COMPMID-3185: Remove padding from NEHeightConcatenateLayerKernel

Change-Id: Ib5de701e0ec0d41eb0d9f7a292e8d2949e8d41e2
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2835
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

1 files changed, 55 insertions, 32 deletions

diff --git a/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp b/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp
index 4f7ae8c0b1..2f130d9aae 100644
--- a/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEHeightConcatenateLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 ARM Limited.
+ * Copyright (c) 2019-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -40,20 +40,6 @@ namespace arm_compute
 {
 namespace
 {
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
-{
-    const unsigned int num_elems_processed_per_iteration = 16 / output->element_size();
-
-    // The window needs to be based on input as we copy all the widths of input
-    Window                 win = calculate_max_window(*input, Steps(num_elems_processed_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);
-
-    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
-    return std::make_pair(err, win);
-}
-
 Status validate_arguments(const ITensorInfo *input, unsigned int height_offset, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
@@ -86,19 +72,16 @@ void NEHeightConcatenateLayerKernel::configure(const ITensor *input, unsigned in
     _height_offset = height_offset;
 
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), output->info());
-    ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
-
-    INEKernel::configure(std::get<1>(win_config));
-
-    // Set output valid region
-    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+    Window      win = calculate_max_window(*output->info(), Steps());
+    Coordinates coord;
+    coord.set_num_dimensions(output->info()->num_dimensions());
+    output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
+    INEKernel::configure(win);
 }
 
 Status NEHeightConcatenateLayerKernel::validate(const ITensorInfo *input, unsigned int height_offset, const ITensorInfo *output)
 {
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, height_offset, output));
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
     return Status{};
 }
 
@@ -111,37 +94,77 @@ void NEHeightConcatenateLayerKernel::run(const Window &window, const ThreadInfo
     // Offset output pointer to the correct position
     uint8_t *output_ptr = _output->buffer() + _output->info()->offset_first_element_in_bytes() + _height_offset * _output->info()->strides_in_bytes()[Window::DimY];
 
+    const auto window_start_x = static_cast<int>(window.x().start());
+    const auto window_end_x   = static_cast<int>(window.x().end()) * static_cast<int>(_output->info()->element_size()); 
+    const int window_step_x = 16;
+
+    Window win{ window };
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+    win.set(Window::DimY, Window::Dimension(0, _input->info()->tensor_shape().y(), 1));
+
     // Create iterators
-    Iterator                       input(_input, window);
-    Iterator                       output(_output, window);
+    Iterator input(_input, win);
+    Iterator output(_output, win);
+
     const DataType                 dt           = _input->info()->data_type();
     const UniformQuantizationInfo &input_qinfo  = _input->info()->quantization_info().uniform();
     const UniformQuantizationInfo &output_qinfo = _output->info()->quantization_info().uniform();
     if(dt == DataType::QASYMM8 && input_qinfo != output_qinfo)
     {
-        execute_window_loop(window, [&](const Coordinates &)
+        execute_window_loop(win, [&](const Coordinates &)
         {
-            vst1q_u8(output_ptr + output.offset(), vquantize(vdequantize(vld1q_u8(input.ptr()), input_qinfo), output_qinfo));
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                vst1q_u8(output_ptr + output.offset() + x, vquantize(vdequantize(vld1q_u8(input.ptr() + x), input_qinfo), output_qinfo));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                *(output_ptr + output.offset() + x) = quantize_qasymm8(dequantize_qasymm8(*(input.ptr() + x), input_qinfo), output_qinfo);
+            }
+
         },
         input, output);
     }
     else if(dt == DataType::QASYMM8_SIGNED && input_qinfo != output_qinfo)
     {
-        execute_window_loop(window, [&](const Coordinates &)
+        execute_window_loop(win, [&](const Coordinates &)
         {
-            vst1q_s8(reinterpret_cast<int8_t *>(output_ptr + output.offset()),
-                     vquantize_signed(vdequantize(vld1q_s8(reinterpret_cast<int8_t *>(input.ptr())), input_qinfo), output_qinfo));
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                vst1q_s8(reinterpret_cast<int8_t *>(output_ptr + output.offset() + x),
+                         vquantize_signed(vdequantize(vld1q_s8(reinterpret_cast<int8_t *>(input.ptr()) + x), input_qinfo), output_qinfo));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                *(output_ptr + output.offset() + x) = quantize_qasymm8_signed(dequantize_qasymm8_signed(*(input.ptr() + x), input_qinfo), output_qinfo);
+            }
         },
         input, output);
     }
     else
     {
-        execute_window_loop(window, [&](const Coordinates &)
+        execute_window_loop(win, [&](const Coordinates &)
         {
             const auto in_ptr  = input.ptr();
             const auto out_ptr = output_ptr + output.offset();
 
-            wrapper::vstore(out_ptr, wrapper::vloadq(in_ptr));
+            int x = window_start_x;
+            for(; x <= (window_end_x - window_step_x); x += window_step_x)
+            {
+                wrapper::vstore(out_ptr + x, wrapper::vloadq(in_ptr + x));
+            }
+
+            // Compute left-over elements
+            for(; x < window_end_x; ++x)
+            {
+                *(out_ptr + x) = *(in_ptr + x);
+            }
         },
         input, output);
     }