commit 94d5051bc56a59d857f8a09560e2da5c0d7894b0
author Sang-Hoon Park <sang-hoon.park@arm.com> Thu Jul 02 10:49:39 2020 +0100
committer Georgios Pinitas <georgios.pinitas@arm.com> Fri Jul 03 17:39:20 2020 +0000
tree 577d9ce10f73608a71c033b495b54457e6c5b1ab
parent 2aad21a900a21f467b3ec6b37420f892f0d80221

COMPMID-3539: Change indexing for nearest neighbor with aligned corners

For nearest neighbor interpolation policy with aligned corners
all of NEON, CL and reference use round() rather than float to
find the nearest integer.

Change-Id: If0360da870e983303bf0424ca1100084084c1efc
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3495
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

src/runtime/NEON/functions/NEScale.cpp

diff --git a/src/runtime/NEON/functions/NEScale.cpp b/src/runtime/NEON/functions/NEScale.cpp
index 28fbab4..eefdfdb 100644
--- a/src/runtime/NEON/functions/NEScale.cpp
+++ b/src/runtime/NEON/functions/NEScale.cpp
@@ -30,6 +30,7 @@
 #include "arm_compute/core/PixelValue.h"
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/Rounding.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 #include "arm_compute/runtime/TensorAllocator.h"
 
@@ -43,7 +44,7 @@
 {
 namespace
 {
-void precompute_dx_dy_offsets(ITensor *dx, ITensor *dy, ITensor *offsets, float wr, float hr, size_t input_element_size, SamplingPolicy sampling_policy)
+void precompute_dx_dy_offsets(ITensor *dx, ITensor *dy, ITensor *offsets, float wr, float hr, size_t input_element_size, SamplingPolicy sampling_policy, bool align_corners)
 {
     ARM_COMPUTE_ERROR_ON(nullptr == offsets);
     ARM_COMPUTE_UNUSED(sampling_policy);
@@ -84,7 +85,8 @@
 
         execute_window_loop(win, [&](const Coordinates & id)
         {
-            const size_t in_xi = std::floor((id.x() + sampling_offset) * wr);
+            const float float_in_xi = (id.x() + sampling_offset) * wr;
+            const auto  in_xi       = static_cast<size_t>(align_corners ? arm_compute::utils::rounding::round_half_away_from_zero(float_in_xi) : std::floor(float_in_xi));
 
             *reinterpret_cast<int32_t *>(offsets_it.ptr()) = in_xi * input_element_size;
         },
@@ -143,7 +145,7 @@
             _offsets.allocator()->allocate();
 
             // Pre-compute offsets for nearest interpolation
-            precompute_dx_dy_offsets(nullptr, nullptr, &_offsets, wr, hr, input_element_size, info.sampling_policy);
+            precompute_dx_dy_offsets(nullptr, nullptr, &_offsets, wr, hr, input_element_size, info.sampling_policy, is_align_corners_used);
             break;
         }
         case InterpolationPolicy::BILINEAR:
@@ -163,7 +165,7 @@
             _dy.allocator()->allocate();
 
             // Pre-compute dx, dy and offsets for bilinear interpolation
-            precompute_dx_dy_offsets(&_dx, &_dy, &_offsets, wr, hr, input_element_size, info.sampling_policy);
+            precompute_dx_dy_offsets(&_dx, &_dy, &_offsets, wr, hr, input_element_size, info.sampling_policy, is_align_corners_used);
             break;
         }
         case InterpolationPolicy::AREA: