arm_compute v18.08
diff --git a/src/runtime/CL/functions/CLScale.cpp b/src/runtime/CL/functions/CLScale.cpp
index cb68481..4ff9763 100644
--- a/src/runtime/CL/functions/CLScale.cpp
+++ b/src/runtime/CL/functions/CLScale.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -27,6 +27,7 @@
#include "arm_compute/core/CL/kernels/CLScaleKernel.h"
#include "arm_compute/core/Error.h"
#include "arm_compute/core/Validate.h"
+#include "arm_compute/runtime/CL/CLScheduler.h"
#include "support/ToolchainSupport.h"
using namespace arm_compute;
@@ -34,7 +35,18 @@
void CLScale::configure(ICLTensor *input, ICLTensor *output, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, SamplingPolicy sampling_policy)
{
auto k = arm_compute::support::cpp14::make_unique<CLScaleKernel>();
- k->configure(input, output, policy, border_mode == BorderMode::UNDEFINED, sampling_policy);
+ k->set_target(CLScheduler::get().target());
+ k->configure(input, output, policy, border_mode, sampling_policy);
_kernel = std::move(k);
+
+ // Tune kernels
+ CLScheduler::get().tune_kernel_static(*_kernel);
+
+ // In the case of NHWC we can't have undefined border mode as this would require to access elements outside z dimension,
+ // so we treat it like border constant.
+ if(border_mode == BorderMode::UNDEFINED && input->info()->data_layout() == DataLayout::NHWC)
+ {
+ border_mode = BorderMode::CONSTANT;
+ }
_border_handler.configure(input, _kernel->border_size(), border_mode, constant_border_value);
}