1 files changed, 8 insertions, 31 deletions

diff --git a/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp
index 8352c94586..b360e9e6be 100644
--- a/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEDepthConcatenateLayerKernel.cpp
@@ -42,18 +42,13 @@ using namespace arm_compute;
 namespace
 {
 template <typename T>
-void depth_concat(const ITensor *in, ITensor *out, std::pair<int, int> start_xy, int depth_offset, const Window &window)
+void depth_concat(const ITensor *in, ITensor *out, int depth_offset, const Window &window)
 {
-    const int start_x = start_xy.first;
-    const int start_y = start_xy.second;
-
     // Offset input
-    const int input_offset_to_first_elements_in_bytes = in->info()->offset_first_element_in_bytes() - start_x * in->info()->strides_in_bytes()[0] - start_y * in->info()->strides_in_bytes()[1];
-    uint8_t *input_ptr                               = in->buffer() + input_offset_to_first_elements_in_bytes;
+    uint8_t *input_ptr = in->buffer() + in->info()->offset_first_element_in_bytes();
 
     // Offset output
-    const unsigned int output_offset_to_first_elements_in_bytes = out->info()->offset_first_element_in_bytes() + depth_offset * out->info()->strides_in_bytes()[2];
-    uint8_t           *output_ptr                               = out->buffer() + output_offset_to_first_elements_in_bytes;
+    uint8_t *output_ptr = out->buffer() + out->info()->offset_first_element_in_bytes() + depth_offset * out->info()->strides_in_bytes()[2];
 
     Iterator input(in, window);
     Iterator output(out, window);
@@ -88,19 +83,13 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsi
 {
     ARM_COMPUTE_UNUSED(depth_offset);
 
-    // Configure kernel window
-    const int left_right = (output->dimension(0) - input->dimension(0)) / 2;
-    const int top_bottom = (output->dimension(1) - input->dimension(1)) / 2;
-
     const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
-    const unsigned int num_elems_read_per_iteration      = 16 / input->element_size();
-    const unsigned int num_rows_read_per_iteration       = 1;
 
     // The window needs to be based on input as we copy all the depths of input
     Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
     win.set(Window::DimZ, Window::Dimension(0, input->tensor_shape().z(), 1));
 
-    AccessWindowRectangle  input_access(input, -left_right, -top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration);
+    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
     AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
     bool                   window_changed = update_window_and_padding(win, input_access, output_access);
     output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
@@ -116,28 +105,18 @@ Status validate_arguments(const ITensorInfo *input, unsigned int depth_offset, c
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
 
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimX) != output->dimension(Window::DimX));
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimY) != output->dimension(Window::DimY));
     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) + depth_offset > output->dimension(2));
-    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) > output->dimension(0));
-    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) > output->dimension(1));
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(3, input, output);
 
-    // The gaps between the two lowest dimensions of input and output need to be divisible by 2
-    // Otherwise it is not clear how the padding should be added onto the input tensor
-    ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(0) - input->dimension(0)) % 2);
-    ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(1) - input->dimension(1)) % 2);
-
     return Status{};
 }
 } // namespace
 
 NEDepthConcatenateLayerKernel::NEDepthConcatenateLayerKernel()
-    : _func(nullptr), _input(nullptr), _output(nullptr), _top_bottom(0), _left_right(0), _depth_offset(0)
-{
-}
-
-BorderSize NEDepthConcatenateLayerKernel::border_size() const
+    : _func(nullptr), _input(nullptr), _output(nullptr), _depth_offset(0)
 {
-    return BorderSize(_top_bottom, _left_right);
 }
 
 void NEDepthConcatenateLayerKernel::configure(const ITensor *input, unsigned int depth_offset, ITensor *output)
@@ -149,8 +128,6 @@ void NEDepthConcatenateLayerKernel::configure(const ITensor *input, unsigned int
     _input        = input;
     _output       = output;
     _depth_offset = depth_offset;
-    _left_right   = (output->info()->dimension(0) - input->info()->dimension(0)) / 2;
-    _top_bottom   = (output->info()->dimension(1) - input->info()->dimension(1)) / 2;
 
     switch(input->info()->data_type())
     {
@@ -190,5 +167,5 @@ void NEDepthConcatenateLayerKernel::run(const Window &window, const ThreadInfo &
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
     ARM_COMPUTE_ERROR_ON(_func == nullptr);
 
-    (*_func)(_input, _output, std::make_pair(_left_right, _top_bottom), _depth_offset, window);
+    (*_func)(_input, _output, _depth_offset, window);
 }