commit a9c4472188abef421adb589e2a6fef52727d465f
author Michalis Spyrou <michalis.spyrou@arm.com> Fri Apr 05 17:18:36 2019 +0100
committer Georgios Pinitas <georgios.pinitas@arm.com> Tue Apr 16 16:02:13 2019 +0000
tree f8f6540b05049074030c32332b5427e826cc58ea
parent 2ec6c1eb6ee77b79e8ab6b97b8cd70bcc4c5589d

COMPMID-2051 Refactor shape_calculator::calculate_concatenate_shape

Change-Id: Ibf316718d11fa975d75f226925747b21c4efd127
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/974
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp

diff --git a/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp
index 36d1b29..6f70efe 100644
--- a/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -38,28 +38,18 @@
 using namespace arm_compute;
 
 GCDepthConcatenateLayerKernel::GCDepthConcatenateLayerKernel()
-    : _input(nullptr), _output(nullptr), _top_bottom(0), _left_right(0), _depth_offset(0)
+    : _input(nullptr), _output(nullptr), _depth_offset(0)
 {
 }
-
-BorderSize GCDepthConcatenateLayerKernel::border_size() const
-{
-    return BorderSize(_top_bottom, _left_right);
-}
-
 void GCDepthConcatenateLayerKernel::configure(const IGCTensor *input, unsigned int depth_offset, IGCTensor *output)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) + depth_offset > output->info()->dimension(2));
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) > output->info()->dimension(0));
-    ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) > output->info()->dimension(1));
-    ARM_COMPUTE_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_ERROR_ON((output->info()->dimension(0) - input->info()->dimension(0)) % 2);
-    ARM_COMPUTE_ERROR_ON((output->info()->dimension(1) - input->info()->dimension(1)) % 2);
+    ARM_COMPUTE_ERROR_ON(input->info()->dimension(Window::DimX) != output->info()->dimension(Window::DimX));
+    ARM_COMPUTE_ERROR_ON(input->info()->dimension(Window::DimY) != output->info()->dimension(Window::DimY));
+    ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) + depth_offset > output->info()->dimension(2));
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(3, input, output);
 
     _input        = input;
     _output       = output;
@@ -73,35 +63,20 @@
     build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1));
     build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1));
 
-    // Configure kernel window
-    _left_right = (output->info()->dimension(0) - input->info()->dimension(0)) / 2;
-    _top_bottom = (output->info()->dimension(1) - input->info()->dimension(1)) / 2;
-
-    build_opts.emplace("#define OFFSET_X " + support::cpp11::to_string(_left_right));
-    build_opts.emplace("#define OFFSET_Y " + support::cpp11::to_string(_top_bottom));
-
     // Create kernel
     _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("concatenate_depth", build_opts));
 
     unsigned int num_elems_processed_per_iteration = 1;
-    unsigned int num_elems_read_per_iteration      = 1;
-    if(input->info()->data_type() == DataType::F32)
-    {
-        num_elems_processed_per_iteration = 1;
-        num_elems_read_per_iteration      = 1;
-    }
-    else if(input->info()->data_type() == DataType::F16)
+    if(input->info()->data_type() == DataType::F16)
     {
         num_elems_processed_per_iteration = 4;
-        num_elems_read_per_iteration      = 4;
     }
-    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->info(), Steps(num_elems_processed_per_iteration));
     win.set(Window::DimZ, Window::Dimension(0, input->info()->tensor_shape().z(), 1));
 
-    AccessWindowRectangle  input_access(input->info(), -_left_right, -_top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration);
+    AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration);
     AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration);
     update_window_and_padding(win, input_access, output_access);
     output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
@@ -118,11 +93,9 @@
 
     _output->set_needs_shifting(true);
 
-    Window slice     = window.first_slice_window_3D();
     Window slice_in  = window.first_slice_window_3D();
     Window slice_out = window.first_slice_window_3D();
 
-    slice.shift(Window::DimX, -(_output->info()->padding()).left);
     slice_out.set(Window::DimZ, Window::Dimension(_depth_offset));
 
     do
@@ -133,7 +106,7 @@
 
         _kernel.update_shader_params();
 
-        enqueue(*this, slice);
+        enqueue(*this, slice_in);
     }
-    while(window.slide_window_slice_3D(slice) && window.slide_window_slice_3D(slice_in));
+    while(window.slide_window_slice_3D(slice_in));
 }