Improve performance of ResizeBilinear:

- use pointer arithmetic instead of accessing data through eigen single-cell
  expressions.
- use only scale_down case, rename to resize_image. This simplifies the code,
  at the cost of slightly worse performance for the scale_up case with one
  channel (note though that special-casing single-channel case would make it 3x
  faster than the AFTER times below).
- pull some common arithmetic out of loops manually, so that more common
  indexing operations don't require multiplication.
- remove some assignments in inner loops.

(the first of these had the largest effect).

Add some more test and benchmark cases. Also fix a typo in the ResizeArea test.
Change: 146795860

1 files changed, 1 insertions, 1 deletions

diff --git a/tensorflow/core/kernels/resize_area_op_test.cc b/tensorflow/core/kernels/resize_area_op_test.cc
index 415bce3cce..cc5244d3a0 100644
--- a/tensorflow/core/kernels/resize_area_op_test.cc
+++ b/tensorflow/core/kernels/resize_area_op_test.cc
@@ -162,7 +162,7 @@ class ResizeAreaOpTest : public OpsTestBase {
       for (int in_h : {1, 3, 5, 8, 100, 233}) {
         for (int target_height : {1, 2, 3, 50, 113}) {
           for (int target_width : {target_height, target_height / 2 + 1}) {
-            RunRandomTest(in_w, in_h, target_height, target_width, channels);
+            RunRandomTest(in_h, in_w, target_height, target_width, channels);
           }
         }
       }