COMPMID-3145: Remove padding from NEScaleKernel

Change-Id: I530b12c6270d7dbeb3ef7af62484842ebcb65925
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4000
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

1 files changed, 16 insertions, 49 deletions

diff --git a/src/runtime/NEON/functions/NEScale.cpp b/src/runtime/NEON/functions/NEScale.cpp
index 424049f24a..2278f07a1c 100644
--- a/src/runtime/NEON/functions/NEScale.cpp
+++ b/src/runtime/NEON/functions/NEScale.cpp
@@ -44,7 +44,7 @@ namespace arm_compute
 {
 namespace
 {
-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)
+void precompute_dx_dy_offsets(ITensor *dx, ITensor *dy, ITensor *offsets, float wr, float hr, SamplingPolicy sampling_policy, bool align_corners)
 {
     ARM_COMPUTE_ERROR_ON(nullptr == offsets);
     ARM_COMPUTE_UNUSED(sampling_policy);
@@ -72,7 +72,7 @@ void precompute_dx_dy_offsets(ITensor *dx, ITensor *dy, ITensor *offsets, float
             const int   in_xi = std::floor(in_x);
             const int   in_yi = std::floor(in_y);
 
-            *reinterpret_cast<int32_t *>(offsets_it.ptr()) = in_xi * static_cast<int>(input_element_size);
+            *reinterpret_cast<int32_t *>(offsets_it.ptr()) = in_xi;
             *reinterpret_cast<float *>(dx_it.ptr())        = in_x - in_xi;
             *reinterpret_cast<float *>(dy_it.ptr())        = in_y - in_yi;
         },
@@ -85,23 +85,17 @@ void precompute_dx_dy_offsets(ITensor *dx, ITensor *dy, ITensor *offsets, float
 
         execute_window_loop(win, [&](const Coordinates & id)
         {
-            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;
+            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;
         },
         offsets_it);
     }
 }
 } // namespace
 
-NEScale::NEScale() // NOLINT
-    : _offsets(),
-      _dx(),
-      _dy(),
-      _scale_kernel(),
-      _border_handler(),
-      _use_padding(true)
+NEScale::NEScale()
+    : _offsets(), _dx(), _dy()
 {
 }
 
@@ -110,7 +104,6 @@ void NEScale::configure(ITensor *input, ITensor *output, const ScaleKernelInfo &
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_ERROR_THROW_ON(NEScale::validate(input->info(), output->info(), info));
 
-    _use_padding                     = info.use_padding;
     const bool is_align_corners_used = info.align_corners && arm_compute::scale_utils::is_align_corners_allowed_sampling_policy(info.sampling_policy);
 
     // Get data layout and width/height indices
@@ -119,18 +112,17 @@ void NEScale::configure(ITensor *input, ITensor *output, const ScaleKernelInfo &
     const int        idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
 
     // Get the tensor shape
-    const TensorShape shape(output->info()->dimension(idx_width), output->info()->dimension(idx_height));
+    TensorShape shape(output->info()->dimension(idx_width));
+    shape.set(1, output->info()->dimension(idx_height), false);
 
     // Compute the ratio between source width/height and destination width/height
     const auto wr = arm_compute::scale_utils::calculate_resize_ratio(input->info()->dimension(idx_width), output->info()->dimension(idx_width), is_align_corners_used);
     const auto hr = arm_compute::scale_utils::calculate_resize_ratio(input->info()->dimension(idx_height), output->info()->dimension(idx_height), is_align_corners_used);
 
-    // Get the element size of the input image
-    const size_t input_element_size = input->info()->element_size();
-
     // Area interpolation behaves as Nearest Neighbour in case of up-sampling
     const auto policy_to_use = (info.interpolation_policy == InterpolationPolicy::AREA && wr <= 1.f && hr <= 1.f) ? InterpolationPolicy::NEAREST_NEIGHBOR : info.interpolation_policy;
 
