commit a2611815f278334c801094d095901d36e111c3f9
author Giorgio Arena <giorgio.arena@arm.com> Fri Jul 21 10:08:48 2017 +0100
committer Anthony Barbier <anthony.barbier@arm.com> Mon Sep 17 14:16:42 2018 +0100
tree 1a2bf86b37614ce5473669b283dc37c708dd4957
parent a629da13544c361ba9580a233109b2418fd6bb73

COMPMID-417 NEON/CL MeanStdDev bugfix using FillBorderKernel

Change-Id: Ic48ba7f69783d0e1e80611264e2bc67d1732436e
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/81293
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

src/core/NEON/kernels/NEMeanStdDevKernel.cpp

diff --git a/src/core/NEON/kernels/NEMeanStdDevKernel.cpp b/src/core/NEON/kernels/NEMeanStdDevKernel.cpp
index 4616203..1eb7e45 100644
--- a/src/core/NEON/kernels/NEMeanStdDevKernel.cpp
+++ b/src/core/NEON/kernels/NEMeanStdDevKernel.cpp
@@ -85,10 +85,15 @@
 } // namespace
 
 NEMeanStdDevKernel::NEMeanStdDevKernel()
-    : _input(nullptr), _mean(nullptr), _stddev(nullptr), _global_sum(nullptr), _global_sum_squared(nullptr), _mtx()
+    : _input(nullptr), _mean(nullptr), _stddev(nullptr), _global_sum(nullptr), _global_sum_squared(nullptr), _mtx(), _border_size(0)
 {
 }
 
+BorderSize NEMeanStdDevKernel::border_size() const
+{
+    return _border_size;
+}
+
 void NEMeanStdDevKernel::configure(const IImage *input, float *mean, uint64_t *global_sum, float *stddev, uint64_t *global_sum_squared)
 {
     ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input);
@@ -105,6 +110,9 @@
 
     constexpr unsigned int num_elems_processed_per_iteration = 16;
 
+    _border_size = BorderSize(std::max(static_cast<int>(num_elems_processed_per_iteration) - static_cast<int>(input->info()->dimension(0)),
+                                       static_cast<int>(input->info()->dimension(0) % num_elems_processed_per_iteration)));
+
     // Configure kernel window
     Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration));