COMPMID-2765 Add support for QASYMM8_SIGNED in NEDeconvolutionLayer

Signed-off-by: Luca Foschiani <luca.foschiani@arm.com>
Change-Id: I8295fadee15311a9ab846aa24c031b82c0b799eb
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2952
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>

1 files changed, 23 insertions, 46 deletions

diff --git a/src/core/NEON/kernels/NEReverseKernel.cpp b/src/core/NEON/kernels/NEReverseKernel.cpp
index 2f584164dc..5a8c446ddd 100644
--- a/src/core/NEON/kernels/NEReverseKernel.cpp
+++ b/src/core/NEON/kernels/NEReverseKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 ARM Limited.
+ * Copyright (c) 2018-2020 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -106,33 +106,20 @@ void run_reverse(const Window &window, const ITensor *input, const ITensor *axis
     }
 
     // Check if we need a left-over loop for the y dimension
-    const int window_step_x            = 16 / input->info()->element_size();
-    const int window_start_x           = window.x().start();
-    const int window_end_x             = std::min(window.x().end(), static_cast<int>(input->info()->dimension(0)));
-    const int window_end_x_multiple_of = ((window_end_x - window_start_x) / window_step_x) * window_step_x;
-    bool      left_over_loop_x         = (((window_end_x - window_start_x) % window_step_x) != 0);
+    const int window_step_x  = 16 / input->info()->element_size();
+    const int window_start_x = window.x().start();
+    const int window_end_x   = window.x().end();
 
-    Window slice = window.first_slice_window_4D();
+    Window win(window);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
 
-    if(left_over_loop_x)
+    Iterator input_it(input, win);
+    execute_window_loop(win, [&](const Coordinates & id)
     {
-        // Check if window_end_y_multiple_of is greater than window_start_y
-        if(window_end_x_multiple_of > window_start_x)
+        int x = window_start_x;
+        for(; x <= (window_end_x - window_step_x); x += window_step_x)
         {
-            slice.set(Window::DimX, Window::Dimension(window_start_x, window_end_x_multiple_of, window_step_x));
-        }
-        else
-        {
-            slice.set(Window::DimX, Window::Dimension(0, 0, 1));
-        }
-    }
-
-    do
-    {
-        Iterator input_it(input, slice);
-        execute_window_loop(slice, [&](const Coordinates & id)
-        {
-            auto in = wrapper::vloadq(reinterpret_cast<T *>(input_it.ptr()));
+            auto in = wrapper::vloadq(reinterpret_cast<T *>(input_it.ptr()) + x);
 
             // Reverse 0 axis
             if(axis_bit & 0x1)
@@ -141,39 +128,29 @@ void run_reverse(const Window &window, const ITensor *input, const ITensor *axis
                 in = wrapper::vcombine(wrapper::vgethigh(in), wrapper::vgetlow(in));
             }
 
-            const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - id.x() - window_step_x : id.x();
+            const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - x - window_step_x : x;
             const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y();
             const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z();
             const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3];
 
             auto out_ptr = reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w)));
             wrapper::vstore(out_ptr, in);
-        },
-        input_it);
+        }
 
-        if(left_over_loop_x)
+        // Compute left-over elements
+        for(; x < window_end_x; ++x)
         {
-            slice.set(Window::DimX, Window::Dimension(window_end_x_multiple_of, window_end_x, 1));
+            const auto in = *(reinterpret_cast<T *>(input_it.ptr()) + x);
 
-            Iterator input_it(input, slice);
-
-            // Compute left-over elements along the y dimension (1x1)
-            execute_window_loop(slice, [&](const Coordinates & id)
-            {
-                const auto in = *reinterpret_cast<T *>(input_it.ptr());
-
-                const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - id.x() - 1 : id.x();
-                const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y();
-                const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z();
-                const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3];
+            const int offset_x = (axis_bit & 0x1) ? output->info()->dimension(0) - x - 1 : x;
+            const int offset_y = (axis_bit & 0x2) ? output->info()->dimension(1) - id.y() - 1 : id.y();
+            const int offset_z = (axis_bit & 0x4) ? output->info()->dimension(2) - id.z() - 1 : id.z();
+            const int offset_w = (axis_bit & 0x8) ? output->info()->dimension(3) - id[3] - 1 : id[3];
 
-                *reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w))) = in;
-            },
-            input_it);
+            *reinterpret_cast<T *>(output->ptr_to_element(Coordinates(offset_x, offset_y, offset_z, offset_w))) = in;
         }
-
-    }
-    while(window.slide_window_slice_4D(slice));
+    },
+    input_it);
 }
 
 void NEReverseKernel::run(const Window &window, const ThreadInfo &info)