+    auto scale_kernel = arm_compute::support::cpp14::make_unique<NEScaleKernel>();
     switch(policy_to_use)
     {
         case InterpolationPolicy::NEAREST_NEIGHBOR:
@@ -138,13 +130,13 @@ void NEScale::configure(ITensor *input, ITensor *output, const ScaleKernelInfo &
             TensorInfo tensor_info_offsets(shape, Format::S32);
             _offsets.allocator()->init(tensor_info_offsets);
 
-            _scale_kernel.configure(input, nullptr, nullptr, &_offsets, output, info);
+            scale_kernel->configure(input, nullptr, nullptr, &_offsets, output, info);
 
             // Allocate once the configure methods have been called
             _offsets.allocator()->allocate();
 
             // Pre-compute offsets for nearest interpolation
-            precompute_dx_dy_offsets(nullptr, nullptr, &_offsets, wr, hr, input_element_size, info.sampling_policy, is_align_corners_used);
+            precompute_dx_dy_offsets(nullptr, nullptr, &_offsets, wr, hr, info.sampling_policy, is_align_corners_used);
             break;
         }
         case InterpolationPolicy::BILINEAR:
@@ -156,7 +148,7 @@ void NEScale::configure(ITensor *input, ITensor *output, const ScaleKernelInfo &
             _dx.allocator()->init(tensor_info_dxdy);
             _dy.allocator()->init(tensor_info_dxdy);
 
-            _scale_kernel.configure(input, &_dx, &_dy, &_offsets, output, info);
+            scale_kernel->configure(input, &_dx, &_dy, &_offsets, output, info);
 
             // Allocate once the configure methods have been called
             _offsets.allocator()->allocate();
@@ -164,27 +156,18 @@ void NEScale::configure(ITensor *input, ITensor *output, const ScaleKernelInfo &
             _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, is_align_corners_used);
+            precompute_dx_dy_offsets(&_dx, &_dy, &_offsets, wr, hr, info.sampling_policy, is_align_corners_used);
             break;
         }
         case InterpolationPolicy::AREA:
         {
-            _scale_kernel.configure(input, nullptr, nullptr, nullptr, output, info);
+            scale_kernel->configure(input, nullptr, nullptr, nullptr, output, info);
             break;
         }
         default:
             ARM_COMPUTE_ERROR("Unsupported interpolation mode");
     }
-    if(info.use_padding)
-    {
-        _border_handler.configure(input, _scale_kernel.border_size(), info.border_mode, info.constant_border_value);
-    }
-}
-
-void NEScale::configure(ITensor *input, ITensor *output, InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, SamplingPolicy sampling_policy, bool use_padding,
-                        bool align_corners)
-{
-    configure(input, output, ScaleKernelInfo{ policy, border_mode, constant_border_value, sampling_policy, use_padding, align_corners });
+    _kernel = std::move(scale_kernel);
 }
 
 Status NEScale::validate(const ITensorInfo *input, const ITensorInfo *output, const ScaleKernelInfo &info)
@@ -225,20 +208,4 @@ Status NEScale::validate(const ITensorInfo *input, const ITensorInfo *output, co
     ARM_COMPUTE_RETURN_ON_ERROR(NEScaleKernel::validate(input->clone().get(), dx, dy, offsets, output->clone().get(), info));
     return Status{};
 }
-
-Status NEScale::validate(const ITensorInfo *input, const ITensorInfo *output, InterpolationPolicy policy,
-                         BorderMode border_mode, PixelValue constant_border_value, SamplingPolicy sampling_policy, bool use_padding, bool align_corners)
-{
-    ARM_COMPUTE_RETURN_ON_ERROR(NEScale::validate(input, output, ScaleKernelInfo{ policy, border_mode, constant_border_value, sampling_policy, use_padding, align_corners }));
-    return Status{};
-}
-
-void NEScale::run()
-{
-    if(_use_padding)
-    {
-        NEScheduler::get().schedule(&_border_handler, Window::DimZ);
-    }
-    NEScheduler::get().schedule(&_scale_kernel, Window::DimY);
-}
 } // namespace arm_compute