Address copilot comments

Author: jgibson2

kernels/portable/cpu/op_grid_sampler_2d.cpp

@@ -194,8 +194,10 @@ void grid_sample_2d_nearest_kernel_impl_nchw(
               y, inp_H, padding_mode, align_corners);
 
           // Get nearest pixel coordinates
-          // Use nearbyint (not round) to match ATen's rounding behavior
-          // see: aten/src/ATen/native/GridSampler.cpp
+          // Use nearbyint (not round) to match ATen's rounding behavior.
+          // nearbyint uses the current rounding mode (typically round-to-even),
+          // which matches PyTorch's (ATen's) behavior. In contrast, round may
+          // not always respect the rounding mode. See: aten/src/ATen/native/GridSampler.cpp
           int64_t ix_nearest = static_cast<int64_t>(std::nearbyint(ix));
           int64_t iy_nearest = static_cast<int64_t>(std::nearbyint(iy));
 
@@ -213,8 +215,8 @@ void grid_sample_2d_nearest_kernel_impl_nchw(
           } else {
             // For border/reflection padding, clip coordinates after rounding
             // Rounding can push coordinates out of bounds even after grid_sampler_compute_source_index
-            ix_nearest = std::max(static_cast<int64_t>(0), std::min(ix_nearest, inp_W - 1));
-            iy_nearest = std::max(static_cast<int64_t>(0), std::min(iy_nearest, inp_H - 1));
+            ix_nearest = clip_coordinates(ix_nearest, inp_W);
+            iy_nearest = clip_coordinates(iy_nearest, inp_H);
             out_val = in_data
                 [in_channel_offset + iy_nearest * in.strides()[2] +
                  ix_nearest * in.strides()[3]];

kernels/portable/cpu/util/grid_sampler_2d_util.h

@@ -17,6 +17,7 @@ namespace executor {
 
 // Ported from aten/src/ATen/native/GridSampler.h
 // note that these need to be in the SAME ORDER as the enum in GridSampler.h
+// as they are mapped to integer values (0, 1, 2) in this order
 enum class GridSamplerInterpolation {Bilinear, Nearest, Bicubic};
 enum class GridSamplerPadding {Zeros, Border, Reflection};
 

kernels/portable/test/op_grid_sampler_2d_test.cpp

@@ -200,6 +200,9 @@ TEST_F(OpGridSampler2dTest, BicubicSimple) {
       out);
 
   // Bicubic at center should be close to 8.5 (average of middle pixels)
+  // Note: The tolerance of 0.5 is intentionally large because the expected value (8.5)
+  // is a rough estimate (average of the middle pixels), not the exact bicubic interpolation result.
+  // Bicubic interpolation can produce values that differ from this average due to its mathematical properties.
   const auto expected = tf.make({1, 1, 1, 1}, {8.5});
   EXPECT_TENSOR_CLOSE_WITH_TOL(out, expected, 0, 0.5);
 }

kernels/portable/test/test_grid_sampler_2d_executorch.py

@@ -87,9 +87,8 @@ def run_executorch_test(
             executorch_output = fwd_method.execute((input_tensor, grid))[0]
 
             # Compare results
-            self.assertEqual(
-                executorch_output.shape,
-                pytorch_output.shape,
+            self.assertTrue(
+                executorch_output.shape == pytorch_output.shape,
                 msg=f"Shape mismatch: ET={executorch_output.shape} vs PT={pytorch_output.shape}",
             )