fix bicubic, change int to floor (#24063)

* change int to floor, test=develop

* fix unittest, test=develop

paddle/fluid/operators/interpolate_op.cu

@@ -553,13 +553,13 @@ __global__ void KeBicubicInterpFw(
     T in_img_idy = align_corners
                        ? static_cast<T>(ratio_h * out_img_idy)
                        : static_cast<T>(ratio_h * (out_img_idy + 0.5) - 0.5);
-    int input_y = static_cast<int>(in_img_idy);
+    int input_y = floorf(in_img_idy);
     const T y_t = in_img_idy - input_y;
 
     T in_img_idx = align_corners
                        ? static_cast<T>(ratio_w * out_img_idx)
                        : static_cast<T>(ratio_w * (out_img_idx + 0.5) - 0.5);
-    int input_x = static_cast<int>(in_img_idx);
+    int input_x = floorf(in_img_idx);
     const T x_t = in_img_idx - input_x;
 
     T coefficients[4];
@@ -665,13 +665,13 @@ __global__ void KeBicubicInterpBw(
     T in_img_idy = align_corners
                        ? static_cast<T>(ratio_h * out_img_idy)
                        : static_cast<T>(ratio_h * (out_img_idy + 0.5) - 0.5);
-    int input_y = static_cast<int>(in_img_idy);
+    int input_y = floorf(in_img_idy);
     const T y_t = in_img_idy - input_y;
 
     T in_img_idx = align_corners
                        ? static_cast<T>(ratio_w * out_img_idx)
                        : static_cast<T>(ratio_w * (out_img_idx + 0.5) - 0.5);
-    int input_x = static_cast<int>(in_img_idx);
+    int input_x = floorf(in_img_idx);
 
     const T x_t = in_img_idx - input_x;
 

paddle/fluid/operators/interpolate_op.h

@@ -389,13 +389,13 @@ static void BicubicInterpolation(const Tensor& input, Tensor* output,
   for (int k = 0; k < out_h; k++) {  // loop for images
     T y_n = align_corners ? static_cast<T>(ratio_h * k)
                           : static_cast<T>(ratio_h * (k + 0.5) - 0.5);
-    int input_y = static_cast<int>(y_n);
+    int input_y = floorf(y_n);
     const T y_t = y_n - input_y;
 
     for (int l = 0; l < out_w; l++) {
       T x_n = align_corners ? static_cast<T>(ratio_w * l)
                             : static_cast<T>(ratio_w * (l + 0.5) - 0.5);
-      int input_x = static_cast<int>(x_n);
+      int input_x = floorf(x_n);
       const T x_t = x_n - input_x;
 
       for (int i = 0; i < n; i++) {    // loop for batches
@@ -625,13 +625,13 @@ static void BicubicInterpolationGrad(const Tensor& output_grad,
   for (int k = 0; k < out_h; k++) {  // loop for images
     T y_n = align_corners ? static_cast<T>(ratio_h * k)
                           : static_cast<T>(ratio_h * (k + 0.5) - 0.5);
-    int input_y = static_cast<int>(y_n);
+    int input_y = floorf(y_n);
     T y_t = y_n - input_y;
 
     for (int l = 0; l < out_w; l++) {
       T x_n = align_corners ? static_cast<T>(ratio_w * l)
                             : static_cast<T>(ratio_w * (l + 0.5) - 0.5);
-      int input_x = static_cast<int>(x_n);
+      int input_x = floorf(x_n);
       T x_t = x_n - input_x;
 
       T x_coeffs[4];

python/paddle/fluid/tests/unittests/test_bicubic_interp_op.py

@@ -88,14 +88,14 @@ def bicubic_interp_np(input,
             h = ratio_h * k
         else:
             h = ratio_h * (k + 0.5) - 0.5
-        input_y = int(h)
+        input_y = np.floor(h)
         y_t = h - input_y
         for l in range(out_w):
             if (align_corners):
                 w = ratio_w * l
             else:
                 w = ratio_w * (l + 0.5) - 0.5
-            input_x = int(w)
+            input_x = np.floor(w)
             x_t = w - input_x
             for i in range(batch_size):
                 for j in range(channel):
@@ -265,7 +265,7 @@ class TestBicubicInterpDataLayout(TestBicubicInterpOp):
 
 class TestBicubicInterpOpAPI(unittest.TestCase):
     def test_case(self):
-
+        np.random.seed(200)
         x_data = np.random.random((2, 3, 6, 6)).astype("float32")
         dim_data = np.array([12]).astype("int32")
         shape_data = np.array([12, 12]).astype("int32")