 					case 1:
 						// X: (2,3) Y: (3) axis=1